Bug Summary

File:nnc/ccv_cnnp_model_addons.c
Warning:line 667, column 15
Value stored to 'ofs' during its initialization is never read

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -cc1 -triple x86_64-unknown-linux-gnu -analyze -disable-free -clear-ast-before-backend -disable-llvm-verifier -discard-value-names -main-file-name ccv_cnnp_model_addons.c -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -setup-static-analyzer -mrelocation-model pic -pic-level 2 -pic-is-pie -mframe-pointer=none -fmath-errno -ffp-contract=on -fno-rounding-math -mconstructor-aliases -funwind-tables=2 -target-cpu x86-64 -target-feature +sse2 -tune-cpu generic -debugger-tuning=gdb -fdebug-compilation-dir=/home/liu/actions-runner/_work/ccv/ccv/lib/nnc -fcoverage-compilation-dir=/home/liu/actions-runner/_work/ccv/ccv/lib/nnc -resource-dir /usr/local/lib/clang/19 -I ../ -I /usr/local/cuda/include -D HAVE_CBLAS -D HAVE_LIBPNG -D HAVE_LIBJPEG -D HAVE_FFTW3 -D HAVE_PTHREAD -D HAVE_LIBLINEAR -D HAVE_TESSERACT -D HAVE_AVCODEC -D HAVE_AVFORMAT -D HAVE_AVUTIL -D HAVE_SWSCALE -D HAVE_SSE2 -D HAVE_GSL -D HAVE_CUDA -D HAVE_CUDNN -D HAVE_NCCL -D USE_SYSTEM_CUB -I /usr/local/include -internal-isystem /usr/local/lib/clang/19/include -internal-isystem /usr/local/include -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/12/../../../../x86_64-linux-gnu/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -O3 -ferror-limit 19 -fgnuc-version=4.2.1 -fskip-odr-check-in-gmf -vectorize-loops -vectorize-slp -analyzer-output=html -faddrsig -D__GCC_HAVE_DWARF2_CFI_ASM=1 -o /home/liu/actions-runner/_work/ccv/ccv/_analyze/2026-10-26-140650-29116-1 -x c ccv_cnnp_model_addons.c
1#include "ccv_nnc.h"
2#include "ccv_nnc_easy.h"
3#include "ccv_nnc_internal.h"
4#include "ccv_internal.h"
5#include "_ccv_cnnp_model.h"
6
7// MARK - Add-on Functions
8
9static int _ccv_cnnp_model_clip_grad_norm_reduce_norm2(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint, const int flags, ccv_nnc_tensor_t* const* const inputs, const int input_size, ccv_nnc_tensor_t* const* const outputs, const int output_size, ccv_nnc_stream_context_t* const stream_context)
10{
11 const int device_id = CCV_TENSOR_GET_DEVICE_ID(inputs[0]->info.type)(((inputs[0]->info.type) & 0xfff00) >> 8);
12 ccv_nnc_tensor_t* const old_norm2 = outputs[1 + device_id * 2];
13 ccv_nnc_tensor_t* const norm2 = outputs[1 + device_id * 2 + 1];
14 const int tensor_count = ccv_nnc_tensor_count(inputs[0]->info);
15 if (tensor_count == 1)
16 ccv_nnc_cmd_exec(CMD_MUL_FORWARD(1)ccv_nnc_cmd(CCV_NNC_MUL_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={1,}}}, 0), hint, flags, TENSOR_LIST(inputs[0], inputs[0])(ccv_nnc_tensor_t* []){inputs[0], inputs[0]}, (1 +1 +1 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(norm2)(ccv_nnc_tensor_t* []){norm2}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context);
17 else {
18 ccv_nnc_cmd_exec(CMD_REDUCE_NORM2_FORWARD()ccv_nnc_cmd(CCV_NNC_REDUCE_NORM2_FORWARD, 0, ((ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.reduce={.count=(1 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1),.axis={}}}
), 0), hint, flags, TENSOR_LIST(inputs[0])(ccv_nnc_tensor_t* []){inputs[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(norm2)(ccv_nnc_tensor_t* []){norm2}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context);
19 ccv_nnc_cmd_exec(CMD_MUL_FORWARD(1)ccv_nnc_cmd(CCV_NNC_MUL_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={1,}}}, 0), hint, flags, TENSOR_LIST(norm2, norm2)(ccv_nnc_tensor_t* []){norm2, norm2}, (1 +1 +1 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(norm2)(ccv_nnc_tensor_t* []){norm2}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context);
20 }
21 ccv_nnc_cmd_exec(CMD_ADD_FORWARD(1, 1)ccv_nnc_cmd(CCV_NNC_ADD_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={1, 1}}}, 0), hint, flags, TENSOR_LIST(old_norm2, norm2)(ccv_nnc_tensor_t* []){old_norm2, norm2}, (1 +1 +1 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(old_norm2)(ccv_nnc_tensor_t* []){old_norm2}, (1 +1 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context);
22 return CCV_NNC_EXEC_SUCCESS;
23}
24
25static ccv_nnc_cmd_vtab_t clip_grad_norm_reduce_norm2_vtab = {
26 .exec = _ccv_cnnp_model_clip_grad_norm_reduce_norm2
27};
28
29static int _ccv_cnnp_model_clip_grad_norm_scatter_norm2(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint, const int flags, ccv_nnc_tensor_t* const* const inputs, const int input_size, ccv_nnc_tensor_t* const* const outputs, const int output_size, ccv_nnc_stream_context_t* const stream_context)
30{
31 const int device_id = CCV_TENSOR_GET_DEVICE_ID(inputs[0]->info.type)(((inputs[0]->info.type) & 0xfff00) >> 8);
32 ccv_nnc_tensor_t* const norm2 = inputs[1 + device_id * 2];
33 ccv_nnc_cmd_exec(CMD_MUL_FORWARD(1)ccv_nnc_cmd(CCV_NNC_MUL_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={1,}}}, 0), hint, flags, TENSOR_LIST(inputs[0], norm2)(ccv_nnc_tensor_t* []){inputs[0], norm2}, (1 +1 +1 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(outputs[0])(ccv_nnc_tensor_t* []){outputs[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context);
34 return CCV_NNC_EXEC_SUCCESS;
35}
36
37static ccv_nnc_cmd_vtab_t clip_grad_norm_scatter_norm2_vtab = {
38 .exec = _ccv_cnnp_model_clip_grad_norm_scatter_norm2
39};
40
41void ccv_cnnp_model_parameters_clip_grad_norm(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t parameters, int norm_type, float max_norm, ccv_nnc_stream_context_t* const stream_context)
42{
43 assert(norm_type == 2)((void) sizeof ((norm_type == 2) ? 1 : 0), __extension__ ({ if
(norm_type == 2) ; else __assert_fail ("norm_type == 2", "ccv_cnnp_model_addons.c"
, 43, __extension__ __PRETTY_FUNCTION__); }));
44 ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
45 assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
(compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model_addons.c"
, 45, __extension__ __PRETTY_FUNCTION__); }));
46 const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count
); typeof (1) _b = (1); (_a > _b) ? _a : _b; });
47 ccv_nnc_tensor_t* norm2[parallel_count * 2];
48 ccv_nnc_tensor_t* max_normt[parallel_count];
49 const int stream_type = model->compiled_data->stream_type;
50 int i;
51 if (stream_type == CCV_STREAM_CONTEXT_GPU)
52 {
53 for (i = 0; i < parallel_count; i++)
54 {
55 ccv_nnc_tensor_param_t info = {
56 .type = CCV_TENSOR_GPU_MEMORY,
57 .format = CCV_TENSOR_FORMAT_NHWC,
58 .datatype = CCV_32F,
59 .dim = {1},
60 };
61 CCV_TENSOR_SET_DEVICE_ID(info.type, i)(info.type) = (((info.type) & ~0xfff00) | (((i) & 0xfff
) << 8));
62 norm2[i * 2] = ccv_nnc_tensor_new(ccv_nnc_xpu_alloc(&compiled_data->xpu_alloc, i, stream_context, ccv_nnc_tensor_data_size(info)), info, 0);
63 norm2[i * 2 + 1] = ccv_nnc_tensor_new(ccv_nnc_xpu_alloc(&compiled_data->xpu_alloc, i, stream_context, ccv_nnc_tensor_data_size(info)), info, 0);
64 max_normt[i] = ccv_nnc_tensor_new(ccv_nnc_xpu_alloc(&compiled_data->xpu_alloc, i, stream_context, ccv_nnc_tensor_data_size(info)), info, 0);
65 }
66 } else {
67 for (i = 0; i < parallel_count; i++)
68 {
69 ccv_nnc_tensor_param_t info = {
70 .type = CCV_TENSOR_CPU_MEMORY,
71 .format = CCV_TENSOR_FORMAT_NHWC,
72 .datatype = CCV_32F,
73 .dim = {1},
74 };
75 norm2[i * 2] = ccv_nnc_tensor_new(0, info, 0);
76 norm2[i * 2 + 1] = ccv_nnc_tensor_new(0, info, 0);
77 max_normt[i] = ccv_nnc_tensor_new(0, info, 0);
78 }
79 }
80 // zero out old norm2.
81 if (parallel_count > 1)
82 {
83 ccv_nnc_stream_context_t* streams[parallel_count];
84 ccv_nnc_stream_signal_t* signal;
85 if (stream_context)
86 signal = ccv_nnc_stream_context_emit_signal_new(stream_context);
87 for (i = 0; i < parallel_count; i++)
88 {
89 const int stream_type = CCV_TENSOR_GET_MEMORY(norm2[i * 2]->info.type)((norm2[i * 2]->info.type) & 0x3) == CCV_TENSOR_GPU_MEMORY ? CCV_STREAM_CONTEXT_GPU : CCV_STREAM_CONTEXT_CPU;
90 const int device_id = CCV_TENSOR_GET_DEVICE_ID(norm2[i * 2]->info.type)(((norm2[i * 2]->info.type) & 0xfff00) >> 8);
91 int type = stream_type;
92 CCV_STREAM_SET_DEVICE_ID(type, device_id)(type) = (((type) & ~0xfff00) | (((device_id) & 0xfff
) << 8));
93 ccv_nnc_stream_context_t* const stream_0 = ccv_cnnp_compiled_data_get_stream(compiled_data, type);
94 // Wait signal to finish.
95 if (stream_context)
96 ccv_nnc_stream_context_wait_signal(stream_0, signal);
97 ccv_nnc_cmd_exec(CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, 0, TENSOR_LIST(norm2[i * 2])(ccv_nnc_tensor_t* []){norm2[i * 2]}, (1 +1 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_0);
98 if (stream_context)
99 {
100 ccv_nnc_stream_signal_t* const signal = ccv_nnc_stream_context_emit_signal_new(stream_0);
101 ccv_nnc_stream_context_wait_signal(stream_context, signal);
102 }
103 streams[i] = stream_0;
104 }
105 // If this should be blocking, blocking it.
106 if (!stream_context)
107 for (i = 0; i < parallel_count; i++)
108 if (streams[i])
109 ccv_nnc_stream_context_wait(streams[i]);
110 } else {
111 ccv_nnc_cmd_exec(CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, 0, TENSOR_LIST(norm2[0])(ccv_nnc_tensor_t* []){norm2[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context);
112 }
113 // Gather norm2.
114 ccv_nnc_cmd_t reduce_cmd = {
115 .cmd = CCV_NNC_CUSTOM_FORWARD,
116 .isa = &clip_grad_norm_reduce_norm2_vtab,
117 };
118 ccv_cnnp_model_parameter_gradients_map(model, parameters, reduce_cmd, ccv_nnc_no_hint, 0, 0, 0, norm2, parallel_count * 2, stream_context);
119 // Now compute max(max_norm / norm2, 1.0).
120 if (parallel_count > 1)
121 {
122 ccv_nnc_stream_context_t* streams[parallel_count];
123 ccv_nnc_stream_signal_t* signal;
124 if (stream_context)
125 signal = ccv_nnc_stream_context_emit_signal_new(stream_context);
126 for (i = 0; i < parallel_count; i++)
127 {
128 const int stream_type = CCV_TENSOR_GET_MEMORY(norm2[i * 2]->info.type)((norm2[i * 2]->info.type) & 0x3) == CCV_TENSOR_GPU_MEMORY ? CCV_STREAM_CONTEXT_GPU : CCV_STREAM_CONTEXT_CPU;
129 const int device_id = CCV_TENSOR_GET_DEVICE_ID(norm2[i * 2]->info.type)(((norm2[i * 2]->info.type) & 0xfff00) >> 8);
130 int type = stream_type;
131 CCV_STREAM_SET_DEVICE_ID(type, device_id)(type) = (((type) & ~0xfff00) | (((device_id) & 0xfff
) << 8));
132 ccv_nnc_stream_context_t* const stream_0 = ccv_cnnp_compiled_data_get_stream(compiled_data, type);
133 // Wait signal to finish.
134 if (stream_context)
135 ccv_nnc_stream_context_wait_signal(stream_0, signal);
136 ccv_nnc_cmd_exec(CMD_EWSQRT_FORWARD()ccv_nnc_cmd(CCV_NNC_EWSQRT_FORWARD, 0, ccv_nnc_cmd_auto, 0), ccv_nnc_no_hint, 0, TENSOR_LIST(norm2[i * 2])(ccv_nnc_tensor_t* []){norm2[i * 2]}, (1 +1 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(norm2[i * 2])(ccv_nnc_tensor_t* []){norm2[i * 2]}, (1 +1 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_0);
137 ccv_nnc_cmd_exec(CMD_SET_FORWARD(max_norm)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={max_norm,}}}, 0), ccv_nnc_no_hint, 0, 0, 0, TENSOR_LIST(max_normt[i])(ccv_nnc_tensor_t* []){max_normt[i]}, (1 +1 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_0);
138 ccv_nnc_cmd_exec(CMD_EWDIV_FORWARD()ccv_nnc_cmd(CCV_NNC_EWDIV_FORWARD, 0, ccv_nnc_cmd_auto, 0), ccv_nnc_no_hint, 0, TENSOR_LIST(max_normt[i], norm2[i * 2])(ccv_nnc_tensor_t* []){max_normt[i], norm2[i * 2]}, (1 +1 +1 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1
), TENSOR_LIST(norm2[i * 2])(ccv_nnc_tensor_t* []){norm2[i * 2]}, (1 +1 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_0);
139 ccv_nnc_cmd_exec(CMD_CLAMP_FORWARD(NAN, 1)ccv_nnc_cmd(CCV_NNC_CLAMP_FORWARD, 0, (ccv_nnc_cmd_param_t){.
size={.dim={1,1,1}},.clamp={.min=(__builtin_nanf ("")),.max=1
}}, 0), ccv_nnc_no_hint, 0, TENSOR_LIST(norm2[i * 2])(ccv_nnc_tensor_t* []){norm2[i * 2]}, (1 +1 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(norm2[i * 2])(ccv_nnc_tensor_t* []){norm2[i * 2]}, (1 +1 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_0);
140 if (stream_context)
141 {
142 ccv_nnc_stream_signal_t* const signal = ccv_nnc_stream_context_emit_signal_new(stream_0);
143 ccv_nnc_stream_context_wait_signal(stream_context, signal);
144 }
145 streams[i] = stream_0;
146 }
147 // If this should be blocking, blocking it.
148 if (!stream_context)
149 for (i = 0; i < parallel_count; i++)
150 if (streams[i])
151 ccv_nnc_stream_context_wait(streams[i]);
152 } else {
153 ccv_nnc_cmd_exec(CMD_EWSQRT_FORWARD()ccv_nnc_cmd(CCV_NNC_EWSQRT_FORWARD, 0, ccv_nnc_cmd_auto, 0), ccv_nnc_no_hint, 0, TENSOR_LIST(norm2[0])(ccv_nnc_tensor_t* []){norm2[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(norm2[0])(ccv_nnc_tensor_t* []){norm2[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context);
154 ccv_nnc_cmd_exec(CMD_SET_FORWARD(max_norm)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={max_norm,}}}, 0), ccv_nnc_no_hint, 0, 0, 0, TENSOR_LIST(max_normt[0])(ccv_nnc_tensor_t* []){max_normt[0]}, (1 +1 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context);
155 ccv_nnc_cmd_exec(CMD_EWDIV_FORWARD()ccv_nnc_cmd(CCV_NNC_EWDIV_FORWARD, 0, ccv_nnc_cmd_auto, 0), ccv_nnc_no_hint, 0, TENSOR_LIST(max_normt[0], norm2[0])(ccv_nnc_tensor_t* []){max_normt[0], norm2[0]}, (1 +1 +1 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(norm2[0])(ccv_nnc_tensor_t* []){norm2[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context);
156 ccv_nnc_cmd_exec(CMD_CLAMP_FORWARD(NAN, 1)ccv_nnc_cmd(CCV_NNC_CLAMP_FORWARD, 0, (ccv_nnc_cmd_param_t){.
size={.dim={1,1,1}},.clamp={.min=(__builtin_nanf ("")),.max=1
}}, 0), ccv_nnc_no_hint, 0, TENSOR_LIST(norm2[0])(ccv_nnc_tensor_t* []){norm2[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(norm2[0])(ccv_nnc_tensor_t* []){norm2[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context);
157 }
158 ccv_nnc_cmd_t scatter_cmd = {
159 .cmd = CCV_NNC_CUSTOM_FORWARD,
160 .isa = &clip_grad_norm_scatter_norm2_vtab,
161 };
162 ccv_cnnp_model_parameter_gradients_map(model, parameters, scatter_cmd, ccv_nnc_no_hint, 0, norm2, parallel_count * 2, 0, 0, stream_context);
163 if (stream_type == CCV_STREAM_CONTEXT_GPU)
164 for (i = 0; i < parallel_count; i++)
165 {
166 ccv_nnc_xpu_free(&compiled_data->xpu_alloc, norm2[i * 2]->data.u8);
167 ccv_nnc_xpu_free(&compiled_data->xpu_alloc, norm2[i * 2 + 1]->data.u8);
168 ccv_nnc_xpu_free(&compiled_data->xpu_alloc, max_normt[i]->data.u8);
169 }
170 for (i = 0; i < parallel_count; i++)
171 {
172 ccv_nnc_tensor_free(norm2[i * 2]);
173 ccv_nnc_tensor_free(norm2[i * 2 + 1]);
174 ccv_nnc_tensor_free(max_normt[i]);
175 }
176}
177
178// MARK - Add-on Functions
179
180static int _ccv_cnnp_model_isnan(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint, const int flags, ccv_nnc_tensor_t* const* const inputs, const int input_size, ccv_nnc_tensor_t* const* const outputs, const int output_size, ccv_nnc_stream_context_t* const stream_context)
181{
182 const int device_id = CCV_TENSOR_GET_DEVICE_ID(inputs[0]->info.type)(((inputs[0]->info.type) & 0xfff00) >> 8);
183 ccv_nnc_tensor_t* const old_isnanr = outputs[1 + device_id * 2];
184 ccv_nnc_tensor_t* const isnanr = outputs[1 + device_id * 2 + 1];
185 ccv_nnc_cmd_t reduce_cmd = CMD_REDUCE_ISNAN_FORWARD()ccv_nnc_cmd(CCV_NNC_REDUCE_ISNAN_FORWARD, 0, ((ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.reduce={.count=(1 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1),.axis={}}}
), 0);
186 reduce_cmd.info.reduce.count = ccv_nnc_tensor_nd(inputs[0]->info.dim);
187 int i;
188 for (i = 0; i < cmd.info.reduce.count; i++)
189 reduce_cmd.info.reduce.axis[i] = i;
190 ccv_nnc_cmd_exec(reduce_cmd, hint, flags, TENSOR_LIST(inputs[0])(ccv_nnc_tensor_t* []){inputs[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(isnanr)(ccv_nnc_tensor_t* []){isnanr}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context);
191 ccv_nnc_cmd_exec(CMD_EWSUM_FORWARD()ccv_nnc_cmd(CCV_NNC_EWSUM_FORWARD, 0, ccv_nnc_cmd_auto, 0), hint, flags, TENSOR_LIST(old_isnanr, isnanr)(ccv_nnc_tensor_t* []){old_isnanr, isnanr}, (1 +1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(old_isnanr)(ccv_nnc_tensor_t* []){old_isnanr}, (1 +1 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context);
192 return CCV_NNC_EXEC_SUCCESS;
193}
194
195static ccv_nnc_cmd_vtab_t reduce_isnan_vtab = {
196 .exec = _ccv_cnnp_model_isnan
197};
198
199int ccv_cnnp_model_parameter_gradients_isnan(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t parameters, ccv_nnc_stream_context_t* const stream_context)
200{
201 ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
202 assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
(compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model_addons.c"
, 202, __extension__ __PRETTY_FUNCTION__); }));
203 const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count
); typeof (1) _b = (1); (_a > _b) ? _a : _b; });
204 ccv_nnc_tensor_t* isnanr[parallel_count * 2];
205 const int stream_type = model->compiled_data->stream_type;
206 int i;
207 if (stream_type == CCV_STREAM_CONTEXT_GPU)
208 {
209 for (i = 0; i < parallel_count; i++)
210 {
211 ccv_nnc_tensor_param_t info = {
212 .type = CCV_TENSOR_GPU_MEMORY,
213 .format = CCV_TENSOR_FORMAT_NHWC,
214 .datatype = CCV_32S,
215 .dim = {1},
216 };
217 CCV_TENSOR_SET_DEVICE_ID(info.type, i)(info.type) = (((info.type) & ~0xfff00) | (((i) & 0xfff
) << 8));
218 isnanr[i * 2] = ccv_nnc_tensor_new(ccv_nnc_xpu_alloc(&compiled_data->xpu_alloc, i, stream_context, ccv_nnc_tensor_data_size(info)), info, 0);
219 isnanr[i * 2 + 1] = ccv_nnc_tensor_new(ccv_nnc_xpu_alloc(&compiled_data->xpu_alloc, i, stream_context, ccv_nnc_tensor_data_size(info)), info, 0);
220 }
221 } else {
222 for (i = 0; i < parallel_count; i++)
223 {
224 ccv_nnc_tensor_param_t info = {
225 .type = CCV_TENSOR_CPU_MEMORY,
226 .format = CCV_TENSOR_FORMAT_NHWC,
227 .datatype = CCV_32S,
228 .dim = {1},
229 };
230 isnanr[i * 2] = ccv_nnc_tensor_new(0, info, 0);
231 isnanr[i * 2 + 1] = ccv_nnc_tensor_new(0, info, 0);
232 }
233 }
234 // zero out old isnanr.
235 if (parallel_count > 1)
236 {
237 ccv_nnc_stream_context_t* streams[parallel_count];
238 ccv_nnc_stream_signal_t* signal;
239 if (stream_context)
240 signal = ccv_nnc_stream_context_emit_signal_new(stream_context);
241 for (i = 0; i < parallel_count; i++)
242 {
243 const int stream_type = CCV_TENSOR_GET_MEMORY(isnanr[i * 2]->info.type)((isnanr[i * 2]->info.type) & 0x3) == CCV_TENSOR_GPU_MEMORY ? CCV_STREAM_CONTEXT_GPU : CCV_STREAM_CONTEXT_CPU;
244 const int device_id = CCV_TENSOR_GET_DEVICE_ID(isnanr[i * 2]->info.type)(((isnanr[i * 2]->info.type) & 0xfff00) >> 8);
245 int type = stream_type;
246 CCV_STREAM_SET_DEVICE_ID(type, device_id)(type) = (((type) & ~0xfff00) | (((device_id) & 0xfff
) << 8));
247 ccv_nnc_stream_context_t* const stream_0 = ccv_cnnp_compiled_data_get_stream(compiled_data, type);
248 // Wait signal to finish.
249 if (stream_context)
250 ccv_nnc_stream_context_wait_signal(stream_0, signal);
251 ccv_nnc_cmd_exec(CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, 0, TENSOR_LIST(isnanr[i * 2])(ccv_nnc_tensor_t* []){isnanr[i * 2]}, (1 +1 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_0);
252 if (stream_context)
253 {
254 ccv_nnc_stream_signal_t* const signal = ccv_nnc_stream_context_emit_signal_new(stream_0);
255 ccv_nnc_stream_context_wait_signal(stream_context, signal);
256 }
257 streams[i] = stream_0;
258 }
259 // If this should be blocking, blocking it.
260 if (!stream_context)
261 for (i = 0; i < parallel_count; i++)
262 if (streams[i])
263 ccv_nnc_stream_context_wait(streams[i]);
264 } else
265 ccv_nnc_cmd_exec(CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, 0, TENSOR_LIST(isnanr[0])(ccv_nnc_tensor_t* []){isnanr[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context);
266 // Gather isnanr.
267 ccv_nnc_cmd_t reduce_cmd = {
268 .cmd = CCV_NNC_CUSTOM_FORWARD,
269 .isa = &reduce_isnan_vtab,
270 };
271 ccv_cnnp_model_parameter_gradients_map(model, parameters, reduce_cmd, ccv_nnc_no_hint, 0, 0, 0, isnanr, parallel_count * 2, stream_context);
272 for (i = 0; i < parallel_count; i++)
273 ccv_nnc_tensor_free(isnanr[i * 2 + 1]);
274 int retval = 0;
275 if (stream_type == CCV_TENSOR_GPU_MEMORY)
276 {
277 ccv_nnc_tensor_param_t info = {
278 .type = CCV_TENSOR_CPU_MEMORY,
279 .format = CCV_TENSOR_FORMAT_NHWC,
280 .datatype = CCV_32S,
281 .dim = {1},
282 };
283 ccv_nnc_tensor_t* checknan = ccv_nnc_tensor_new(0, info, 0);
284 for (i = 0; i < parallel_count; i++)
285 {
286 ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD()ccv_nnc_cmd(CCV_NNC_DATA_TRANSFER_FORWARD, 0, ccv_nnc_cmd_auto
, 0), ccv_nnc_no_hint, 0, TENSOR_LIST(isnanr[i * 2])(ccv_nnc_tensor_t* []){isnanr[i * 2]}, (1 +1 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(checknan)(ccv_nnc_tensor_t* []){checknan}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), 0);
287 if (checknan->data.i32[0] > 0)
288 {
289 retval = 1;
290 break;
291 }
292 }
293 ccv_nnc_tensor_free(checknan);
294 } else {
295 for (i = 0; i < parallel_count; i++)
296 if (isnanr[i * 2]->data.i32[0] > 0)
297 {
298 retval = 1;
299 break;
300 }
301 }
302 for (i = 0; i < parallel_count; i++)
303 ccv_nnc_tensor_free(isnanr[i * 2]);
304 return retval;
305}
306
307// MARK - Core Layers
308
309static void _ccv_cnnp_sum_build(ccv_cnnp_model_t* const self, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
310{
311 PRINT(CCV_CLI_VERBOSE, "[cnnp_sum_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_sum_build] -\n"); fflush(stdout); } } while (
0);
312 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 312, __extension__ __PRETTY_FUNCTION__
); }));
313 outputs[0] = ccv_nnc_tensor_symbol_new(graph, ccv_nnc_tensor_symbol_params(graph, inputs[0]), 0);
314 ccv_nnc_graph_exec_symbol_new(graph, CMD_EWSUM_FORWARD()ccv_nnc_cmd(CCV_NNC_EWSUM_FORWARD, 0, ccv_nnc_cmd_auto, 0), inputs, input_size, outputs, output_size, 0);
315}
316
317static ccv_cnnp_model_t* _ccv_cnnp_sum_copy(const ccv_cnnp_model_t* const self, void* const context);
318
319static const ccv_cnnp_model_vtab_t ccv_cnnp_sum_isa = {
320 .build = _ccv_cnnp_sum_build,
321 .copy = _ccv_cnnp_sum_copy,
322};
323
324typedef struct {
325 ccv_cnnp_model_t super;
326 ccv_nnc_tensor_symbol_t output;
327} ccv_cnnp_model_sum_t;
328
329ccv_cnnp_model_t* ccv_cnnp_sum(const char* const name)
330{
331 ccv_cnnp_model_sum_t* const model_sum = (ccv_cnnp_model_sum_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_sum_t));
332 model_sum->super.isa = &ccv_cnnp_sum_isa;
333 model_sum->super.input_size = 0;
334 model_sum->super.outputs = &model_sum->output;
335 model_sum->super.output_size = 1;
336 ccv_cnnp_model_copy_name(&model_sum->super, name);
337 return (ccv_cnnp_model_t*)model_sum;
338}
339
340static ccv_cnnp_model_t* _ccv_cnnp_sum_copy(const ccv_cnnp_model_t* const self, void* const context)
341{
342 return ccv_cnnp_sum(self->name);
343}
344
345typedef struct {
346 ccv_cnnp_model_t super;
347 int axis;
348 ccv_nnc_tensor_symbol_t output;
349} ccv_cnnp_model_concat_t;
350
351static void _ccv_cnnp_concat_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
352{
353 const ccv_cnnp_model_concat_t* const self = (const ccv_cnnp_model_concat_t*)super;
354 PRINT(CCV_CLI_VERBOSE, "[cnnp_concat_build] 1. -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_concat_build] 1. -\n"); fflush(stdout); } } while
(0);
355 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 355, __extension__ __PRETTY_FUNCTION__
); }));
356 ccv_nnc_tensor_param_t output_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
357 int i, j;
358 if (output_params.dim[0] == 0)
359 for (i = 1; i < input_size; i++)
360 {
361 output_params = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
362 if (output_params.dim[0] != 0)
363 break;
364 }
365 const int nd = ccv_nnc_tensor_nd(output_params.dim);
366 const int axis = self->axis;
367 if (nd > 0)
368 {
369 assert(axis < nd)((void) sizeof ((axis < nd) ? 1 : 0), __extension__ ({ if (
axis < nd) ; else __assert_fail ("axis < nd", "ccv_cnnp_model_addons.c"
, 369, __extension__ __PRETTY_FUNCTION__); }));
370 output_params.dim[axis] = 0;
371 } else
372 for (i = 0; i < CCV_NNC_MAX_DIM_ALLOC(12); i++)
373 output_params.dim[i] = 0;
374 int input_is_contiguous = 1;
375 for (i = 0; i < input_size; i++)
376 {
377 const ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
378 const int input_nd = ccv_nnc_tensor_nd(input_params.dim);
379 if (input_nd == 0)
380 {
381 PRINT(CCV_CLI_VERBOSE, "[cnnp_concat_build] %d. input[%d]: -\n", i + 2, i)do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_concat_build] %d. input[%d]: -\n", i + 2, i)
; fflush(stdout); } } while (0);
382 input_is_contiguous = 0;
383 continue;
384 }
385 if (CCV_CLI_OUTPUT_LEVEL_IS(CCV_CLI_VERBOSE)(CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
386 {
387 PRINT(CCV_CLI_VERBOSE, "[cnnp_concat_build] %d. input[%d]: (%d", i + 2, i, input_params.dim[0])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_concat_build] %d. input[%d]: (%d", i + 2, i,
input_params.dim[0]); fflush(stdout); } } while (0);
388 int i;
389 for (i = 1; i < CCV_NNC_MAX_DIM_ALLOC(12) && input_params.dim[i] > 0; i++)
390 PRINT(CCV_CLI_VERBOSE, ", %d", input_params.dim[i])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf(", %d", input_params.dim[i]); fflush(stdout); } } while
(0);
391 PRINT(CCV_CLI_VERBOSE, ")\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf(")\n"); fflush(stdout); } } while (0);
392 }
393 assert(input_nd == nd)((void) sizeof ((input_nd == nd) ? 1 : 0), __extension__ ({ if
(input_nd == nd) ; else __assert_fail ("input_nd == nd", "ccv_cnnp_model_addons.c"
, 393, __extension__ __PRETTY_FUNCTION__); }));
394 for (j = 0; j < nd; j++)
395 if (j != axis)
396 { assert(input_params.dim[j] == output_params.dim[j])((void) sizeof ((input_params.dim[j] == output_params.dim[j])
? 1 : 0), __extension__ ({ if (input_params.dim[j] == output_params
.dim[j]) ; else __assert_fail ("input_params.dim[j] == output_params.dim[j]"
, "ccv_cnnp_model_addons.c", 396, __extension__ __PRETTY_FUNCTION__
); })); }
397 output_params.dim[axis] += input_params.dim[axis];
398 }
399 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
400 int ofs[CCV_NNC_MAX_DIM_ALLOC(12)] = {};
401 int stride[CCV_NNC_MAX_DIM_ALLOC(12)] = {};
402 ccv_nnc_tensor_get_stride(output_params.dim, stride);
403 if (input_is_contiguous)
404 {
405 ccv_nnc_tensor_symbol_t aliases[input_size];
406 for (i = 0; i < input_size; i++)
407 {
408 const ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
409 aliases[i] = ccv_nnc_tensor_symbol_alias_new(graph, outputs[0], ofs, stride, input_params, 0);
410 ofs[axis] += input_params.dim[axis];
411 }
412 // Format transform is more flexible.
413 ccv_nnc_graph_exec_symbol_new(graph, CMD_FORMAT_TRANSFORM_FORWARD()ccv_nnc_cmd(CCV_NNC_FORMAT_TRANSFORM_FORWARD, 0, ccv_nnc_cmd_auto
, 0), inputs, input_size, aliases, input_size, "concat");
414 } else {
415 ccv_nnc_tensor_symbol_t aliases[input_size];
416 for (i = 0; i < input_size; i++)
417 {
418 const ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
419 if (input_params.dim[0] == 0)
420 {
421 // Create a new alias anyway, but not going to use it, in this way, the alias count will match during absorb.
422 aliases[i] = ccv_nnc_tensor_symbol_alias_new(graph, outputs[0], ofs, stride, input_params, 0);
423 continue;
424 }
425 aliases[i] = ccv_nnc_tensor_symbol_alias_new(graph, outputs[0], ofs, stride, input_params, 0);
426 ofs[axis] += input_params.dim[axis];
427 }
428 // Format transform is more flexible.
429 ccv_nnc_graph_exec_symbol_new(graph, CMD_FORMAT_TRANSFORM_FORWARD()ccv_nnc_cmd(CCV_NNC_FORMAT_TRANSFORM_FORWARD, 0, ccv_nnc_cmd_auto
, 0), inputs, input_size, aliases, input_size, "concat");
430 }
431}
432
433static ccv_cnnp_model_t* _ccv_cnnp_concat_copy(const ccv_cnnp_model_t* const self, void* const context);
434
435static const ccv_cnnp_model_vtab_t ccv_cnnp_concat_isa = {
436 .build = _ccv_cnnp_concat_build,
437 .copy = _ccv_cnnp_concat_copy,
438};
439
440ccv_cnnp_model_t* ccv_cnnp_concat(const int axis, const char* const name)
441{
442 ccv_cnnp_model_concat_t* const model_concat = (ccv_cnnp_model_concat_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_concat_t));
443 model_concat->super.isa = &ccv_cnnp_concat_isa;
444 model_concat->super.input_size = 0;
445 model_concat->super.outputs = &model_concat->output;
446 model_concat->super.output_size = 1;
447 model_concat->axis = axis;
448 ccv_cnnp_model_copy_name(&model_concat->super, name);
449 return (ccv_cnnp_model_t*)model_concat;
450}
451
452static ccv_cnnp_model_t* _ccv_cnnp_concat_copy(const ccv_cnnp_model_t* const super, void* const context)
453{
454 const ccv_cnnp_model_concat_t* const self = (const ccv_cnnp_model_concat_t*)super;
455 return ccv_cnnp_concat(self->axis, self->super.name);
456}
457
458typedef struct {
459 ccv_cnnp_model_t super;
460 int axis;
461 ccv_nnc_tensor_symbol_t outputs[1];
462} ccv_cnnp_model_chunk_t;
463
464static void _ccv_cnnp_chunk_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
465{
466 const ccv_cnnp_model_concat_t* const self = (const ccv_cnnp_model_concat_t*)super;
467 PRINT(CCV_CLI_VERBOSE, "[cnnp_chunk_build] 1. axis: %d\n", self->axis)do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_chunk_build] 1. axis: %d\n", self->axis);
fflush(stdout); } } while (0);
468 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 468, __extension__ __PRETTY_FUNCTION__); }));
469 const ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
470 if (CCV_CLI_OUTPUT_LEVEL_IS(CCV_CLI_VERBOSE)(CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
471 {
472 PRINT(CCV_CLI_VERBOSE, "[cnnp_chunk_build] 2. input: (%d", input_params.dim[0])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_chunk_build] 2. input: (%d", input_params.dim
[0]); fflush(stdout); } } while (0);
473 int i;
474 for (i = 1; i < CCV_NNC_MAX_DIM_ALLOC(12) && input_params.dim[i] > 0; i++)
475 PRINT(CCV_CLI_VERBOSE, ", %d", input_params.dim[i])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf(", %d", input_params.dim[i]); fflush(stdout); } } while
(0);
476 PRINT(CCV_CLI_VERBOSE, ")\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf(")\n"); fflush(stdout); } } while (0);
477 }
478 ccv_nnc_tensor_param_t output_params = input_params;
479 int i;
480 const int nd = ccv_nnc_tensor_nd(output_params.dim);
481 const int axis = self->axis;
482 assert(axis < nd)((void) sizeof ((axis < nd) ? 1 : 0), __extension__ ({ if (
axis < nd) ; else __assert_fail ("axis < nd", "ccv_cnnp_model_addons.c"
, 482, __extension__ __PRETTY_FUNCTION__); }));
483 const int n = self->super.output_size;
484 assert(n == output_size)((void) sizeof ((n == output_size) ? 1 : 0), __extension__ ({
if (n == output_size) ; else __assert_fail ("n == output_size"
, "ccv_cnnp_model_addons.c", 484, __extension__ __PRETTY_FUNCTION__
); }));
485 assert(output_params.dim[axis] % n == 0)((void) sizeof ((output_params.dim[axis] % n == 0) ? 1 : 0), __extension__
({ if (output_params.dim[axis] % n == 0) ; else __assert_fail
("output_params.dim[axis] % n == 0", "ccv_cnnp_model_addons.c"
, 485, __extension__ __PRETTY_FUNCTION__); }));
486 output_params.dim[axis] = output_params.dim[axis] / n;
487 int ofs[CCV_NNC_MAX_DIM_ALLOC(12)] = {};
488 int stride[CCV_NNC_MAX_DIM_ALLOC(12)] = {};
489 ccv_nnc_tensor_get_stride(input_params.dim, stride);
490 ccv_nnc_tensor_symbol_t to = ccv_nnc_tensor_symbol_alias_to(graph, inputs[0]);
491 if (to.d == CCV_NNC_NO_TENSOR_SYMBOL) // If we are not reshape an alias, it is straightforward.
492 {
493 for (i = 0; i < output_size; i++)
494 {
495 outputs[i] = ccv_nnc_tensor_symbol_alias_new(graph, inputs[0], ofs, stride, output_params, 0);
496 ofs[axis] += output_params.dim[axis];
497 }
498 } else {
499 // Otherwise, we need to check if it is permute. For permute, we cannot do alias directly.
500 // We need to first materialize the permute and then run reshape on top of it, otherwise it will be wrong.
501 int old_stride[CCV_NNC_MAX_DIM_ALLOC(12)];
502 ccv_nnc_tensor_symbol_alias_params(graph, inputs[0], 0, old_stride);
503 // We identify permute by checking if the stride is not in descending order.
504 // This also covered "permute" through reshape, rather than using ccv_cnnp_permute directly.
505 int i, no_permute = 1;
506 for (i = 1; no_permute && i < nd; i++)
507 if (old_stride[i - 1] < old_stride[i])
508 no_permute = 0;
509 if (no_permute)
510 { // Just straightforward reshape if there is no no permute.
511 for (i = 0; i < output_size; i++)
512 {
513 outputs[i] = ccv_nnc_tensor_symbol_alias_new(graph, inputs[0], ofs, old_stride, output_params, 0);
514 ofs[axis] += output_params.dim[axis];
515 }
516 } else {
517 // Otherwise, we first do format transform to plain tensor and then do reshape.
518 ccv_nnc_tensor_symbol_t permuted = ccv_nnc_tensor_symbol_new(graph, input_params, 0);
519 ccv_nnc_graph_exec_symbol_new(graph, CMD_FORMAT_TRANSFORM_FORWARD()ccv_nnc_cmd(CCV_NNC_FORMAT_TRANSFORM_FORWARD, 0, ccv_nnc_cmd_auto
, 0), TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(permuted)(const ccv_nnc_tensor_symbol_t []){permuted}, (1 +1 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "reshape");
520 for (i = 0; i < output_size; i++)
521 {
522 outputs[i] = ccv_nnc_tensor_symbol_alias_new(graph, permuted, ofs, stride, output_params, 0);
523 ofs[axis] += output_params.dim[axis];
524 }
525 }
526 }
527}
528
529static ccv_cnnp_model_t* _ccv_cnnp_chunk_copy(const ccv_cnnp_model_t* const self, void* const context);
530
531static const ccv_cnnp_model_vtab_t ccv_cnnp_chunk_isa = {
532 .build = _ccv_cnnp_chunk_build,
533 .copy = _ccv_cnnp_chunk_copy,
534};
535
536ccv_cnnp_model_t* ccv_cnnp_chunk(const int n, const int axis, const char* const name)
537{
538 assert(n >= 1)((void) sizeof ((n >= 1) ? 1 : 0), __extension__ ({ if (n >=
1) ; else __assert_fail ("n >= 1", "ccv_cnnp_model_addons.c"
, 538, __extension__ __PRETTY_FUNCTION__); }));
539 ccv_cnnp_model_chunk_t* const model_chunk = (ccv_cnnp_model_chunk_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_chunk_t) + sizeof(ccv_nnc_tensor_symbol_t) * (n - 1));
540 model_chunk->super.isa = &ccv_cnnp_chunk_isa;
541 model_chunk->super.input_size = 1;
542 model_chunk->super.outputs = model_chunk->outputs;
543 model_chunk->super.output_size = n;
544 model_chunk->axis = axis;
545 ccv_cnnp_model_copy_name(&model_chunk->super, name);
546 return (ccv_cnnp_model_t*)model_chunk;
547}
548
549static ccv_cnnp_model_t* _ccv_cnnp_chunk_copy(const ccv_cnnp_model_t* const super, void* const context)
550{
551 const ccv_cnnp_model_chunk_t* const self = (const ccv_cnnp_model_chunk_t*)super;
552 return ccv_cnnp_chunk(self->super.output_size, self->axis, self->super.name);
553}
554
555typedef struct {
556 ccv_cnnp_model_t super;
557 ccv_nnc_tensor_symbol_t output;
558 int format;
559 int dim[CCV_NNC_MAX_DIM_ALLOC(12)];
560 int ofs[CCV_NNC_MAX_DIM_ALLOC(12)];
561 int stride[CCV_NNC_MAX_DIM_ALLOC(12)];
562} ccv_cnnp_model_reshape_t;
563
564static void _ccv_cnnp_reshape_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
565{
566 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 566, __extension__ __PRETTY_FUNCTION__); }));
567 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 567, __extension__ __PRETTY_FUNCTION__
); }));
568 ccv_cnnp_model_reshape_t* const self = (ccv_cnnp_model_reshape_t*)super;
569 if (CCV_CLI_OUTPUT_LEVEL_IS(CCV_CLI_VERBOSE)(CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
570 {
571 PRINT(CCV_CLI_VERBOSE, "[cnnp_reshape_build] 1. dim: (%d", self->dim[0])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_reshape_build] 1. dim: (%d", self->dim[0]
); fflush(stdout); } } while (0);
572 int i;
573 for (i = 1; i < CCV_NNC_MAX_DIM_ALLOC(12) && self->dim[i] > 0; i++)
574 PRINT(CCV_CLI_VERBOSE, ", %d", self->dim[i])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf(", %d", self->dim[i]); fflush(stdout); } } while
(0);
575 const int count = i;
576 PRINT(CCV_CLI_VERBOSE, "), ofs: (%d", self->ofs[0])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("), ofs: (%d", self->ofs[0]); fflush(stdout); } }
while (0);
577 for (i = 1; i < count; i++)
578 PRINT(CCV_CLI_VERBOSE, ", %d", self->ofs[i])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf(", %d", self->ofs[i]); fflush(stdout); } } while
(0);
579 PRINT(CCV_CLI_VERBOSE, "), stride: (%d", self->stride[0])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("), stride: (%d", self->stride[0]); fflush(stdout
); } } while (0);
580 for (i = 1; i < count; i++)
581 PRINT(CCV_CLI_VERBOSE, ", %d", self->stride[i])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf(", %d", self->stride[i]); fflush(stdout); } } while
(0);
582 PRINT(CCV_CLI_VERBOSE, ")\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf(")\n"); fflush(stdout); } } while (0);
583 }
584 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
585 if (params.dim[0] == 0)
586 {
587 int ofs[CCV_NNC_MAX_DIM_ALLOC(12)] = {};
588 int stride[CCV_NNC_MAX_DIM_ALLOC(12)] = {};
589 if (ccv_nnc_tensor_symbol_alias_params(graph, inputs[0], ofs, stride) < 0)
590 ccv_nnc_tensor_get_stride(params.dim, stride);
591 outputs[0] = ccv_nnc_tensor_symbol_alias_new(graph, inputs[0], ofs, stride, params, 0);
592 return;
593 }
594 int dim[CCV_NNC_MAX_DIM_ALLOC(12)];
595 memcpy(dim, self->dim, sizeof(dim));
596 int i, auto_idx = -1;
597 size_t known = 1;
598 const size_t tensor_count = ccv_nnc_tensor_count(params);
599 for (i = 0; i < CCV_NNC_MAX_DIM_ALLOC(12) && dim[i]; i++)
600 if (dim[i] == -1)
601 auto_idx = i;
602 else
603 known *= dim[i];
604 if (auto_idx >= 0)
605 {
606 assert(known > 0 && tensor_count % known == 0)((void) sizeof ((known > 0 && tensor_count % known
== 0) ? 1 : 0), __extension__ ({ if (known > 0 &&
tensor_count % known == 0) ; else __assert_fail ("known > 0 && tensor_count % known == 0"
, "ccv_cnnp_model_addons.c", 606, __extension__ __PRETTY_FUNCTION__
); }));
607 dim[auto_idx] = tensor_count / known;
608 }
609 if (CCV_CLI_OUTPUT_LEVEL_IS(CCV_CLI_VERBOSE)(CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
610 {
611 PRINT(CCV_CLI_VERBOSE, "[cnnp_reshape_build] 2. input: (%d", params.dim[0])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_reshape_build] 2. input: (%d", params.dim[0]
); fflush(stdout); } } while (0);
612 int i;
613 for (i = 1; i < CCV_NNC_MAX_DIM_ALLOC(12) && params.dim[i] > 0; i++)
614 PRINT(CCV_CLI_VERBOSE, ", %d", params.dim[i])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf(", %d", params.dim[i]); fflush(stdout); } } while (
0);
615 PRINT(CCV_CLI_VERBOSE, ")\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf(")\n"); fflush(stdout); } } while (0);
616 }
617 assert(ccv_nnc_dimension_count(dim) <= ccv_nnc_tensor_count(params))((void) sizeof ((ccv_nnc_dimension_count(dim) <= ccv_nnc_tensor_count
(params)) ? 1 : 0), __extension__ ({ if (ccv_nnc_dimension_count
(dim) <= ccv_nnc_tensor_count(params)) ; else __assert_fail
("ccv_nnc_dimension_count(dim) <= ccv_nnc_tensor_count(params)"
, "ccv_cnnp_model_addons.c", 617, __extension__ __PRETTY_FUNCTION__
); }));
618 ccv_nnc_tensor_symbol_t to = ccv_nnc_tensor_symbol_alias_to(graph, inputs[0]);
619 int stride_from_dim[CCV_NNC_MAX_DIM_ALLOC(12)];
620 if (to.d == CCV_NNC_NO_TENSOR_SYMBOL) // If we are not reshape an alias, it is straightforward.
621 {
622 memcpy(params.dim, dim, sizeof(params.dim));
623 int* stride;
624 if (self->stride[0] == 0)
625 {
626 ccv_nnc_tensor_get_stride(dim, stride_from_dim);
627 stride = stride_from_dim;
628 } else
629 stride = self->stride;
630 if (self->format > 0)
631 params.format = self->format;
632 outputs[0] = ccv_nnc_tensor_symbol_alias_new(graph, inputs[0], self->ofs, stride, params, 0);
633 } else {
634 // Otherwise, we need to check if it is permute. For permute, we cannot do alias directly.
635 // We need to first materialize the permute and then run reshape on top of it, otherwise it will be wrong.
636 int old_ofs[CCV_NNC_MAX_DIM_ALLOC(12)];
637 int old_stride[CCV_NNC_MAX_DIM_ALLOC(12)];
638 ccv_nnc_tensor_symbol_alias_params(graph, inputs[0], old_ofs, old_stride);
639 // We identify permute by checking if the stride is not in descending order.
640 // This also covered "permute" through reshape, rather than using ccv_cnnp_permute directly.
641 const int nd = ccv_nnc_tensor_nd(params.dim);
642 const int new_nd = ccv_nnc_tensor_nd(dim);
643 int i, no_permute = 1;
644 // If the new dim has different nd, or we actually have a stride, we need to check if it is no permute or not.
645 if (new_nd != nd || (self->stride[0] != 0 && memcmp(self->stride, old_stride, sizeof(self->stride))))
646 for (i = 1; no_permute && i < nd; i++)
647 if (old_stride[i - 1] < old_stride[i])
648 no_permute = 0;
649 if (no_permute)
650 { // Just straightforward reshape if there is no no permute.
651 memcpy(params.dim, dim, sizeof(params.dim));
652 int* stride;
653 if (self->stride[0] == 0)
654 {
655 if (new_nd != nd) // Cannot use old stride.
656 {
657 ccv_nnc_tensor_get_stride(dim, stride_from_dim);
658 stride = stride_from_dim;
659 } else
660 stride = old_stride;
661 } else
662 stride = self->stride;
663 if (self->format > 0)
664 params.format = self->format;
665 // tensor_symbol_alias_new flattens an alias to its root tensor. Preserve the
666 // parent view's starting position, expressing it in the new strides if needed.
667 const int* ofs = self->ofs;
Value stored to 'ofs' during its initialization is never read
668 int new_ofs[CCV_NNC_MAX_DIM_ALLOC(12)];
669 if (new_nd == nd && memcmp(stride, old_stride, sizeof(old_stride)) == 0)
670 {
671 for (i = 0; i < CCV_NNC_MAX_DIM_ALLOC(12); i++)
672 new_ofs[i] = old_ofs[i] + self->ofs[i];
673 ofs = new_ofs;
674 } else {
675 size_t offset = 0;
676 for (i = 0; i < nd; i++)
677 offset += (size_t)old_ofs[i] * old_stride[i];
678 memcpy(new_ofs, self->ofs, sizeof(new_ofs));
679 for (i = 0; i < new_nd; i++)
680 {
681 assert(stride[i] > 0)((void) sizeof ((stride[i] > 0) ? 1 : 0), __extension__ ({
if (stride[i] > 0) ; else __assert_fail ("stride[i] > 0"
, "ccv_cnnp_model_addons.c", 681, __extension__ __PRETTY_FUNCTION__
); }));
682 assert(offset / stride[i] <= INT_MAX - self->ofs[i])((void) sizeof ((offset / stride[i] <= 2147483647 - self->
ofs[i]) ? 1 : 0), __extension__ ({ if (offset / stride[i] <=
2147483647 - self->ofs[i]) ; else __assert_fail ("offset / stride[i] <= INT_MAX - self->ofs[i]"
, "ccv_cnnp_model_addons.c", 682, __extension__ __PRETTY_FUNCTION__
); }));
683 new_ofs[i] += (int)(offset / stride[i]);
684 offset %= stride[i];
685 }
686 // The caller must copy explicitly if the new strides cannot represent the offset.
687 assert(offset == 0)((void) sizeof ((offset == 0) ? 1 : 0), __extension__ ({ if (
offset == 0) ; else __assert_fail ("offset == 0", "ccv_cnnp_model_addons.c"
, 687, __extension__ __PRETTY_FUNCTION__); }));
688 ofs = new_ofs;
689 }
690 outputs[0] = ccv_nnc_tensor_symbol_alias_new(graph, inputs[0], ofs, stride, params, 0);
691 } else {
692 // Otherwise, we first do format transform to plain tensor and then do reshape.
693 ccv_nnc_tensor_symbol_t permuted = ccv_nnc_tensor_symbol_new(graph, params, 0);
694 ccv_nnc_graph_exec_symbol_new(graph, CMD_FORMAT_TRANSFORM_FORWARD()ccv_nnc_cmd(CCV_NNC_FORMAT_TRANSFORM_FORWARD, 0, ccv_nnc_cmd_auto
, 0), TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(permuted)(const ccv_nnc_tensor_symbol_t []){permuted}, (1 +1 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "reshape");
695 memcpy(params.dim, dim, sizeof(params.dim));
696 int* stride;
697 if (self->stride[0] == 0)
698 {
699 ccv_nnc_tensor_get_stride(dim, stride_from_dim);
700 stride = stride_from_dim;
701 } else
702 stride = self->stride;
703 if (self->format > 0)
704 params.format = self->format;
705 // And then we create alias against the permuted one.
706 outputs[0] = ccv_nnc_tensor_symbol_alias_new(graph, permuted, self->ofs, stride, params, 0);
707 }
708 }
709}
710
711static ccv_cnnp_model_t* _ccv_cnnp_reshape_copy(const ccv_cnnp_model_t* const super, void* const context);
712
713static const ccv_cnnp_model_vtab_t ccv_cnnp_reshape_isa = {
714 .build = _ccv_cnnp_reshape_build,
715 .copy = _ccv_cnnp_reshape_copy,
716};
717
718ccv_cnnp_model_t* ccv_cnnp_reshape(const int format, const int dim[CCV_NNC_MAX_DIM_ALLOC(12)], const int ofs[CCV_NNC_MAX_DIM_ALLOC(12)], const int stride[CCV_NNC_MAX_DIM_ALLOC(12)], const char* const name)
719{
720 ccv_cnnp_model_reshape_t* const model_reshape = (ccv_cnnp_model_reshape_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_reshape_t));
721 model_reshape->super.isa = &ccv_cnnp_reshape_isa;
722 model_reshape->super.input_size = 1;
723 model_reshape->super.outputs = &model_reshape->output;
724 model_reshape->super.output_size = 1;
725 ccv_cnnp_model_copy_name(&model_reshape->super, name);
726 model_reshape->format = format;
727 memcpy(model_reshape->dim, dim, sizeof(model_reshape->dim));
728 memcpy(model_reshape->ofs, ofs, sizeof(model_reshape->ofs));
729 if (stride[0] != 0)
730 memcpy(model_reshape->stride, stride, sizeof(model_reshape->stride));
731 return (ccv_cnnp_model_t*)model_reshape;
732}
733
734static ccv_cnnp_model_t* _ccv_cnnp_reshape_copy(const ccv_cnnp_model_t* const super, void* const context)
735{
736 const ccv_cnnp_model_reshape_t* const self = (const ccv_cnnp_model_reshape_t*)super;
737 return ccv_cnnp_reshape(self->format, self->dim, self->ofs, self->stride, self->super.name);
738}
739
740typedef struct {
741 ccv_cnnp_model_t super;
742 ccv_nnc_tensor_symbol_t output;
743 int type;
744 int begin[CCV_NNC_MAX_DIM_ALLOC(12)];
745 int end[CCV_NNC_MAX_DIM_ALLOC(12)];
746} ccv_cnnp_model_pad_t;
747
748static void _ccv_cnnp_pad_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
749{
750 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 750, __extension__ __PRETTY_FUNCTION__); }));
751 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 751, __extension__ __PRETTY_FUNCTION__
); }));
752 ccv_cnnp_model_pad_t* const self = (ccv_cnnp_model_pad_t*)super;
753 PRINT(CCV_CLI_VERBOSE, "[cnnp_pad_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_pad_build] -\n"); fflush(stdout); } } while (
0);
754 const ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
755 const int nd = ccv_nnc_tensor_nd(input_params.dim);
756 ccv_nnc_tensor_param_t params = input_params;
757 int i;
758 for (i = 0 ; i < nd; i++)
759 params.dim[i] += self->begin[i] + self->end[i];
760 const ccv_nnc_tensor_symbol_t padded = ccv_nnc_tensor_symbol_new(graph, params, 0);
761 ccv_nnc_cmd_t pad = CMD_PAD_FORWARD(self->type, (), ())ccv_nnc_cmd(CCV_NNC_PAD_FORWARD, 0, ((ccv_nnc_cmd_param_t){.size
={.dim={}},.pad={.type=self->type,.end={}}}), 0);
762 memcpy(pad.info.size.dim, self->begin, sizeof(pad.info.size.dim));
763 memcpy(pad.info.pad.end, self->end, sizeof(pad.info.pad.end));
764 ccv_nnc_graph_exec_symbol_new(graph, pad, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(padded)(const ccv_nnc_tensor_symbol_t []){padded}, (1 +1 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "pad");
765 outputs[0] = padded;
766}
767
768static ccv_cnnp_model_t* _ccv_cnnp_pad_copy(const ccv_cnnp_model_t* const super, void* const context);
769
770static const ccv_cnnp_model_vtab_t ccv_cnnp_pad_isa = {
771 .build = _ccv_cnnp_pad_build,
772 .copy = _ccv_cnnp_pad_copy,
773};
774
775ccv_cnnp_model_t* ccv_cnnp_pad(const int type, const int begin[CCV_NNC_MAX_DIM_ALLOC(12)], const int end[CCV_NNC_MAX_DIM_ALLOC(12)], const char* const name)
776{
777 ccv_cnnp_model_pad_t* const model_pad = (ccv_cnnp_model_pad_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_pad_t));
778 model_pad->super.isa = &ccv_cnnp_pad_isa;
779 model_pad->super.input_size = 1;
780 model_pad->super.outputs = &model_pad->output;
781 model_pad->super.output_size = 1;
782 ccv_cnnp_model_copy_name(&model_pad->super, name);
783 model_pad->type = type;
784 memcpy(model_pad->begin, begin, sizeof(model_pad->begin));
785 memcpy(model_pad->end, end, sizeof(model_pad->end));
786 return (ccv_cnnp_model_t*)model_pad;
787}
788
789static ccv_cnnp_model_t* _ccv_cnnp_pad_copy(const ccv_cnnp_model_t* const super, void* const context)
790{
791 const ccv_cnnp_model_pad_t* const self = (const ccv_cnnp_model_pad_t*)super;
792 return ccv_cnnp_pad(self->type, self->begin, self->end, self->super.name);
793}
794
795typedef struct {
796 ccv_cnnp_model_t super;
797 ccv_nnc_tensor_symbol_t output;
798} ccv_cnnp_model_identity_t;
799
800static void _ccv_cnnp_identity_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
801{
802 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 802, __extension__ __PRETTY_FUNCTION__); }));
803 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 803, __extension__ __PRETTY_FUNCTION__
); }));
804 PRINT(CCV_CLI_VERBOSE, "[cnnp_identity_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_identity_build] -\n"); fflush(stdout); } } while
(0);
805 outputs[0] = inputs[0];
806}
807
808static ccv_cnnp_model_t* _ccv_cnnp_identity_copy(const ccv_cnnp_model_t* const super, void* const context);
809
810static const ccv_cnnp_model_vtab_t ccv_cnnp_identity_isa = {
811 .build = _ccv_cnnp_identity_build,
812 .copy = _ccv_cnnp_identity_copy,
813};
814
815ccv_cnnp_model_t* ccv_cnnp_identity(const char* const name)
816{
817 ccv_cnnp_model_identity_t* const model_identity = (ccv_cnnp_model_identity_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_identity_t));
818 model_identity->super.isa = &ccv_cnnp_identity_isa;
819 model_identity->super.input_size = 1;
820 model_identity->super.outputs = &model_identity->output;
821 model_identity->super.output_size = 1;
822 ccv_cnnp_model_copy_name(&model_identity->super, name);
823 return (ccv_cnnp_model_t*)model_identity;
824}
825
826static ccv_cnnp_model_t* _ccv_cnnp_identity_copy(const ccv_cnnp_model_t* const super, void* const context)
827{
828 const ccv_cnnp_model_identity_t* const self = (const ccv_cnnp_model_identity_t*)super;
829 return ccv_cnnp_identity(self->super.name);
830}
831
832typedef struct {
833 ccv_cnnp_model_t super;
834 ccv_nnc_tensor_symbol_t output;
835 int index[CCV_NNC_MAX_DIM_ALLOC(12)];
836} ccv_cnnp_model_permute_t;
837
838static void _ccv_cnnp_permute_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
839{
840 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 840, __extension__ __PRETTY_FUNCTION__); }));
841 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 841, __extension__ __PRETTY_FUNCTION__
); }));
842 ccv_cnnp_model_permute_t* const self = (ccv_cnnp_model_permute_t*)super;
843 PRINT(CCV_CLI_VERBOSE, "[cnnp_permute_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_permute_build] -\n"); fflush(stdout); } } while
(0);
844 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
845 ccv_nnc_tensor_symbol_t to = ccv_nnc_tensor_symbol_alias_to(graph, inputs[0]);
846 const int nd = ccv_nnc_tensor_nd(params.dim);
847 int input_dim[CCV_NNC_MAX_DIM_ALLOC(12)];
848 memcpy(input_dim, params.dim, sizeof(params.dim));
849 int input_stride[CCV_NNC_MAX_DIM_ALLOC(12)] = {};
850 int output_stride[CCV_NNC_MAX_DIM_ALLOC(12)] = {};
851 if (to.d == CCV_NNC_NO_TENSOR_SYMBOL) // If it is not an alias. Find stride and permute.
852 {
853 ccv_nnc_tensor_get_stride(input_dim, input_stride);
854 int i;
855 for (i = 0; i < nd; i++)
856 {
857 const int idx = self->index[i];
858 assert(idx >= 0 && idx < nd)((void) sizeof ((idx >= 0 && idx < nd) ? 1 : 0)
, __extension__ ({ if (idx >= 0 && idx < nd) ; else
__assert_fail ("idx >= 0 && idx < nd", "ccv_cnnp_model_addons.c"
, 858, __extension__ __PRETTY_FUNCTION__); }));
859 params.dim[i] = input_dim[idx];
860 output_stride[i] = input_stride[idx];
861 }
862 outputs[0] = ccv_nnc_tensor_symbol_alias_new(graph, inputs[0], ccv_nnc_no_ofs, output_stride, params, 0);
863 } else {
864 // if it is an alias, we can get the stride from it and use that.
865 int input_ofs[CCV_NNC_MAX_DIM_ALLOC(12)];
866 ccv_nnc_tensor_symbol_alias_params(graph, inputs[0], input_ofs, input_stride);
867 assert(input_stride[0] != 0)((void) sizeof ((input_stride[0] != 0) ? 1 : 0), __extension__
({ if (input_stride[0] != 0) ; else __assert_fail ("input_stride[0] != 0"
, "ccv_cnnp_model_addons.c", 867, __extension__ __PRETTY_FUNCTION__
); }));
868 int output_ofs[CCV_NNC_MAX_DIM_ALLOC(12)] = {};
869 int i;
870 for (i = 0; i < nd; i++)
871 {
872 const int idx = self->index[i];
873 assert(idx >= 0 && idx < nd)((void) sizeof ((idx >= 0 && idx < nd) ? 1 : 0)
, __extension__ ({ if (idx >= 0 && idx < nd) ; else
__assert_fail ("idx >= 0 && idx < nd", "ccv_cnnp_model_addons.c"
, 873, __extension__ __PRETTY_FUNCTION__); }));
874 params.dim[i] = input_dim[idx];
875 output_stride[i] = input_stride[idx];
876 output_ofs[i] = input_ofs[idx];
877 }
878 outputs[0] = ccv_nnc_tensor_symbol_alias_new(graph, inputs[0], output_ofs, output_stride, params, 0);
879 }
880}
881
882static ccv_cnnp_model_t* _ccv_cnnp_permute_copy(const ccv_cnnp_model_t* const super, void* const context);
883
884static const ccv_cnnp_model_vtab_t ccv_cnnp_permute_isa = {
885 .build = _ccv_cnnp_permute_build,
886 .copy = _ccv_cnnp_permute_copy,
887};
888
889ccv_cnnp_model_t* ccv_cnnp_permute(const int index[CCV_NNC_MAX_DIM_ALLOC(12)], const char* const name)
890{
891 ccv_cnnp_model_permute_t* const model_permute = (ccv_cnnp_model_permute_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_permute_t));
892 model_permute->super.isa = &ccv_cnnp_permute_isa;
893 model_permute->super.input_size = 1;
894 model_permute->super.outputs = &model_permute->output;
895 model_permute->super.output_size = 1;
896 ccv_cnnp_model_copy_name(&model_permute->super, name);
897 memcpy(model_permute->index, index, sizeof(model_permute->index));
898 return (ccv_cnnp_model_t*)model_permute;
899}
900
901static ccv_cnnp_model_t* _ccv_cnnp_permute_copy(const ccv_cnnp_model_t* const super, void* const context)
902{
903 const ccv_cnnp_model_permute_t* const self = (const ccv_cnnp_model_permute_t*)super;
904 return ccv_cnnp_permute(self->index, self->super.name);
905}
906
907typedef struct {
908 ccv_cnnp_model_t super;
909 int index;
910 ccv_nnc_tensor_symbol_t output;
911} ccv_cnnp_model_extract_t;
912
913static void _ccv_cnnp_extract_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
914{
915 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 915, __extension__ __PRETTY_FUNCTION__
); }));
916 ccv_cnnp_model_extract_t* const self = (ccv_cnnp_model_extract_t*)super;
917 PRINT(CCV_CLI_VERBOSE, "[cnnp_extract_build] index: %d\n", self->index)do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_extract_build] index: %d\n", self->index)
; fflush(stdout); } } while (0);
918 outputs[0] = inputs[self->index];
919}
920
921static ccv_cnnp_model_t* _ccv_cnnp_extract_copy(const ccv_cnnp_model_t* const self, void* const context);
922
923static const ccv_cnnp_model_vtab_t ccv_cnnp_extract_isa = {
924 .build = _ccv_cnnp_extract_build,
925 .copy = _ccv_cnnp_extract_copy,
926};
927
928ccv_cnnp_model_t* ccv_cnnp_extract(const int index, const char* const name)
929{
930 ccv_cnnp_model_extract_t* const model_extract = (ccv_cnnp_model_extract_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_extract_t));
931 model_extract->index = index;
932 model_extract->super.isa = &ccv_cnnp_extract_isa;
933 model_extract->super.input_size = 0;
934 model_extract->super.outputs = &model_extract->output;
935 model_extract->super.output_size = 1;
936 ccv_cnnp_model_copy_name(&model_extract->super, name);
937 return (ccv_cnnp_model_t*)model_extract;
938}
939
940static ccv_cnnp_model_t* _ccv_cnnp_extract_copy(const ccv_cnnp_model_t* const super, void* const context)
941{
942 ccv_cnnp_model_extract_t* const self = (ccv_cnnp_model_extract_t*)super;
943 return ccv_cnnp_extract(self->index, self->super.name);
944}
945
946typedef struct {
947 ccv_cnnp_model_t super;
948 ccv_nnc_tensor_symbol_t output;
949} ccv_cnnp_model_flatten_t;
950
951static void _ccv_cnnp_flatten_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
952{
953 PRINT(CCV_CLI_VERBOSE, "[cnnp_flatten_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_flatten_build] -\n"); fflush(stdout); } } while
(0);
954 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 954, __extension__ __PRETTY_FUNCTION__); }));
955 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 955, __extension__ __PRETTY_FUNCTION__
); }));
956 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
957 ccv_nnc_tensor_param_t output_params = params;
958 memset(output_params.dim, 0, sizeof(output_params.dim));
959 output_params.dim[0] = ccv_nnc_tensor_get_n(params);
960 assert(output_params.dim[0] > 0)((void) sizeof ((output_params.dim[0] > 0) ? 1 : 0), __extension__
({ if (output_params.dim[0] > 0) ; else __assert_fail ("output_params.dim[0] > 0"
, "ccv_cnnp_model_addons.c", 960, __extension__ __PRETTY_FUNCTION__
); }));
961 output_params.dim[1] = ccv_nnc_tensor_count(params) / output_params.dim[0];
962 int stride[CCV_NNC_MAX_DIM_ALLOC(12)] = {};
963 ccv_nnc_tensor_get_stride(output_params.dim, stride);
964 outputs[0] = ccv_nnc_tensor_symbol_alias_new(graph, inputs[0], DIM_ALLOC()(int [(12)]){}, stride, output_params, 0);
965}
966
967static ccv_cnnp_model_t* _ccv_cnnp_flatten_copy(const ccv_cnnp_model_t* const self, void* const context);
968
969static const ccv_cnnp_model_vtab_t ccv_cnnp_flatten_isa = {
970 .build = _ccv_cnnp_flatten_build,
971 .copy = _ccv_cnnp_flatten_copy,
972};
973
974ccv_cnnp_model_t* ccv_cnnp_flatten(const char* const name)
975{
976 ccv_cnnp_model_flatten_t* const model_flatten = (ccv_cnnp_model_flatten_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_flatten_t));
977 model_flatten->super.isa = &ccv_cnnp_flatten_isa;
978 model_flatten->super.input_size = 1;
979 model_flatten->super.outputs = &model_flatten->output;
980 model_flatten->super.output_size = 1;
981 ccv_cnnp_model_copy_name(&model_flatten->super, name);
982 return (ccv_cnnp_model_t*)model_flatten;
983}
984
985static ccv_cnnp_model_t* _ccv_cnnp_flatten_copy(const ccv_cnnp_model_t* const self, void* const context)
986{
987 return ccv_cnnp_flatten(self->name);
988}
989
990// MARK - Batch Norm Layer
991
992typedef struct {
993 ccv_cnnp_model_t super;
994 ccv_nnc_tensor_symbol_t output;
995 ccv_nnc_tensor_symbol_t bias;
996 ccv_nnc_tensor_symbol_t scale;
997 ccv_nnc_graph_exec_symbol_t batch_norm;
998 ccv_nnc_cmd_param_t params;
999 ccv_array_t* zero_inits;
1000 ccv_array_t* retainables;
1001} ccv_cnnp_model_batch_norm_t;
1002
1003static void _ccv_cnnp_batch_norm_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
1004{
1005 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 1005, __extension__ __PRETTY_FUNCTION__); }));
1006 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 1006, __extension__ __PRETTY_FUNCTION__
); }));
1007 ccv_cnnp_model_batch_norm_t* const self = (ccv_cnnp_model_batch_norm_t*)super;
1008 PRINT(CCV_CLI_VERBOSE, "[cnnp_batch_norm_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_batch_norm_build] -\n"); fflush(stdout); } }
while (0);
1009 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1010 const int nd = ccv_nnc_tensor_nd(params.dim);
1011 ccv_nnc_tensor_param_t bias_params = params;
1012 memset(bias_params.dim, 0, sizeof(bias_params.dim));
1013 // If the accuracy is not enough, bump it to 32-bit floating point.
1014 if (bias_params.datatype != CCV_32F && bias_params.datatype != CCV_64F)
1015 bias_params.datatype = CCV_32F;
1016 bias_params.dim[0] = nd > 1 ? ccv_nnc_tensor_get_c(params) : params.dim[0];
1017 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, params, 0);
1018 // Both scale and bias are shared between if this model is reused.
1019 if (!self->scale.graph)
1020 self->scale = ccv_nnc_tensor_symbol_new(graph, bias_params, "scale");
1021 if (!self->bias.graph)
1022 self->bias = ccv_nnc_tensor_symbol_new(graph, bias_params, "bias");
1023 const ccv_nnc_tensor_symbol_t scale = ccv_cnnp_model_get_symbol(super, self->scale);
1024 const ccv_nnc_tensor_symbol_t bias = ccv_cnnp_model_get_symbol(super, self->bias);
1025 const ccv_nnc_tensor_symbol_t mean = ccv_nnc_tensor_symbol_new(graph, bias_params, "mean");
1026 const ccv_nnc_tensor_symbol_t var = ccv_nnc_tensor_symbol_new(graph, bias_params, "var");
1027 // Otherwise, notice mean, var, saved_mean, saved_inv_std are not reused.
1028 if (!self->zero_inits)
1029 self->zero_inits = ccv_array_new(sizeof(ccv_nnc_tensor_symbol_t), 0, 0);
1030 ccv_array_push(self->zero_inits, &mean);
1031 ccv_array_push(self->zero_inits, &var);
1032 const ccv_nnc_tensor_symbol_t out_mean = ccv_nnc_tensor_symbol_new(graph, bias_params, "out_mean");
1033 const ccv_nnc_tensor_symbol_t out_var = ccv_nnc_tensor_symbol_new(graph, bias_params, "out_var");
1034 if (!self->retainables)
1035 self->retainables = ccv_array_new(sizeof(ccv_nnc_tensor_symbol_t), 0, 0);
1036 ccv_array_push(self->retainables, &out_mean);
1037 ccv_array_push(self->retainables, &out_var);
1038 const ccv_nnc_tensor_symbol_t saved_mean = ccv_nnc_tensor_symbol_new(graph, bias_params, "saved_mean");
1039 const ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(graph, bias_params, "saved_inv_std");
1040 const int hw = ccv_nnc_tensor_hw(params, ccv_nnc_tensor_nd(params.dim), CCV_NNC_MAX_DIM(2));
1041 ccv_nnc_cmd_param_t batch_norm = self->params;
1042 batch_norm.bnorm.count = hw >= 0 ? CCV_NNC_MAX_DIM(2) + 1 : 1;
1043 int i;
1044 batch_norm.bnorm.axis[0] = (params.format == CCV_TENSOR_FORMAT_CHWN) ? 3 : 0;
1045 if (hw >= 0)
1046 for (i = 0; i < CCV_NNC_MAX_DIM(2); i++)
1047 batch_norm.bnorm.axis[i + 1] = i + hw;
1048 self->params = batch_norm;
1049 self->batch_norm = ccv_nnc_graph_exec_symbol_new(graph, ccv_nnc_cmd(CCV_NNC_BATCH_NORM_FORWARD, 0, batch_norm, 0), TENSOR_SYMBOL_LIST(inputs[0], scale, bias, mean, var)(const ccv_nnc_tensor_symbol_t []){inputs[0], scale, bias, mean
, var}, (1 +1 +1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(output, out_mean, out_var, saved_mean, saved_inv_std)(const ccv_nnc_tensor_symbol_t []){output, out_mean, out_var,
saved_mean, saved_inv_std}, (1 +1 +1 +1 +1 +1 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "batch_norm");
1050 outputs[0] = output;
1051}
1052
1053static void _ccv_cnnp_batch_norm_init_states(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_cnnp_state_initializer_f initializer, void* const context)
1054{
1055 ccv_cnnp_model_batch_norm_t* const self = (ccv_cnnp_model_batch_norm_t*)super;
1056 if (self->scale.graph)
1057 initializer(context, CMD_RANDOM_UNIFORM_FORWARD(0, 1)ccv_nnc_cmd(CCV_NNC_RANDOM_UNIFORM_FORWARD, 0, (ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.blas={.a={0, 1}}}, 0), ccv_nnc_no_hint, 0, 0, self->scale);
1058 if (self->bias.graph)
1059 initializer(context, CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, self->bias);
1060 int i;
1061 if (self->zero_inits)
1062 for (i = 0; i < self->zero_inits->rnum; i++)
1063 initializer(context, CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, *(ccv_nnc_tensor_symbol_t*)ccv_array_get(self->zero_inits, i)((void*)(((char*)((self->zero_inits)->data)) + (size_t)
(self->zero_inits)->rsize * (size_t)(i))));
1064}
1065
1066static void _ccv_cnnp_batch_norm_add_to_parameter(ccv_cnnp_model_t* const super, const ccv_cnnp_add_to_array_f add_to_array, void* const parameters, const int is_trainable)
1067{
1068 ccv_cnnp_model_batch_norm_t* const self = (ccv_cnnp_model_batch_norm_t*)super;
1069 if (self->scale.graph)
1070 add_to_array(parameters, self->scale, is_trainable);
1071 if (self->bias.graph)
1072 add_to_array(parameters, self->bias, is_trainable);
1073}
1074
1075static void _ccv_cnnp_batch_norm_add_to_output(ccv_cnnp_model_t* const super, const ccv_cnnp_add_to_array_f add_to_array, void* const outputs)
1076{
1077 ccv_cnnp_model_batch_norm_t* const self = (ccv_cnnp_model_batch_norm_t*)super;
1078 int i;
1079 if (self->retainables)
1080 for (i = 0; i < self->retainables->rnum; i++)
1081 {
1082 const ccv_nnc_tensor_symbol_t symbol = *(ccv_nnc_tensor_symbol_t*)ccv_array_get(self->retainables, i)((void*)(((char*)((self->retainables)->data)) + (size_t
)(self->retainables)->rsize * (size_t)(i)));
1083 add_to_array(outputs, symbol, 0);
1084 }
1085}
1086
1087static void _ccv_cnnp_batch_norm_set_is_test(ccv_cnnp_model_t* const super, const int is_test, const ccv_cnnp_cmd_updater_f updater, void* const context)
1088{
1089 ccv_cnnp_model_batch_norm_t* const self = (ccv_cnnp_model_batch_norm_t*)super;
1090 if (self->batch_norm.graph)
1091 {
1092 self->params.bnorm.is_test = is_test;
1093 updater(context, self->batch_norm, ccv_nnc_cmd(CCV_NNC_BATCH_NORM_FORWARD, 0, self->params, 0), ccv_nnc_no_hint);
1094 }
1095}
1096
1097static void _ccv_cnnp_batch_norm_deinit(ccv_cnnp_model_t* const super)
1098{
1099 ccv_cnnp_model_batch_norm_t* const self = (ccv_cnnp_model_batch_norm_t*)super;
1100 if (self->zero_inits)
1101 ccv_array_free(self->zero_inits);
1102 if (self->retainables)
1103 ccv_array_free(self->retainables);
1104}
1105
1106static ccv_cnnp_model_t* _ccv_cnnp_batch_norm_copy(const ccv_cnnp_model_t* const super, void* const context);
1107
1108static const ccv_cnnp_model_vtab_t ccv_cnnp_batch_norm_isa = {
1109 .build = _ccv_cnnp_batch_norm_build,
1110 .init_states = _ccv_cnnp_batch_norm_init_states,
1111 .add_to_parameter = _ccv_cnnp_batch_norm_add_to_parameter,
1112 .add_to_output = _ccv_cnnp_batch_norm_add_to_output,
1113 .copy = _ccv_cnnp_batch_norm_copy,
1114 .set_is_test = _ccv_cnnp_batch_norm_set_is_test,
1115 .deinit = _ccv_cnnp_batch_norm_deinit,
1116};
1117
1118ccv_cnnp_model_t* ccv_cnnp_batch_norm(const float momentum, const float epsilon, const int is_trainable, const char* const name)
1119{
1120 ccv_cnnp_model_batch_norm_t* const model_batch_norm = (ccv_cnnp_model_batch_norm_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_batch_norm_t));
1121 model_batch_norm->super.isa = &ccv_cnnp_batch_norm_isa;
1122 model_batch_norm->super.input_size = 1;
1123 model_batch_norm->super.outputs = &model_batch_norm->output;
1124 model_batch_norm->super.output_size = 1;
1125 model_batch_norm->super.is_trainable = is_trainable;
1126 ccv_cnnp_model_copy_name(&model_batch_norm->super, name);
1127 model_batch_norm->scale.d = CCV_NNC_NO_TENSOR_SYMBOL;
1128 model_batch_norm->scale.graph = 0;
1129 model_batch_norm->bias.d = CCV_NNC_NO_TENSOR_SYMBOL;
1130 model_batch_norm->bias.graph = 0;
1131 model_batch_norm->params.bnorm.momentum = momentum;
1132 model_batch_norm->params.bnorm.epsilon = epsilon;
1133 return (ccv_cnnp_model_t*)model_batch_norm;
1134}
1135
1136static ccv_cnnp_model_t* _ccv_cnnp_batch_norm_copy(const ccv_cnnp_model_t* const super, void* const context)
1137{
1138 const ccv_cnnp_model_batch_norm_t* const self = (const ccv_cnnp_model_batch_norm_t*)super;
1139 return ccv_cnnp_batch_norm(self->params.bnorm.momentum, self->params.bnorm.epsilon, self->super.is_trainable, self->super.name);
1140}
1141
1142// MARK - Convolution Layer
1143
1144typedef struct {
1145 ccv_cnnp_model_t super;
1146 ccv_nnc_tensor_symbol_t output;
1147 ccv_nnc_tensor_symbol_t weights;
1148 ccv_nnc_tensor_symbol_t bias;
1149 int groups;
1150 int filters;
1151 int kdim[CCV_NNC_MAX_DIM_ALLOC(12)];
1152 int dilation[CCV_NNC_MAX_DIM_ALLOC(12)];
1153 int no_bias;
1154 int format;
1155 ccv_nnc_hint_t hint;
1156} ccv_cnnp_model_convolution_t;
1157
1158static void _ccv_cnnp_convolution_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
1159{
1160 ccv_cnnp_model_convolution_t* const self = (ccv_cnnp_model_convolution_t*)super;
1161 PRINT(CCV_CLI_VERBOSE, "[cnnp_convolution_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_convolution_build] -\n"); fflush(stdout); } }
while (0);
1162 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 1162, __extension__ __PRETTY_FUNCTION__); }));
1163 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 1163, __extension__ __PRETTY_FUNCTION__
); }));
1164 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1165 int i;
1166 const int k_nd = ccv_nnc_tensor_nd(self->kdim);
1167 const int nd = k_nd + 2;
1168 ccv_nnc_tensor_param_t weights_params = params;
1169 if (self->format)
1170 weights_params.format = self->format;
1171 ccv_nnc_tensor_set_n(&weights_params, self->filters);
1172 const int a_nd = ccv_nnc_tensor_nd(params.dim);
1173 int c;
1174 switch (params.format)
1175 {
1176 case CCV_TENSOR_FORMAT_NHWC:
1177 c = params.dim[a_nd - 1];
1178 break;
1179 case CCV_TENSOR_FORMAT_NCHW:
1180 if (a_nd == k_nd + 1)
1181 c = params.dim[0];
1182 else
1183 c = params.dim[a_nd <= 1 ? 0 : 1];
1184 break;
1185 case CCV_TENSOR_FORMAT_CHWN:
1186 c = params.dim[0];
1187 break;
1188 }
1189 assert(c % self->groups == 0)((void) sizeof ((c % self->groups == 0) ? 1 : 0), __extension__
({ if (c % self->groups == 0) ; else __assert_fail ("c % self->groups == 0"
, "ccv_cnnp_model_addons.c", 1189, __extension__ __PRETTY_FUNCTION__
); }));
1190 ccv_nnc_tensor_set_c(&weights_params, nd, c / self->groups);
1191 int hw = -1;
1192 if (weights_params.format == CCV_TENSOR_FORMAT_NHWC || weights_params.format == CCV_TENSOR_FORMAT_CHWN)
1193 hw = 1;
1194 else if (weights_params.format == CCV_TENSOR_FORMAT_NCHW)
1195 hw = 2;
1196 assert(hw >= 0)((void) sizeof ((hw >= 0) ? 1 : 0), __extension__ ({ if (hw
>= 0) ; else __assert_fail ("hw >= 0", "ccv_cnnp_model_addons.c"
, 1196, __extension__ __PRETTY_FUNCTION__); }));
1197 for (i = 0; i < k_nd; i++)
1198 weights_params.dim[i + hw] = self->kdim[i];
1199 if (!self->weights.graph)
1200 self->weights = ccv_nnc_tensor_symbol_new(graph, weights_params, "weights");
1201 assert(self->weights.graph == graph)((void) sizeof ((self->weights.graph == graph) ? 1 : 0), __extension__
({ if (self->weights.graph == graph) ; else __assert_fail
("self->weights.graph == graph", "ccv_cnnp_model_addons.c"
, 1201, __extension__ __PRETTY_FUNCTION__); }));
1202 const ccv_nnc_tensor_symbol_t weights = ccv_cnnp_model_get_symbol(super, self->weights);
1203 ccv_nnc_tensor_param_t bias_params = params;
1204 if (self->format)
1205 bias_params.format = self->format;
1206 memset(bias_params.dim, 0, sizeof(bias_params.dim));
1207 bias_params.dim[0] = self->filters;
1208 ccv_nnc_cmd_t cmd = CMD_CONVOLUTION_FORWARD(self->groups, self->filters)ccv_nnc_cmd(CCV_NNC_CONVOLUTION_FORWARD, 0, ((ccv_nnc_cmd_param_t
){.size={.dim={}},.convolution={.count=self->filters,.groups
=self->groups}}), 0);
1209 for (i = 0; i < k_nd; i++)
1210 cmd.info.size.dim[i] = self->kdim[i];
1211 cmd.info.size.dim[k_nd] = c;
1212 memcpy(cmd.info.convolution.dilation, self->dilation, sizeof(self->dilation));
1213 ccv_nnc_tensor_param_t output_params;
1214 // Dilate weight size based on the dilation factor.
1215 for (i = 0; i < k_nd; i++)
1216 weights_params.dim[i + hw] = (self->kdim[i] - 1) * ccv_max(self->dilation[i], 1)({ typeof (self->dilation[i]) _a = (self->dilation[i]);
typeof (1) _b = (1); (_a > _b) ? _a : _b; }) + 1;
1217 ccv_nnc_hint_tensor_auto(cmd, (ccv_nnc_tensor_param_t []){
1218 params,
1219 weights_params,
1220 bias_params,
1221 }, 3, self->hint, &output_params, 1);
1222 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1223 ccv_nnc_graph_exec_symbol_t convolution;
1224 if (self->no_bias)
1225 convolution = ccv_nnc_graph_exec_symbol_new(graph, cmd, TENSOR_SYMBOL_LIST(inputs[0], weights)(const ccv_nnc_tensor_symbol_t []){inputs[0], weights}, (1 +1
+1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 -1), TENSOR_SYMBOL_LIST(output)(const ccv_nnc_tensor_symbol_t []){output}, (1 +1 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "convolution");
1226 else {
1227 if (!self->bias.graph)
1228 self->bias = ccv_nnc_tensor_symbol_new(graph, bias_params, "bias");
1229 const ccv_nnc_tensor_symbol_t bias = ccv_cnnp_model_get_symbol(super, self->bias);
1230 convolution = ccv_nnc_graph_exec_symbol_new(graph, cmd, TENSOR_SYMBOL_LIST(inputs[0], weights, bias)(const ccv_nnc_tensor_symbol_t []){inputs[0], weights, bias},
(1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 -1), TENSOR_SYMBOL_LIST(output)(const ccv_nnc_tensor_symbol_t []){output}, (1 +1 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "convolution");
1231 }
1232 ccv_nnc_graph_exec_symbol_set_hint(graph, convolution, self->hint);
1233 outputs[0] = output;
1234}
1235
1236static void _ccv_cnnp_convolution_init_states(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_cnnp_state_initializer_f initializer, void* const context)
1237{
1238 ccv_cnnp_model_convolution_t* const self = (ccv_cnnp_model_convolution_t*)super;
1239 const ccv_nnc_tensor_param_t weight_params = ccv_nnc_tensor_symbol_params(graph, self->weights);
1240 const int n = ccv_max(ccv_nnc_tensor_get_n(weight_params), 1)({ typeof (ccv_nnc_tensor_get_n(weight_params)) _a = (ccv_nnc_tensor_get_n
(weight_params)); typeof (1) _b = (1); (_a > _b) ? _a : _b
; });
1241 const int count = ccv_nnc_tensor_count(weight_params);
1242 const float std = sqrtf(2) / sqrtf(count / n);
1243 const float bound = sqrtf(3) * std;
1244 initializer(context, CMD_RANDOM_UNIFORM_FORWARD(-bound, bound)ccv_nnc_cmd(CCV_NNC_RANDOM_UNIFORM_FORWARD, 0, (ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.blas={.a={-bound, bound}}}, 0), ccv_nnc_no_hint, 0, 0, self->weights);
1245 if (self->bias.graph)
1246 initializer(context, CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, self->bias);
1247}
1248
1249static void _ccv_cnnp_convolution_add_to_parameter(ccv_cnnp_model_t* const super, const ccv_cnnp_add_to_array_f add_to_array, void* const parameters, const int is_trainable)
1250{
1251 ccv_cnnp_model_convolution_t* const self = (ccv_cnnp_model_convolution_t*)super;
1252 add_to_array(parameters, self->weights, is_trainable);
1253 if (self->bias.graph)
1254 add_to_array(parameters, self->bias, is_trainable);
1255}
1256
1257static ccv_cnnp_model_t* _ccv_cnnp_convolution_copy(const ccv_cnnp_model_t* const super, void* const context);
1258
1259static const ccv_cnnp_model_vtab_t ccv_cnnp_convolution_isa = {
1260 .build = _ccv_cnnp_convolution_build,
1261 .init_states = _ccv_cnnp_convolution_init_states,
1262 .add_to_parameter = _ccv_cnnp_convolution_add_to_parameter,
1263 .copy = _ccv_cnnp_convolution_copy,
1264};
1265
1266ccv_cnnp_model_t* ccv_cnnp_convolution(const int groups, const int filters, const int kdim[CCV_NNC_MAX_DIM_ALLOC(12)], const int dilation[CCV_NNC_MAX_DIM_ALLOC(12)], const int no_bias, ccv_nnc_hint_t hint, const int format, const int is_trainable, const char* const name)
1267{
1268 ccv_cnnp_model_convolution_t* const model_convolution = (ccv_cnnp_model_convolution_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_convolution_t));
1269 model_convolution->super.isa = &ccv_cnnp_convolution_isa;
1270 model_convolution->super.input_size = 1;
1271 model_convolution->super.outputs = &model_convolution->output;
1272 model_convolution->super.output_size = 1;
1273 model_convolution->super.is_trainable = is_trainable;
1274 ccv_cnnp_model_copy_name(&model_convolution->super, name);
1275 model_convolution->weights.d = CCV_NNC_NO_TENSOR_SYMBOL;
1276 model_convolution->weights.graph = 0;
1277 model_convolution->bias.d = CCV_NNC_NO_TENSOR_SYMBOL;
1278 model_convolution->bias.graph = 0;
1279 model_convolution->groups = groups;
1280 model_convolution->filters = filters;
1281 memcpy(model_convolution->kdim, kdim, sizeof(model_convolution->kdim));
1282 memcpy(model_convolution->dilation, dilation, sizeof(model_convolution->dilation));
1283 model_convolution->no_bias = no_bias;
1284 model_convolution->hint = hint;
1285 model_convolution->format = format;
1286 return (ccv_cnnp_model_t*)model_convolution;
1287}
1288
1289static ccv_cnnp_model_t* _ccv_cnnp_convolution_copy(const ccv_cnnp_model_t* const super, void* const context)
1290{
1291 ccv_cnnp_model_convolution_t* const self = (ccv_cnnp_model_convolution_t*)super;
1292 return ccv_cnnp_convolution(self->groups, self->filters, self->kdim, self->dilation, self->no_bias, self->hint, self->format, self->super.is_trainable, self->super.name);
1293}
1294
1295// MARK - Convolution Transpose Layer
1296
1297typedef struct {
1298 ccv_cnnp_model_t super;
1299 ccv_nnc_tensor_symbol_t output;
1300 ccv_nnc_tensor_symbol_t weights;
1301 ccv_nnc_tensor_symbol_t bias;
1302 int groups;
1303 int filters;
1304 int kdim[CCV_NNC_MAX_DIM_ALLOC(12)];
1305 int dilation[CCV_NNC_MAX_DIM_ALLOC(12)];
1306 int output_padding;
1307 int no_bias;
1308 int format;
1309 ccv_nnc_hint_t hint;
1310} ccv_cnnp_model_convolution_transpose_t;
1311
1312static void _ccv_cnnp_convolution_transpose_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
1313{
1314 ccv_cnnp_model_convolution_transpose_t* const self = (ccv_cnnp_model_convolution_transpose_t*)super;
1315 PRINT(CCV_CLI_VERBOSE, "[cnnp_convolution_transpose_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_convolution_transpose_build] -\n"); fflush(stdout
); } } while (0);
1316 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 1316, __extension__ __PRETTY_FUNCTION__); }));
1317 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 1317, __extension__ __PRETTY_FUNCTION__
); }));
1318 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1319 int i;
1320 const int nd = CCV_NNC_MAX_DIM(2) + 2;
1321 ccv_nnc_tensor_param_t weights_params = params;
1322 if (self->format)
1323 weights_params.format = self->format;
1324 const int c = ccv_nnc_tensor_get_c(params);
1325 ccv_nnc_tensor_set_n(&weights_params, c);
1326 assert(c % self->groups == 0)((void) sizeof ((c % self->groups == 0) ? 1 : 0), __extension__
({ if (c % self->groups == 0) ; else __assert_fail ("c % self->groups == 0"
, "ccv_cnnp_model_addons.c", 1326, __extension__ __PRETTY_FUNCTION__
); }));
1327 ccv_nnc_tensor_set_c(&weights_params, nd, self->filters / self->groups);
1328 const int hw = ccv_nnc_tensor_hw(weights_params, nd, CCV_NNC_MAX_DIM(2));
1329 assert(hw >= 0)((void) sizeof ((hw >= 0) ? 1 : 0), __extension__ ({ if (hw
>= 0) ; else __assert_fail ("hw >= 0", "ccv_cnnp_model_addons.c"
, 1329, __extension__ __PRETTY_FUNCTION__); }));
1330 for (i = 0; i < CCV_NNC_MAX_DIM(2); i++)
1331 weights_params.dim[i + hw] = self->kdim[i];
1332 if (!self->weights.graph)
1333 self->weights = ccv_nnc_tensor_symbol_new(graph, weights_params, "weights");
1334 assert(self->weights.graph == graph)((void) sizeof ((self->weights.graph == graph) ? 1 : 0), __extension__
({ if (self->weights.graph == graph) ; else __assert_fail
("self->weights.graph == graph", "ccv_cnnp_model_addons.c"
, 1334, __extension__ __PRETTY_FUNCTION__); }));
1335 const ccv_nnc_tensor_symbol_t weights = ccv_cnnp_model_get_symbol(super, self->weights);
1336 ccv_nnc_tensor_param_t bias_params = params;
1337 if (self->format)
1338 bias_params.format = self->format;
1339 memset(bias_params.dim, 0, sizeof(bias_params.dim));
1340 bias_params.dim[0] = self->filters;
1341 ccv_nnc_cmd_t cmd = CMD_CONVOLUTION_TRANSPOSE_FORWARD(self->groups, self->filters, self->output_padding)ccv_nnc_cmd(CCV_NNC_CONVOLUTION_TRANSPOSE_FORWARD, 0, ((ccv_nnc_cmd_param_t
){.size={.dim={}},.convolution_transpose={.count=self->filters
,.groups=self->groups,.output_padding=self->output_padding
}}), 0);
1342 for (i = 0; i < CCV_NNC_MAX_DIM(2); i++)
1343 cmd.info.size.dim[i] = self->kdim[i];
1344 cmd.info.size.dim[CCV_NNC_MAX_DIM(2)] = c;
1345 memcpy(cmd.info.convolution_transpose.dilation, self->dilation, sizeof(self->dilation));
1346 ccv_nnc_tensor_param_t output_params;
1347 // Dilate weight size based on the dilation factor.
1348 for (i = 0; i < CCV_NNC_MAX_DIM(2); i++)
1349 weights_params.dim[i + hw] = (self->kdim[i] - 1) * ccv_max(self->dilation[i], 1)({ typeof (self->dilation[i]) _a = (self->dilation[i]);
typeof (1) _b = (1); (_a > _b) ? _a : _b; }) + 1;
1350 ccv_nnc_hint_tensor_auto(cmd, (ccv_nnc_tensor_param_t []){
1351 params,
1352 weights_params,
1353 bias_params,
1354 }, 3, self->hint, &output_params, 1);
1355 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1356 ccv_nnc_graph_exec_symbol_t convolution_transpose;
1357 if (self->no_bias)
1358 convolution_transpose = ccv_nnc_graph_exec_symbol_new(graph, cmd, TENSOR_SYMBOL_LIST(inputs[0], weights)(const ccv_nnc_tensor_symbol_t []){inputs[0], weights}, (1 +1
+1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 -1), TENSOR_SYMBOL_LIST(output)(const ccv_nnc_tensor_symbol_t []){output}, (1 +1 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "convolution_transpose");
1359 else {
1360 if (!self->bias.graph)
1361 self->bias = ccv_nnc_tensor_symbol_new(graph, bias_params, "bias");
1362 const ccv_nnc_tensor_symbol_t bias = ccv_cnnp_model_get_symbol(super, self->bias);
1363 convolution_transpose = ccv_nnc_graph_exec_symbol_new(graph, cmd, TENSOR_SYMBOL_LIST(inputs[0], weights, bias)(const ccv_nnc_tensor_symbol_t []){inputs[0], weights, bias},
(1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 -1), TENSOR_SYMBOL_LIST(output)(const ccv_nnc_tensor_symbol_t []){output}, (1 +1 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "convolution_transpose");
1364 }
1365 ccv_nnc_graph_exec_symbol_set_hint(graph, convolution_transpose, self->hint);
1366 outputs[0] = output;
1367}
1368
1369static void _ccv_cnnp_convolution_transpose_init_states(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_cnnp_state_initializer_f initializer, void* const context)
1370{
1371 ccv_cnnp_model_convolution_transpose_t* const self = (ccv_cnnp_model_convolution_transpose_t*)super;
1372 const ccv_nnc_tensor_param_t weight_params = ccv_nnc_tensor_symbol_params(graph, self->weights);
1373 const int n = ccv_max(ccv_nnc_tensor_get_n(weight_params), 1)({ typeof (ccv_nnc_tensor_get_n(weight_params)) _a = (ccv_nnc_tensor_get_n
(weight_params)); typeof (1) _b = (1); (_a > _b) ? _a : _b
; });
1374 const int count = ccv_nnc_tensor_count(weight_params);
1375 const float std = sqrtf(2) / sqrtf(count / n);
1376 const float bound = sqrtf(3) * std;
1377 initializer(context, CMD_RANDOM_UNIFORM_FORWARD(-bound, bound)ccv_nnc_cmd(CCV_NNC_RANDOM_UNIFORM_FORWARD, 0, (ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.blas={.a={-bound, bound}}}, 0), ccv_nnc_no_hint, 0, 0, self->weights);
1378 if (self->bias.graph)
1379 initializer(context, CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, self->bias);
1380}
1381
1382static void _ccv_cnnp_convolution_transpose_add_to_parameter(ccv_cnnp_model_t* const super, const ccv_cnnp_add_to_array_f add_to_array, void* const parameters, const int is_trainable)
1383{
1384 ccv_cnnp_model_convolution_transpose_t* const self = (ccv_cnnp_model_convolution_transpose_t*)super;
1385 add_to_array(parameters, self->weights, is_trainable);
1386 if (self->bias.graph)
1387 add_to_array(parameters, self->bias, is_trainable);
1388}
1389
1390static ccv_cnnp_model_t* _ccv_cnnp_convolution_transpose_copy(const ccv_cnnp_model_t* const super, void* const context);
1391
1392static const ccv_cnnp_model_vtab_t ccv_cnnp_convolution_transpose_isa = {
1393 .build = _ccv_cnnp_convolution_transpose_build,
1394 .init_states = _ccv_cnnp_convolution_transpose_init_states,
1395 .add_to_parameter = _ccv_cnnp_convolution_transpose_add_to_parameter,
1396 .copy = _ccv_cnnp_convolution_transpose_copy,
1397};
1398
1399ccv_cnnp_model_t* ccv_cnnp_convolution_transpose(const int groups, const int filters, const int kdim[CCV_NNC_MAX_DIM_ALLOC(12)], const int dilation[CCV_NNC_MAX_DIM_ALLOC(12)], const int output_padding, const int no_bias, ccv_nnc_hint_t hint, const int format, const int is_trainable, const char* const name)
1400{
1401 ccv_cnnp_model_convolution_transpose_t* const model_convolution_transpose = (ccv_cnnp_model_convolution_transpose_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_convolution_transpose_t));
1402 model_convolution_transpose->super.isa = &ccv_cnnp_convolution_transpose_isa;
1403 model_convolution_transpose->super.input_size = 1;
1404 model_convolution_transpose->super.outputs = &model_convolution_transpose->output;
1405 model_convolution_transpose->super.output_size = 1;
1406 model_convolution_transpose->super.is_trainable = is_trainable;
1407 ccv_cnnp_model_copy_name(&model_convolution_transpose->super, name);
1408 model_convolution_transpose->weights.d = CCV_NNC_NO_TENSOR_SYMBOL;
1409 model_convolution_transpose->weights.graph = 0;
1410 model_convolution_transpose->bias.d = CCV_NNC_NO_TENSOR_SYMBOL;
1411 model_convolution_transpose->bias.graph = 0;
1412 model_convolution_transpose->groups = groups;
1413 model_convolution_transpose->filters = filters;
1414 memcpy(model_convolution_transpose->kdim, kdim, sizeof(model_convolution_transpose->kdim));
1415 memcpy(model_convolution_transpose->dilation, dilation, sizeof(model_convolution_transpose->dilation));
1416 model_convolution_transpose->output_padding = output_padding;
1417 model_convolution_transpose->no_bias = no_bias;
1418 model_convolution_transpose->hint = hint;
1419 model_convolution_transpose->format = format;
1420 return (ccv_cnnp_model_t*)model_convolution_transpose;
1421}
1422
1423static ccv_cnnp_model_t* _ccv_cnnp_convolution_transpose_copy(const ccv_cnnp_model_t* const super, void* const context)
1424{
1425 ccv_cnnp_model_convolution_transpose_t* const self = (ccv_cnnp_model_convolution_transpose_t*)super;
1426 return ccv_cnnp_convolution_transpose(self->groups, self->filters, self->kdim, self->dilation, self->output_padding, self->no_bias, self->hint, self->format, self->super.is_trainable, self->super.name);
1427}
1428
1429// MARK - Dense Layer
1430
1431typedef struct {
1432 ccv_cnnp_model_t super;
1433 ccv_nnc_tensor_symbol_t output;
1434 ccv_nnc_tensor_symbol_t weights;
1435 ccv_nnc_tensor_symbol_t bias;
1436 int count;
1437 int no_bias;
1438 int flags;
1439} ccv_cnnp_model_dense_t;
1440
1441static void _ccv_cnnp_dense_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
1442{
1443 ccv_cnnp_model_dense_t* const self = (ccv_cnnp_model_dense_t*)super;
1444 PRINT(CCV_CLI_VERBOSE, "[cnnp_dense_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_dense_build] -\n"); fflush(stdout); } } while
(0);
1445 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 1445, __extension__ __PRETTY_FUNCTION__); }));
1446 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 1446, __extension__ __PRETTY_FUNCTION__
); }));
1447 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1448 ccv_nnc_tensor_param_t weights_params = params;
1449 memset(weights_params.dim, 0, sizeof(weights_params.dim));
1450 if (params.dim[0] != 0)
1451 {
1452 weights_params.dim[0] = self->count;
1453 weights_params.dim[1] = params.dim[ccv_nnc_tensor_nd(params.dim) - 1];
1454 }
1455 if (!self->weights.graph)
1456 self->weights = ccv_nnc_tensor_symbol_new(graph, weights_params, "weights");
1457 assert(self->weights.graph == graph)((void) sizeof ((self->weights.graph == graph) ? 1 : 0), __extension__
({ if (self->weights.graph == graph) ; else __assert_fail
("self->weights.graph == graph", "ccv_cnnp_model_addons.c"
, 1457, __extension__ __PRETTY_FUNCTION__); }));
1458 const ccv_nnc_tensor_symbol_t weights = ccv_cnnp_model_get_symbol(super, self->weights);
1459 ccv_nnc_tensor_param_t bias_params = params;
1460 memset(bias_params.dim, 0, sizeof(bias_params.dim));
1461 bias_params.dim[0] = self->count;
1462 ccv_nnc_cmd_t cmd = {0};
1463 cmd.cmd = CCV_NNC_GEMM_FORWARD;
1464 cmd.info.blas.a[0] = 1;
1465 cmd.info.blas.a[1] = 1;
1466 cmd.info.blas.transpose_b[0] = 0;
1467 cmd.info.blas.transpose_b[1] = 1;
1468 cmd.info.blas.flags = self->flags;
1469 ccv_nnc_tensor_param_t output_params;
1470 ccv_nnc_hint_tensor_auto(cmd, (ccv_nnc_tensor_param_t []){
1471 params,
1472 weights_params,
1473 bias_params,
1474 }, 3, ccv_nnc_no_hint, &output_params, 1);
1475 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1476 if (self->no_bias)
1477 ccv_nnc_graph_exec_symbol_new(graph, cmd, TENSOR_SYMBOL_LIST(inputs[0], weights)(const ccv_nnc_tensor_symbol_t []){inputs[0], weights}, (1 +1
+1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 -1), TENSOR_SYMBOL_LIST(output)(const ccv_nnc_tensor_symbol_t []){output}, (1 +1 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "dense");
1478 else {
1479 if (!self->bias.graph)
1480 self->bias = ccv_nnc_tensor_symbol_new(graph, bias_params, "bias");
1481 const ccv_nnc_tensor_symbol_t bias = ccv_cnnp_model_get_symbol(super, self->bias);
1482 ccv_nnc_graph_exec_symbol_new(graph, cmd, TENSOR_SYMBOL_LIST(inputs[0], weights, bias)(const ccv_nnc_tensor_symbol_t []){inputs[0], weights, bias},
(1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 -1), TENSOR_SYMBOL_LIST(output)(const ccv_nnc_tensor_symbol_t []){output}, (1 +1 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "dense");
1483 }
1484 outputs[0] = output;
1485}
1486
1487static void _ccv_cnnp_dense_init_states(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_cnnp_state_initializer_f initializer, void* const context)
1488{
1489 ccv_cnnp_model_dense_t* const self = (ccv_cnnp_model_dense_t*)super;
1490 const ccv_nnc_tensor_param_t weight_params = ccv_nnc_tensor_symbol_params(graph, self->weights);
1491 const int c = weight_params.dim[1];
1492 const float std = sqrtf(2) / sqrtf(c);
1493 const float bound = sqrtf(3) * std;
1494 initializer(context, CMD_RANDOM_UNIFORM_FORWARD(-bound, bound)ccv_nnc_cmd(CCV_NNC_RANDOM_UNIFORM_FORWARD, 0, (ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.blas={.a={-bound, bound}}}, 0), ccv_nnc_no_hint, 0, 0, self->weights);
1495 if (self->bias.graph)
1496 initializer(context, CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, self->bias);
1497}
1498
1499static void _ccv_cnnp_dense_add_to_parameter(ccv_cnnp_model_t* const super, const ccv_cnnp_add_to_array_f add_to_array, void* const parameters, const int is_trainable)
1500{
1501 ccv_cnnp_model_dense_t* const self = (ccv_cnnp_model_dense_t*)super;
1502 add_to_array(parameters, self->weights, is_trainable);
1503 if (self->bias.graph)
1504 add_to_array(parameters, self->bias, is_trainable);
1505}
1506
1507static ccv_cnnp_model_t* _ccv_cnnp_dense_copy(const ccv_cnnp_model_t* const super, void* const context);
1508
1509static const ccv_cnnp_model_vtab_t ccv_cnnp_dense_isa = {
1510 .build = _ccv_cnnp_dense_build,
1511 .init_states = _ccv_cnnp_dense_init_states,
1512 .add_to_parameter = _ccv_cnnp_dense_add_to_parameter,
1513 .copy = _ccv_cnnp_dense_copy,
1514};
1515
1516ccv_cnnp_model_t* ccv_cnnp_dense(const int count, const int no_bias, const int flags, const int is_trainable, const char* const name)
1517{
1518 ccv_cnnp_model_dense_t* const model_dense = (ccv_cnnp_model_dense_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_dense_t));
1519 model_dense->super.isa = &ccv_cnnp_dense_isa;
1520 model_dense->super.input_size = 1;
1521 model_dense->super.outputs = &model_dense->output;
1522 model_dense->super.output_size = 1;
1523 model_dense->super.is_trainable = is_trainable;
1524 ccv_cnnp_model_copy_name(&model_dense->super, name);
1525 model_dense->weights.d = CCV_NNC_NO_TENSOR_SYMBOL;
1526 model_dense->weights.graph = 0;
1527 model_dense->bias.d = CCV_NNC_NO_TENSOR_SYMBOL;
1528 model_dense->bias.graph = 0;
1529 model_dense->count = count;
1530 model_dense->no_bias = no_bias;
1531 model_dense->flags = flags;
1532 return (ccv_cnnp_model_t*)model_dense;
1533}
1534
1535static ccv_cnnp_model_t* _ccv_cnnp_dense_copy(const ccv_cnnp_model_t* const super, void* const context)
1536{
1537 const ccv_cnnp_model_dense_t* const self = (const ccv_cnnp_model_dense_t*)super;
1538 return ccv_cnnp_dense(self->count, self->no_bias, self->flags, self->super.is_trainable, self->super.name);
1539}
1540
1541// MARK - Pool Layers
1542
1543typedef struct {
1544 ccv_cnnp_model_t super;
1545 ccv_nnc_tensor_symbol_t output;
1546 int kdim[CCV_NNC_MAX_DIM_ALLOC(12)];
1547 ccv_nnc_hint_t hint;
1548} ccv_cnnp_model_pool_t;
1549
1550static void _ccv_cnnp_max_pool_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
1551{
1552 ccv_cnnp_model_pool_t* const self = (ccv_cnnp_model_pool_t*)super;
1553 PRINT(CCV_CLI_VERBOSE, "[cnnp_max_pool_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_max_pool_build] -\n"); fflush(stdout); } } while
(0);
1554 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 1554, __extension__ __PRETTY_FUNCTION__); }));
1555 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 1555, __extension__ __PRETTY_FUNCTION__
); }));
1556 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1557 const int hw = ccv_nnc_tensor_hw(params, ccv_nnc_tensor_nd(params.dim), CCV_NNC_MAX_DIM(2));
1558 ccv_nnc_cmd_t cmd;
1559 if (hw >= 0 && self->kdim[0] == 0 && self->kdim[1] == 0)
1560 cmd = CMD_MAX_POOL_FORWARD(params.dim[hw], params.dim[hw + 1])ccv_nnc_cmd(CCV_NNC_MAX_POOL_FORWARD, 0, ((ccv_nnc_cmd_param_t
){.size={.dim={params.dim[hw], params.dim[hw + 1],1}}}), 0);
1561 else
1562 cmd = CMD_MAX_POOL_FORWARD(self->kdim[0], self->kdim[1])ccv_nnc_cmd(CCV_NNC_MAX_POOL_FORWARD, 0, ((ccv_nnc_cmd_param_t
){.size={.dim={self->kdim[0], self->kdim[1],1}}}), 0);
1563 ccv_nnc_tensor_param_t output_params;
1564 ccv_nnc_hint_tensor_auto(cmd, &params, 1, self->hint, &output_params, 1);
1565 const ccv_nnc_tensor_symbol_t pool_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1566 const ccv_nnc_graph_exec_symbol_t exec = ccv_nnc_graph_exec_symbol_new(graph, cmd, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(pool_output)(const ccv_nnc_tensor_symbol_t []){pool_output}, (1 +1 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "max_pool");
1567 ccv_nnc_graph_exec_symbol_set_hint(graph, exec, self->hint);
1568 outputs[0] = pool_output;
1569}
1570
1571static ccv_cnnp_model_t* _ccv_cnnp_max_pool_copy(const ccv_cnnp_model_t* const super, void* const context);
1572
1573static const ccv_cnnp_model_vtab_t ccv_cnnp_max_pool_isa = {
1574 .build = _ccv_cnnp_max_pool_build,
1575 .copy = _ccv_cnnp_max_pool_copy,
1576};
1577
1578ccv_cnnp_model_t* ccv_cnnp_max_pool(const int kdim[CCV_NNC_MAX_DIM_ALLOC(12)], const ccv_nnc_hint_t hint, const char* const name)
1579{
1580 ccv_cnnp_model_pool_t* const model_pool = (ccv_cnnp_model_pool_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_pool_t));
1581 model_pool->super.isa = &ccv_cnnp_max_pool_isa;
1582 model_pool->super.input_size = 1;
1583 model_pool->super.outputs = &model_pool->output;
1584 model_pool->super.output_size = 1;
1585 ccv_cnnp_model_copy_name(&model_pool->super, name);
1586 memcpy(model_pool->kdim, kdim, sizeof(model_pool->kdim));
1587 model_pool->hint = hint;
1588 return (ccv_cnnp_model_t*)model_pool;
1589}
1590
1591static ccv_cnnp_model_t* _ccv_cnnp_max_pool_copy(const ccv_cnnp_model_t* const super, void* const context)
1592{
1593 const ccv_cnnp_model_pool_t* const self = (const ccv_cnnp_model_pool_t*)super;
1594 return ccv_cnnp_max_pool(self->kdim, self->hint, self->super.name);
1595}
1596
1597static void _ccv_cnnp_average_pool_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
1598{
1599 ccv_cnnp_model_pool_t* const self = (ccv_cnnp_model_pool_t*)super;
1600 PRINT(CCV_CLI_VERBOSE, "[cnnp_average_pool_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_average_pool_build] -\n"); fflush(stdout); }
} while (0);
1601 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 1601, __extension__ __PRETTY_FUNCTION__); }));
1602 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 1602, __extension__ __PRETTY_FUNCTION__
); }));
1603 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1604 const int hw = ccv_nnc_tensor_hw(params, ccv_nnc_tensor_nd(params.dim), CCV_NNC_MAX_DIM(2));
1605 ccv_nnc_cmd_t cmd;
1606 if (hw >= 0 && self->kdim[0] == 0 && self->kdim[1] == 0)
1607 cmd = CMD_AVERAGE_POOL_FORWARD(params.dim[hw], params.dim[hw + 1])ccv_nnc_cmd(CCV_NNC_AVERAGE_POOL_FORWARD, 0, ((ccv_nnc_cmd_param_t
){.size={.dim={params.dim[hw], params.dim[hw + 1],1}}}), 0);
1608 else
1609 cmd = CMD_AVERAGE_POOL_FORWARD(self->kdim[0], self->kdim[1])ccv_nnc_cmd(CCV_NNC_AVERAGE_POOL_FORWARD, 0, ((ccv_nnc_cmd_param_t
){.size={.dim={self->kdim[0], self->kdim[1],1}}}), 0);
1610 ccv_nnc_tensor_param_t output_params;
1611 ccv_nnc_hint_tensor_auto(cmd, &params, 1, self->hint, &output_params, 1);
1612 const ccv_nnc_tensor_symbol_t pool_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1613 const ccv_nnc_graph_exec_symbol_t exec = ccv_nnc_graph_exec_symbol_new(graph, cmd, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(pool_output)(const ccv_nnc_tensor_symbol_t []){pool_output}, (1 +1 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "average_pool");
1614 ccv_nnc_graph_exec_symbol_set_hint(graph, exec, self->hint);
1615 outputs[0] = pool_output;
1616}
1617
1618static ccv_cnnp_model_t* _ccv_cnnp_average_pool_copy(const ccv_cnnp_model_t* const super, void* const context);
1619
1620static const ccv_cnnp_model_vtab_t ccv_cnnp_average_pool_isa = {
1621 .build = _ccv_cnnp_average_pool_build,
1622 .copy = _ccv_cnnp_average_pool_copy,
1623};
1624
1625ccv_cnnp_model_t* ccv_cnnp_average_pool(const int kdim[CCV_NNC_MAX_DIM_ALLOC(12)], const ccv_nnc_hint_t hint, const char* const name)
1626{
1627 ccv_cnnp_model_pool_t* const model_pool = (ccv_cnnp_model_pool_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_pool_t));
1628 model_pool->super.isa = &ccv_cnnp_average_pool_isa;
1629 model_pool->super.input_size = 1;
1630 model_pool->super.outputs = &model_pool->output;
1631 model_pool->super.output_size = 1;
1632 ccv_cnnp_model_copy_name(&model_pool->super, name);
1633 memcpy(model_pool->kdim, kdim, sizeof(model_pool->kdim));
1634 model_pool->hint = hint;
1635 return (ccv_cnnp_model_t*)model_pool;
1636}
1637
1638static ccv_cnnp_model_t* _ccv_cnnp_average_pool_copy(const ccv_cnnp_model_t* const super, void* const context)
1639{
1640 const ccv_cnnp_model_pool_t* const self = (const ccv_cnnp_model_pool_t*)super;
1641 return ccv_cnnp_average_pool(self->kdim, self->hint, self->super.name);
1642}
1643
1644// MARK - RELU Layer
1645
1646typedef struct {
1647 ccv_cnnp_model_t super;
1648 ccv_nnc_tensor_symbol_t output;
1649} ccv_cnnp_model_relu_t;
1650
1651static void _ccv_cnnp_relu_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
1652{
1653 PRINT(CCV_CLI_VERBOSE, "[cnnp_relu_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_relu_build] -\n"); fflush(stdout); } } while
(0);
1654 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 1654, __extension__ __PRETTY_FUNCTION__); }));
1655 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 1655, __extension__ __PRETTY_FUNCTION__
); }));
1656 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1657 ccv_nnc_tensor_param_t output_params;
1658 const ccv_nnc_cmd_t relu = CMD_RELU_FORWARD()ccv_nnc_cmd(CCV_NNC_RELU_FORWARD, 0, ccv_nnc_cmd_auto, 0);
1659 ccv_nnc_hint_tensor_auto(relu, (ccv_nnc_tensor_param_t []){
1660 params,
1661 }, 1, ccv_nnc_no_hint, &output_params, 1);
1662 const ccv_nnc_tensor_symbol_t relu_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1663 ccv_nnc_graph_exec_symbol_new(graph, relu, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(relu_output)(const ccv_nnc_tensor_symbol_t []){relu_output}, (1 +1 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "relu");
1664 outputs[0] = relu_output;
1665}
1666
1667static ccv_cnnp_model_t* _ccv_cnnp_relu_copy(const ccv_cnnp_model_t* const self, void* const context);
1668
1669static const ccv_cnnp_model_vtab_t ccv_cnnp_relu_isa = {
1670 .build = _ccv_cnnp_relu_build,
1671 .copy = _ccv_cnnp_relu_copy,
1672};
1673
1674ccv_cnnp_model_t* ccv_cnnp_relu(const char* const name)
1675{
1676 ccv_cnnp_model_relu_t* const model_relu = (ccv_cnnp_model_relu_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_relu_t));
1677 model_relu->super.isa = &ccv_cnnp_relu_isa;
1678 model_relu->super.input_size = 1;
1679 model_relu->super.outputs = &model_relu->output;
1680 model_relu->super.output_size = 1;
1681 ccv_cnnp_model_copy_name(&model_relu->super, name);
1682 return (ccv_cnnp_model_t*)model_relu;
1683}
1684
1685static ccv_cnnp_model_t* _ccv_cnnp_relu_copy(const ccv_cnnp_model_t* const self, void* const context)
1686{
1687 return ccv_cnnp_relu(self->name);
1688}
1689
1690// MARK - Sigmoid Layer
1691
1692typedef struct {
1693 ccv_cnnp_model_t super;
1694 ccv_nnc_tensor_symbol_t output;
1695} ccv_cnnp_model_sigmoid_t;
1696
1697static void _ccv_cnnp_sigmoid_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
1698{
1699 PRINT(CCV_CLI_VERBOSE, "[cnnp_sigmoid_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_sigmoid_build] -\n"); fflush(stdout); } } while
(0);
1700 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 1700, __extension__ __PRETTY_FUNCTION__); }));
1701 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 1701, __extension__ __PRETTY_FUNCTION__
); }));
1702 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1703 ccv_nnc_tensor_param_t output_params;
1704 const ccv_nnc_cmd_t sigmoid = CMD_SIGMOID_FORWARD()ccv_nnc_cmd(CCV_NNC_SIGMOID_FORWARD, 0, ccv_nnc_cmd_auto, 0);
1705 ccv_nnc_hint_tensor_auto(sigmoid, (ccv_nnc_tensor_param_t []){
1706 params,
1707 }, 1, ccv_nnc_no_hint, &output_params, 1);
1708 const ccv_nnc_tensor_symbol_t sigmoid_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1709 ccv_nnc_graph_exec_symbol_new(graph, sigmoid, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(sigmoid_output)(const ccv_nnc_tensor_symbol_t []){sigmoid_output}, (1 +1 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1
), "sigmoid");
1710 outputs[0] = sigmoid_output;
1711}
1712
1713static ccv_cnnp_model_t* _ccv_cnnp_sigmoid_copy(const ccv_cnnp_model_t* const self, void* const context);
1714
1715static const ccv_cnnp_model_vtab_t ccv_cnnp_sigmoid_isa = {
1716 .build = _ccv_cnnp_sigmoid_build,
1717 .copy = _ccv_cnnp_sigmoid_copy,
1718};
1719
1720ccv_cnnp_model_t* ccv_cnnp_sigmoid(const char* const name)
1721{
1722 ccv_cnnp_model_sigmoid_t* const model_sigmoid = (ccv_cnnp_model_sigmoid_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_sigmoid_t));
1723 model_sigmoid->super.isa = &ccv_cnnp_sigmoid_isa;
1724 model_sigmoid->super.input_size = 1;
1725 model_sigmoid->super.outputs = &model_sigmoid->output;
1726 model_sigmoid->super.output_size = 1;
1727 ccv_cnnp_model_copy_name(&model_sigmoid->super, name);
1728 return (ccv_cnnp_model_t*)model_sigmoid;
1729}
1730
1731static ccv_cnnp_model_t* _ccv_cnnp_sigmoid_copy(const ccv_cnnp_model_t* const self, void* const context)
1732{
1733 return ccv_cnnp_sigmoid(self->name);
1734}
1735
1736// MARK - Tanh Layer
1737
1738typedef struct {
1739 ccv_cnnp_model_t super;
1740 ccv_nnc_tensor_symbol_t output;
1741 float soft_cap;
1742} ccv_cnnp_model_tanh_t;
1743
1744static void _ccv_cnnp_tanh_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
1745{
1746 PRINT(CCV_CLI_VERBOSE, "[cnnp_tanh_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_tanh_build] -\n"); fflush(stdout); } } while
(0);
1747 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 1747, __extension__ __PRETTY_FUNCTION__); }));
1748 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 1748, __extension__ __PRETTY_FUNCTION__
); }));
1749 ccv_cnnp_model_tanh_t* const self = (ccv_cnnp_model_tanh_t*)super;
1750 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1751 ccv_nnc_tensor_param_t output_params;
1752 const ccv_nnc_cmd_t tanh = CMD_TANH_FORWARD(self->soft_cap)ccv_nnc_cmd(CCV_NNC_TANH_FORWARD, 0, ((ccv_nnc_cmd_param_t){.
size={.dim={1,1,1}},.tanh={.soft_cap=self->soft_cap}}), 0);
1753 ccv_nnc_hint_tensor_auto(tanh, (ccv_nnc_tensor_param_t []){
1754 params,
1755 }, 1, ccv_nnc_no_hint, &output_params, 1);
1756 const ccv_nnc_tensor_symbol_t tanh_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1757 ccv_nnc_graph_exec_symbol_new(graph, tanh, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(tanh_output)(const ccv_nnc_tensor_symbol_t []){tanh_output}, (1 +1 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "tanh");
1758 outputs[0] = tanh_output;
1759}
1760
1761static ccv_cnnp_model_t* _ccv_cnnp_tanh_copy(const ccv_cnnp_model_t* const self, void* const context);
1762
1763static const ccv_cnnp_model_vtab_t ccv_cnnp_tanh_isa = {
1764 .build = _ccv_cnnp_tanh_build,
1765 .copy = _ccv_cnnp_tanh_copy,
1766};
1767
1768ccv_cnnp_model_t* ccv_cnnp_tanh(const float soft_cap, const char* const name)
1769{
1770 ccv_cnnp_model_tanh_t* const model_tanh = (ccv_cnnp_model_tanh_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_tanh_t));
1771 model_tanh->super.isa = &ccv_cnnp_tanh_isa;
1772 model_tanh->super.input_size = 1;
1773 model_tanh->super.outputs = &model_tanh->output;
1774 model_tanh->super.output_size = 1;
1775 model_tanh->soft_cap = soft_cap;
1776 ccv_cnnp_model_copy_name(&model_tanh->super, name);
1777 return (ccv_cnnp_model_t*)model_tanh;
1778}
1779
1780static ccv_cnnp_model_t* _ccv_cnnp_tanh_copy(const ccv_cnnp_model_t* const super, void* const context)
1781{
1782 const ccv_cnnp_model_tanh_t* const self = (const ccv_cnnp_model_tanh_t*)super;
1783 return ccv_cnnp_tanh(self->soft_cap, self->super.name);
1784}
1785
1786// MARK - Exp Layer
1787
1788typedef struct {
1789 ccv_cnnp_model_t super;
1790 ccv_nnc_tensor_symbol_t output;
1791} ccv_cnnp_model_exp_t;
1792
1793static void _ccv_cnnp_exp_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
1794{
1795 PRINT(CCV_CLI_VERBOSE, "[cnnp_exp_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_exp_build] -\n"); fflush(stdout); } } while (
0);
1796 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 1796, __extension__ __PRETTY_FUNCTION__); }));
1797 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 1797, __extension__ __PRETTY_FUNCTION__
); }));
1798 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1799 ccv_nnc_tensor_param_t output_params;
1800 const ccv_nnc_cmd_t exp = CMD_EWEXP_FORWARD()ccv_nnc_cmd(CCV_NNC_EWEXP_FORWARD, 0, ccv_nnc_cmd_auto, 0);
1801 ccv_nnc_hint_tensor_auto(exp, (ccv_nnc_tensor_param_t []){
1802 params,
1803 }, 1, ccv_nnc_no_hint, &output_params, 1);
1804 const ccv_nnc_tensor_symbol_t exp_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1805 ccv_nnc_graph_exec_symbol_new(graph, exp, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(exp_output)(const ccv_nnc_tensor_symbol_t []){exp_output}, (1 +1 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "exp");
1806 outputs[0] = exp_output;
1807}
1808
1809static ccv_cnnp_model_t* _ccv_cnnp_exp_copy(const ccv_cnnp_model_t* const self, void* const context);
1810
1811static const ccv_cnnp_model_vtab_t ccv_cnnp_exp_isa = {
1812 .build = _ccv_cnnp_exp_build,
1813 .copy = _ccv_cnnp_exp_copy,
1814};
1815
1816ccv_cnnp_model_t* ccv_cnnp_exp(const char* const name)
1817{
1818 ccv_cnnp_model_exp_t* const model_exp = (ccv_cnnp_model_exp_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_exp_t));
1819 model_exp->super.isa = &ccv_cnnp_exp_isa;
1820 model_exp->super.input_size = 1;
1821 model_exp->super.outputs = &model_exp->output;
1822 model_exp->super.output_size = 1;
1823 ccv_cnnp_model_copy_name(&model_exp->super, name);
1824 return (ccv_cnnp_model_t*)model_exp;
1825}
1826
1827static ccv_cnnp_model_t* _ccv_cnnp_exp_copy(const ccv_cnnp_model_t* const self, void* const context)
1828{
1829 return ccv_cnnp_exp(self->name);
1830}
1831
1832// MARK - Softplus Layer
1833
1834typedef struct {
1835 ccv_cnnp_model_t super;
1836 ccv_nnc_tensor_symbol_t output;
1837} ccv_cnnp_model_softplus_t;
1838
1839static void _ccv_cnnp_softplus_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
1840{
1841 PRINT(CCV_CLI_VERBOSE, "[cnnp_softplus_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_softplus_build] -\n"); fflush(stdout); } } while
(0);
1842 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 1842, __extension__ __PRETTY_FUNCTION__); }));
1843 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 1843, __extension__ __PRETTY_FUNCTION__
); }));
1844 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1845 ccv_nnc_tensor_param_t output_params;
1846 const ccv_nnc_cmd_t softplus = CMD_EWSOFTPLUS_FORWARD()ccv_nnc_cmd(CCV_NNC_EWSOFTPLUS_FORWARD, 0, ccv_nnc_cmd_auto, 0
);
1847 ccv_nnc_hint_tensor_auto(softplus, (ccv_nnc_tensor_param_t []){
1848 params,
1849 }, 1, ccv_nnc_no_hint, &output_params, 1);
1850 const ccv_nnc_tensor_symbol_t softplus_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1851 ccv_nnc_graph_exec_symbol_new(graph, softplus, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(softplus_output)(const ccv_nnc_tensor_symbol_t []){softplus_output}, (1 +1 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -
1), "softplus");
1852 outputs[0] = softplus_output;
1853}
1854
1855static ccv_cnnp_model_t* _ccv_cnnp_softplus_copy(const ccv_cnnp_model_t* const self, void* const context);
1856
1857static const ccv_cnnp_model_vtab_t ccv_cnnp_softplus_isa = {
1858 .build = _ccv_cnnp_softplus_build,
1859 .copy = _ccv_cnnp_softplus_copy,
1860};
1861
1862ccv_cnnp_model_t* ccv_cnnp_softplus(const char* const name)
1863{
1864 ccv_cnnp_model_softplus_t* const model_softplus = (ccv_cnnp_model_softplus_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_softplus_t));
1865 model_softplus->super.isa = &ccv_cnnp_softplus_isa;
1866 model_softplus->super.input_size = 1;
1867 model_softplus->super.outputs = &model_softplus->output;
1868 model_softplus->super.output_size = 1;
1869 ccv_cnnp_model_copy_name(&model_softplus->super, name);
1870 return (ccv_cnnp_model_t*)model_softplus;
1871}
1872
1873static ccv_cnnp_model_t* _ccv_cnnp_softplus_copy(const ccv_cnnp_model_t* const self, void* const context)
1874{
1875 return ccv_cnnp_softplus(self->name);
1876}
1877
1878// MARK - Swish Layer
1879
1880typedef struct {
1881 ccv_cnnp_model_t super;
1882 ccv_nnc_tensor_symbol_t output;
1883 float beta;
1884} ccv_cnnp_model_swish_t;
1885
1886static void _ccv_cnnp_swish_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
1887{
1888 PRINT(CCV_CLI_VERBOSE, "[cnnp_swish_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_swish_build] -\n"); fflush(stdout); } } while
(0);
1889 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 1889, __extension__ __PRETTY_FUNCTION__); }));
1890 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 1890, __extension__ __PRETTY_FUNCTION__
); }));
1891 ccv_cnnp_model_swish_t* const self = (ccv_cnnp_model_swish_t*)super;
1892 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1893 ccv_nnc_tensor_param_t output_params;
1894 const ccv_nnc_cmd_t swish = CMD_SWISH_FORWARD(self->beta)ccv_nnc_cmd(CCV_NNC_SWISH_FORWARD, 0, ((ccv_nnc_cmd_param_t){
.size={.dim={1,1,1}},.swish={.beta=self->beta}}), 0);
1895 ccv_nnc_hint_tensor_auto(swish, (ccv_nnc_tensor_param_t []){
1896 params,
1897 }, 1, ccv_nnc_no_hint, &output_params, 1);
1898 const ccv_nnc_tensor_symbol_t swish_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1899 ccv_nnc_graph_exec_symbol_new(graph, swish, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(swish_output)(const ccv_nnc_tensor_symbol_t []){swish_output}, (1 +1 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "swish");
1900 outputs[0] = swish_output;
1901}
1902
1903static ccv_cnnp_model_t* _ccv_cnnp_swish_copy(const ccv_cnnp_model_t* const self, void* const context);
1904
1905static const ccv_cnnp_model_vtab_t ccv_cnnp_swish_isa = {
1906 .build = _ccv_cnnp_swish_build,
1907 .copy = _ccv_cnnp_swish_copy,
1908};
1909
1910ccv_cnnp_model_t* ccv_cnnp_swish(const float beta, const char* const name)
1911{
1912 ccv_cnnp_model_swish_t* const model_swish = (ccv_cnnp_model_swish_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_swish_t));
1913 model_swish->super.isa = &ccv_cnnp_swish_isa;
1914 model_swish->super.input_size = 1;
1915 model_swish->super.outputs = &model_swish->output;
1916 model_swish->super.output_size = 1;
1917 model_swish->beta = beta;
1918 ccv_cnnp_model_copy_name(&model_swish->super, name);
1919 return (ccv_cnnp_model_t*)model_swish;
1920}
1921
1922static ccv_cnnp_model_t* _ccv_cnnp_swish_copy(const ccv_cnnp_model_t* const self, void* const context)
1923{
1924 const ccv_cnnp_model_swish_t* const swish = (const ccv_cnnp_model_swish_t*)self;
1925 return ccv_cnnp_swish(swish->beta, self->name);
1926}
1927
1928// MARK - Swish Mul Layer
1929
1930typedef struct {
1931 ccv_cnnp_model_t super;
1932 ccv_nnc_tensor_symbol_t output;
1933 float beta;
1934 float scale;
1935} ccv_cnnp_model_swish_mul_t;
1936
1937static void _ccv_cnnp_swish_mul_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
1938{
1939 PRINT(CCV_CLI_VERBOSE, "[cnnp_swish_mul_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_swish_mul_build] -\n"); fflush(stdout); } } while
(0);
1940 assert(input_size == 2)((void) sizeof ((input_size == 2) ? 1 : 0), __extension__ ({ if
(input_size == 2) ; else __assert_fail ("input_size == 2", "ccv_cnnp_model_addons.c"
, 1940, __extension__ __PRETTY_FUNCTION__); }));
1941 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 1941, __extension__ __PRETTY_FUNCTION__
); }));
1942 const ccv_cnnp_model_swish_mul_t* const self = (const ccv_cnnp_model_swish_mul_t*)super;
1943 ccv_nnc_tensor_param_t input_params[2];
1944 int i;
1945 for (i = 0; i < 2; i++)
1946 input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
1947 ccv_nnc_tensor_param_t output_params;
1948 const ccv_nnc_cmd_t swish_mul = CMD_SWISH_MUL_FORWARD(self->beta, self->scale)ccv_nnc_cmd(CCV_NNC_SWISH_MUL_FORWARD, 0, ((ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.swish_mul={.beta=self->beta,.scale
=self->scale}}), 0);
1949 ccv_nnc_hint_tensor_auto(swish_mul, input_params, 2, ccv_nnc_no_hint, &output_params, 1);
1950 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1951 ccv_nnc_graph_exec_symbol_new(graph, swish_mul, inputs, input_size, outputs, output_size, "swish_mul");
1952}
1953
1954static ccv_cnnp_model_t* _ccv_cnnp_swish_mul_copy(const ccv_cnnp_model_t* const self, void* const context);
1955
1956static const ccv_cnnp_model_vtab_t ccv_cnnp_swish_mul_isa = {
1957 .build = _ccv_cnnp_swish_mul_build,
1958 .copy = _ccv_cnnp_swish_mul_copy,
1959};
1960
1961ccv_cnnp_model_t* ccv_cnnp_swish_mul(const float beta, const float scale, const char* const name)
1962{
1963 ccv_cnnp_model_swish_mul_t* const model_swish_mul = (ccv_cnnp_model_swish_mul_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_swish_mul_t));
1964 model_swish_mul->super.isa = &ccv_cnnp_swish_mul_isa;
1965 model_swish_mul->super.input_size = 2;
1966 model_swish_mul->super.outputs = &model_swish_mul->output;
1967 model_swish_mul->super.output_size = 1;
1968 model_swish_mul->beta = beta;
1969 model_swish_mul->scale = scale;
1970 ccv_cnnp_model_copy_name(&model_swish_mul->super, name);
1971 return (ccv_cnnp_model_t*)model_swish_mul;
1972}
1973
1974static ccv_cnnp_model_t* _ccv_cnnp_swish_mul_copy(const ccv_cnnp_model_t* const super, void* const context)
1975{
1976 const ccv_cnnp_model_swish_mul_t* const self = (const ccv_cnnp_model_swish_mul_t*)super;
1977 return ccv_cnnp_swish_mul(self->beta, self->scale, self->super.name);
1978}
1979
1980// MARK - GELU Layer
1981
1982typedef struct {
1983 ccv_cnnp_model_t super;
1984 ccv_nnc_tensor_symbol_t output;
1985 int tanh;
1986} ccv_cnnp_model_gelu_t;
1987
1988static void _ccv_cnnp_gelu_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
1989{
1990 PRINT(CCV_CLI_VERBOSE, "[cnnp_gelu_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_gelu_build] -\n"); fflush(stdout); } } while
(0);
1991 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 1991, __extension__ __PRETTY_FUNCTION__); }));
1992 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 1992, __extension__ __PRETTY_FUNCTION__
); }));
1993 ccv_cnnp_model_gelu_t* const self = (ccv_cnnp_model_gelu_t*)super;
1994 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1995 ccv_nnc_tensor_param_t output_params;
1996 const ccv_nnc_cmd_t gelu = CMD_GELU_FORWARD(self->tanh)ccv_nnc_cmd(CCV_NNC_GELU_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.gelu={.tanh=self->tanh}}, 0);
1997 ccv_nnc_hint_tensor_auto(gelu, (ccv_nnc_tensor_param_t []){
1998 params,
1999 }, 1, ccv_nnc_no_hint, &output_params, 1);
2000 const ccv_nnc_tensor_symbol_t gelu_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2001 ccv_nnc_graph_exec_symbol_new(graph, gelu, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(gelu_output)(const ccv_nnc_tensor_symbol_t []){gelu_output}, (1 +1 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "gelu");
2002 outputs[0] = gelu_output;
2003}
2004
2005static ccv_cnnp_model_t* _ccv_cnnp_gelu_copy(const ccv_cnnp_model_t* const self, void* const context);
2006
2007static const ccv_cnnp_model_vtab_t ccv_cnnp_gelu_isa = {
2008 .build = _ccv_cnnp_gelu_build,
2009 .copy = _ccv_cnnp_gelu_copy,
2010};
2011
2012ccv_cnnp_model_t* ccv_cnnp_gelu(const int tanh, const char* const name)
2013{
2014 ccv_cnnp_model_gelu_t* const model_gelu = (ccv_cnnp_model_gelu_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_gelu_t));
2015 model_gelu->super.isa = &ccv_cnnp_gelu_isa;
2016 model_gelu->super.input_size = 1;
2017 model_gelu->super.outputs = &model_gelu->output;
2018 model_gelu->super.output_size = 1;
2019 model_gelu->tanh = tanh;
2020 ccv_cnnp_model_copy_name(&model_gelu->super, name);
2021 return (ccv_cnnp_model_t*)model_gelu;
2022}
2023
2024static ccv_cnnp_model_t* _ccv_cnnp_gelu_copy(const ccv_cnnp_model_t* const super, void* const context)
2025{
2026 ccv_cnnp_model_gelu_t* const self = (ccv_cnnp_model_gelu_t*)super;
2027 return ccv_cnnp_gelu(self->tanh, self->super.name);
2028}
2029
2030// MARK - Leaky ReLU Layer
2031
2032typedef struct {
2033 ccv_cnnp_model_t super;
2034 ccv_nnc_tensor_symbol_t output;
2035 float negative_slope;
2036} ccv_cnnp_model_leaky_relu_t;
2037
2038static void _ccv_cnnp_leaky_relu_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
2039{
2040 PRINT(CCV_CLI_VERBOSE, "[cnnp_leaky_relu_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_leaky_relu_build] -\n"); fflush(stdout); } }
while (0);
2041 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 2041, __extension__ __PRETTY_FUNCTION__); }));
2042 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 2042, __extension__ __PRETTY_FUNCTION__
); }));
2043 ccv_cnnp_model_leaky_relu_t* const self = (ccv_cnnp_model_leaky_relu_t*)super;
2044 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2045 ccv_nnc_tensor_param_t output_params;
2046 const ccv_nnc_cmd_t leaky_relu = CMD_LEAKY_RELU_FORWARD(self->negative_slope)ccv_nnc_cmd(CCV_NNC_LEAKY_RELU_FORWARD, 0, (ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.leaky_relu={.negative_slope=self->
negative_slope}}, 0);
2047 ccv_nnc_hint_tensor_auto(leaky_relu, (ccv_nnc_tensor_param_t []){
2048 params,
2049 }, 1, ccv_nnc_no_hint, &output_params, 1);
2050 const ccv_nnc_tensor_symbol_t leaky_relu_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2051 ccv_nnc_graph_exec_symbol_new(graph, leaky_relu, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(leaky_relu_output)(const ccv_nnc_tensor_symbol_t []){leaky_relu_output}, (1 +1 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
-1), "leaky_relu");
2052 outputs[0] = leaky_relu_output;
2053}
2054
2055static ccv_cnnp_model_t* _ccv_cnnp_leaky_relu_copy(const ccv_cnnp_model_t* const self, void* const context);
2056
2057static const ccv_cnnp_model_vtab_t ccv_cnnp_leaky_relu_isa = {
2058 .build = _ccv_cnnp_leaky_relu_build,
2059 .copy = _ccv_cnnp_leaky_relu_copy,
2060};
2061
2062ccv_cnnp_model_t* ccv_cnnp_leaky_relu(const float negative_slope, const char* const name)
2063{
2064 ccv_cnnp_model_leaky_relu_t* const model_leaky_relu = (ccv_cnnp_model_leaky_relu_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_leaky_relu_t));
2065 model_leaky_relu->super.isa = &ccv_cnnp_leaky_relu_isa;
2066 model_leaky_relu->super.input_size = 1;
2067 model_leaky_relu->super.outputs = &model_leaky_relu->output;
2068 model_leaky_relu->super.output_size = 1;
2069 model_leaky_relu->negative_slope = negative_slope;
2070 ccv_cnnp_model_copy_name(&model_leaky_relu->super, name);
2071 return (ccv_cnnp_model_t*)model_leaky_relu;
2072}
2073
2074static ccv_cnnp_model_t* _ccv_cnnp_leaky_relu_copy(const ccv_cnnp_model_t* const super, void* const context)
2075{
2076 ccv_cnnp_model_leaky_relu_t* const self = (ccv_cnnp_model_leaky_relu_t*)super;
2077 return ccv_cnnp_leaky_relu(self->negative_slope, self->super.name);
2078}
2079
2080// MARK - Softmax Layer
2081
2082typedef struct {
2083 ccv_cnnp_model_t super;
2084 ccv_nnc_tensor_symbol_t output;
2085} ccv_cnnp_model_softmax_t;
2086
2087static void _ccv_cnnp_softmax_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
2088{
2089 PRINT(CCV_CLI_VERBOSE, "[cnnp_softmax_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_softmax_build] -\n"); fflush(stdout); } } while
(0);
2090 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 2090, __extension__ __PRETTY_FUNCTION__); }));
2091 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 2091, __extension__ __PRETTY_FUNCTION__
); }));
2092 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2093 ccv_nnc_tensor_param_t output_params;
2094 const ccv_nnc_cmd_t softmax = CMD_SOFTMAX_FORWARD()ccv_nnc_cmd(CCV_NNC_SOFTMAX_FORWARD, 0, ccv_nnc_cmd_auto, 0);
2095 ccv_nnc_hint_tensor_auto(softmax, (ccv_nnc_tensor_param_t []){
2096 params,
2097 }, 1, ccv_nnc_no_hint, &output_params, 1);
2098 const ccv_nnc_tensor_symbol_t softmax_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2099 ccv_nnc_graph_exec_symbol_new(graph, softmax, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(softmax_output)(const ccv_nnc_tensor_symbol_t []){softmax_output}, (1 +1 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1
), "softmax");
2100 outputs[0] = softmax_output;
2101}
2102
2103static ccv_cnnp_model_t* _ccv_cnnp_softmax_copy(const ccv_cnnp_model_t* const self, void* const context);
2104
2105static const ccv_cnnp_model_vtab_t ccv_cnnp_softmax_isa = {
2106 .build = _ccv_cnnp_softmax_build,
2107 .copy = _ccv_cnnp_softmax_copy,
2108};
2109
2110ccv_cnnp_model_t* ccv_cnnp_softmax(const char* const name)
2111{
2112 ccv_cnnp_model_softmax_t* const model_softmax = (ccv_cnnp_model_softmax_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_softmax_t));
2113 model_softmax->super.isa = &ccv_cnnp_softmax_isa;
2114 model_softmax->super.input_size = 1;
2115 model_softmax->super.outputs = &model_softmax->output;
2116 model_softmax->super.output_size = 1;
2117 ccv_cnnp_model_copy_name(&model_softmax->super, name);
2118 return (ccv_cnnp_model_t*)model_softmax;
2119}
2120
2121static ccv_cnnp_model_t* _ccv_cnnp_softmax_copy(const ccv_cnnp_model_t* const self, void* const context)
2122{
2123 return ccv_cnnp_softmax(self->name);
2124}
2125
2126// MARK - Add Layer
2127
2128typedef struct {
2129 ccv_cnnp_model_t super;
2130 float p;
2131 float q;
2132 ccv_nnc_tensor_symbol_t output;
2133} ccv_cnnp_model_add_t;
2134
2135static void _ccv_cnnp_add_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
2136{
2137 PRINT(CCV_CLI_VERBOSE, "[cnnp_add_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_add_build] -\n"); fflush(stdout); } } while (
0);
2138 const ccv_cnnp_model_add_t* const self = (const ccv_cnnp_model_add_t*)super;
2139 assert(input_size == 2)((void) sizeof ((input_size == 2) ? 1 : 0), __extension__ ({ if
(input_size == 2) ; else __assert_fail ("input_size == 2", "ccv_cnnp_model_addons.c"
, 2139, __extension__ __PRETTY_FUNCTION__); }));
2140 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 2140, __extension__ __PRETTY_FUNCTION__
); }));
2141 ccv_nnc_tensor_param_t input_params[2];
2142 int i;
2143 for (i = 0; i < 2; i++)
2144 input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
2145 ccv_nnc_tensor_param_t output_params;
2146 const ccv_nnc_cmd_t add = CMD_ADD_FORWARD(self->p, self->q)ccv_nnc_cmd(CCV_NNC_ADD_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={self->p, self->q}}}, 0);
2147 ccv_nnc_hint_tensor_auto(add, input_params, 2, ccv_nnc_no_hint, &output_params, 1);
2148 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2149 ccv_nnc_graph_exec_symbol_new(graph, add, inputs, input_size, outputs, output_size, "add");
2150}
2151
2152static ccv_cnnp_model_t* _ccv_cnnp_add_copy(const ccv_cnnp_model_t* const self, void* const context);
2153
2154static const ccv_cnnp_model_vtab_t ccv_cnnp_add_isa = {
2155 .build = _ccv_cnnp_add_build,
2156 .copy = _ccv_cnnp_add_copy,
2157};
2158
2159ccv_cnnp_model_t* ccv_cnnp_add(const float p, const float q, const char* const name)
2160{
2161 ccv_cnnp_model_add_t* const model_add = (ccv_cnnp_model_add_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_add_t));
2162 model_add->super.isa = &ccv_cnnp_add_isa;
2163 model_add->super.input_size = 2;
2164 model_add->super.outputs = &model_add->output;
2165 model_add->super.output_size = 1;
2166 model_add->p = p;
2167 model_add->q = q;
2168 ccv_cnnp_model_copy_name(&model_add->super, name);
2169 return (ccv_cnnp_model_t*)model_add;
2170}
2171
2172static ccv_cnnp_model_t* _ccv_cnnp_add_copy(const ccv_cnnp_model_t* const super, void* const context)
2173{
2174 const ccv_cnnp_model_add_t* const self = (const ccv_cnnp_model_add_t*)super;
2175 return ccv_cnnp_add(self->p, self->q, self->super.name);
2176}
2177
2178// MARK - Mul Layer
2179
2180typedef struct {
2181 ccv_cnnp_model_t super;
2182 ccv_nnc_tensor_symbol_t output;
2183 float p;
2184} ccv_cnnp_model_mul_t;
2185
2186static void _ccv_cnnp_mul_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
2187{
2188 PRINT(CCV_CLI_VERBOSE, "[cnnp_mul_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_mul_build] -\n"); fflush(stdout); } } while (
0);
2189 const ccv_cnnp_model_mul_t* const self = (const ccv_cnnp_model_mul_t*)super;
2190 assert(input_size == 2)((void) sizeof ((input_size == 2) ? 1 : 0), __extension__ ({ if
(input_size == 2) ; else __assert_fail ("input_size == 2", "ccv_cnnp_model_addons.c"
, 2190, __extension__ __PRETTY_FUNCTION__); }));
2191 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 2191, __extension__ __PRETTY_FUNCTION__
); }));
2192 ccv_nnc_tensor_param_t input_params[2];
2193 int i;
2194 for (i = 0; i < 2; i++)
2195 input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
2196 ccv_nnc_tensor_param_t output_params;
2197 const ccv_nnc_cmd_t mul = CMD_MUL_FORWARD(self->p)ccv_nnc_cmd(CCV_NNC_MUL_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={self->p,}}}, 0);
2198 ccv_nnc_hint_tensor_auto(mul, input_params, 2, ccv_nnc_no_hint, &output_params, 1);
2199 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2200 ccv_nnc_graph_exec_symbol_new(graph, mul, inputs, input_size, outputs, output_size, "mul");
2201}
2202
2203static ccv_cnnp_model_t* _ccv_cnnp_mul_copy(const ccv_cnnp_model_t* const self, void* const context);
2204
2205static const ccv_cnnp_model_vtab_t ccv_cnnp_mul_isa = {
2206 .build = _ccv_cnnp_mul_build,
2207 .copy = _ccv_cnnp_mul_copy,
2208};
2209
2210ccv_cnnp_model_t* ccv_cnnp_mul(const float p, const char* const name)
2211{
2212 ccv_cnnp_model_mul_t* const model_mul = (ccv_cnnp_model_mul_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_mul_t));
2213 model_mul->super.isa = &ccv_cnnp_mul_isa;
2214 model_mul->super.input_size = 2;
2215 model_mul->super.outputs = &model_mul->output;
2216 model_mul->super.output_size = 1;
2217 model_mul->p = p;
2218 ccv_cnnp_model_copy_name(&model_mul->super, name);
2219 return (ccv_cnnp_model_t*)model_mul;
2220}
2221
2222static ccv_cnnp_model_t* _ccv_cnnp_mul_copy(const ccv_cnnp_model_t* const super, void* const context)
2223{
2224 const ccv_cnnp_model_mul_t* const self = (const ccv_cnnp_model_mul_t*)super;
2225 return ccv_cnnp_mul(self->p, self->super.name);
2226}
2227
2228// MARK - Scalar Mul Layer
2229
2230typedef struct {
2231 ccv_cnnp_model_t super;
2232 ccv_nnc_tensor_symbol_t output;
2233 float a;
2234} ccv_cnnp_model_scalar_mul_t;
2235
2236static void _ccv_cnnp_scalar_mul_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
2237{
2238 PRINT(CCV_CLI_VERBOSE, "[cnnp_scalar_mul_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_scalar_mul_build] -\n"); fflush(stdout); } }
while (0);
2239 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 2239, __extension__ __PRETTY_FUNCTION__); }));
2240 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 2240, __extension__ __PRETTY_FUNCTION__
); }));
2241 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2242 ccv_nnc_tensor_param_t output_params;
2243 ccv_cnnp_model_scalar_mul_t* const self = (ccv_cnnp_model_scalar_mul_t*)super;
2244 const ccv_nnc_cmd_t scalar_mul = CMD_SCALAR_MUL_FORWARD(self->a)ccv_nnc_cmd(CCV_NNC_SCALAR_MUL_FORWARD, 0, (ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.blas={.a={self->a,}}}, 0);
2245 ccv_nnc_hint_tensor_auto(scalar_mul, (ccv_nnc_tensor_param_t []){
2246 params,
2247 }, 1, ccv_nnc_no_hint, &output_params, 1);
2248 const ccv_nnc_tensor_symbol_t scalar_mul_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2249 ccv_nnc_graph_exec_symbol_new(graph, scalar_mul, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(scalar_mul_output)(const ccv_nnc_tensor_symbol_t []){scalar_mul_output}, (1 +1 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
-1), "scalar_mul");
2250 outputs[0] = scalar_mul_output;
2251}
2252
2253static ccv_cnnp_model_t* _ccv_cnnp_scalar_mul_copy(const ccv_cnnp_model_t* const super, void* const context);
2254
2255static const ccv_cnnp_model_vtab_t ccv_cnnp_scalar_mul_isa = {
2256 .build = _ccv_cnnp_scalar_mul_build,
2257 .copy = _ccv_cnnp_scalar_mul_copy,
2258};
2259
2260ccv_cnnp_model_t* ccv_cnnp_scalar_mul(const float a, const char* const name)
2261{
2262 ccv_cnnp_model_scalar_mul_t* const model_scalar_mul = (ccv_cnnp_model_scalar_mul_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_scalar_mul_t));
2263 model_scalar_mul->super.isa = &ccv_cnnp_scalar_mul_isa;
2264 model_scalar_mul->super.input_size = 1;
2265 model_scalar_mul->super.outputs = &model_scalar_mul->output;
2266 model_scalar_mul->super.output_size = 1;
2267 model_scalar_mul->a = a;
2268 ccv_cnnp_model_copy_name(&model_scalar_mul->super, name);
2269 return (ccv_cnnp_model_t*)model_scalar_mul;
2270}
2271
2272static ccv_cnnp_model_t* _ccv_cnnp_scalar_mul_copy(const ccv_cnnp_model_t* const super, void* const context)
2273{
2274 const ccv_cnnp_model_scalar_mul_t* const self = (const ccv_cnnp_model_scalar_mul_t*)super;
2275 return ccv_cnnp_scalar_mul(self->a, self->super.name);
2276}
2277
2278// MARK - Div Layer
2279
2280typedef struct {
2281 ccv_cnnp_model_t super;
2282 ccv_nnc_tensor_symbol_t output;
2283 int reciprocal;
2284} ccv_cnnp_model_div_t;
2285
2286static void _ccv_cnnp_div_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
2287{
2288 const ccv_cnnp_model_div_t* const self = (const ccv_cnnp_model_div_t*)super;
2289 PRINT(CCV_CLI_VERBOSE, "[cnnp_div_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_div_build] -\n"); fflush(stdout); } } while (
0);
2290 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 2290, __extension__ __PRETTY_FUNCTION__
); }));
2291 ccv_nnc_tensor_param_t input_params[2];
2292 int i;
2293 ccv_nnc_tensor_param_t output_params;
2294 const ccv_nnc_cmd_t div = CMD_EWDIV_FORWARD()ccv_nnc_cmd(CCV_NNC_EWDIV_FORWARD, 0, ccv_nnc_cmd_auto, 0);
2295 if (self->reciprocal)
2296 {
2297 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 2297, __extension__ __PRETTY_FUNCTION__); }));
2298 input_params[0] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2299 input_params[1] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2300 ccv_nnc_hint_tensor_auto(div, input_params, 2, ccv_nnc_no_hint, &output_params, 1);
2301 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2302 ccv_nnc_graph_exec_symbol_new(graph, div, TENSOR_SYMBOL_LIST(NO_TENSOR_SYMBOL, inputs[0])(const ccv_nnc_tensor_symbol_t []){(const ccv_nnc_tensor_symbol_t
){.d = CCV_NNC_NO_TENSOR_SYMBOL}, inputs[0]}, (1 +1 +1 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), outputs, output_size, "div");
2303 } else {
2304 assert(input_size == 2)((void) sizeof ((input_size == 2) ? 1 : 0), __extension__ ({ if
(input_size == 2) ; else __assert_fail ("input_size == 2", "ccv_cnnp_model_addons.c"
, 2304, __extension__ __PRETTY_FUNCTION__); }));
2305 for (i = 0; i < 2; i++)
2306 input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
2307 ccv_nnc_hint_tensor_auto(div, input_params, input_size, ccv_nnc_no_hint, &output_params, 1);
2308 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2309 ccv_nnc_graph_exec_symbol_new(graph, div, inputs, input_size, outputs, output_size, "div");
2310 }
2311}
2312
2313static ccv_cnnp_model_t* _ccv_cnnp_div_copy(const ccv_cnnp_model_t* const self, void* const context);
2314
2315static const ccv_cnnp_model_vtab_t ccv_cnnp_div_isa = {
2316 .build = _ccv_cnnp_div_build,
2317 .copy = _ccv_cnnp_div_copy,
2318};
2319
2320ccv_cnnp_model_t* ccv_cnnp_div(const int reciprocal, const char* const name)
2321{
2322 ccv_cnnp_model_div_t* const model_div = (ccv_cnnp_model_div_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_div_t));
2323 model_div->super.isa = &ccv_cnnp_div_isa;
2324 model_div->super.input_size = reciprocal ? 1 : 2;
2325 model_div->super.outputs = &model_div->output;
2326 model_div->super.output_size = 1;
2327 model_div->reciprocal = reciprocal;
2328 ccv_cnnp_model_copy_name(&model_div->super, name);
2329 return (ccv_cnnp_model_t*)model_div;
2330}
2331
2332static ccv_cnnp_model_t* _ccv_cnnp_div_copy(const ccv_cnnp_model_t* const super, void* const context)
2333{
2334 const ccv_cnnp_model_div_t* const self = (const ccv_cnnp_model_div_t*)super;
2335 return ccv_cnnp_div(self->reciprocal, self->super.name);
2336}
2337
2338// MARK - Sqrt Layer
2339
2340typedef struct {
2341 ccv_cnnp_model_t super;
2342 ccv_nnc_tensor_symbol_t output;
2343} ccv_cnnp_model_sqrt_t;
2344
2345static void _ccv_cnnp_sqrt_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
2346{
2347 PRINT(CCV_CLI_VERBOSE, "[cnnp_sqrt_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_sqrt_build] -\n"); fflush(stdout); } } while
(0);
2348 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 2348, __extension__ __PRETTY_FUNCTION__
); }));
2349 ccv_nnc_tensor_param_t input_params[1];
2350 ccv_nnc_tensor_param_t output_params;
2351 const ccv_nnc_cmd_t sqrt = CMD_EWSQRT_FORWARD()ccv_nnc_cmd(CCV_NNC_EWSQRT_FORWARD, 0, ccv_nnc_cmd_auto, 0);
2352 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 2352, __extension__ __PRETTY_FUNCTION__); }));
2353 input_params[0] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2354 ccv_nnc_hint_tensor_auto(sqrt, input_params, 1, ccv_nnc_no_hint, &output_params, 1);
2355 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2356 ccv_nnc_graph_exec_symbol_new(graph, sqrt, inputs, 1, outputs, output_size, "sqrt");
2357}
2358
2359static ccv_cnnp_model_t* _ccv_cnnp_sqrt_copy(const ccv_cnnp_model_t* const self, void* const context);
2360
2361static const ccv_cnnp_model_vtab_t ccv_cnnp_sqrt_isa = {
2362 .build = _ccv_cnnp_sqrt_build,
2363 .copy = _ccv_cnnp_sqrt_copy,
2364};
2365
2366ccv_cnnp_model_t* ccv_cnnp_sqrt(const char* const name)
2367{
2368 ccv_cnnp_model_sqrt_t* const model_sqrt = (ccv_cnnp_model_sqrt_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_sqrt_t));
2369 model_sqrt->super.isa = &ccv_cnnp_sqrt_isa;
2370 model_sqrt->super.input_size = 1;
2371 model_sqrt->super.outputs = &model_sqrt->output;
2372 model_sqrt->super.output_size = 1;
2373 ccv_cnnp_model_copy_name(&model_sqrt->super, name);
2374 return (ccv_cnnp_model_t*)model_sqrt;
2375}
2376
2377static ccv_cnnp_model_t* _ccv_cnnp_sqrt_copy(const ccv_cnnp_model_t* const super, void* const context)
2378{
2379 const ccv_cnnp_model_sqrt_t* const self = (const ccv_cnnp_model_sqrt_t*)super;
2380 return ccv_cnnp_sqrt(self->super.name);
2381}
2382
2383// MARK - Signed Sqrt Layer
2384
2385typedef struct {
2386 ccv_cnnp_model_t super;
2387 ccv_nnc_tensor_symbol_t output;
2388 float minimum_magnitude;
2389} ccv_cnnp_model_signed_sqrt_t;
2390
2391static void _ccv_cnnp_signed_sqrt_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
2392{
2393 PRINT(CCV_CLI_VERBOSE, "[cnnp_signed_sqrt_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_signed_sqrt_build] -\n"); fflush(stdout); } }
while (0);
2394 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 2394, __extension__ __PRETTY_FUNCTION__); }));
2395 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 2395, __extension__ __PRETTY_FUNCTION__
); }));
2396 const ccv_cnnp_model_signed_sqrt_t* const self = (const ccv_cnnp_model_signed_sqrt_t*)super;
2397 const ccv_nnc_cmd_t cmd = CMD_SIGNED_SQRT_FORWARD(self->minimum_magnitude)ccv_nnc_cmd(CCV_NNC_SIGNED_SQRT_FORWARD, 0, ((ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.signed_sqrt={.minimum_magnitude=(self
->minimum_magnitude)}}), 0);
2398 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2399 ccv_nnc_tensor_param_t output_params;
2400 ccv_nnc_hint_tensor_auto(cmd, &params, 1, ccv_nnc_no_hint, &output_params, 1);
2401 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2402 ccv_nnc_graph_exec_symbol_new(graph, cmd, inputs, input_size, outputs, output_size, "signed_sqrt");
2403}
2404
2405static ccv_cnnp_model_t* _ccv_cnnp_signed_sqrt_copy(const ccv_cnnp_model_t* const self, void* const context);
2406
2407static const ccv_cnnp_model_vtab_t ccv_cnnp_signed_sqrt_isa = {
2408 .build = _ccv_cnnp_signed_sqrt_build,
2409 .copy = _ccv_cnnp_signed_sqrt_copy,
2410};
2411
2412ccv_cnnp_model_t* ccv_cnnp_signed_sqrt(const float minimum_magnitude, const char* const name)
2413{
2414 assert(minimum_magnitude > 0 && isfinite(minimum_magnitude))((void) sizeof ((minimum_magnitude > 0 && __builtin_isfinite
(minimum_magnitude)) ? 1 : 0), __extension__ ({ if (minimum_magnitude
> 0 && __builtin_isfinite (minimum_magnitude)) ; else
__assert_fail ("minimum_magnitude > 0 && isfinite(minimum_magnitude)"
, "ccv_cnnp_model_addons.c", 2414, __extension__ __PRETTY_FUNCTION__
); }));
2415 ccv_cnnp_model_signed_sqrt_t* const model = (ccv_cnnp_model_signed_sqrt_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_signed_sqrt_t));
2416 model->super.isa = &ccv_cnnp_signed_sqrt_isa;
2417 model->super.input_size = 1;
2418 model->super.outputs = &model->output;
2419 model->super.output_size = 1;
2420 model->minimum_magnitude = minimum_magnitude;
2421 ccv_cnnp_model_copy_name(&model->super, name);
2422 return (ccv_cnnp_model_t*)model;
2423}
2424
2425static ccv_cnnp_model_t* _ccv_cnnp_signed_sqrt_copy(const ccv_cnnp_model_t* const super, void* const context)
2426{
2427 const ccv_cnnp_model_signed_sqrt_t* const self = (const ccv_cnnp_model_signed_sqrt_t*)super;
2428 return ccv_cnnp_signed_sqrt(self->minimum_magnitude, self->super.name);
2429}
2430
2431// MARK - Log Layer
2432
2433typedef struct {
2434 ccv_cnnp_model_t super;
2435 ccv_nnc_tensor_symbol_t output;
2436} ccv_cnnp_model_log_t;
2437
2438static void _ccv_cnnp_log_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
2439{
2440 PRINT(CCV_CLI_VERBOSE, "[cnnp_log_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_log_build] -\n"); fflush(stdout); } } while (
0);
2441 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 2441, __extension__ __PRETTY_FUNCTION__
); }));
2442 ccv_nnc_tensor_param_t input_params[1];
2443 ccv_nnc_tensor_param_t output_params;
2444 const ccv_nnc_cmd_t log = CMD_EWLOG_FORWARD()ccv_nnc_cmd(CCV_NNC_EWLOG_FORWARD, 0, ccv_nnc_cmd_auto, 0);
2445 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 2445, __extension__ __PRETTY_FUNCTION__); }));
2446 input_params[0] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2447 ccv_nnc_hint_tensor_auto(log, input_params, 1, ccv_nnc_no_hint, &output_params, 1);
2448 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2449 ccv_nnc_graph_exec_symbol_new(graph, log, inputs, 1, outputs, output_size, "log");
2450}
2451
2452static ccv_cnnp_model_t* _ccv_cnnp_log_copy(const ccv_cnnp_model_t* const self, void* const context);
2453
2454static const ccv_cnnp_model_vtab_t ccv_cnnp_log_isa = {
2455 .build = _ccv_cnnp_log_build,
2456 .copy = _ccv_cnnp_log_copy,
2457};
2458
2459ccv_cnnp_model_t* ccv_cnnp_log(const char* const name)
2460{
2461 ccv_cnnp_model_log_t* const model_log = (ccv_cnnp_model_log_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_log_t));
2462 model_log->super.isa = &ccv_cnnp_log_isa;
2463 model_log->super.input_size = 1;
2464 model_log->super.outputs = &model_log->output;
2465 model_log->super.output_size = 1;
2466 ccv_cnnp_model_copy_name(&model_log->super, name);
2467 return (ccv_cnnp_model_t*)model_log;
2468}
2469
2470static ccv_cnnp_model_t* _ccv_cnnp_log_copy(const ccv_cnnp_model_t* const super, void* const context)
2471{
2472 return ccv_cnnp_log(super->name);
2473}
2474
2475// MARK - Pow Layer
2476
2477typedef struct {
2478 ccv_cnnp_model_t super;
2479 ccv_nnc_tensor_symbol_t output;
2480 ccv_nnc_cmd_param_t params;
2481} ccv_cnnp_model_pow_t;
2482
2483static void _ccv_cnnp_pow_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
2484{
2485 ccv_cnnp_model_pow_t* const self = (ccv_cnnp_model_pow_t*)super;
2486 PRINT(CCV_CLI_VERBOSE, "[cnnp_pow_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_pow_build] -\n"); fflush(stdout); } } while (
0);
2487 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 2487, __extension__ __PRETTY_FUNCTION__); }));
2488 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 2488, __extension__ __PRETTY_FUNCTION__
); }));
2489 ccv_nnc_tensor_param_t input_params[1];
2490 input_params[0] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2491 ccv_nnc_tensor_param_t output_params;
2492 const ccv_nnc_cmd_t pow = ccv_nnc_cmd(CCV_NNC_EWPOW_FORWARD, 0, self->params, 0);
2493 ccv_nnc_hint_tensor_auto(pow, input_params, 1, ccv_nnc_no_hint, &output_params, 1);
2494 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2495 ccv_nnc_graph_exec_symbol_new(graph, pow, inputs, input_size, outputs, output_size, "pow");
2496}
2497
2498static ccv_cnnp_model_t* _ccv_cnnp_pow_copy(const ccv_cnnp_model_t* const self, void* const context);
2499
2500static const ccv_cnnp_model_vtab_t ccv_cnnp_pow_isa = {
2501 .build = _ccv_cnnp_pow_build,
2502 .copy = _ccv_cnnp_pow_copy,
2503};
2504
2505ccv_cnnp_model_t* ccv_cnnp_pow(const float exponent, const char* const name)
2506{
2507 ccv_cnnp_model_pow_t* const model_pow = (ccv_cnnp_model_pow_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_pow_t));
2508 model_pow->super.isa = &ccv_cnnp_pow_isa;
2509 model_pow->super.input_size = 1;
2510 model_pow->super.outputs = &model_pow->output;
2511 model_pow->super.output_size = 1;
2512 model_pow->params = (ccv_nnc_cmd_param_t){
2513 .size = {
2514 .dim = { 1, 1, 1 }
2515 },
2516 .pow = {
2517 .exponent = exponent,
2518 },
2519 };
2520 ccv_cnnp_model_copy_name(&model_pow->super, name);
2521 return (ccv_cnnp_model_t*)model_pow;
2522}
2523
2524static ccv_cnnp_model_t* _ccv_cnnp_pow_copy(const ccv_cnnp_model_t* const super, void* const context)
2525{
2526 const ccv_cnnp_model_pow_t* const self = (const ccv_cnnp_model_pow_t*)super;
2527 return ccv_cnnp_pow(self->params.pow.exponent, super->name);
2528}
2529
2530// MARK - Sin Layer
2531
2532typedef struct {
2533 ccv_cnnp_model_t super;
2534 ccv_nnc_tensor_symbol_t output;
2535} ccv_cnnp_model_sin_t;
2536
2537static void _ccv_cnnp_sin_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
2538{
2539 PRINT(CCV_CLI_VERBOSE, "[cnnp_sin_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_sin_build] -\n"); fflush(stdout); } } while (
0);
2540 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 2540, __extension__ __PRETTY_FUNCTION__
); }));
2541 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 2541, __extension__ __PRETTY_FUNCTION__); }));
2542 ccv_nnc_tensor_param_t input_params[1];
2543 ccv_nnc_tensor_param_t output_params;
2544 const ccv_nnc_cmd_t sin = CMD_EWSIN_FORWARD()ccv_nnc_cmd(CCV_NNC_EWSIN_FORWARD, 0, ccv_nnc_cmd_auto, 0);
2545 input_params[0] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2546 ccv_nnc_hint_tensor_auto(sin, input_params, 1, ccv_nnc_no_hint, &output_params, 1);
2547 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2548 ccv_nnc_graph_exec_symbol_new(graph, sin, inputs, 1, outputs, output_size, "sin");
2549}
2550
2551static ccv_cnnp_model_t* _ccv_cnnp_sin_copy(const ccv_cnnp_model_t* const self, void* const context);
2552
2553static const ccv_cnnp_model_vtab_t ccv_cnnp_sin_isa = {
2554 .build = _ccv_cnnp_sin_build,
2555 .copy = _ccv_cnnp_sin_copy,
2556};
2557
2558ccv_cnnp_model_t* ccv_cnnp_sin(const char* const name)
2559{
2560 ccv_cnnp_model_sin_t* const model_sin = (ccv_cnnp_model_sin_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_sin_t));
2561 model_sin->super.isa = &ccv_cnnp_sin_isa;
2562 model_sin->super.input_size = 1;
2563 model_sin->super.outputs = &model_sin->output;
2564 model_sin->super.output_size = 1;
2565 ccv_cnnp_model_copy_name(&model_sin->super, name);
2566 return (ccv_cnnp_model_t*)model_sin;
2567}
2568
2569static ccv_cnnp_model_t* _ccv_cnnp_sin_copy(const ccv_cnnp_model_t* const super, void* const context)
2570{
2571 return ccv_cnnp_sin(super->name);
2572}
2573
2574// MARK - Cos Layer
2575
2576typedef struct {
2577 ccv_cnnp_model_t super;
2578 ccv_nnc_tensor_symbol_t output;
2579} ccv_cnnp_model_cos_t;
2580
2581static void _ccv_cnnp_cos_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
2582{
2583 PRINT(CCV_CLI_VERBOSE, "[cnnp_cos_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_cos_build] -\n"); fflush(stdout); } } while (
0);
2584 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 2584, __extension__ __PRETTY_FUNCTION__
); }));
2585 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 2585, __extension__ __PRETTY_FUNCTION__); }));
2586 ccv_nnc_tensor_param_t input_params[1];
2587 ccv_nnc_tensor_param_t output_params;
2588 const ccv_nnc_cmd_t cos = CMD_EWCOS_FORWARD()ccv_nnc_cmd(CCV_NNC_EWCOS_FORWARD, 0, ccv_nnc_cmd_auto, 0);
2589 input_params[0] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2590 ccv_nnc_hint_tensor_auto(cos, input_params, 1, ccv_nnc_no_hint, &output_params, 1);
2591 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2592 ccv_nnc_graph_exec_symbol_new(graph, cos, inputs, 1, outputs, output_size, "cos");
2593}
2594
2595static ccv_cnnp_model_t* _ccv_cnnp_cos_copy(const ccv_cnnp_model_t* const self, void* const context);
2596
2597static const ccv_cnnp_model_vtab_t ccv_cnnp_cos_isa = {
2598 .build = _ccv_cnnp_cos_build,
2599 .copy = _ccv_cnnp_cos_copy,
2600};
2601
2602ccv_cnnp_model_t* ccv_cnnp_cos(const char* const name)
2603{
2604 ccv_cnnp_model_cos_t* const model_cos = (ccv_cnnp_model_cos_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_cos_t));
2605 model_cos->super.isa = &ccv_cnnp_cos_isa;
2606 model_cos->super.input_size = 1;
2607 model_cos->super.outputs = &model_cos->output;
2608 model_cos->super.output_size = 1;
2609 ccv_cnnp_model_copy_name(&model_cos->super, name);
2610 return (ccv_cnnp_model_t*)model_cos;
2611}
2612
2613static ccv_cnnp_model_t* _ccv_cnnp_cos_copy(const ccv_cnnp_model_t* const super, void* const context)
2614{
2615 return ccv_cnnp_cos(super->name);
2616}
2617
2618// MARK - Rotate Half Layer
2619
2620typedef struct {
2621 ccv_cnnp_model_t super;
2622 ccv_nnc_tensor_symbol_t output;
2623} ccv_cnnp_model_rotate_half_t;
2624
2625static void _ccv_cnnp_rotate_half_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
2626{
2627 PRINT(CCV_CLI_VERBOSE, "[cnnp_rotate_half_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_rotate_half_build] -\n"); fflush(stdout); } }
while (0);
2628 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 2628, __extension__ __PRETTY_FUNCTION__); }));
2629 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 2629, __extension__ __PRETTY_FUNCTION__
); }));
2630 ccv_nnc_tensor_param_t input_params[1];
2631 ccv_nnc_tensor_param_t output_params;
2632 const ccv_nnc_cmd_t rotate_half = CMD_ROTATE_HALF_FORWARD()ccv_nnc_cmd(CCV_NNC_ROTATE_HALF_FORWARD, 0, (ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}}}, 0);
2633 input_params[0] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2634 ccv_nnc_hint_tensor_auto(rotate_half, input_params, 1, ccv_nnc_no_hint, &output_params, 1);
2635 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2636 ccv_nnc_graph_exec_symbol_new(graph, rotate_half, inputs, 1, outputs, output_size, "rotate_half");
2637}
2638
2639static ccv_cnnp_model_t* _ccv_cnnp_rotate_half_copy(const ccv_cnnp_model_t* const self, void* const context);
2640
2641static const ccv_cnnp_model_vtab_t ccv_cnnp_rotate_half_isa = {
2642 .build = _ccv_cnnp_rotate_half_build,
2643 .copy = _ccv_cnnp_rotate_half_copy,
2644};
2645
2646ccv_cnnp_model_t* ccv_cnnp_rotate_half(const char* const name)
2647{
2648 ccv_cnnp_model_rotate_half_t* const model_rotate_half = (ccv_cnnp_model_rotate_half_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_rotate_half_t));
2649 model_rotate_half->super.isa = &ccv_cnnp_rotate_half_isa;
2650 model_rotate_half->super.input_size = 1;
2651 model_rotate_half->super.outputs = &model_rotate_half->output;
2652 model_rotate_half->super.output_size = 1;
2653 ccv_cnnp_model_copy_name(&model_rotate_half->super, name);
2654 return (ccv_cnnp_model_t*)model_rotate_half;
2655}
2656
2657static ccv_cnnp_model_t* _ccv_cnnp_rotate_half_copy(const ccv_cnnp_model_t* const super, void* const context)
2658{
2659 return ccv_cnnp_rotate_half(super->name);
2660}
2661
2662// MARK - Walsh-Hadamard Transform Layer
2663
2664typedef struct {
2665 ccv_cnnp_model_t super;
2666 ccv_nnc_tensor_symbol_t output;
2667 float scale;
2668} ccv_cnnp_model_walsh_hadamard_transform_t;
2669
2670static void _ccv_cnnp_walsh_hadamard_transform_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
2671{
2672 PRINT(CCV_CLI_VERBOSE, "[cnnp_walsh_hadamard_transform_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_walsh_hadamard_transform_build] -\n"); fflush
(stdout); } } while (0);
2673 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 2673, __extension__ __PRETTY_FUNCTION__); }));
2674 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 2674, __extension__ __PRETTY_FUNCTION__
); }));
2675 const ccv_cnnp_model_walsh_hadamard_transform_t* const self = (const ccv_cnnp_model_walsh_hadamard_transform_t*)super;
2676 ccv_nnc_tensor_param_t input_params[1];
2677 ccv_nnc_tensor_param_t output_params;
2678 const ccv_nnc_cmd_t walsh_hadamard_transform = CMD_WALSH_HADAMARD_TRANSFORM_FORWARD(self->scale)ccv_nnc_cmd(CCV_NNC_WALSH_HADAMARD_TRANSFORM_FORWARD, 0, ((ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.walsh_hadamard_transform={.scale=(self
->scale)}}), 0);
2679 input_params[0] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2680 ccv_nnc_hint_tensor_auto(walsh_hadamard_transform, input_params, 1, ccv_nnc_no_hint, &output_params, 1);
2681 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2682 ccv_nnc_graph_exec_symbol_new(graph, walsh_hadamard_transform, inputs, 1, outputs, output_size, "walsh_hadamard_transform");
2683}
2684
2685static ccv_cnnp_model_t* _ccv_cnnp_walsh_hadamard_transform_copy(const ccv_cnnp_model_t* const self, void* const context);
2686
2687static const ccv_cnnp_model_vtab_t ccv_cnnp_walsh_hadamard_transform_isa = {
2688 .build = _ccv_cnnp_walsh_hadamard_transform_build,
2689 .copy = _ccv_cnnp_walsh_hadamard_transform_copy,
2690};
2691
2692ccv_cnnp_model_t* ccv_cnnp_walsh_hadamard_transform(const float scale, const char* const name)
2693{
2694 ccv_cnnp_model_walsh_hadamard_transform_t* const model_walsh_hadamard_transform = (ccv_cnnp_model_walsh_hadamard_transform_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_walsh_hadamard_transform_t));
2695 model_walsh_hadamard_transform->super.isa = &ccv_cnnp_walsh_hadamard_transform_isa;
2696 model_walsh_hadamard_transform->super.input_size = 1;
2697 model_walsh_hadamard_transform->super.outputs = &model_walsh_hadamard_transform->output;
2698 model_walsh_hadamard_transform->super.output_size = 1;
2699 model_walsh_hadamard_transform->scale = scale;
2700 ccv_cnnp_model_copy_name(&model_walsh_hadamard_transform->super, name);
2701 return (ccv_cnnp_model_t*)model_walsh_hadamard_transform;
2702}
2703
2704static ccv_cnnp_model_t* _ccv_cnnp_walsh_hadamard_transform_copy(const ccv_cnnp_model_t* const super, void* const context)
2705{
2706 const ccv_cnnp_model_walsh_hadamard_transform_t* const self = (const ccv_cnnp_model_walsh_hadamard_transform_t*)super;
2707 return ccv_cnnp_walsh_hadamard_transform(self->scale, self->super.name);
2708}
2709
2710// MARK - Hyper Connection Layer
2711
2712typedef struct {
2713 ccv_cnnp_model_t super;
2714 ccv_nnc_tensor_symbol_t outputs[3];
2715 int count;
2716 int sinkhorn_iterations;
2717 float epsilon;
2718 int operation;
2719} ccv_cnnp_model_hyper_connection_t;
2720
2721static void _ccv_cnnp_hyper_connection_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
2722{
2723 PRINT(CCV_CLI_VERBOSE, "[cnnp_hyper_connection_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_hyper_connection_build] -\n"); fflush(stdout
); } } while (0);
2724 const ccv_cnnp_model_hyper_connection_t* const self = (const ccv_cnnp_model_hyper_connection_t*)super;
2725 assert(input_size == (self->operation == 0 ? 3 : 4))((void) sizeof ((input_size == (self->operation == 0 ? 3 :
4)) ? 1 : 0), __extension__ ({ if (input_size == (self->operation
== 0 ? 3 : 4)) ; else __assert_fail ("input_size == (self->operation == 0 ? 3 : 4)"
, "ccv_cnnp_model_addons.c", 2725, __extension__ __PRETTY_FUNCTION__
); }));
2726 assert(output_size == (self->operation == 2 ? 1 : 3))((void) sizeof ((output_size == (self->operation == 2 ? 1 :
3)) ? 1 : 0), __extension__ ({ if (output_size == (self->
operation == 2 ? 1 : 3)) ; else __assert_fail ("output_size == (self->operation == 2 ? 1 : 3)"
, "ccv_cnnp_model_addons.c", 2726, __extension__ __PRETTY_FUNCTION__
); }));
2727 ccv_nnc_tensor_param_t input_params[4];
2728 int i;
2729 for (i = 0; i < input_size; i++)
2730 input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
2731 ccv_nnc_tensor_param_t output_params[3];
2732 const ccv_nnc_cmd_t hyper_connection = CMD_HYPER_CONNECTION_FORWARD(self->count, self->sinkhorn_iterations, self->epsilon)ccv_nnc_cmd(CCV_NNC_HYPER_CONNECTION_FORWARD, 0, ((ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.hyper_connection={.count=(self->count
),.sinkhorn_iterations=(self->sinkhorn_iterations),.epsilon
=(self->epsilon)}}), 0);
2733 ccv_nnc_hint_tensor_auto(hyper_connection, input_params, input_size, ccv_nnc_no_hint, output_params, output_size);
2734 for (i = 0; i < output_size; i++)
2735 outputs[i] = ccv_nnc_tensor_symbol_new(graph, output_params[i], 0);
2736 ccv_nnc_graph_exec_symbol_new(graph, hyper_connection, inputs, input_size, outputs, output_size, "hyper_connection");
2737}
2738
2739static ccv_cnnp_model_t* _ccv_cnnp_hyper_connection_copy(const ccv_cnnp_model_t* const self, void* const context);
2740
2741static const ccv_cnnp_model_vtab_t ccv_cnnp_hyper_connection_isa = {
2742 .build = _ccv_cnnp_hyper_connection_build,
2743 .copy = _ccv_cnnp_hyper_connection_copy,
2744};
2745
2746ccv_cnnp_model_t* ccv_cnnp_hyper_connection(const int count, const int sinkhorn_iterations, const float epsilon, const int operation, const char* const name)
2747{
2748 assert(count > 0 && count <= 16)((void) sizeof ((count > 0 && count <= 16) ? 1 :
0), __extension__ ({ if (count > 0 && count <=
16) ; else __assert_fail ("count > 0 && count <= 16"
, "ccv_cnnp_model_addons.c", 2748, __extension__ __PRETTY_FUNCTION__
); }));
2749 assert(sinkhorn_iterations > 0)((void) sizeof ((sinkhorn_iterations > 0) ? 1 : 0), __extension__
({ if (sinkhorn_iterations > 0) ; else __assert_fail ("sinkhorn_iterations > 0"
, "ccv_cnnp_model_addons.c", 2749, __extension__ __PRETTY_FUNCTION__
); }));
2750 assert(epsilon >= 0)((void) sizeof ((epsilon >= 0) ? 1 : 0), __extension__ ({ if
(epsilon >= 0) ; else __assert_fail ("epsilon >= 0", "ccv_cnnp_model_addons.c"
, 2750, __extension__ __PRETTY_FUNCTION__); }));
2751 assert(operation >= 0 && operation <= 2)((void) sizeof ((operation >= 0 && operation <=
2) ? 1 : 0), __extension__ ({ if (operation >= 0 &&
operation <= 2) ; else __assert_fail ("operation >= 0 && operation <= 2"
, "ccv_cnnp_model_addons.c", 2751, __extension__ __PRETTY_FUNCTION__
); }));
2752 ccv_cnnp_model_hyper_connection_t* const model_hyper_connection = (ccv_cnnp_model_hyper_connection_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_hyper_connection_t));
2753 model_hyper_connection->super.isa = &ccv_cnnp_hyper_connection_isa;
2754 model_hyper_connection->super.input_size = operation == 0 ? 3 : 4;
2755 model_hyper_connection->super.outputs = model_hyper_connection->outputs;
2756 model_hyper_connection->super.output_size = operation == 2 ? 1 : 3;
2757 model_hyper_connection->count = count;
2758 model_hyper_connection->sinkhorn_iterations = sinkhorn_iterations;
2759 model_hyper_connection->epsilon = epsilon;
2760 model_hyper_connection->operation = operation;
2761 ccv_cnnp_model_copy_name(&model_hyper_connection->super, name);
2762 return (ccv_cnnp_model_t*)model_hyper_connection;
2763}
2764
2765static ccv_cnnp_model_t* _ccv_cnnp_hyper_connection_copy(const ccv_cnnp_model_t* const super, void* const context)
2766{
2767 const ccv_cnnp_model_hyper_connection_t* const self = (const ccv_cnnp_model_hyper_connection_t*)super;
2768 return ccv_cnnp_hyper_connection(self->count, self->sinkhorn_iterations, self->epsilon, self->operation, self->super.name);
2769}
2770
2771// MARK - Gated Delta Layer
2772
2773typedef struct {
2774 ccv_cnnp_model_t super;
2775 ccv_nnc_tensor_symbol_t outputs[2];
2776 int log_decay;
2777 int state_checkpoint_count;
2778} ccv_cnnp_model_gated_delta_t;
2779
2780static void _ccv_cnnp_gated_delta_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
2781{
2782 PRINT(CCV_CLI_VERBOSE, "[cnnp_gated_delta_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_gated_delta_build] -\n"); fflush(stdout); } }
while (0);
2783 ccv_cnnp_model_gated_delta_t* const self = (ccv_cnnp_model_gated_delta_t*)super;
2784 assert(input_size == 6)((void) sizeof ((input_size == 6) ? 1 : 0), __extension__ ({ if
(input_size == 6) ; else __assert_fail ("input_size == 6", "ccv_cnnp_model_addons.c"
, 2784, __extension__ __PRETTY_FUNCTION__); }));
2785 assert(output_size == 2)((void) sizeof ((output_size == 2) ? 1 : 0), __extension__ ({
if (output_size == 2) ; else __assert_fail ("output_size == 2"
, "ccv_cnnp_model_addons.c", 2785, __extension__ __PRETTY_FUNCTION__
); }));
2786 ccv_nnc_tensor_param_t input_params[6];
2787 int i;
2788 for (i = 0; i < 6; i++)
2789 input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
2790 ccv_nnc_tensor_param_t output_params[2];
2791 const ccv_nnc_cmd_t gated_delta = CMD_GATED_DELTA_FORWARD(self->log_decay, self->state_checkpoint_count)ccv_nnc_cmd(CCV_NNC_GATED_DELTA_FORWARD, 0, ((ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.gated_delta={.log_decay=(self->log_decay
),.state_checkpoint_count=(self->state_checkpoint_count)}}
), 0);
2792 ccv_nnc_hint_tensor_auto(gated_delta, input_params, 6, ccv_nnc_no_hint, output_params, 2);
2793 for (i = 0; i < 2; i++)
2794 outputs[i] = ccv_nnc_tensor_symbol_new(graph, output_params[i], 0);
2795 ccv_nnc_graph_exec_symbol_new(graph, gated_delta, inputs, input_size, outputs, output_size, "gated_delta");
2796}
2797
2798static ccv_cnnp_model_t* _ccv_cnnp_gated_delta_copy(const ccv_cnnp_model_t* const self, void* const context);
2799
2800static const ccv_cnnp_model_vtab_t ccv_cnnp_gated_delta_isa = {
2801 .build = _ccv_cnnp_gated_delta_build,
2802 .copy = _ccv_cnnp_gated_delta_copy,
2803};
2804
2805ccv_cnnp_model_t* ccv_cnnp_gated_delta(const int log_decay, const int state_checkpoint_count, const char* const name)
2806{
2807 ccv_cnnp_model_gated_delta_t* const model_gated_delta = (ccv_cnnp_model_gated_delta_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_gated_delta_t));
2808 model_gated_delta->super.isa = &ccv_cnnp_gated_delta_isa;
2809 model_gated_delta->super.input_size = 6;
2810 model_gated_delta->super.outputs = model_gated_delta->outputs;
2811 model_gated_delta->super.output_size = 2;
2812 model_gated_delta->log_decay = log_decay;
2813 model_gated_delta->state_checkpoint_count = state_checkpoint_count;
2814 ccv_cnnp_model_copy_name(&model_gated_delta->super, name);
2815 return (ccv_cnnp_model_t*)model_gated_delta;
2816}
2817
2818static ccv_cnnp_model_t* _ccv_cnnp_gated_delta_copy(const ccv_cnnp_model_t* const super, void* const context)
2819{
2820 const ccv_cnnp_model_gated_delta_t* const self = (const ccv_cnnp_model_gated_delta_t*)super;
2821 return ccv_cnnp_gated_delta(self->log_decay, self->state_checkpoint_count, super->name);
2822}
2823
2824// MARK - Cmul Layer
2825
2826typedef struct {
2827 ccv_cnnp_model_t super;
2828 ccv_nnc_tensor_symbol_t output;
2829 int conjugate;
2830} ccv_cnnp_model_cmul_t;
2831
2832static void _ccv_cnnp_cmul_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
2833{
2834 PRINT(CCV_CLI_VERBOSE, "[cnnp_cmul_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_cmul_build] -\n"); fflush(stdout); } } while
(0);
2835 assert(input_size == 2)((void) sizeof ((input_size == 2) ? 1 : 0), __extension__ ({ if
(input_size == 2) ; else __assert_fail ("input_size == 2", "ccv_cnnp_model_addons.c"
, 2835, __extension__ __PRETTY_FUNCTION__); }));
2836 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 2836, __extension__ __PRETTY_FUNCTION__
); }));
2837 ccv_nnc_tensor_param_t input_params[2];
2838 int i;
2839 for (i = 0; i < 2; i++)
2840 input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
2841 ccv_nnc_tensor_param_t output_params;
2842 ccv_nnc_cmd_t mul = CMD_CMUL_FORWARD()ccv_nnc_cmd(CCV_NNC_CMUL_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}}}, 0);
2843 mul.info.cmul.conjugate = ((ccv_cnnp_model_cmul_t*)super)->conjugate;
2844 ccv_nnc_hint_tensor_auto(mul, input_params, 2, ccv_nnc_no_hint, &output_params, 1);
2845 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2846 ccv_nnc_graph_exec_symbol_new(graph, mul, inputs, input_size, outputs, output_size, "cmul");
2847}
2848
2849static ccv_cnnp_model_t* _ccv_cnnp_cmul_copy(const ccv_cnnp_model_t* const self, void* const context);
2850
2851static const ccv_cnnp_model_vtab_t ccv_cnnp_cmul_isa = {
2852 .build = _ccv_cnnp_cmul_build,
2853 .copy = _ccv_cnnp_cmul_copy,
2854};
2855
2856ccv_cnnp_model_t* ccv_cnnp_cmul(const int conjugate, const char* const name)
2857{
2858 assert(conjugate == 0 || conjugate == 1)((void) sizeof ((conjugate == 0 || conjugate == 1) ? 1 : 0), __extension__
({ if (conjugate == 0 || conjugate == 1) ; else __assert_fail
("conjugate == 0 || conjugate == 1", "ccv_cnnp_model_addons.c"
, 2858, __extension__ __PRETTY_FUNCTION__); }));
2859 ccv_cnnp_model_cmul_t* const model_cmul = (ccv_cnnp_model_cmul_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_cmul_t));
2860 model_cmul->super.isa = &ccv_cnnp_cmul_isa;
2861 model_cmul->super.input_size = 2;
2862 model_cmul->super.outputs = &model_cmul->output;
2863 model_cmul->super.output_size = 1;
2864 model_cmul->conjugate = conjugate;
2865 ccv_cnnp_model_copy_name(&model_cmul->super, name);
2866 return (ccv_cnnp_model_t*)model_cmul;
2867}
2868
2869static ccv_cnnp_model_t* _ccv_cnnp_cmul_copy(const ccv_cnnp_model_t* const super, void* const context)
2870{
2871 const ccv_cnnp_model_cmul_t* const self = (const ccv_cnnp_model_cmul_t*)super;
2872 return ccv_cnnp_cmul(self->conjugate, super->name);
2873}
2874
2875// MARK - Transpose Layer
2876
2877typedef struct {
2878 ccv_cnnp_model_t super;
2879 ccv_nnc_tensor_symbol_t output;
2880 int transpose[2];
2881} ccv_cnnp_model_transpose_t;
2882
2883static void _ccv_cnnp_transpose_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
2884{
2885 ccv_cnnp_model_transpose_t* const self = (ccv_cnnp_model_transpose_t*)super;
2886 PRINT(CCV_CLI_VERBOSE, "[cnnp_transpose_build] (%d, %d)\n", self->transpose[0], self->transpose[1])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_transpose_build] (%d, %d)\n", self->transpose
[0], self->transpose[1]); fflush(stdout); } } while (0);
2887 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 2887, __extension__ __PRETTY_FUNCTION__); }));
2888 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 2888, __extension__ __PRETTY_FUNCTION__
); }));
2889 if (self->transpose[0] == self->transpose[1])
2890 {
2891 outputs[0] = inputs[0];
2892 return;
2893 }
2894 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2895 ccv_nnc_tensor_param_t output_params;
2896 const ccv_nnc_cmd_t transpose = CMD_TRANSPOSE_FORWARD(self->transpose[0], self->transpose[1])ccv_nnc_cmd(CCV_NNC_TRANSPOSE_FORWARD, 0, ((ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.transpose={.axis={self->transpose[
0], self->transpose[1]}}}), 0);
2897 ccv_nnc_hint_tensor_auto(transpose, (ccv_nnc_tensor_param_t []){
2898 params,
2899 }, 1, ccv_nnc_no_hint, &output_params, 1);
2900 const ccv_nnc_tensor_symbol_t transpose_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2901 ccv_nnc_graph_exec_symbol_new(graph, transpose, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(transpose_output)(const ccv_nnc_tensor_symbol_t []){transpose_output}, (1 +1 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
-1), "transpose");
2902 outputs[0] = transpose_output;
2903}
2904
2905static ccv_cnnp_model_t* _ccv_cnnp_transpose_copy(const ccv_cnnp_model_t* const super, void* const context);
2906
2907static const ccv_cnnp_model_vtab_t ccv_cnnp_transpose_isa = {
2908 .build = _ccv_cnnp_transpose_build,
2909 .copy = _ccv_cnnp_transpose_copy,
2910};
2911
2912ccv_cnnp_model_t* ccv_cnnp_transpose(const int axis_a, const int axis_b, const char* const name)
2913{
2914 ccv_cnnp_model_transpose_t* const model_transpose = (ccv_cnnp_model_transpose_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_transpose_t));
2915 model_transpose->super.isa = &ccv_cnnp_transpose_isa;
2916 model_transpose->super.input_size = 1;
2917 model_transpose->super.outputs = &model_transpose->output;
2918 model_transpose->super.output_size = 1;
2919 model_transpose->transpose[0] = axis_a;
2920 model_transpose->transpose[1] = axis_b;
2921 ccv_cnnp_model_copy_name(&model_transpose->super, name);
2922 return (ccv_cnnp_model_t*)model_transpose;
2923}
2924
2925static ccv_cnnp_model_t* _ccv_cnnp_transpose_copy(const ccv_cnnp_model_t* const super, void* const context)
2926{
2927 const ccv_cnnp_model_transpose_t* const self = (const ccv_cnnp_model_transpose_t*)super;
2928 return ccv_cnnp_transpose(self->transpose[0], self->transpose[1], self->super.name);
2929}
2930
2931// MARK - Layer Norm Layer
2932
2933typedef struct {
2934 ccv_cnnp_model_t super;
2935 ccv_nnc_tensor_symbol_t output;
2936 ccv_nnc_tensor_symbol_t bias;
2937 ccv_nnc_tensor_symbol_t scale;
2938 ccv_nnc_cmd_param_t params;
2939} ccv_cnnp_model_layer_norm_t;
2940
2941static void _ccv_cnnp_layer_norm_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
2942{
2943 PRINT(CCV_CLI_VERBOSE, "[cnnp_layer_norm_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_layer_norm_build] -\n"); fflush(stdout); } }
while (0);
2944 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 2944, __extension__ __PRETTY_FUNCTION__); }));
2945 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 2945, __extension__ __PRETTY_FUNCTION__
); }));
2946 ccv_cnnp_model_layer_norm_t* const self = (ccv_cnnp_model_layer_norm_t*)super;
2947 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2948 ccv_nnc_tensor_param_t bias_params = params;
2949 const int nd = ccv_nnc_tensor_nd(params.dim);
2950 int i;
2951 for (i = 0; i < nd; i++)
2952 bias_params.dim[i] = 1;
2953 for (i = 0; i < self->params.lnorm.count; i++)
2954 bias_params.dim[self->params.lnorm.axis[i]] = params.dim[self->params.lnorm.axis[i]];
2955 if (self->params.lnorm.elementwise_affine)
2956 {
2957 // Both scale and bias are shared between if this model is reused.
2958 if (!self->scale.graph)
2959 self->scale = ccv_nnc_tensor_symbol_new(graph, bias_params, "scale");
2960 if (!self->bias.graph)
2961 self->bias = ccv_nnc_tensor_symbol_new(graph, bias_params, "bias");
2962 }
2963 const ccv_nnc_tensor_symbol_t scale = self->params.lnorm.elementwise_affine ? ccv_cnnp_model_get_symbol(super, self->scale) : NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
2964 const ccv_nnc_tensor_symbol_t bias = self->params.lnorm.elementwise_affine ? ccv_cnnp_model_get_symbol(super, self->bias) : NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
2965 const ccv_nnc_cmd_t layer_norm = ccv_nnc_cmd(CCV_NNC_LAYER_NORM_FORWARD, 0, self->params, 0);
2966 ccv_nnc_tensor_param_t output_params[3];
2967 if (self->params.lnorm.elementwise_affine)
2968 ccv_nnc_hint_tensor_auto(layer_norm, (ccv_nnc_tensor_param_t []){
2969 params,
2970 bias_params,
2971 bias_params,
2972 }, 3, ccv_nnc_no_hint, output_params, 3);
2973 else
2974 ccv_nnc_hint_tensor_auto(layer_norm, (ccv_nnc_tensor_param_t []){
2975 params,
2976 }, 1, ccv_nnc_no_hint, output_params, 3);
2977 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0);
2978 const ccv_nnc_tensor_symbol_t saved_mean = ccv_nnc_tensor_symbol_new(graph, output_params[1], "saved_mean");
2979 const ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(graph, output_params[2], "saved_inv_std");
2980 if (self->params.lnorm.elementwise_affine)
2981 ccv_nnc_graph_exec_symbol_new(graph, layer_norm, TENSOR_SYMBOL_LIST(inputs[0], scale, bias)(const ccv_nnc_tensor_symbol_t []){inputs[0], scale, bias}, (
1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 -1), TENSOR_SYMBOL_LIST(output, saved_mean, saved_inv_std)(const ccv_nnc_tensor_symbol_t []){output, saved_mean, saved_inv_std
}, (1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 -1), "layer_norm");
2982 else
2983 ccv_nnc_graph_exec_symbol_new(graph, layer_norm, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(output, saved_mean, saved_inv_std)(const ccv_nnc_tensor_symbol_t []){output, saved_mean, saved_inv_std
}, (1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 -1), "layer_norm");
2984 outputs[0] = output;
2985}
2986
2987static void _ccv_cnnp_layer_norm_init_states(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_cnnp_state_initializer_f initializer, void* const context)
2988{
2989 ccv_cnnp_model_layer_norm_t* const self = (ccv_cnnp_model_layer_norm_t*)super;
2990 if (self->scale.graph)
2991 initializer(context, CMD_SET_FORWARD(1)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={1,}}}, 0), ccv_nnc_no_hint, 0, 0, self->scale);
2992 if (self->bias.graph)
2993 initializer(context, CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, self->bias);
2994}
2995
2996static void _ccv_cnnp_layer_norm_add_to_parameter(ccv_cnnp_model_t* const super, const ccv_cnnp_add_to_array_f add_to_array, void* const parameters, const int is_trainable)
2997{
2998 ccv_cnnp_model_layer_norm_t* const self = (ccv_cnnp_model_layer_norm_t*)super;
2999 if (self->scale.graph)
3000 add_to_array(parameters, self->scale, is_trainable);
3001 if (self->bias.graph)
3002 add_to_array(parameters, self->bias, is_trainable);
3003}
3004
3005static ccv_cnnp_model_t* _ccv_cnnp_layer_norm_copy(const ccv_cnnp_model_t* const super, void* const context);
3006
3007static const ccv_cnnp_model_vtab_t ccv_cnnp_layer_norm_isa = {
3008 .build = _ccv_cnnp_layer_norm_build,
3009 .init_states = _ccv_cnnp_layer_norm_init_states,
3010 .add_to_parameter = _ccv_cnnp_layer_norm_add_to_parameter,
3011 .copy = _ccv_cnnp_layer_norm_copy,
3012};
3013
3014ccv_cnnp_model_t* ccv_cnnp_layer_norm(const float epsilon, const int axis[CCV_NNC_MAX_DIM_ALLOC(12)], const int axis_count, const int elementwise_affine, const float scale, const int is_trainable, const char* const name)
3015{
3016 ccv_cnnp_model_layer_norm_t* const model_layer_norm = (ccv_cnnp_model_layer_norm_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_layer_norm_t));
3017 model_layer_norm->super.isa = &ccv_cnnp_layer_norm_isa;
3018 model_layer_norm->super.input_size = 1;
3019 model_layer_norm->super.outputs = &model_layer_norm->output;
3020 model_layer_norm->super.output_size = 1;
3021 model_layer_norm->super.is_trainable = is_trainable;
3022 ccv_cnnp_model_copy_name(&model_layer_norm->super, name);
3023 model_layer_norm->scale.d = CCV_NNC_NO_TENSOR_SYMBOL;
3024 model_layer_norm->scale.graph = 0;
3025 model_layer_norm->bias.d = CCV_NNC_NO_TENSOR_SYMBOL;
3026 model_layer_norm->bias.graph = 0;
3027 model_layer_norm->params.lnorm.epsilon = epsilon;
3028 model_layer_norm->params.lnorm.scale = scale;
3029 model_layer_norm->params.lnorm.count = axis_count;
3030 model_layer_norm->params.lnorm.elementwise_affine = elementwise_affine;
3031 memcpy(model_layer_norm->params.lnorm.axis, axis, sizeof(int) * axis_count);
3032 return (ccv_cnnp_model_t*)model_layer_norm;
3033}
3034
3035static ccv_cnnp_model_t* _ccv_cnnp_layer_norm_copy(const ccv_cnnp_model_t* const super, void* const context)
3036{
3037 const ccv_cnnp_model_layer_norm_t* const self = (const ccv_cnnp_model_layer_norm_t*)super;
3038 return ccv_cnnp_layer_norm(self->params.lnorm.epsilon, self->params.lnorm.axis, self->params.lnorm.count, self->params.lnorm.elementwise_affine, self->params.lnorm.scale, self->super.is_trainable, self->super.name);
3039}
3040
3041// MARK - Group Norm Layer
3042
3043typedef struct {
3044 ccv_cnnp_model_t super;
3045 ccv_nnc_tensor_symbol_t output;
3046 ccv_nnc_tensor_symbol_t bias;
3047 ccv_nnc_tensor_symbol_t scale;
3048 ccv_nnc_cmd_param_t params;
3049} ccv_cnnp_model_group_norm_t;
3050
3051static void _ccv_cnnp_group_norm_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
3052{
3053 PRINT(CCV_CLI_VERBOSE, "[cnnp_group_norm_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_group_norm_build] -\n"); fflush(stdout); } }
while (0);
3054 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 3054, __extension__ __PRETTY_FUNCTION__); }));
3055 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 3055, __extension__ __PRETTY_FUNCTION__
); }));
3056 ccv_cnnp_model_group_norm_t* const self = (ccv_cnnp_model_group_norm_t*)super;
3057 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3058 ccv_nnc_tensor_param_t bias_params = params;
3059 const int nd = ccv_nnc_tensor_nd(params.dim);
3060 int i;
3061 for (i = 0; i < nd; i++)
3062 bias_params.dim[i] = 1;
3063 bias_params.dim[self->params.gnorm.group_axis] = params.dim[self->params.gnorm.group_axis];
3064 if (self->params.gnorm.elementwise_affine)
3065 {
3066 // Both scale and bias are shared between if this model is reused.
3067 if (!self->scale.graph)
3068 self->scale = ccv_nnc_tensor_symbol_new(graph, bias_params, "scale");
3069 if (!self->bias.graph)
3070 self->bias = ccv_nnc_tensor_symbol_new(graph, bias_params, "bias");
3071 }
3072 const ccv_nnc_tensor_symbol_t scale = self->params.gnorm.elementwise_affine ? ccv_cnnp_model_get_symbol(super, self->scale) : NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
3073 const ccv_nnc_tensor_symbol_t bias = self->params.gnorm.elementwise_affine ? ccv_cnnp_model_get_symbol(super, self->bias) : NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
3074 const ccv_nnc_cmd_t group_norm = ccv_nnc_cmd(CCV_NNC_GROUP_NORM_FORWARD, 0, self->params, 0);
3075 ccv_nnc_tensor_param_t output_params[3];
3076 if (self->params.gnorm.elementwise_affine)
3077 ccv_nnc_hint_tensor_auto(group_norm, (ccv_nnc_tensor_param_t []){
3078 params,
3079 bias_params,
3080 bias_params,
3081 }, 3, ccv_nnc_no_hint, output_params, 3);
3082 else
3083 ccv_nnc_hint_tensor_auto(group_norm, (ccv_nnc_tensor_param_t []){
3084 params,
3085 }, 1, ccv_nnc_no_hint, output_params, 3);
3086 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0);
3087 const ccv_nnc_tensor_symbol_t saved_mean = ccv_nnc_tensor_symbol_new(graph, output_params[1], "saved_mean");
3088 const ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(graph, output_params[2], "saved_inv_std");
3089 if (self->params.gnorm.elementwise_affine)
3090 ccv_nnc_graph_exec_symbol_new(graph, group_norm, TENSOR_SYMBOL_LIST(inputs[0], scale, bias)(const ccv_nnc_tensor_symbol_t []){inputs[0], scale, bias}, (
1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 -1), TENSOR_SYMBOL_LIST(output, saved_mean, saved_inv_std)(const ccv_nnc_tensor_symbol_t []){output, saved_mean, saved_inv_std
}, (1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 -1), "group_norm");
3091 else
3092 ccv_nnc_graph_exec_symbol_new(graph, group_norm, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(output, saved_mean, saved_inv_std)(const ccv_nnc_tensor_symbol_t []){output, saved_mean, saved_inv_std
}, (1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 -1), "group_norm");
3093 outputs[0] = output;
3094}
3095
3096static void _ccv_cnnp_group_norm_init_states(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_cnnp_state_initializer_f initializer, void* const context)
3097{
3098 ccv_cnnp_model_group_norm_t* const self = (ccv_cnnp_model_group_norm_t*)super;
3099 if (self->scale.graph)
3100 initializer(context, CMD_SET_FORWARD(1)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={1,}}}, 0), ccv_nnc_no_hint, 0, 0, self->scale);
3101 if (self->bias.graph)
3102 initializer(context, CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, self->bias);
3103}
3104
3105static void _ccv_cnnp_group_norm_add_to_parameter(ccv_cnnp_model_t* const super, const ccv_cnnp_add_to_array_f add_to_array, void* const parameters, const int is_trainable)
3106{
3107 ccv_cnnp_model_group_norm_t* const self = (ccv_cnnp_model_group_norm_t*)super;
3108 if (self->scale.graph)
3109 add_to_array(parameters, self->scale, is_trainable);
3110 if (self->bias.graph)
3111 add_to_array(parameters, self->bias, is_trainable);
3112}
3113
3114static ccv_cnnp_model_t* _ccv_cnnp_group_norm_copy(const ccv_cnnp_model_t* const super, void* const context);
3115
3116static const ccv_cnnp_model_vtab_t ccv_cnnp_group_norm_isa = {
3117 .build = _ccv_cnnp_group_norm_build,
3118 .init_states = _ccv_cnnp_group_norm_init_states,
3119 .add_to_parameter = _ccv_cnnp_group_norm_add_to_parameter,
3120 .copy = _ccv_cnnp_group_norm_copy,
3121};
3122
3123ccv_cnnp_model_t* ccv_cnnp_group_norm(const int group_axis, const int groups, const float epsilon, const int reduce_axis[CCV_NNC_MAX_DIM_ALLOC(12)], const int axis_count, const int elementwise_affine, const int is_trainable, const char* const name)
3124{
3125 ccv_cnnp_model_group_norm_t* const model_group_norm = (ccv_cnnp_model_group_norm_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_group_norm_t));
3126 model_group_norm->super.isa = &ccv_cnnp_group_norm_isa;
3127 model_group_norm->super.input_size = 1;
3128 model_group_norm->super.outputs = &model_group_norm->output;
3129 model_group_norm->super.output_size = 1;
3130 model_group_norm->super.is_trainable = is_trainable;
3131 ccv_cnnp_model_copy_name(&model_group_norm->super, name);
3132 model_group_norm->scale.d = CCV_NNC_NO_TENSOR_SYMBOL;
3133 model_group_norm->scale.graph = 0;
3134 model_group_norm->bias.d = CCV_NNC_NO_TENSOR_SYMBOL;
3135 model_group_norm->bias.graph = 0;
3136 model_group_norm->params.gnorm.group_axis = group_axis;
3137 model_group_norm->params.gnorm.groups = groups;
3138 model_group_norm->params.gnorm.epsilon = epsilon;
3139 model_group_norm->params.gnorm.reduce_count = axis_count;
3140 model_group_norm->params.gnorm.elementwise_affine = elementwise_affine;
3141 memcpy(model_group_norm->params.gnorm.reduce_axis, reduce_axis, sizeof(int) * axis_count);
3142 return (ccv_cnnp_model_t*)model_group_norm;
3143}
3144
3145static ccv_cnnp_model_t* _ccv_cnnp_group_norm_copy(const ccv_cnnp_model_t* const super, void* const context)
3146{
3147 const ccv_cnnp_model_group_norm_t* const self = (const ccv_cnnp_model_group_norm_t*)super;
3148 return ccv_cnnp_group_norm(self->params.gnorm.group_axis, self->params.gnorm.groups, self->params.gnorm.epsilon, self->params.gnorm.reduce_axis, self->params.gnorm.reduce_count, self->params.gnorm.elementwise_affine, self->super.is_trainable, self->super.name);
3149}
3150
3151// MARK - RMSNorm Layer
3152
3153typedef struct {
3154 ccv_cnnp_model_t super;
3155 ccv_nnc_tensor_symbol_t output;
3156 ccv_nnc_tensor_symbol_t scale;
3157 ccv_nnc_cmd_param_t params;
3158} ccv_cnnp_model_rmsnorm_t;
3159
3160static void _ccv_cnnp_rmsnorm_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
3161{
3162 PRINT(CCV_CLI_VERBOSE, "[cnnp_rmsnorm_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_rmsnorm_build] -\n"); fflush(stdout); } } while
(0);
3163 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 3163, __extension__ __PRETTY_FUNCTION__); }));
3164 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 3164, __extension__ __PRETTY_FUNCTION__
); }));
3165 ccv_cnnp_model_rmsnorm_t* const self = (ccv_cnnp_model_rmsnorm_t*)super;
3166 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3167 ccv_nnc_tensor_param_t scale_params = params;
3168 const int nd = ccv_nnc_tensor_nd(params.dim);
3169 int i;
3170 for (i = 0; i < nd; i++)
3171 scale_params.dim[i] = 1;
3172 for (i = 0; i < self->params.rmsnorm.count; i++)
3173 scale_params.dim[self->params.rmsnorm.axis[i]] = params.dim[self->params.rmsnorm.axis[i]];
3174 // Both scale and bias are shared between if this model is reused.
3175 if (self->params.rmsnorm.elementwise_affine)
3176 {
3177 if (!self->scale.graph)
3178 self->scale = ccv_nnc_tensor_symbol_new(graph, scale_params, "scale");
3179 }
3180 const ccv_nnc_tensor_symbol_t scale = self->params.rmsnorm.elementwise_affine ? ccv_cnnp_model_get_symbol(super, self->scale) : NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
3181 const ccv_nnc_cmd_t rmsnorm = ccv_nnc_cmd(CCV_NNC_RMSNORM_FORWARD, 0, self->params, 0);
3182 ccv_nnc_tensor_param_t output_params[2];
3183 if (self->params.rmsnorm.elementwise_affine)
3184 ccv_nnc_hint_tensor_auto(rmsnorm, (ccv_nnc_tensor_param_t []){
3185 params,
3186 scale_params,
3187 }, 2, ccv_nnc_no_hint, output_params, 2);
3188 else
3189 ccv_nnc_hint_tensor_auto(rmsnorm, (ccv_nnc_tensor_param_t []){
3190 params,
3191 }, 1, ccv_nnc_no_hint, output_params, 2);
3192 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0);
3193 const ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(graph, output_params[1], "saved_inv_std");
3194 if (self->params.rmsnorm.elementwise_affine)
3195 ccv_nnc_graph_exec_symbol_new(graph, rmsnorm, TENSOR_SYMBOL_LIST(inputs[0], scale)(const ccv_nnc_tensor_symbol_t []){inputs[0], scale}, (1 +1 +
1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
-1), TENSOR_SYMBOL_LIST(output, saved_inv_std)(const ccv_nnc_tensor_symbol_t []){output, saved_inv_std}, (1
+1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 -1), "rmsnorm");
3196 else
3197 ccv_nnc_graph_exec_symbol_new(graph, rmsnorm, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(output, saved_inv_std)(const ccv_nnc_tensor_symbol_t []){output, saved_inv_std}, (1
+1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 -1), "rmsnorm");
3198 outputs[0] = output;
3199}
3200
3201static void _ccv_cnnp_rmsnorm_init_states(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_cnnp_state_initializer_f initializer, void* const context)
3202{
3203 ccv_cnnp_model_rmsnorm_t* const self = (ccv_cnnp_model_rmsnorm_t*)super;
3204 if (self->scale.graph)
3205 initializer(context, CMD_SET_FORWARD(1)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={1,}}}, 0), ccv_nnc_no_hint, 0, 0, self->scale);
3206}
3207
3208static void _ccv_cnnp_rmsnorm_add_to_parameter(ccv_cnnp_model_t* const super, const ccv_cnnp_add_to_array_f add_to_array, void* const parameters, const int is_trainable)
3209{
3210 ccv_cnnp_model_rmsnorm_t* const self = (ccv_cnnp_model_rmsnorm_t*)super;
3211 if (self->scale.graph)
3212 add_to_array(parameters, self->scale, is_trainable);
3213}
3214
3215static ccv_cnnp_model_t* _ccv_cnnp_rmsnorm_copy(const ccv_cnnp_model_t* const super, void* const context);
3216
3217static const ccv_cnnp_model_vtab_t ccv_cnnp_rmsnorm_isa = {
3218 .build = _ccv_cnnp_rmsnorm_build,
3219 .init_states = _ccv_cnnp_rmsnorm_init_states,
3220 .add_to_parameter = _ccv_cnnp_rmsnorm_add_to_parameter,
3221 .copy = _ccv_cnnp_rmsnorm_copy,
3222};
3223
3224ccv_cnnp_model_t* ccv_cnnp_rmsnorm(const float epsilon, const int axis[CCV_NNC_MAX_DIM_ALLOC(12)], const int axis_count, const int elementwise_affine, const float scale, const int is_trainable, const char* const name)
3225{
3226 ccv_cnnp_model_rmsnorm_t* const model_rmsnorm = (ccv_cnnp_model_rmsnorm_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_rmsnorm_t));
3227 model_rmsnorm->super.isa = &ccv_cnnp_rmsnorm_isa;
3228 model_rmsnorm->super.input_size = 1;
3229 model_rmsnorm->super.outputs = &model_rmsnorm->output;
3230 model_rmsnorm->super.output_size = 1;
3231 model_rmsnorm->super.is_trainable = is_trainable;
3232 ccv_cnnp_model_copy_name(&model_rmsnorm->super, name);
3233 model_rmsnorm->scale.d = CCV_NNC_NO_TENSOR_SYMBOL;
3234 model_rmsnorm->scale.graph = 0;
3235 model_rmsnorm->params.rmsnorm.epsilon = epsilon;
3236 model_rmsnorm->params.rmsnorm.scale = scale;
3237 model_rmsnorm->params.rmsnorm.count = axis_count;
3238 model_rmsnorm->params.rmsnorm.elementwise_affine = elementwise_affine;
3239 memcpy(model_rmsnorm->params.lnorm.axis, axis, sizeof(int) * axis_count);
3240 return (ccv_cnnp_model_t*)model_rmsnorm;
3241}
3242
3243static ccv_cnnp_model_t* _ccv_cnnp_rmsnorm_copy(const ccv_cnnp_model_t* const super, void* const context)
3244{
3245 const ccv_cnnp_model_rmsnorm_t* const self = (const ccv_cnnp_model_rmsnorm_t*)super;
3246 return ccv_cnnp_rmsnorm(self->params.rmsnorm.epsilon, self->params.rmsnorm.axis, self->params.rmsnorm.count, self->params.rmsnorm.elementwise_affine, self->params.rmsnorm.scale, self->super.is_trainable, self->super.name);
3247}
3248
3249// MARK - RMSNorm CMul Layer
3250
3251typedef struct {
3252 ccv_cnnp_model_t super;
3253 ccv_nnc_tensor_symbol_t output;
3254 ccv_nnc_tensor_symbol_t scale;
3255 ccv_nnc_cmd_param_t params;
3256} ccv_cnnp_model_rmsnorm_cmul_t;
3257
3258static void _ccv_cnnp_rmsnorm_cmul_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
3259{
3260 PRINT(CCV_CLI_VERBOSE, "[cnnp_rmsnorm_cmul_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_rmsnorm_cmul_build] -\n"); fflush(stdout); }
} while (0);
3261 assert(input_size == 2)((void) sizeof ((input_size == 2) ? 1 : 0), __extension__ ({ if
(input_size == 2) ; else __assert_fail ("input_size == 2", "ccv_cnnp_model_addons.c"
, 3261, __extension__ __PRETTY_FUNCTION__); }));
3262 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 3262, __extension__ __PRETTY_FUNCTION__
); }));
3263 ccv_cnnp_model_rmsnorm_cmul_t* const self = (ccv_cnnp_model_rmsnorm_cmul_t*)super;
3264 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3265 const ccv_nnc_tensor_param_t rotation_params = ccv_nnc_tensor_symbol_params(graph, inputs[1]);
3266 ccv_nnc_tensor_param_t scale_params = params;
3267 const int nd = ccv_nnc_tensor_nd(params.dim);
3268 int i;
3269 for (i = 0; i < nd; i++)
3270 scale_params.dim[i] = 1;
3271 for (i = 0; i < self->params.rmsnorm_cmul.count; i++)
3272 scale_params.dim[self->params.rmsnorm_cmul.axis[i]] = params.dim[self->params.rmsnorm_cmul.axis[i]];
3273 if (self->params.rmsnorm_cmul.elementwise_affine)
3274 {
3275 if (!self->scale.graph)
3276 self->scale = ccv_nnc_tensor_symbol_new(graph, scale_params, "scale");
3277 }
3278 const ccv_nnc_tensor_symbol_t scale = self->params.rmsnorm_cmul.elementwise_affine ? ccv_cnnp_model_get_symbol(super, self->scale) : NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
3279 const ccv_nnc_cmd_t rmsnorm_cmul = ccv_nnc_cmd(CCV_NNC_RMSNORM_CMUL_FORWARD, 0, self->params, 0);
3280 ccv_nnc_tensor_param_t output_params;
3281 if (self->params.rmsnorm_cmul.elementwise_affine)
3282 ccv_nnc_hint_tensor_auto(rmsnorm_cmul, (ccv_nnc_tensor_param_t []){
3283 params,
3284 rotation_params,
3285 scale_params,
3286 }, 3, ccv_nnc_no_hint, &output_params, 1);
3287 else
3288 ccv_nnc_hint_tensor_auto(rmsnorm_cmul, (ccv_nnc_tensor_param_t []){
3289 params,
3290 rotation_params,
3291 }, 2, ccv_nnc_no_hint, &output_params, 1);
3292 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3293 if (self->params.rmsnorm_cmul.elementwise_affine)
3294 ccv_nnc_graph_exec_symbol_new(graph, rmsnorm_cmul, TENSOR_SYMBOL_LIST(inputs[0], inputs[1], scale)(const ccv_nnc_tensor_symbol_t []){inputs[0], inputs[1], scale
}, (1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 -1), outputs, output_size, "rmsnorm_cmul");
3295 else
3296 ccv_nnc_graph_exec_symbol_new(graph, rmsnorm_cmul, inputs, input_size, outputs, output_size, "rmsnorm_cmul");
3297}
3298
3299static void _ccv_cnnp_rmsnorm_cmul_init_states(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_cnnp_state_initializer_f initializer, void* const context)
3300{
3301 ccv_cnnp_model_rmsnorm_cmul_t* const self = (ccv_cnnp_model_rmsnorm_cmul_t*)super;
3302 if (self->scale.graph)
3303 initializer(context, CMD_SET_FORWARD(1)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={1,}}}, 0), ccv_nnc_no_hint, 0, 0, self->scale);
3304}
3305
3306static void _ccv_cnnp_rmsnorm_cmul_add_to_parameter(ccv_cnnp_model_t* const super, const ccv_cnnp_add_to_array_f add_to_array, void* const parameters, const int is_trainable)
3307{
3308 ccv_cnnp_model_rmsnorm_cmul_t* const self = (ccv_cnnp_model_rmsnorm_cmul_t*)super;
3309 if (self->scale.graph)
3310 add_to_array(parameters, self->scale, is_trainable);
3311}
3312
3313static ccv_cnnp_model_t* _ccv_cnnp_rmsnorm_cmul_copy(const ccv_cnnp_model_t* const super, void* const context);
3314
3315static const ccv_cnnp_model_vtab_t ccv_cnnp_rmsnorm_cmul_isa = {
3316 .build = _ccv_cnnp_rmsnorm_cmul_build,
3317 .init_states = _ccv_cnnp_rmsnorm_cmul_init_states,
3318 .add_to_parameter = _ccv_cnnp_rmsnorm_cmul_add_to_parameter,
3319 .copy = _ccv_cnnp_rmsnorm_cmul_copy,
3320};
3321
3322ccv_cnnp_model_t* ccv_cnnp_rmsnorm_cmul(const float epsilon, const int axis[CCV_NNC_MAX_DIM_ALLOC(12)], const int axis_count, const int elementwise_affine, const int is_trainable, const char* const name)
3323{
3324 ccv_cnnp_model_rmsnorm_cmul_t* const model_rmsnorm_cmul = (ccv_cnnp_model_rmsnorm_cmul_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_rmsnorm_cmul_t));
3325 model_rmsnorm_cmul->super.isa = &ccv_cnnp_rmsnorm_cmul_isa;
3326 model_rmsnorm_cmul->super.input_size = 2;
3327 model_rmsnorm_cmul->super.outputs = &model_rmsnorm_cmul->output;
3328 model_rmsnorm_cmul->super.output_size = 1;
3329 model_rmsnorm_cmul->super.is_trainable = is_trainable;
3330 ccv_cnnp_model_copy_name(&model_rmsnorm_cmul->super, name);
3331 model_rmsnorm_cmul->scale.d = CCV_NNC_NO_TENSOR_SYMBOL;
3332 model_rmsnorm_cmul->scale.graph = 0;
3333 model_rmsnorm_cmul->params.rmsnorm_cmul.epsilon = epsilon;
3334 model_rmsnorm_cmul->params.rmsnorm_cmul.count = axis_count;
3335 model_rmsnorm_cmul->params.rmsnorm_cmul.elementwise_affine = elementwise_affine;
3336 memcpy(model_rmsnorm_cmul->params.rmsnorm_cmul.axis, axis, sizeof(int) * axis_count);
3337 return (ccv_cnnp_model_t*)model_rmsnorm_cmul;
3338}
3339
3340static ccv_cnnp_model_t* _ccv_cnnp_rmsnorm_cmul_copy(const ccv_cnnp_model_t* const super, void* const context)
3341{
3342 const ccv_cnnp_model_rmsnorm_cmul_t* const self = (const ccv_cnnp_model_rmsnorm_cmul_t*)super;
3343 return ccv_cnnp_rmsnorm_cmul(self->params.rmsnorm_cmul.epsilon, self->params.rmsnorm_cmul.axis, self->params.rmsnorm_cmul.count, self->params.rmsnorm_cmul.elementwise_affine, self->super.is_trainable, self->super.name);
3344}
3345
3346// MARK - RMSNorm Gated Layer
3347
3348typedef struct {
3349 ccv_cnnp_model_t super;
3350 ccv_nnc_tensor_symbol_t output;
3351 ccv_nnc_tensor_symbol_t scale;
3352 ccv_nnc_cmd_param_t params;
3353} ccv_cnnp_model_rmsnorm_gated_t;
3354
3355static void _ccv_cnnp_rmsnorm_gated_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
3356{
3357 PRINT(CCV_CLI_VERBOSE, "[cnnp_rmsnorm_gated_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_rmsnorm_gated_build] -\n"); fflush(stdout); }
} while (0);
3358 assert(input_size == 2)((void) sizeof ((input_size == 2) ? 1 : 0), __extension__ ({ if
(input_size == 2) ; else __assert_fail ("input_size == 2", "ccv_cnnp_model_addons.c"
, 3358, __extension__ __PRETTY_FUNCTION__); }));
3359 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 3359, __extension__ __PRETTY_FUNCTION__
); }));
3360 ccv_cnnp_model_rmsnorm_gated_t* const self = (ccv_cnnp_model_rmsnorm_gated_t*)super;
3361 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3362 const ccv_nnc_tensor_param_t gate_params = ccv_nnc_tensor_symbol_params(graph, inputs[1]);
3363 ccv_nnc_tensor_param_t scale_params = params;
3364 const int nd = ccv_nnc_tensor_nd(params.dim);
3365 int i;
3366 for (i = 0; i < nd; i++)
3367 scale_params.dim[i] = 1;
3368 for (i = 0; i < self->params.rmsnorm_gated.count; i++)
3369 scale_params.dim[self->params.rmsnorm_gated.axis[i]] = params.dim[self->params.rmsnorm_gated.axis[i]];
3370 if (self->params.rmsnorm_gated.elementwise_affine)
3371 {
3372 if (!self->scale.graph)
3373 self->scale = ccv_nnc_tensor_symbol_new(graph, scale_params, "scale");
3374 }
3375 const ccv_nnc_tensor_symbol_t scale = self->params.rmsnorm_gated.elementwise_affine ? ccv_cnnp_model_get_symbol(super, self->scale) : NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
3376 const ccv_nnc_cmd_t rmsnorm_gated = ccv_nnc_cmd(CCV_NNC_RMSNORM_GATED_FORWARD, 0, self->params, 0);
3377 ccv_nnc_tensor_param_t output_params;
3378 if (self->params.rmsnorm_gated.elementwise_affine)
3379 ccv_nnc_hint_tensor_auto(rmsnorm_gated, (ccv_nnc_tensor_param_t []){
3380 params,
3381 gate_params,
3382 scale_params,
3383 }, 3, ccv_nnc_no_hint, &output_params, 1);
3384 else
3385 ccv_nnc_hint_tensor_auto(rmsnorm_gated, (ccv_nnc_tensor_param_t []){
3386 params,
3387 gate_params,
3388 }, 2, ccv_nnc_no_hint, &output_params, 1);
3389 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3390 if (self->params.rmsnorm_gated.elementwise_affine)
3391 ccv_nnc_graph_exec_symbol_new(graph, rmsnorm_gated, TENSOR_SYMBOL_LIST(inputs[0], inputs[1], scale)(const ccv_nnc_tensor_symbol_t []){inputs[0], inputs[1], scale
}, (1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 -1), outputs, output_size, "rmsnorm_gated");
3392 else
3393 ccv_nnc_graph_exec_symbol_new(graph, rmsnorm_gated, inputs, input_size, outputs, output_size, "rmsnorm_gated");
3394}
3395
3396static void _ccv_cnnp_rmsnorm_gated_init_states(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_cnnp_state_initializer_f initializer, void* const context)
3397{
3398 ccv_cnnp_model_rmsnorm_gated_t* const self = (ccv_cnnp_model_rmsnorm_gated_t*)super;
3399 if (self->scale.graph)
3400 initializer(context, CMD_SET_FORWARD(1)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={1,}}}, 0), ccv_nnc_no_hint, 0, 0, self->scale);
3401}
3402
3403static void _ccv_cnnp_rmsnorm_gated_add_to_parameter(ccv_cnnp_model_t* const super, const ccv_cnnp_add_to_array_f add_to_array, void* const parameters, const int is_trainable)
3404{
3405 ccv_cnnp_model_rmsnorm_gated_t* const self = (ccv_cnnp_model_rmsnorm_gated_t*)super;
3406 if (self->scale.graph)
3407 add_to_array(parameters, self->scale, is_trainable);
3408}
3409
3410static ccv_cnnp_model_t* _ccv_cnnp_rmsnorm_gated_copy(const ccv_cnnp_model_t* const super, void* const context);
3411
3412static const ccv_cnnp_model_vtab_t ccv_cnnp_rmsnorm_gated_isa = {
3413 .build = _ccv_cnnp_rmsnorm_gated_build,
3414 .init_states = _ccv_cnnp_rmsnorm_gated_init_states,
3415 .add_to_parameter = _ccv_cnnp_rmsnorm_gated_add_to_parameter,
3416 .copy = _ccv_cnnp_rmsnorm_gated_copy,
3417};
3418
3419ccv_cnnp_model_t* ccv_cnnp_rmsnorm_gated(const float epsilon, const int axis[CCV_NNC_MAX_DIM_ALLOC(12)], const int axis_count, const int elementwise_affine, const int sigmoid_gate, const int is_trainable, const char* const name)
3420{
3421 ccv_cnnp_model_rmsnorm_gated_t* const model_rmsnorm_gated = (ccv_cnnp_model_rmsnorm_gated_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_rmsnorm_gated_t));
3422 model_rmsnorm_gated->super.isa = &ccv_cnnp_rmsnorm_gated_isa;
3423 model_rmsnorm_gated->super.input_size = 2;
3424 model_rmsnorm_gated->super.outputs = &model_rmsnorm_gated->output;
3425 model_rmsnorm_gated->super.output_size = 1;
3426 model_rmsnorm_gated->super.is_trainable = is_trainable;
3427 ccv_cnnp_model_copy_name(&model_rmsnorm_gated->super, name);
3428 model_rmsnorm_gated->scale.d = CCV_NNC_NO_TENSOR_SYMBOL;
3429 model_rmsnorm_gated->scale.graph = 0;
3430 model_rmsnorm_gated->params.rmsnorm_gated.epsilon = epsilon;
3431 model_rmsnorm_gated->params.rmsnorm_gated.count = axis_count;
3432 model_rmsnorm_gated->params.rmsnorm_gated.elementwise_affine = elementwise_affine;
3433 model_rmsnorm_gated->params.rmsnorm_gated.sigmoid_gate = sigmoid_gate;
3434 memcpy(model_rmsnorm_gated->params.rmsnorm_gated.axis, axis, sizeof(int) * axis_count);
3435 return (ccv_cnnp_model_t*)model_rmsnorm_gated;
3436}
3437
3438static ccv_cnnp_model_t* _ccv_cnnp_rmsnorm_gated_copy(const ccv_cnnp_model_t* const super, void* const context)
3439{
3440 const ccv_cnnp_model_rmsnorm_gated_t* const self = (const ccv_cnnp_model_rmsnorm_gated_t*)super;
3441 return ccv_cnnp_rmsnorm_gated(self->params.rmsnorm_gated.epsilon, self->params.rmsnorm_gated.axis, self->params.rmsnorm_gated.count, self->params.rmsnorm_gated.elementwise_affine, self->params.rmsnorm_gated.sigmoid_gate, self->super.is_trainable, self->super.name);
3442}
3443
3444// MARK - Batched Matrix Mul Layer
3445
3446typedef struct {
3447 ccv_cnnp_model_t super;
3448 ccv_nnc_tensor_symbol_t output;
3449 int transpose_a[2];
3450 int transpose_b[2];
3451 int flags;
3452} ccv_cnnp_model_matmul_t;
3453
3454static void _ccv_cnnp_matmul_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
3455{
3456 PRINT(CCV_CLI_VERBOSE, "[cnnp_matmul_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_matmul_build] -\n"); fflush(stdout); } } while
(0);
3457 assert(input_size == 2)((void) sizeof ((input_size == 2) ? 1 : 0), __extension__ ({ if
(input_size == 2) ; else __assert_fail ("input_size == 2", "ccv_cnnp_model_addons.c"
, 3457, __extension__ __PRETTY_FUNCTION__); }));
3458 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 3458, __extension__ __PRETTY_FUNCTION__
); }));
3459 ccv_cnnp_model_matmul_t* const self = (ccv_cnnp_model_matmul_t*)super;
3460 ccv_nnc_tensor_param_t a_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3461 ccv_nnc_tensor_param_t b_params = ccv_nnc_tensor_symbol_params(graph, inputs[1]);
3462 ccv_nnc_tensor_param_t output_params;
3463 ccv_nnc_cmd_t matmul = CMD_GEMM_FORWARD(self->transpose_a, self->transpose_b)ccv_nnc_cmd(CCV_NNC_GEMM_FORWARD, 0, ((ccv_nnc_cmd_param_t){.
size={.dim={1,1,1}},.blas={.a={1,1},.transpose_a={self->transpose_a
[0],self->transpose_a[1]},.transpose_b={self->transpose_b
[0],self->transpose_b[1]},}}), 0);
3464 matmul.info.blas.flags = self->flags;
3465 ccv_nnc_hint_tensor_auto(matmul, (ccv_nnc_tensor_param_t []){
3466 a_params,
3467 b_params,
3468 }, 2, ccv_nnc_no_hint, &output_params, 1);
3469 const ccv_nnc_tensor_symbol_t matmul_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3470 ccv_nnc_graph_exec_symbol_new(graph, matmul, inputs, input_size, TENSOR_SYMBOL_LIST(matmul_output)(const ccv_nnc_tensor_symbol_t []){matmul_output}, (1 +1 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1
), "matmul");
3471 outputs[0] = matmul_output;
3472}
3473
3474static ccv_cnnp_model_t* _ccv_cnnp_matmul_copy(const ccv_cnnp_model_t* const super, void* const context);
3475
3476static const ccv_cnnp_model_vtab_t ccv_cnnp_matmul_isa = {
3477 .build = _ccv_cnnp_matmul_build,
3478 .copy = _ccv_cnnp_matmul_copy,
3479};
3480
3481ccv_cnnp_model_t* ccv_cnnp_matmul(const int transpose_a[2], const int transpose_b[2], const int flags, const char* const name)
3482{
3483 ccv_cnnp_model_matmul_t* const model_matmul = (ccv_cnnp_model_matmul_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_matmul_t));
3484 model_matmul->super.isa = &ccv_cnnp_matmul_isa;
3485 model_matmul->super.input_size = 2;
3486 model_matmul->super.outputs = &model_matmul->output;
3487 model_matmul->super.output_size = 1;
3488 model_matmul->transpose_a[0] = transpose_a[0];
3489 model_matmul->transpose_a[1] = transpose_a[1];
3490 model_matmul->transpose_b[0] = transpose_b[0];
3491 model_matmul->transpose_b[1] = transpose_b[1];
3492 model_matmul->flags = flags;
3493 ccv_cnnp_model_copy_name(&model_matmul->super, name);
3494 return (ccv_cnnp_model_t*)model_matmul;
3495}
3496
3497static ccv_cnnp_model_t* _ccv_cnnp_matmul_copy(const ccv_cnnp_model_t* const super, void* const context)
3498{
3499 const ccv_cnnp_model_matmul_t* const self = (const ccv_cnnp_model_matmul_t*)super;
3500 return ccv_cnnp_matmul(self->transpose_a, self->transpose_b, self->flags, self->super.name);
3501}
3502
3503// MARK - Dropout Layer
3504
3505typedef struct {
3506 ccv_cnnp_model_t super;
3507 ccv_nnc_tensor_symbol_t output;
3508 ccv_nnc_graph_exec_symbol_t dropout;
3509 float p;
3510 int entirety;
3511} ccv_cnnp_model_dropout_t;
3512
3513static void _ccv_cnnp_dropout_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
3514{
3515 PRINT(CCV_CLI_VERBOSE, "[cnnp_dropout_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_dropout_build] -\n"); fflush(stdout); } } while
(0);
3516 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 3516, __extension__ __PRETTY_FUNCTION__); }));
3517 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 3517, __extension__ __PRETTY_FUNCTION__
); }));
3518 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3519 ccv_nnc_tensor_param_t output_params[2];
3520 ccv_cnnp_model_dropout_t* const self = (ccv_cnnp_model_dropout_t*)super;
3521 const ccv_nnc_cmd_t dropout = CMD_DROPOUT_FORWARD(self->p, self->entirety)ccv_nnc_cmd(CCV_NNC_DROPOUT_FORWARD, 0, ((ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.dropout={.p=self->p,.entirety=self
->entirety}}), 0);
3522 ccv_nnc_hint_tensor_auto(dropout, (ccv_nnc_tensor_param_t []){
3523 params,
3524 }, 1, ccv_nnc_no_hint, output_params, 2);
3525 const ccv_nnc_tensor_symbol_t dropout_output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0);
3526 const ccv_nnc_tensor_symbol_t mask = ccv_nnc_tensor_symbol_new(graph, output_params[1], "mask");
3527 self->dropout = ccv_nnc_graph_exec_symbol_new(graph, dropout, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(dropout_output, mask)(const ccv_nnc_tensor_symbol_t []){dropout_output, mask}, (1 +
1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 -1), "dropout");
3528 outputs[0] = dropout_output;
3529}
3530
3531static void _ccv_cnnp_dropout_set_is_test(ccv_cnnp_model_t* const super, const int is_test, const ccv_cnnp_cmd_updater_f updater, void* const context)
3532{
3533 ccv_cnnp_model_dropout_t* const self = (ccv_cnnp_model_dropout_t*)super;
3534 if (self->dropout.graph)
3535 {
3536 if (is_test)
3537 // During test, the dropout is not applied. Data transfer is perfect because if these are the same tensor, it will skip.
3538 updater(context, self->dropout, CMD_DATA_TRANSFER_FORWARD()ccv_nnc_cmd(CCV_NNC_DATA_TRANSFER_FORWARD, 0, ccv_nnc_cmd_auto
, 0), ccv_nnc_no_hint);
3539 else
3540 updater(context, self->dropout, CMD_DROPOUT_FORWARD(self->p, self->entirety)ccv_nnc_cmd(CCV_NNC_DROPOUT_FORWARD, 0, ((ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.dropout={.p=self->p,.entirety=self
->entirety}}), 0), ccv_nnc_no_hint);
3541 }
3542}
3543
3544static ccv_cnnp_model_t* _ccv_cnnp_dropout_copy(const ccv_cnnp_model_t* const super, void* const context);
3545
3546static const ccv_cnnp_model_vtab_t ccv_cnnp_dropout_isa = {
3547 .build = _ccv_cnnp_dropout_build,
3548 .set_is_test = _ccv_cnnp_dropout_set_is_test,
3549 .copy = _ccv_cnnp_dropout_copy,
3550};
3551
3552ccv_cnnp_model_t* ccv_cnnp_dropout(const float p, const int entirety, const char* const name)
3553{
3554 ccv_cnnp_model_dropout_t* const model_dropout = (ccv_cnnp_model_dropout_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_dropout_t));
3555 model_dropout->super.isa = &ccv_cnnp_dropout_isa;
3556 model_dropout->super.input_size = 1;
3557 model_dropout->super.outputs = &model_dropout->output;
3558 model_dropout->super.output_size = 1;
3559 model_dropout->p = p;
3560 model_dropout->entirety = entirety;
3561 ccv_cnnp_model_copy_name(&model_dropout->super, name);
3562 return (ccv_cnnp_model_t*)model_dropout;
3563}
3564
3565static ccv_cnnp_model_t* _ccv_cnnp_dropout_copy(const ccv_cnnp_model_t* const super, void* const context)
3566{
3567 const ccv_cnnp_model_dropout_t* const self = (const ccv_cnnp_model_dropout_t*)super;
3568 return ccv_cnnp_dropout(self->p, self->entirety, self->super.name);
3569}
3570
3571// MARK - Masked Fill Layer
3572
3573typedef struct {
3574 ccv_cnnp_model_t super;
3575 ccv_nnc_tensor_symbol_t output;
3576 float eq;
3577 float fill;
3578} ccv_cnnp_model_masked_fill_t;
3579
3580static void _ccv_cnnp_masked_fill_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
3581{
3582 PRINT(CCV_CLI_VERBOSE, "[cnnp_masked_fill_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_masked_fill_build] -\n"); fflush(stdout); } }
while (0);
3583 assert(input_size == 2)((void) sizeof ((input_size == 2) ? 1 : 0), __extension__ ({ if
(input_size == 2) ; else __assert_fail ("input_size == 2", "ccv_cnnp_model_addons.c"
, 3583, __extension__ __PRETTY_FUNCTION__); }));
3584 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 3584, __extension__ __PRETTY_FUNCTION__
); }));
3585 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3586 ccv_cnnp_model_masked_fill_t* const self = (ccv_cnnp_model_masked_fill_t*)super;
3587 const ccv_nnc_tensor_symbol_t masked_fill_output = ccv_nnc_tensor_symbol_new(graph, params, 0);
3588 ccv_nnc_graph_exec_symbol_new(graph, CMD_MASKED_FILL_FORWARD(self->eq, self->fill)ccv_nnc_cmd(CCV_NNC_MASKED_FILL_FORWARD, 0, (ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.blas={.a={self->eq, self->fill}
}}, 0), TENSOR_SYMBOL_LIST(inputs[0], inputs[1])(const ccv_nnc_tensor_symbol_t []){inputs[0], inputs[1]}, (1 +
1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 -1), TENSOR_SYMBOL_LIST(masked_fill_output)(const ccv_nnc_tensor_symbol_t []){masked_fill_output}, (1 +1
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 -1), "masked_fill");
3589 outputs[0] = masked_fill_output;
3590}
3591
3592static ccv_cnnp_model_t* _ccv_cnnp_masked_fill_copy(const ccv_cnnp_model_t* const super, void* const context);
3593
3594static const ccv_cnnp_model_vtab_t ccv_cnnp_masked_fill_isa = {
3595 .build = _ccv_cnnp_masked_fill_build,
3596 .copy = _ccv_cnnp_masked_fill_copy,
3597};
3598
3599ccv_cnnp_model_t* ccv_cnnp_masked_fill(const float eq, const float fill, const char* const name)
3600{
3601 ccv_cnnp_model_masked_fill_t* const model_masked_fill = (ccv_cnnp_model_masked_fill_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_masked_fill_t));
3602 model_masked_fill->super.isa = &ccv_cnnp_masked_fill_isa;
3603 model_masked_fill->super.input_size = 2;
3604 model_masked_fill->super.outputs = &model_masked_fill->output;
3605 model_masked_fill->super.output_size = 1;
3606 model_masked_fill->eq = eq;
3607 model_masked_fill->fill = fill;
3608 ccv_cnnp_model_copy_name(&model_masked_fill->super, name);
3609 return (ccv_cnnp_model_t*)model_masked_fill;
3610}
3611
3612static ccv_cnnp_model_t* _ccv_cnnp_masked_fill_copy(const ccv_cnnp_model_t* const super, void* const context)
3613{
3614 const ccv_cnnp_model_masked_fill_t* const self = (const ccv_cnnp_model_masked_fill_t*)super;
3615 return ccv_cnnp_masked_fill(self->eq, self->fill, self->super.name);
3616}
3617
3618// MARK - Index Select Layer
3619
3620typedef struct {
3621 ccv_cnnp_model_t super;
3622 ccv_nnc_tensor_symbol_t output;
3623} ccv_cnnp_model_index_select_t;
3624
3625static void _ccv_cnnp_index_select_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
3626{
3627 PRINT(CCV_CLI_VERBOSE, "[cnnp_index_select_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_index_select_build] -\n"); fflush(stdout); }
} while (0);
3628 assert(input_size == 2)((void) sizeof ((input_size == 2) ? 1 : 0), __extension__ ({ if
(input_size == 2) ; else __assert_fail ("input_size == 2", "ccv_cnnp_model_addons.c"
, 3628, __extension__ __PRETTY_FUNCTION__); }));
3629 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 3629, __extension__ __PRETTY_FUNCTION__
); }));
3630 const ccv_nnc_tensor_param_t vocab_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3631 const ccv_nnc_tensor_param_t index_params = ccv_nnc_tensor_symbol_params(graph, inputs[1]);
3632 ccv_nnc_tensor_param_t output_params;
3633 const ccv_nnc_cmd_t index_select = CMD_INDEX_SELECT_FORWARD()ccv_nnc_cmd(CCV_NNC_INDEX_SELECT_FORWARD, 0, ccv_nnc_cmd_auto
, 0);
3634 ccv_nnc_hint_tensor_auto(index_select, (ccv_nnc_tensor_param_t []){
3635 vocab_params,
3636 index_params,
3637 }, 2, ccv_nnc_no_hint, &output_params, 1);
3638 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3639 ccv_nnc_graph_exec_symbol_new(graph, index_select, TENSOR_SYMBOL_LIST(inputs[0], inputs[1])(const ccv_nnc_tensor_symbol_t []){inputs[0], inputs[1]}, (1 +
1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 -1), TENSOR_SYMBOL_LIST(output)(const ccv_nnc_tensor_symbol_t []){output}, (1 +1 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "index_select");
3640 outputs[0] = output;
3641}
3642
3643static ccv_cnnp_model_t* _ccv_cnnp_index_select_copy(const ccv_cnnp_model_t* const super, void* const context);
3644
3645static const ccv_cnnp_model_vtab_t ccv_cnnp_index_select_isa = {
3646 .build = _ccv_cnnp_index_select_build,
3647 .copy = _ccv_cnnp_index_select_copy,
3648};
3649
3650ccv_cnnp_model_t* ccv_cnnp_index_select(const char* const name)
3651{
3652 ccv_cnnp_model_index_select_t* const model_index_select = (ccv_cnnp_model_index_select_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_index_select_t));
3653 model_index_select->super.isa = &ccv_cnnp_index_select_isa;
3654 model_index_select->super.input_size = 2;
3655 model_index_select->super.outputs = &model_index_select->output;
3656 model_index_select->super.output_size = 1;
3657 ccv_cnnp_model_copy_name(&model_index_select->super, name);
3658 return (ccv_cnnp_model_t*)model_index_select;
3659}
3660
3661static ccv_cnnp_model_t* _ccv_cnnp_index_select_copy(const ccv_cnnp_model_t* const super, void* const context)
3662{
3663 ccv_cnnp_model_index_select_t* const self = (ccv_cnnp_model_index_select_t*)super;
3664 return ccv_cnnp_index_select(self->super.name);
3665}
3666
3667// MARK - Embedding Layer
3668
3669typedef struct {
3670 ccv_cnnp_model_t super;
3671 ccv_nnc_tensor_symbol_t output;
3672 ccv_nnc_tensor_symbol_t vocab;
3673 int datatype;
3674 int vocab_size;
3675 int embed_size;
3676} ccv_cnnp_model_embedding_t;
3677
3678static void _ccv_cnnp_embedding_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
3679{
3680 ccv_cnnp_model_embedding_t* const self = (ccv_cnnp_model_embedding_t*)super;
3681 PRINT(CCV_CLI_VERBOSE, "[cnnp_embedding_build] vocab_size: %d, embed_size: %d\n", self->vocab_size, self->embed_size)do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_embedding_build] vocab_size: %d, embed_size: %d\n"
, self->vocab_size, self->embed_size); fflush(stdout); }
} while (0);
3682 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 3682, __extension__ __PRETTY_FUNCTION__); }));
3683 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 3683, __extension__ __PRETTY_FUNCTION__
); }));
3684 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3685 ccv_nnc_tensor_param_t vocab_params = params;
3686 memset(vocab_params.dim, 0, sizeof(vocab_params.dim));
3687 vocab_params.datatype = self->datatype;
3688 vocab_params.dim[0] = self->vocab_size;
3689 vocab_params.dim[1] = self->embed_size;
3690 if (!self->vocab.graph)
3691 self->vocab = ccv_nnc_tensor_symbol_new(graph, vocab_params, "vocab");
3692 assert(self->vocab.graph == graph)((void) sizeof ((self->vocab.graph == graph) ? 1 : 0), __extension__
({ if (self->vocab.graph == graph) ; else __assert_fail (
"self->vocab.graph == graph", "ccv_cnnp_model_addons.c", 3692
, __extension__ __PRETTY_FUNCTION__); }));
3693 const ccv_nnc_tensor_symbol_t vocab = ccv_cnnp_model_get_symbol(super, self->vocab);
3694 ccv_nnc_tensor_param_t output_params;
3695 const ccv_nnc_cmd_t embedding = CMD_INDEX_SELECT_FORWARD()ccv_nnc_cmd(CCV_NNC_INDEX_SELECT_FORWARD, 0, ccv_nnc_cmd_auto
, 0);
3696 ccv_nnc_hint_tensor_auto(embedding, (ccv_nnc_tensor_param_t []){
3697 vocab_params,
3698 params,
3699 }, 2, ccv_nnc_no_hint, &output_params, 1);
3700 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3701 ccv_nnc_graph_exec_symbol_new(graph, embedding, TENSOR_SYMBOL_LIST(vocab, inputs[0])(const ccv_nnc_tensor_symbol_t []){vocab, inputs[0]}, (1 +1 +
1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
-1), TENSOR_SYMBOL_LIST(output)(const ccv_nnc_tensor_symbol_t []){output}, (1 +1 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "embedding");
3702 outputs[0] = output;
3703}
3704
3705static void _ccv_cnnp_embedding_init_states(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_cnnp_state_initializer_f initializer, void* const context)
3706{
3707 ccv_cnnp_model_embedding_t* const self = (ccv_cnnp_model_embedding_t*)super;
3708 const float std = sqrtf(2) / sqrtf(self->vocab_size + self->embed_size);
3709 const float bound = sqrtf(3) * std;
3710 initializer(context, CMD_RANDOM_UNIFORM_FORWARD(-bound, bound)ccv_nnc_cmd(CCV_NNC_RANDOM_UNIFORM_FORWARD, 0, (ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.blas={.a={-bound, bound}}}, 0), ccv_nnc_no_hint, 0, 0, self->vocab);
3711}
3712
3713static void _ccv_cnnp_embedding_add_to_parameter(ccv_cnnp_model_t* const super, const ccv_cnnp_add_to_array_f add_to_array, void* const parameters, const int is_trainable)
3714{
3715 ccv_cnnp_model_embedding_t* const self = (ccv_cnnp_model_embedding_t*)super;
3716 add_to_array(parameters, self->vocab, is_trainable);
3717}
3718
3719static ccv_cnnp_model_t* _ccv_cnnp_embedding_copy(const ccv_cnnp_model_t* const super, void* const context);
3720
3721static const ccv_cnnp_model_vtab_t ccv_cnnp_embedding_isa = {
3722 .build = _ccv_cnnp_embedding_build,
3723 .init_states = _ccv_cnnp_embedding_init_states,
3724 .add_to_parameter = _ccv_cnnp_embedding_add_to_parameter,
3725 .copy = _ccv_cnnp_embedding_copy,
3726};
3727
3728ccv_cnnp_model_t* ccv_cnnp_embedding(const int datatype, const int vocab_size, const int embed_size, const int is_trainable, const char* const name)
3729{
3730 ccv_cnnp_model_embedding_t* const model_embedding = (ccv_cnnp_model_embedding_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_embedding_t));
3731 model_embedding->super.isa = &ccv_cnnp_embedding_isa;
3732 model_embedding->super.input_size = 1;
3733 model_embedding->super.outputs = &model_embedding->output;
3734 model_embedding->super.output_size = 1;
3735 model_embedding->super.is_trainable = is_trainable;
3736 ccv_cnnp_model_copy_name(&model_embedding->super, name);
3737 model_embedding->vocab.d = CCV_NNC_NO_TENSOR_SYMBOL;
3738 model_embedding->vocab.graph = 0;
3739 assert(datatype == CCV_32F || datatype == CCV_16F || datatype == CCV_16BF || datatype == CCV_32S)((void) sizeof ((datatype == CCV_32F || datatype == CCV_16F ||
datatype == CCV_16BF || datatype == CCV_32S) ? 1 : 0), __extension__
({ if (datatype == CCV_32F || datatype == CCV_16F || datatype
== CCV_16BF || datatype == CCV_32S) ; else __assert_fail ("datatype == CCV_32F || datatype == CCV_16F || datatype == CCV_16BF || datatype == CCV_32S"
, "ccv_cnnp_model_addons.c", 3739, __extension__ __PRETTY_FUNCTION__
); }));
3740 model_embedding->datatype = datatype;
3741 assert(vocab_size > 0)((void) sizeof ((vocab_size > 0) ? 1 : 0), __extension__ (
{ if (vocab_size > 0) ; else __assert_fail ("vocab_size > 0"
, "ccv_cnnp_model_addons.c", 3741, __extension__ __PRETTY_FUNCTION__
); }));
3742 model_embedding->vocab_size = vocab_size;
3743 assert(embed_size > 0)((void) sizeof ((embed_size > 0) ? 1 : 0), __extension__ (
{ if (embed_size > 0) ; else __assert_fail ("embed_size > 0"
, "ccv_cnnp_model_addons.c", 3743, __extension__ __PRETTY_FUNCTION__
); }));
3744 model_embedding->embed_size = embed_size;
3745 return (ccv_cnnp_model_t*)model_embedding;
3746}
3747
3748static ccv_cnnp_model_t* _ccv_cnnp_embedding_copy(const ccv_cnnp_model_t* const super, void* const context)
3749{
3750 ccv_cnnp_model_embedding_t* const self = (ccv_cnnp_model_embedding_t*)super;
3751 return ccv_cnnp_embedding(self->datatype, self->vocab_size, self->embed_size, self->super.is_trainable, self->super.name);
3752}
3753
3754// MARK - Pool Layers
3755
3756typedef struct {
3757 ccv_cnnp_model_t super;
3758 ccv_nnc_tensor_symbol_t output;
3759 int type;
3760 float width_scale;
3761 float height_scale;
3762 int align_corners;
3763} ccv_cnnp_model_upsample_t;
3764
3765static void _ccv_cnnp_upsample_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
3766{
3767 PRINT(CCV_CLI_VERBOSE, "[cnnp_upsample_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_upsample_build] -\n"); fflush(stdout); } } while
(0);
3768 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 3768, __extension__ __PRETTY_FUNCTION__); }));
3769 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 3769, __extension__ __PRETTY_FUNCTION__
); }));
3770 ccv_cnnp_model_upsample_t* const self = (ccv_cnnp_model_upsample_t*)super;
3771 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3772 ccv_nnc_cmd_t cmd = CMD_UPSAMPLE_FORWARD(self->type, self->width_scale, self->height_scale, self->align_corners)ccv_nnc_cmd(CCV_NNC_UPSAMPLE_FORWARD, 0, ((ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.upsample={.type=self->type,.width_scale
=self->width_scale,.height_scale=self->height_scale,.align_corners
=self->align_corners}}), 0);
3773 ccv_nnc_tensor_param_t output_params;
3774 ccv_nnc_hint_tensor_auto(cmd, &params, 1, ccv_nnc_no_hint, &output_params, 1);
3775 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3776 ccv_nnc_graph_exec_symbol_new(graph, cmd, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(output)(const ccv_nnc_tensor_symbol_t []){output}, (1 +1 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "upsample");
3777 outputs[0] = output;
3778}
3779
3780static ccv_cnnp_model_t* _ccv_cnnp_upsample_copy(const ccv_cnnp_model_t* const super, void* const context);
3781
3782static const ccv_cnnp_model_vtab_t ccv_cnnp_upsample_isa = {
3783 .build = _ccv_cnnp_upsample_build,
3784 .copy = _ccv_cnnp_upsample_copy,
3785};
3786
3787ccv_cnnp_model_t* ccv_cnnp_upsample(const int type, const float width_scale, const float height_scale, const int align_corners, const char* const name)
3788{
3789 ccv_cnnp_model_upsample_t* const model_upsample = (ccv_cnnp_model_upsample_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_upsample_t));
3790 model_upsample->super.isa = &ccv_cnnp_upsample_isa;
3791 model_upsample->super.input_size = 1;
3792 model_upsample->super.outputs = &model_upsample->output;
3793 model_upsample->super.output_size = 1;
3794 ccv_cnnp_model_copy_name(&model_upsample->super, name);
3795 assert(type == CCV_NNC_UPSAMPLE_NEAREST || type == CCV_NNC_UPSAMPLE_BILINEAR)((void) sizeof ((type == CCV_NNC_UPSAMPLE_NEAREST || type == CCV_NNC_UPSAMPLE_BILINEAR
) ? 1 : 0), __extension__ ({ if (type == CCV_NNC_UPSAMPLE_NEAREST
|| type == CCV_NNC_UPSAMPLE_BILINEAR) ; else __assert_fail (
"type == CCV_NNC_UPSAMPLE_NEAREST || type == CCV_NNC_UPSAMPLE_BILINEAR"
, "ccv_cnnp_model_addons.c", 3795, __extension__ __PRETTY_FUNCTION__
); }));
3796 model_upsample->type = type;
3797 model_upsample->width_scale = width_scale;
3798 model_upsample->height_scale = height_scale;
3799 model_upsample->align_corners = align_corners;
3800 return (ccv_cnnp_model_t*)model_upsample;
3801}
3802
3803static ccv_cnnp_model_t* _ccv_cnnp_upsample_copy(const ccv_cnnp_model_t* const super, void* const context)
3804{
3805 const ccv_cnnp_model_upsample_t* const self = (const ccv_cnnp_model_upsample_t*)super;
3806 return ccv_cnnp_upsample(self->type, self->width_scale, self->height_scale, self->align_corners, self->super.name);
3807}
3808
3809// MARK - Reduce Sum Layer
3810
3811typedef struct {
3812 ccv_cnnp_model_t super;
3813 int axis[CCV_NNC_MAX_DIM_ALLOC(12)];
3814 int count;
3815 ccv_nnc_tensor_symbol_t output;
3816} ccv_cnnp_model_reduce_sum_t;
3817
3818static void _ccv_cnnp_reduce_sum_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
3819{
3820 PRINT(CCV_CLI_VERBOSE, "[cnnp_reduce_sum_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_reduce_sum_build] -\n"); fflush(stdout); } }
while (0);
3821 const ccv_cnnp_model_reduce_sum_t* const self = (const ccv_cnnp_model_reduce_sum_t*)super;
3822 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 3822, __extension__ __PRETTY_FUNCTION__); }));
3823 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 3823, __extension__ __PRETTY_FUNCTION__
); }));
3824 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3825 ccv_nnc_tensor_param_t output_params;
3826 ccv_nnc_cmd_t reduce_sum = CMD_REDUCE_SUM_FORWARD()ccv_nnc_cmd(CCV_NNC_REDUCE_SUM_FORWARD, 0, ((ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.reduce={.count=(1 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1),.axis={}}}
), 0);
3827 int i;
3828 for (i = 0; i < self->count; i++)
3829 reduce_sum.info.reduce.axis[i] = self->axis[i];
3830 reduce_sum.info.reduce.count = self->count;
3831 ccv_nnc_hint_tensor_auto(reduce_sum, &input_params, 1, ccv_nnc_no_hint, &output_params, 1);
3832 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3833 ccv_nnc_graph_exec_symbol_new(graph, reduce_sum, inputs, input_size, outputs, output_size, "reduce_sum");
3834}
3835
3836static ccv_cnnp_model_t* _ccv_cnnp_reduce_sum_copy(const ccv_cnnp_model_t* const self, void* const context);
3837
3838static const ccv_cnnp_model_vtab_t ccv_cnnp_reduce_sum_isa = {
3839 .build = _ccv_cnnp_reduce_sum_build,
3840 .copy = _ccv_cnnp_reduce_sum_copy,
3841};
3842
3843ccv_cnnp_model_t* ccv_cnnp_reduce_sum(const int* const axis, const int axis_count, const char* const name)
3844{
3845 ccv_cnnp_model_reduce_sum_t* const model_reduce_sum = (ccv_cnnp_model_reduce_sum_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_reduce_sum_t));
3846 model_reduce_sum->super.isa = &ccv_cnnp_reduce_sum_isa;
3847 model_reduce_sum->super.input_size = 1;
3848 model_reduce_sum->super.outputs = &model_reduce_sum->output;
3849 model_reduce_sum->super.output_size = 1;
3850 ccv_cnnp_model_copy_name(&model_reduce_sum->super, name);
3851 assert(axis_count <= CCV_NNC_MAX_DIM_ALLOC)((void) sizeof ((axis_count <= (12)) ? 1 : 0), __extension__
({ if (axis_count <= (12)) ; else __assert_fail ("axis_count <= CCV_NNC_MAX_DIM_ALLOC"
, "ccv_cnnp_model_addons.c", 3851, __extension__ __PRETTY_FUNCTION__
); }));
3852 int i;
3853 for (i = 0; i < axis_count; i++)
3854 model_reduce_sum->axis[i] = axis[i];
3855 model_reduce_sum->count = axis_count;
3856 return (ccv_cnnp_model_t*)model_reduce_sum;
3857}
3858
3859static ccv_cnnp_model_t* _ccv_cnnp_reduce_sum_copy(const ccv_cnnp_model_t* const super, void* const context)
3860{
3861 const ccv_cnnp_model_reduce_sum_t* const self = (const ccv_cnnp_model_reduce_sum_t*)super;
3862 return ccv_cnnp_reduce_sum(self->axis, self->count, self->super.name);
3863}
3864
3865// MARK - Reduce Mean Layer
3866
3867typedef struct {
3868 ccv_cnnp_model_t super;
3869 int axis[CCV_NNC_MAX_DIM_ALLOC(12)];
3870 int count;
3871 ccv_nnc_tensor_symbol_t output;
3872} ccv_cnnp_model_reduce_mean_t;
3873
3874static void _ccv_cnnp_reduce_mean_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
3875{
3876 PRINT(CCV_CLI_VERBOSE, "[cnnp_reduce_mean_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_reduce_mean_build] -\n"); fflush(stdout); } }
while (0);
3877 const ccv_cnnp_model_reduce_mean_t* const self = (const ccv_cnnp_model_reduce_mean_t*)super;
3878 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 3878, __extension__ __PRETTY_FUNCTION__); }));
3879 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 3879, __extension__ __PRETTY_FUNCTION__
); }));
3880 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3881 ccv_nnc_tensor_param_t output_params;
3882 ccv_nnc_cmd_t reduce_mean = CMD_REDUCE_MEAN_FORWARD()ccv_nnc_cmd(CCV_NNC_REDUCE_MEAN_FORWARD, 0, ((ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.reduce={.count=(1 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1),.axis={}}}
), 0);
3883 int i;
3884 for (i = 0; i < self->count; i++)
3885 reduce_mean.info.reduce.axis[i] = self->axis[i];
3886 reduce_mean.info.reduce.count = self->count;
3887 ccv_nnc_hint_tensor_auto(reduce_mean, &input_params, 1, ccv_nnc_no_hint, &output_params, 1);
3888 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3889 ccv_nnc_graph_exec_symbol_new(graph, reduce_mean, inputs, input_size, outputs, output_size, "reduce_mean");
3890}
3891
3892static ccv_cnnp_model_t* _ccv_cnnp_reduce_mean_copy(const ccv_cnnp_model_t* const self, void* const context);
3893
3894static const ccv_cnnp_model_vtab_t ccv_cnnp_reduce_mean_isa = {
3895 .build = _ccv_cnnp_reduce_mean_build,
3896 .copy = _ccv_cnnp_reduce_mean_copy,
3897};
3898
3899ccv_cnnp_model_t* ccv_cnnp_reduce_mean(const int* const axis, const int axis_count, const char* const name)
3900{
3901 ccv_cnnp_model_reduce_mean_t* const model_reduce_mean = (ccv_cnnp_model_reduce_mean_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_reduce_mean_t));
3902 model_reduce_mean->super.isa = &ccv_cnnp_reduce_mean_isa;
3903 model_reduce_mean->super.input_size = 1;
3904 model_reduce_mean->super.outputs = &model_reduce_mean->output;
3905 model_reduce_mean->super.output_size = 1;
3906 ccv_cnnp_model_copy_name(&model_reduce_mean->super, name);
3907 assert(axis_count <= CCV_NNC_MAX_DIM_ALLOC)((void) sizeof ((axis_count <= (12)) ? 1 : 0), __extension__
({ if (axis_count <= (12)) ; else __assert_fail ("axis_count <= CCV_NNC_MAX_DIM_ALLOC"
, "ccv_cnnp_model_addons.c", 3907, __extension__ __PRETTY_FUNCTION__
); }));
3908 int i;
3909 for (i = 0; i < axis_count; i++)
3910 model_reduce_mean->axis[i] = axis[i];
3911 model_reduce_mean->count = axis_count;
3912 return (ccv_cnnp_model_t*)model_reduce_mean;
3913}
3914
3915static ccv_cnnp_model_t* _ccv_cnnp_reduce_mean_copy(const ccv_cnnp_model_t* const super, void* const context)
3916{
3917 const ccv_cnnp_model_reduce_mean_t* const self = (const ccv_cnnp_model_reduce_mean_t*)super;
3918 return ccv_cnnp_reduce_mean(self->axis, self->count, self->super.name);
3919}
3920
3921// MARK - Reduce Max Layer
3922
3923typedef struct {
3924 ccv_cnnp_model_t super;
3925 int axis[CCV_NNC_MAX_DIM_ALLOC(12)];
3926 int count;
3927 ccv_nnc_tensor_symbol_t output;
3928} ccv_cnnp_model_reduce_max_t;
3929
3930static void _ccv_cnnp_reduce_max_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
3931{
3932 PRINT(CCV_CLI_VERBOSE, "[cnnp_reduce_max_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_reduce_max_build] -\n"); fflush(stdout); } }
while (0);
3933 const ccv_cnnp_model_reduce_max_t* const self = (const ccv_cnnp_model_reduce_max_t*)super;
3934 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 3934, __extension__ __PRETTY_FUNCTION__); }));
3935 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 3935, __extension__ __PRETTY_FUNCTION__
); }));
3936 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3937 ccv_nnc_tensor_param_t output_params;
3938 ccv_nnc_cmd_t reduce_max = CMD_REDUCE_MAX_FORWARD()ccv_nnc_cmd(CCV_NNC_REDUCE_MAX_FORWARD, 0, ((ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.reduce={.count=(1 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1),.axis={}}}
), 0);
3939 int i;
3940 for (i = 0; i < self->count; i++)
3941 reduce_max.info.reduce.axis[i] = self->axis[i];
3942 reduce_max.info.reduce.count = self->count;
3943 ccv_nnc_hint_tensor_auto(reduce_max, &input_params, 1, ccv_nnc_no_hint, &output_params, 1);
3944 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3945 ccv_nnc_graph_exec_symbol_new(graph, reduce_max, inputs, input_size, outputs, output_size, "reduce_max");
3946}
3947
3948static ccv_cnnp_model_t* _ccv_cnnp_reduce_max_copy(const ccv_cnnp_model_t* const self, void* const context);
3949
3950static const ccv_cnnp_model_vtab_t ccv_cnnp_reduce_max_isa = {
3951 .build = _ccv_cnnp_reduce_max_build,
3952 .copy = _ccv_cnnp_reduce_max_copy,
3953};
3954
3955ccv_cnnp_model_t* ccv_cnnp_reduce_max(const int* const axis, const int axis_count, const char* const name)
3956{
3957 ccv_cnnp_model_reduce_max_t* const model_reduce_max = (ccv_cnnp_model_reduce_max_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_reduce_max_t));
3958 model_reduce_max->super.isa = &ccv_cnnp_reduce_max_isa;
3959 model_reduce_max->super.input_size = 1;
3960 model_reduce_max->super.outputs = &model_reduce_max->output;
3961 model_reduce_max->super.output_size = 1;
3962 ccv_cnnp_model_copy_name(&model_reduce_max->super, name);
3963 assert(axis_count <= CCV_NNC_MAX_DIM_ALLOC)((void) sizeof ((axis_count <= (12)) ? 1 : 0), __extension__
({ if (axis_count <= (12)) ; else __assert_fail ("axis_count <= CCV_NNC_MAX_DIM_ALLOC"
, "ccv_cnnp_model_addons.c", 3963, __extension__ __PRETTY_FUNCTION__
); }));
3964 int i;
3965 for (i = 0; i < axis_count; i++)
3966 model_reduce_max->axis[i] = axis[i];
3967 model_reduce_max->count = axis_count;
3968 return (ccv_cnnp_model_t*)model_reduce_max;
3969}
3970
3971static ccv_cnnp_model_t* _ccv_cnnp_reduce_max_copy(const ccv_cnnp_model_t* const super, void* const context)
3972{
3973 const ccv_cnnp_model_reduce_max_t* const self = (const ccv_cnnp_model_reduce_max_t*)super;
3974 return ccv_cnnp_reduce_max(self->axis, self->count, self->super.name);
3975}
3976
3977// MARK - Reduce Min Layer
3978
3979typedef struct {
3980 ccv_cnnp_model_t super;
3981 int axis[CCV_NNC_MAX_DIM_ALLOC(12)];
3982 int count;
3983 ccv_nnc_tensor_symbol_t output;
3984} ccv_cnnp_model_reduce_min_t;
3985
3986static void _ccv_cnnp_reduce_min_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
3987{
3988 PRINT(CCV_CLI_VERBOSE, "[cnnp_reduce_min_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_reduce_min_build] -\n"); fflush(stdout); } }
while (0);
3989 const ccv_cnnp_model_reduce_min_t* const self = (const ccv_cnnp_model_reduce_min_t*)super;
3990 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 3990, __extension__ __PRETTY_FUNCTION__); }));
3991 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 3991, __extension__ __PRETTY_FUNCTION__
); }));
3992 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3993 ccv_nnc_tensor_param_t output_params;
3994 ccv_nnc_cmd_t reduce_min = CMD_REDUCE_MIN_FORWARD()ccv_nnc_cmd(CCV_NNC_REDUCE_MIN_FORWARD, 0, ((ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.reduce={.count=(1 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1),.axis={}}}
), 0);
3995 int i;
3996 for (i = 0; i < self->count; i++)
3997 reduce_min.info.reduce.axis[i] = self->axis[i];
3998 reduce_min.info.reduce.count = self->count;
3999 ccv_nnc_hint_tensor_auto(reduce_min, &input_params, 1, ccv_nnc_no_hint, &output_params, 1);
4000 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
4001 ccv_nnc_graph_exec_symbol_new(graph, reduce_min, inputs, input_size, outputs, output_size, "reduce_min");
4002}
4003
4004static ccv_cnnp_model_t* _ccv_cnnp_reduce_min_copy(const ccv_cnnp_model_t* const self, void* const context);
4005
4006static const ccv_cnnp_model_vtab_t ccv_cnnp_reduce_min_isa = {
4007 .build = _ccv_cnnp_reduce_min_build,
4008 .copy = _ccv_cnnp_reduce_min_copy,
4009};
4010
4011ccv_cnnp_model_t* ccv_cnnp_reduce_min(const int* const axis, const int axis_count, const char* const name)
4012{
4013 ccv_cnnp_model_reduce_min_t* const model_reduce_min = (ccv_cnnp_model_reduce_min_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_reduce_min_t));
4014 model_reduce_min->super.isa = &ccv_cnnp_reduce_min_isa;
4015 model_reduce_min->super.input_size = 1;
4016 model_reduce_min->super.outputs = &model_reduce_min->output;
4017 model_reduce_min->super.output_size = 1;
4018 ccv_cnnp_model_copy_name(&model_reduce_min->super, name);
4019 assert(axis_count <= CCV_NNC_MAX_DIM_ALLOC)((void) sizeof ((axis_count <= (12)) ? 1 : 0), __extension__
({ if (axis_count <= (12)) ; else __assert_fail ("axis_count <= CCV_NNC_MAX_DIM_ALLOC"
, "ccv_cnnp_model_addons.c", 4019, __extension__ __PRETTY_FUNCTION__
); }));
4020 int i;
4021 for (i = 0; i < axis_count; i++)
4022 model_reduce_min->axis[i] = axis[i];
4023 model_reduce_min->count = axis_count;
4024 return (ccv_cnnp_model_t*)model_reduce_min;
4025}
4026
4027static ccv_cnnp_model_t* _ccv_cnnp_reduce_min_copy(const ccv_cnnp_model_t* const super, void* const context)
4028{
4029 const ccv_cnnp_model_reduce_min_t* const self = (const ccv_cnnp_model_reduce_min_t*)super;
4030 return ccv_cnnp_reduce_min(self->axis, self->count, self->super.name);
4031}
4032
4033// MARK - Reduce Norm2 Layer
4034
4035typedef struct {
4036 ccv_cnnp_model_t super;
4037 int axis[CCV_NNC_MAX_DIM_ALLOC(12)];
4038 int count;
4039 ccv_nnc_tensor_symbol_t output;
4040} ccv_cnnp_model_reduce_norm2_t;
4041
4042static void _ccv_cnnp_reduce_norm2_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
4043{
4044 const ccv_cnnp_model_reduce_norm2_t* const self = (const ccv_cnnp_model_reduce_norm2_t*)super;
4045 PRINT(CCV_CLI_VERBOSE, "[cnnp_reduce_norm2_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_reduce_norm2_build] -\n"); fflush(stdout); }
} while (0);
4046 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 4046, __extension__ __PRETTY_FUNCTION__); }));
4047 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 4047, __extension__ __PRETTY_FUNCTION__
); }));
4048 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4049 ccv_nnc_tensor_param_t output_params;
4050 ccv_nnc_cmd_t reduce_norm2 = CMD_REDUCE_NORM2_FORWARD()ccv_nnc_cmd(CCV_NNC_REDUCE_NORM2_FORWARD, 0, ((ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.reduce={.count=(1 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1),.axis={}}}
), 0);
4051 int i;
4052 for (i = 0; i < self->count; i++)
4053 reduce_norm2.info.reduce.axis[i] = self->axis[i];
4054 reduce_norm2.info.reduce.count = self->count;
4055 ccv_nnc_hint_tensor_auto(reduce_norm2, &input_params, 1, ccv_nnc_no_hint, &output_params, 1);
4056 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
4057 ccv_nnc_graph_exec_symbol_new(graph, reduce_norm2, inputs, input_size, outputs, output_size, "reduce_norm2");
4058}
4059
4060static ccv_cnnp_model_t* _ccv_cnnp_reduce_norm2_copy(const ccv_cnnp_model_t* const self, void* const context);
4061
4062static const ccv_cnnp_model_vtab_t ccv_cnnp_reduce_norm2_isa = {
4063 .build = _ccv_cnnp_reduce_norm2_build,
4064 .copy = _ccv_cnnp_reduce_norm2_copy,
4065};
4066
4067ccv_cnnp_model_t* ccv_cnnp_reduce_norm2(const int* const axis, const int axis_count, const char* const name)
4068{
4069 ccv_cnnp_model_reduce_norm2_t* const model_reduce_norm2 = (ccv_cnnp_model_reduce_norm2_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_reduce_norm2_t));
4070 model_reduce_norm2->super.isa = &ccv_cnnp_reduce_norm2_isa;
4071 model_reduce_norm2->super.input_size = 1;
4072 model_reduce_norm2->super.outputs = &model_reduce_norm2->output;
4073 model_reduce_norm2->super.output_size = 1;
4074 ccv_cnnp_model_copy_name(&model_reduce_norm2->super, name);
4075 assert(axis_count <= CCV_NNC_MAX_DIM_ALLOC)((void) sizeof ((axis_count <= (12)) ? 1 : 0), __extension__
({ if (axis_count <= (12)) ; else __assert_fail ("axis_count <= CCV_NNC_MAX_DIM_ALLOC"
, "ccv_cnnp_model_addons.c", 4075, __extension__ __PRETTY_FUNCTION__
); }));
4076 int i;
4077 for (i = 0; i < axis_count; i++)
4078 model_reduce_norm2->axis[i] = axis[i];
4079 model_reduce_norm2->count = axis_count;
4080 return (ccv_cnnp_model_t*)model_reduce_norm2;
4081}
4082
4083static ccv_cnnp_model_t* _ccv_cnnp_reduce_norm2_copy(const ccv_cnnp_model_t* const super, void* const context)
4084{
4085 const ccv_cnnp_model_reduce_norm2_t* const self = (const ccv_cnnp_model_reduce_norm2_t*)super;
4086 return ccv_cnnp_reduce_norm2(self->axis, self->count, self->super.name);
4087}
4088
4089// MARK - Argmax Layer
4090
4091typedef struct {
4092 ccv_cnnp_model_t super;
4093 int axis;
4094 ccv_nnc_tensor_symbol_t output;
4095} ccv_cnnp_model_argmax_t;
4096
4097static void _ccv_cnnp_argmax_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
4098{
4099 const ccv_cnnp_model_argmax_t* const self = (const ccv_cnnp_model_argmax_t*)super;
4100 PRINT(CCV_CLI_VERBOSE, "[cnnp_argmax_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_argmax_build] -\n"); fflush(stdout); } } while
(0);
4101 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 4101, __extension__ __PRETTY_FUNCTION__); }));
4102 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 4102, __extension__ __PRETTY_FUNCTION__
); }));
4103 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4104 ccv_nnc_tensor_param_t output_params;
4105 ccv_nnc_cmd_t argmax = CMD_ARGMAX_FORWARD()ccv_nnc_cmd(CCV_NNC_ARGMAX_FORWARD, 0, ((ccv_nnc_cmd_param_t)
{.size={.dim={1,1,1}},.reduce={.count=(1 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1),.axis={}}})
, 0);
4106 argmax.info.reduce.axis[0] = self->axis;
4107 argmax.info.reduce.count = 1;
4108 ccv_nnc_hint_tensor_auto(argmax, &input_params, 1, ccv_nnc_no_hint, &output_params, 1);
4109 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
4110 ccv_nnc_graph_exec_symbol_new(graph, argmax, inputs, input_size, outputs, output_size, "argmax");
4111}
4112
4113static ccv_cnnp_model_t* _ccv_cnnp_argmax_copy(const ccv_cnnp_model_t* const self, void* const context);
4114
4115static const ccv_cnnp_model_vtab_t ccv_cnnp_argmax_isa = {
4116 .build = _ccv_cnnp_argmax_build,
4117 .copy = _ccv_cnnp_argmax_copy,
4118};
4119
4120ccv_cnnp_model_t* ccv_cnnp_argmax(const int axis, const char* const name)
4121{
4122 ccv_cnnp_model_argmax_t* const model_argmax = (ccv_cnnp_model_argmax_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_argmax_t));
4123 model_argmax->super.isa = &ccv_cnnp_argmax_isa;
4124 model_argmax->super.input_size = 1;
4125 model_argmax->super.outputs = &model_argmax->output;
4126 model_argmax->super.output_size = 1;
4127 ccv_cnnp_model_copy_name(&model_argmax->super, name);
4128 model_argmax->axis = axis;
4129 return (ccv_cnnp_model_t*)model_argmax;
4130}
4131
4132static ccv_cnnp_model_t* _ccv_cnnp_argmax_copy(const ccv_cnnp_model_t* const super, void* const context)
4133{
4134 const ccv_cnnp_model_argmax_t* const self = (const ccv_cnnp_model_argmax_t*)super;
4135 return ccv_cnnp_argmax(self->axis, self->super.name);
4136}
4137
4138// MARK - Argmin Layer
4139
4140typedef struct {
4141 ccv_cnnp_model_t super;
4142 int axis;
4143 ccv_nnc_tensor_symbol_t output;
4144} ccv_cnnp_model_argmin_t;
4145
4146static void _ccv_cnnp_argmin_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
4147{
4148 const ccv_cnnp_model_argmin_t* const self = (const ccv_cnnp_model_argmin_t*)super;
4149 PRINT(CCV_CLI_VERBOSE, "[cnnp_argmin_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_argmin_build] -\n"); fflush(stdout); } } while
(0);
4150 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 4150, __extension__ __PRETTY_FUNCTION__); }));
4151 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 4151, __extension__ __PRETTY_FUNCTION__
); }));
4152 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4153 ccv_nnc_tensor_param_t output_params;
4154 ccv_nnc_cmd_t argmin = CMD_ARGMIN_FORWARD()ccv_nnc_cmd(CCV_NNC_ARGMIN_FORWARD, 0, ((ccv_nnc_cmd_param_t)
{.size={.dim={1,1,1}},.reduce={.count=(1 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1),.axis={}}})
, 0);
4155 argmin.info.reduce.axis[0] = self->axis;
4156 argmin.info.reduce.count = 1;
4157 ccv_nnc_hint_tensor_auto(argmin, &input_params, 1, ccv_nnc_no_hint, &output_params, 1);
4158 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
4159 ccv_nnc_graph_exec_symbol_new(graph, argmin, inputs, input_size, outputs, output_size, "argmin");
4160}
4161
4162static ccv_cnnp_model_t* _ccv_cnnp_argmin_copy(const ccv_cnnp_model_t* const self, void* const context);
4163
4164static const ccv_cnnp_model_vtab_t ccv_cnnp_argmin_isa = {
4165 .build = _ccv_cnnp_argmin_build,
4166 .copy = _ccv_cnnp_argmin_copy,
4167};
4168
4169ccv_cnnp_model_t* ccv_cnnp_argmin(const int axis, const char* const name)
4170{
4171 ccv_cnnp_model_argmin_t* const model_argmin = (ccv_cnnp_model_argmin_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_argmin_t));
4172 model_argmin->super.isa = &ccv_cnnp_argmin_isa;
4173 model_argmin->super.input_size = 1;
4174 model_argmin->super.outputs = &model_argmin->output;
4175 model_argmin->super.output_size = 1;
4176 ccv_cnnp_model_copy_name(&model_argmin->super, name);
4177 model_argmin->axis = axis;
4178 return (ccv_cnnp_model_t*)model_argmin;
4179}
4180
4181static ccv_cnnp_model_t* _ccv_cnnp_argmin_copy(const ccv_cnnp_model_t* const super, void* const context)
4182{
4183 const ccv_cnnp_model_argmin_t* const self = (const ccv_cnnp_model_argmin_t*)super;
4184 return ccv_cnnp_argmin(self->axis, self->super.name);
4185}
4186
4187// MARK - Min Layer
4188
4189typedef struct {
4190 ccv_cnnp_model_t super;
4191 ccv_nnc_tensor_symbol_t output;
4192} ccv_cnnp_model_min_t;
4193
4194static void _ccv_cnnp_min_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
4195{
4196 PRINT(CCV_CLI_VERBOSE, "[cnnp_min_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_min_build] -\n"); fflush(stdout); } } while (
0);
4197 assert(input_size == 2)((void) sizeof ((input_size == 2) ? 1 : 0), __extension__ ({ if
(input_size == 2) ; else __assert_fail ("input_size == 2", "ccv_cnnp_model_addons.c"
, 4197, __extension__ __PRETTY_FUNCTION__); }));
4198 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 4198, __extension__ __PRETTY_FUNCTION__
); }));
4199 ccv_nnc_tensor_param_t input_params[2];
4200 int i;
4201 for (i = 0; i < 2; i++)
4202 input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
4203 ccv_nnc_tensor_param_t output_params;
4204 const ccv_nnc_cmd_t min = CMD_MIN_FORWARD()ccv_nnc_cmd(CCV_NNC_MIN_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}}}, 0);
4205 ccv_nnc_hint_tensor_auto(min, input_params, 2, ccv_nnc_no_hint, &output_params, 1);
4206 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
4207 ccv_nnc_graph_exec_symbol_new(graph, min, inputs, input_size, outputs, output_size, "min");
4208}
4209
4210static ccv_cnnp_model_t* _ccv_cnnp_min_copy(const ccv_cnnp_model_t* const self, void* const context);
4211
4212static const ccv_cnnp_model_vtab_t ccv_cnnp_min_isa = {
4213 .build = _ccv_cnnp_min_build,
4214 .copy = _ccv_cnnp_min_copy,
4215};
4216
4217ccv_cnnp_model_t* ccv_cnnp_min(const char* const name)
4218{
4219 ccv_cnnp_model_min_t* const model_min = (ccv_cnnp_model_min_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_min_t));
4220 model_min->super.isa = &ccv_cnnp_min_isa;
4221 model_min->super.input_size = 2;
4222 model_min->super.outputs = &model_min->output;
4223 model_min->super.output_size = 1;
4224 ccv_cnnp_model_copy_name(&model_min->super, name);
4225 return (ccv_cnnp_model_t*)model_min;
4226}
4227
4228static ccv_cnnp_model_t* _ccv_cnnp_min_copy(const ccv_cnnp_model_t* const super, void* const context)
4229{
4230 const ccv_cnnp_model_min_t* const self = (const ccv_cnnp_model_min_t*)super;
4231 return ccv_cnnp_min(self->super.name);
4232}
4233
4234// MARK - Max Layer
4235
4236typedef struct {
4237 ccv_cnnp_model_t super;
4238 ccv_nnc_tensor_symbol_t output;
4239} ccv_cnnp_model_max_t;
4240
4241static void _ccv_cnnp_max_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
4242{
4243 PRINT(CCV_CLI_VERBOSE, "[cnnp_max_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_max_build] -\n"); fflush(stdout); } } while (
0);
4244 assert(input_size == 2)((void) sizeof ((input_size == 2) ? 1 : 0), __extension__ ({ if
(input_size == 2) ; else __assert_fail ("input_size == 2", "ccv_cnnp_model_addons.c"
, 4244, __extension__ __PRETTY_FUNCTION__); }));
4245 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 4245, __extension__ __PRETTY_FUNCTION__
); }));
4246 ccv_nnc_tensor_param_t input_params[2];
4247 int i;
4248 for (i = 0; i < 2; i++)
4249 input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
4250 ccv_nnc_tensor_param_t output_params;
4251 const ccv_nnc_cmd_t max = CMD_MAX_FORWARD()ccv_nnc_cmd(CCV_NNC_MAX_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}}}, 0);
4252 ccv_nnc_hint_tensor_auto(max, input_params, 2, ccv_nnc_no_hint, &output_params, 1);
4253 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
4254 ccv_nnc_graph_exec_symbol_new(graph, max, inputs, input_size, outputs, output_size, "max");
4255}
4256
4257static ccv_cnnp_model_t* _ccv_cnnp_max_copy(const ccv_cnnp_model_t* const self, void* const context);
4258
4259static const ccv_cnnp_model_vtab_t ccv_cnnp_max_isa = {
4260 .build = _ccv_cnnp_max_build,
4261 .copy = _ccv_cnnp_max_copy,
4262};
4263
4264ccv_cnnp_model_t* ccv_cnnp_max(const char* const name)
4265{
4266 ccv_cnnp_model_max_t* const model_max = (ccv_cnnp_model_max_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_max_t));
4267 model_max->super.isa = &ccv_cnnp_max_isa;
4268 model_max->super.input_size = 2;
4269 model_max->super.outputs = &model_max->output;
4270 model_max->super.output_size = 1;
4271 ccv_cnnp_model_copy_name(&model_max->super, name);
4272 return (ccv_cnnp_model_t*)model_max;
4273}
4274
4275static ccv_cnnp_model_t* _ccv_cnnp_max_copy(const ccv_cnnp_model_t* const super, void* const context)
4276{
4277 const ccv_cnnp_model_max_t* const self = (const ccv_cnnp_model_max_t*)super;
4278 return ccv_cnnp_max(self->super.name);
4279}
4280
4281// MARK - LSTM Layer
4282
4283typedef struct {
4284 ccv_cnnp_model_t super;
4285 int masked;
4286 ccv_nnc_tensor_symbol_t output;
4287 ccv_nnc_tensor_symbol_t weights;
4288 ccv_nnc_tensor_symbol_t reserves;
4289 ccv_nnc_cmd_param_t params;
4290 ccv_nnc_graph_exec_symbol_t lstm;
4291} ccv_cnnp_model_lstm_t;
4292
4293static int _ccv_cnnp_lstm_weight_dim(int bidirectional, int num_layers, int input_size, int hidden_size, int proj_size, int bias)
4294{
4295 const int D = !!bidirectional + 1;
4296 if (hidden_size == proj_size)
4297 return (num_layers * (bias ? 8 : 0) + (num_layers - 1) * (hidden_size * 4 * D + hidden_size * 4) + input_size * 4 + hidden_size * 4) * D;
4298 else
4299 return (num_layers * (bias ? 8 : 0) + (num_layers - 1) * (proj_size * 4 * D + proj_size * 4) + (proj_size * 4 + input_size * 4) + num_layers * proj_size) * D;
4300}
4301
4302static void _ccv_cnnp_lstm_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
4303{
4304 ccv_cnnp_model_lstm_t* const self = (ccv_cnnp_model_lstm_t*)super;
4305 PRINT(CCV_CLI_VERBOSE, "[cnnp_lstm_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_lstm_build] -\n"); fflush(stdout); } } while
(0);
4306 assert(input_size == self->super.input_size)((void) sizeof ((input_size == self->super.input_size) ? 1
: 0), __extension__ ({ if (input_size == self->super.input_size
) ; else __assert_fail ("input_size == self->super.input_size"
, "ccv_cnnp_model_addons.c", 4306, __extension__ __PRETTY_FUNCTION__
); }));
4307 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 4307, __extension__ __PRETTY_FUNCTION__
); }));
4308 const int proj_size = self->params.rnn.proj_size == 0 ? self->params.rnn.hidden_size : self->params.rnn.proj_size;
4309 ccv_nnc_tensor_param_t input_params[5];
4310 input_params[0]= ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4311 if (input_size == 2)
4312 input_params[1] = ccv_nnc_tensor_symbol_params(graph, inputs[1]);
4313 input_params[4] = input_params[0];
4314 memset(input_params[4].dim, 0, sizeof(input_params[4].dim));
4315 const int x_nd = ccv_nnc_tensor_nd(input_params[0].dim);
4316 const int feature_count = input_params[0].dim[x_nd - 1];
4317 input_params[4].dim[0] = _ccv_cnnp_lstm_weight_dim(self->params.rnn.bidirectional, self->params.rnn.num_layers, feature_count, self->params.rnn.hidden_size, proj_size, self->params.rnn.bias);
4318 input_params[4].dim[1] = self->params.rnn.hidden_size;
4319 const ccv_nnc_cmd_t lstm = ccv_nnc_cmd(CCV_NNC_LSTM_FORWARD, 0, self->params, 0);
4320 ccv_nnc_tensor_param_t output_params[4];
4321 ccv_nnc_hint_tensor_auto(lstm, input_params, 5, ccv_nnc_no_hint, output_params, 4);
4322 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0);
4323 if (!self->weights.graph)
4324 self->weights = ccv_nnc_tensor_symbol_new(graph, input_params[4], "weights");
4325 assert(self->weights.graph == graph)((void) sizeof ((self->weights.graph == graph) ? 1 : 0), __extension__
({ if (self->weights.graph == graph) ; else __assert_fail
("self->weights.graph == graph", "ccv_cnnp_model_addons.c"
, 4325, __extension__ __PRETTY_FUNCTION__); }));
4326 const ccv_nnc_tensor_symbol_t weights = ccv_cnnp_model_get_symbol(super, self->weights);
4327 if (!self->reserves.graph)
4328 self->reserves = ccv_nnc_tensor_symbol_new(graph, output_params[3], "reserves");
4329 assert(self->reserves.graph == graph)((void) sizeof ((self->reserves.graph == graph) ? 1 : 0), __extension__
({ if (self->reserves.graph == graph) ; else __assert_fail
("self->reserves.graph == graph", "ccv_cnnp_model_addons.c"
, 4329, __extension__ __PRETTY_FUNCTION__); }));
4330 const ccv_nnc_tensor_symbol_t reserves = ccv_cnnp_model_get_symbol(super, self->reserves);
4331 const ccv_nnc_tensor_symbol_t mask = input_size == 2 ? inputs[1] : NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
4332 self->lstm = ccv_nnc_graph_exec_symbol_new(graph, lstm, TENSOR_SYMBOL_LIST(inputs[0], mask, NO_TENSOR_SYMBOL, NO_TENSOR_SYMBOL, weights)(const ccv_nnc_tensor_symbol_t []){inputs[0], mask, (const ccv_nnc_tensor_symbol_t
){.d = CCV_NNC_NO_TENSOR_SYMBOL}, (const ccv_nnc_tensor_symbol_t
){.d = CCV_NNC_NO_TENSOR_SYMBOL}, weights}, (1 +1 +1 +1 +1 +1
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(outputs[0], NO_TENSOR_SYMBOL, NO_TENSOR_SYMBOL, reserves)(const ccv_nnc_tensor_symbol_t []){outputs[0], (const ccv_nnc_tensor_symbol_t
){.d = CCV_NNC_NO_TENSOR_SYMBOL}, (const ccv_nnc_tensor_symbol_t
){.d = CCV_NNC_NO_TENSOR_SYMBOL}, reserves}, (1 +1 +1 +1 +1 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "lstm");
4333}
4334
4335static void _ccv_cnnp_lstm_init_states(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_cnnp_state_initializer_f initializer, void* const context)
4336{
4337 ccv_cnnp_model_lstm_t* const self = (ccv_cnnp_model_lstm_t*)super;
4338 if (self->weights.graph)
4339 {
4340 const float stdv = 1.0 / sqrt(self->params.rnn.hidden_size);
4341 initializer(context, CMD_RANDOM_UNIFORM_FORWARD(-stdv, stdv)ccv_nnc_cmd(CCV_NNC_RANDOM_UNIFORM_FORWARD, 0, (ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.blas={.a={-stdv, stdv}}}, 0), ccv_nnc_no_hint, 0, 0, self->weights);
4342 }
4343}
4344
4345static void _ccv_cnnp_lstm_add_to_parameter(ccv_cnnp_model_t* const super, const ccv_cnnp_add_to_array_f add_to_array, void* const parameters, const int is_trainable)
4346{
4347 ccv_cnnp_model_lstm_t* const self = (ccv_cnnp_model_lstm_t*)super;
4348 if (self->weights.graph)
4349 add_to_array(parameters, self->weights, is_trainable);
4350}
4351
4352static void _ccv_cnnp_lstm_set_is_test(ccv_cnnp_model_t* const super, const int is_test, const ccv_cnnp_cmd_updater_f updater, void* const context)
4353{
4354 ccv_cnnp_model_lstm_t* const self = (ccv_cnnp_model_lstm_t*)super;
4355 if (self->lstm.graph)
4356 {
4357 self->params.rnn.is_test = is_test;
4358 updater(context, self->lstm, ccv_nnc_cmd(CCV_NNC_LSTM_FORWARD, 0, self->params, 0), ccv_nnc_no_hint);
4359 }
4360}
4361
4362static ccv_cnnp_model_t* _ccv_cnnp_lstm_copy(const ccv_cnnp_model_t* const self, void* const context);
4363
4364static const ccv_cnnp_model_vtab_t ccv_cnnp_lstm_isa = {
4365 .build = _ccv_cnnp_lstm_build,
4366 .init_states = _ccv_cnnp_lstm_init_states,
4367 .add_to_parameter = _ccv_cnnp_lstm_add_to_parameter,
4368 .copy = _ccv_cnnp_lstm_copy,
4369 .set_is_test = _ccv_cnnp_lstm_set_is_test,
4370};
4371
4372ccv_cnnp_model_t* ccv_cnnp_lstm(const int masked, const int hidden_size, const int proj_size, const int num_layers, const int bias, const int batch_first, const int bidirectional, const float dropout, const int is_trainable, const char* const name)
4373{
4374 ccv_cnnp_model_lstm_t* const model_lstm = (ccv_cnnp_model_lstm_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_lstm_t));
4375 model_lstm->super.isa = &ccv_cnnp_lstm_isa;
4376 model_lstm->super.input_size = masked ? 2 : 1;
4377 model_lstm->super.outputs = &model_lstm->output;
4378 model_lstm->super.output_size = 1;
4379 model_lstm->super.is_trainable = is_trainable;
4380 ccv_cnnp_model_copy_name(&model_lstm->super, name);
4381 model_lstm->masked = masked;
4382 model_lstm->weights.d = CCV_NNC_NO_TENSOR_SYMBOL;
4383 model_lstm->weights.graph = 0;
4384 model_lstm->params.rnn.hidden_size = hidden_size;
4385 model_lstm->params.rnn.proj_size = proj_size;
4386 model_lstm->params.rnn.num_layers = num_layers;
4387 model_lstm->params.rnn.bias = bias;
4388 model_lstm->params.rnn.batch_first = batch_first;
4389 model_lstm->params.rnn.bidirectional = bidirectional;
4390 model_lstm->params.rnn.dropout = dropout;
4391 return (ccv_cnnp_model_t*)model_lstm;
4392}
4393
4394static ccv_cnnp_model_t* _ccv_cnnp_lstm_copy(const ccv_cnnp_model_t* const super, void* const context)
4395{
4396 const ccv_cnnp_model_lstm_t* const self = (const ccv_cnnp_model_lstm_t*)super;
4397 return ccv_cnnp_lstm(self->masked, self->params.rnn.hidden_size, self->params.rnn.proj_size, self->params.rnn.num_layers, self->params.rnn.bias, self->params.rnn.batch_first, self->params.rnn.bidirectional, self->params.rnn.dropout, self->super.is_trainable, self->super.name);
4398}
4399
4400/// MARK - Datatype conversion layer.
4401
4402typedef struct {
4403 ccv_cnnp_model_t super;
4404 ccv_nnc_tensor_symbol_t output;
4405 int datatype;
4406 int ref_to_last;
4407} ccv_cnnp_model_datatype_conversion_t;
4408
4409static void _ccv_cnnp_datatype_conversion_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
4410{
4411 ccv_cnnp_model_datatype_conversion_t* const self = (ccv_cnnp_model_datatype_conversion_t*)super;
4412 PRINT(CCV_CLI_VERBOSE, "[cnnp_datatype_conversion_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_datatype_conversion_build] -\n"); fflush(stdout
); } } while (0);
4413 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4414 if (self->ref_to_last)
4415 {
4416 assert(input_size > 1)((void) sizeof ((input_size > 1) ? 1 : 0), __extension__ (
{ if (input_size > 1) ; else __assert_fail ("input_size > 1"
, "ccv_cnnp_model_addons.c", 4416, __extension__ __PRETTY_FUNCTION__
); }));
4417 const ccv_nnc_tensor_param_t last_params = ccv_nnc_tensor_symbol_params(graph, inputs[input_size - 1]);
4418 params.datatype = last_params.datatype;
4419 } else
4420 params.datatype = self->datatype;
4421 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 4421, __extension__ __PRETTY_FUNCTION__
); }));
4422 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4423 ccv_nnc_graph_exec_symbol_new(graph, CMD_DATATYPE_CONVERSION_FORWARD()ccv_nnc_cmd(CCV_NNC_DATATYPE_CONVERSION_FORWARD, 0, ccv_nnc_cmd_auto
, 0), inputs, output_size /* intentional */, outputs, output_size, 0);
4424}
4425
4426static ccv_cnnp_model_t* _ccv_cnnp_datatype_conversion_copy(const ccv_cnnp_model_t* const self, void* const context);
4427
4428static const ccv_cnnp_model_vtab_t ccv_cnnp_datatype_conversion_isa = {
4429 .build = _ccv_cnnp_datatype_conversion_build,
4430 .copy = _ccv_cnnp_datatype_conversion_copy,
4431};
4432
4433ccv_cnnp_model_t* ccv_cnnp_datatype_conversion(const int datatype, const int ref_to_last, const char* const name)
4434{
4435 ccv_cnnp_model_datatype_conversion_t* const model_datatype_conversion = (ccv_cnnp_model_datatype_conversion_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_datatype_conversion_t));
4436 model_datatype_conversion->super.isa = &ccv_cnnp_datatype_conversion_isa;
4437 model_datatype_conversion->super.input_size = 0;
4438 model_datatype_conversion->super.outputs = &model_datatype_conversion->output;
4439 model_datatype_conversion->super.output_size = 1;
4440 model_datatype_conversion->datatype = datatype;
4441 model_datatype_conversion->ref_to_last = ref_to_last;
4442 ccv_cnnp_model_copy_name(&model_datatype_conversion->super, name);
4443 return (ccv_cnnp_model_t*)model_datatype_conversion;
4444}
4445
4446static ccv_cnnp_model_t* _ccv_cnnp_datatype_conversion_copy(const ccv_cnnp_model_t* const super, void* const context)
4447{
4448 ccv_cnnp_model_datatype_conversion_t* const self = (ccv_cnnp_model_datatype_conversion_t*)super;
4449 return ccv_cnnp_datatype_conversion(self->datatype, self->ref_to_last, self->super.name);
4450}
4451
4452/// MARK - Clamp layer.
4453
4454typedef struct {
4455 ccv_cnnp_model_t super;
4456 ccv_nnc_tensor_symbol_t output;
4457 float min;
4458 float max;
4459} ccv_cnnp_model_clamp_t;
4460
4461static void _ccv_cnnp_clamp_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
4462{
4463 ccv_cnnp_model_clamp_t* const self = (ccv_cnnp_model_clamp_t*)super;
4464 PRINT(CCV_CLI_VERBOSE, "[cnnp_clamp_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_clamp_build] -\n"); fflush(stdout); } } while
(0);
4465 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4466 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 4466, __extension__ __PRETTY_FUNCTION__
); }));
4467 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4468 ccv_nnc_graph_exec_symbol_new(graph, CMD_CLAMP_FORWARD(self->min, self->max)ccv_nnc_cmd(CCV_NNC_CLAMP_FORWARD, 0, (ccv_nnc_cmd_param_t){.
size={.dim={1,1,1}},.clamp={.min=self->min,.max=self->max
}}, 0), inputs, output_size /* intentional */, outputs, output_size, 0);
4469}
4470
4471static ccv_cnnp_model_t* _ccv_cnnp_clamp_copy(const ccv_cnnp_model_t* const self, void* const context);
4472
4473static const ccv_cnnp_model_vtab_t ccv_cnnp_clamp_isa = {
4474 .build = _ccv_cnnp_clamp_build,
4475 .copy = _ccv_cnnp_clamp_copy,
4476};
4477
4478ccv_cnnp_model_t* ccv_cnnp_clamp(const float min, const float max, const char* const name)
4479{
4480 ccv_cnnp_model_clamp_t* const model_clamp = (ccv_cnnp_model_clamp_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_clamp_t));
4481 model_clamp->super.isa = &ccv_cnnp_clamp_isa;
4482 model_clamp->super.input_size = 0;
4483 model_clamp->super.outputs = &model_clamp->output;
4484 model_clamp->super.output_size = 1;
4485 model_clamp->min = min;
4486 model_clamp->max = max;
4487 ccv_cnnp_model_copy_name(&model_clamp->super, name);
4488 return (ccv_cnnp_model_t*)model_clamp;
4489}
4490
4491static ccv_cnnp_model_t* _ccv_cnnp_clamp_copy(const ccv_cnnp_model_t* const super, void* const context)
4492{
4493 ccv_cnnp_model_clamp_t* const self = (ccv_cnnp_model_clamp_t*)super;
4494 return ccv_cnnp_clamp(self->min, self->max, self->super.name);
4495}
4496
4497// MARK - Parameter Layer
4498
4499typedef struct {
4500 ccv_cnnp_model_t super;
4501 float init_bound;
4502 ccv_nnc_tensor_symbol_t weights;
4503 ccv_nnc_tensor_param_t weights_params;
4504 ccv_nnc_tensor_symbol_t output;
4505} ccv_cnnp_model_parameter_t;
4506
4507static void _ccv_cnnp_parameter_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
4508{
4509 PRINT(CCV_CLI_VERBOSE, "[cnnp_parameter_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_parameter_build] -\n"); fflush(stdout); } } while
(0);
4510 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 4510, __extension__ __PRETTY_FUNCTION__
); }));
4511 ccv_cnnp_model_parameter_t* const self = (ccv_cnnp_model_parameter_t*)super;
4512 if (!self->weights.graph)
4513 self->weights = ccv_nnc_tensor_symbol_new(graph, self->weights_params, "weights");
4514 assert(self->weights.graph == graph)((void) sizeof ((self->weights.graph == graph) ? 1 : 0), __extension__
({ if (self->weights.graph == graph) ; else __assert_fail
("self->weights.graph == graph", "ccv_cnnp_model_addons.c"
, 4514, __extension__ __PRETTY_FUNCTION__); }));
4515 outputs[0] = ccv_cnnp_model_get_symbol(super, self->weights);
4516}
4517
4518static void _ccv_cnnp_parameter_init_states(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_cnnp_state_initializer_f initializer, void* const context)
4519{
4520 ccv_cnnp_model_parameter_t* const self = (ccv_cnnp_model_parameter_t*)super;
4521 if (self->init_bound > 0)
4522 initializer(context, CMD_RANDOM_UNIFORM_FORWARD(-self->init_bound, self->init_bound)ccv_nnc_cmd(CCV_NNC_RANDOM_UNIFORM_FORWARD, 0, (ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.blas={.a={-self->init_bound, self->
init_bound}}}, 0), ccv_nnc_no_hint, 0, 0, self->weights);
4523 else
4524 initializer(context, CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, self->weights);
4525}
4526
4527static void _ccv_cnnp_parameter_add_to_parameter(ccv_cnnp_model_t* const super, const ccv_cnnp_add_to_array_f add_to_array, void* const parameters, const int is_trainable)
4528{
4529 ccv_cnnp_model_parameter_t* const self = (ccv_cnnp_model_parameter_t*)super;
4530 add_to_array(parameters, self->weights, is_trainable);
4531}
4532
4533static ccv_cnnp_model_t* _ccv_cnnp_parameter_copy(const ccv_cnnp_model_t* const super, void* const context);
4534
4535static const ccv_cnnp_model_vtab_t ccv_cnnp_parameter_isa = {
4536 .build = _ccv_cnnp_parameter_build,
4537 .init_states = _ccv_cnnp_parameter_init_states,
4538 .add_to_parameter = _ccv_cnnp_parameter_add_to_parameter,
4539 .copy = _ccv_cnnp_parameter_copy,
4540};
4541
4542ccv_cnnp_model_t* ccv_cnnp_parameter(const ccv_nnc_tensor_param_t params, const float init_bound, const int is_trainable, const char* const name)
4543{
4544 ccv_cnnp_model_parameter_t* const model_parameter = (ccv_cnnp_model_parameter_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_parameter_t));
4545 model_parameter->super.isa = &ccv_cnnp_parameter_isa;
4546 model_parameter->super.input_size = 0;
4547 model_parameter->super.outputs = &model_parameter->output;
4548 model_parameter->super.output_size = 1;
4549 model_parameter->super.is_trainable = is_trainable;
4550 ccv_cnnp_model_copy_name(&model_parameter->super, name);
4551 model_parameter->weights.d = CCV_NNC_NO_TENSOR_SYMBOL;
4552 model_parameter->weights.graph = 0;
4553 model_parameter->weights_params = params;
4554 return (ccv_cnnp_model_t*)model_parameter;
4555}
4556
4557static ccv_cnnp_model_t* _ccv_cnnp_parameter_copy(const ccv_cnnp_model_t* const super, void* const context)
4558{
4559 const ccv_cnnp_model_parameter_t* const self = (const ccv_cnnp_model_parameter_t*)super;
4560 return ccv_cnnp_parameter(self->weights_params, self->init_bound, self->super.is_trainable, self->super.name);
4561}
4562
4563// MARK - Scalar Layer
4564
4565typedef struct {
4566 ccv_cnnp_model_t super;
4567 int type;
4568 int format;
4569 int datatype;
4570 float value;
4571 ccv_nnc_tensor_symbol_t output;
4572} ccv_cnnp_model_scalar_t;
4573
4574static void _ccv_cnnp_scalar_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
4575{
4576 PRINT(CCV_CLI_VERBOSE, "[cnnp_scalar_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_scalar_build] -\n"); fflush(stdout); } } while
(0);
4577 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 4577, __extension__ __PRETTY_FUNCTION__
); }));
4578 ccv_cnnp_model_scalar_t* const self = (ccv_cnnp_model_scalar_t*)super;
4579 ccv_nnc_tensor_param_t params = {
4580 .type = self->type,
4581 .format = self->format,
4582 .datatype = self->datatype,
4583 .dim = {
4584 1
4585 }
4586 };
4587 if (input_size > 0)
4588 {
4589 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4590 params.type = input_params.type;
4591 params.format = input_params.format;
4592 params.datatype = input_params.datatype;
4593 }
4594 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4595 ccv_nnc_graph_exec_symbol_new(graph, CMD_SET_FORWARD(self->value)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={self->value,}}}, 0), 0, 0, outputs, 1, 0);
4596}
4597
4598static ccv_cnnp_model_t* _ccv_cnnp_scalar_copy(const ccv_cnnp_model_t* const super, void* const context);
4599
4600static const ccv_cnnp_model_vtab_t ccv_cnnp_scalar_isa = {
4601 .build = _ccv_cnnp_scalar_build,
4602 .copy = _ccv_cnnp_scalar_copy,
4603};
4604
4605ccv_cnnp_model_t* ccv_cnnp_scalar(const int type, const int format, const int datatype, const float value, const char* const name)
4606{
4607 ccv_cnnp_model_scalar_t* const model_scalar = (ccv_cnnp_model_scalar_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_scalar_t));
4608 model_scalar->super.isa = &ccv_cnnp_scalar_isa;
4609 model_scalar->super.input_size = 0;
4610 model_scalar->super.outputs = &model_scalar->output;
4611 model_scalar->super.output_size = 1;
4612 ccv_cnnp_model_copy_name(&model_scalar->super, name);
4613 model_scalar->type = type;
4614 model_scalar->format = format;
4615 model_scalar->datatype = datatype;
4616 model_scalar->value = value;
4617 return (ccv_cnnp_model_t*)model_scalar;
4618}
4619
4620static ccv_cnnp_model_t* _ccv_cnnp_scalar_copy(const ccv_cnnp_model_t* const super, void* const context)
4621{
4622 const ccv_cnnp_model_scalar_t* const self = (const ccv_cnnp_model_scalar_t*)super;
4623 return ccv_cnnp_scalar(self->type, self->format, self->datatype, self->value, self->super.name);
4624}
4625
4626// MARK - Variable Layer
4627
4628typedef struct {
4629 ccv_cnnp_model_t super;
4630 ccv_nnc_tensor_param_t params;
4631 ccv_nnc_tensor_symbol_t output;
4632} ccv_cnnp_model_variable_t;
4633
4634static void _ccv_cnnp_variable_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
4635{
4636 PRINT(CCV_CLI_VERBOSE, "[cnnp_variable_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_variable_build] -\n"); fflush(stdout); } } while
(0);
4637 assert(input_size == 0)((void) sizeof ((input_size == 0) ? 1 : 0), __extension__ ({ if
(input_size == 0) ; else __assert_fail ("input_size == 0", "ccv_cnnp_model_addons.c"
, 4637, __extension__ __PRETTY_FUNCTION__); }));
4638 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 4638, __extension__ __PRETTY_FUNCTION__
); }));
4639 ccv_cnnp_model_variable_t* const self = (ccv_cnnp_model_variable_t*)super;
4640 outputs[0] = ccv_nnc_tensor_symbol_new(graph, self->params, 0);
4641}
4642
4643static ccv_cnnp_model_t* _ccv_cnnp_variable_copy(const ccv_cnnp_model_t* const super, void* const context);
4644
4645static const ccv_cnnp_model_vtab_t ccv_cnnp_variable_isa = {
4646 .build = _ccv_cnnp_variable_build,
4647 .copy = _ccv_cnnp_variable_copy,
4648};
4649
4650ccv_cnnp_model_t* ccv_cnnp_variable(const ccv_nnc_tensor_param_t params, const char* const name)
4651{
4652 ccv_cnnp_model_variable_t* const model_variable = (ccv_cnnp_model_variable_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_variable_t));
4653 model_variable->super.isa = &ccv_cnnp_variable_isa;
4654 model_variable->super.input_size = 0;
4655 model_variable->super.outputs = &model_variable->output;
4656 model_variable->super.output_size = 1;
4657 ccv_cnnp_model_copy_name(&model_variable->super, name);
4658 model_variable->params = params;
4659 return (ccv_cnnp_model_t*)model_variable;
4660}
4661
4662static ccv_cnnp_model_t* _ccv_cnnp_variable_copy(const ccv_cnnp_model_t* const super, void* const context)
4663{
4664 const ccv_cnnp_model_variable_t* const self = (const ccv_cnnp_model_variable_t*)super;
4665 return ccv_cnnp_variable(self->params, self->super.name);
4666}
4667
4668// MARK - Set Layer
4669
4670typedef struct {
4671 ccv_cnnp_model_t super;
4672 float value;
4673 ccv_nnc_tensor_symbol_t output;
4674} ccv_cnnp_model_set_t;
4675
4676static void _ccv_cnnp_set_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
4677{
4678 PRINT(CCV_CLI_VERBOSE, "[cnnp_set_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_set_build] -\n"); fflush(stdout); } } while (
0);
4679 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 4679, __extension__ __PRETTY_FUNCTION__); }));
4680 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 4680, __extension__ __PRETTY_FUNCTION__
); }));
4681 ccv_cnnp_model_set_t* const self = (ccv_cnnp_model_set_t*)super;
4682 outputs[0] = inputs[0];
4683 ccv_nnc_graph_exec_symbol_new(graph, CMD_SET_FORWARD(self->value)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={self->value,}}}, 0), 0, 0, outputs, 1, "set");
4684}
4685
4686static ccv_cnnp_model_t* _ccv_cnnp_set_copy(const ccv_cnnp_model_t* const super, void* const context);
4687
4688static const ccv_cnnp_model_vtab_t ccv_cnnp_set_isa = {
4689 .build = _ccv_cnnp_set_build,
4690 .copy = _ccv_cnnp_set_copy,
4691};
4692
4693ccv_cnnp_model_t* ccv_cnnp_set(const float value, const char* const name)
4694{
4695 ccv_cnnp_model_set_t* const model_set = (ccv_cnnp_model_set_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_set_t));
4696 model_set->super.isa = &ccv_cnnp_set_isa;
4697 model_set->super.input_size = 1;
4698 model_set->super.outputs = &model_set->output;
4699 model_set->super.output_size = 1;
4700 ccv_cnnp_model_copy_name(&model_set->super, name);
4701 model_set->value = value;
4702 return (ccv_cnnp_model_t*)model_set;
4703}
4704
4705static ccv_cnnp_model_t* _ccv_cnnp_set_copy(const ccv_cnnp_model_t* const super, void* const context)
4706{
4707 const ccv_cnnp_model_set_t* const self = (const ccv_cnnp_model_set_t*)super;
4708 return ccv_cnnp_set(self->value, self->super.name);
4709}
4710
4711// MARK - Send Layer
4712
4713typedef struct {
4714 ccv_cnnp_model_t super;
4715 int device_id;
4716 ccv_nnc_tensor_symbol_t output;
4717} ccv_cnnp_model_send_t;
4718
4719static void _ccv_cnnp_send_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
4720{
4721 ccv_cnnp_model_send_t* const self = (ccv_cnnp_model_send_t*)super;
4722 PRINT(CCV_CLI_VERBOSE, "[cnnp_send_build] - device_id: %d\n", self->device_id)do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_send_build] - device_id: %d\n", self->device_id
); fflush(stdout); } } while (0);
4723 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 4723, __extension__ __PRETTY_FUNCTION__); }));
4724 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 4724, __extension__ __PRETTY_FUNCTION__
); }));
4725 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4726 params.type = (params.type & ~0x3) | CCV_TENSOR_GPU_MEMORY;
4727 CCV_TENSOR_SET_DEVICE_ID(params.type, self->device_id)(params.type) = (((params.type) & ~0xfff00) | (((self->
device_id) & 0xfff) << 8));
4728 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4729 ccv_nnc_graph_exec_symbol_new(graph, CMD_DATA_TRANSFER_FORWARD()ccv_nnc_cmd(CCV_NNC_DATA_TRANSFER_FORWARD, 0, ccv_nnc_cmd_auto
, 0), inputs, 1, outputs, 1, "send");
4730}
4731
4732static ccv_cnnp_model_t* _ccv_cnnp_send_copy(const ccv_cnnp_model_t* const super, void* const context);
4733
4734static const ccv_cnnp_model_vtab_t ccv_cnnp_send_isa = {
4735 .build = _ccv_cnnp_send_build,
4736 .copy = _ccv_cnnp_send_copy,
4737};
4738
4739ccv_cnnp_model_t* ccv_cnnp_send(const int device_id, const char* const name)
4740{
4741 assert(device_id >= 0)((void) sizeof ((device_id >= 0) ? 1 : 0), __extension__ (
{ if (device_id >= 0) ; else __assert_fail ("device_id >= 0"
, "ccv_cnnp_model_addons.c", 4741, __extension__ __PRETTY_FUNCTION__
); }));
4742 ccv_cnnp_model_send_t* const model_send = (ccv_cnnp_model_send_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_send_t));
4743 model_send->super.isa = &ccv_cnnp_send_isa;
4744 model_send->super.input_size = 1;
4745 model_send->super.outputs = &model_send->output;
4746 model_send->super.output_size = 1;
4747 ccv_cnnp_model_copy_name(&model_send->super, name);
4748 model_send->device_id = device_id;
4749 return (ccv_cnnp_model_t*)model_send;
4750}
4751
4752static ccv_cnnp_model_t* _ccv_cnnp_send_copy(const ccv_cnnp_model_t* const super, void* const context)
4753{
4754 const ccv_cnnp_model_send_t* const self = (const ccv_cnnp_model_send_t*)super;
4755 return ccv_cnnp_send(self->device_id, self->super.name);
4756}
4757
4758// MARK - Move Layer
4759
4760typedef struct {
4761 ccv_cnnp_model_t super;
4762 ccv_nnc_tensor_symbol_t output;
4763} ccv_cnnp_model_move_t;
4764
4765static void _ccv_cnnp_move_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
4766{
4767 PRINT(CCV_CLI_VERBOSE, "[cnnp_move_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_move_build] -\n"); fflush(stdout); } } while
(0);
4768 assert(input_size == 2)((void) sizeof ((input_size == 2) ? 1 : 0), __extension__ ({ if
(input_size == 2) ; else __assert_fail ("input_size == 2", "ccv_cnnp_model_addons.c"
, 4768, __extension__ __PRETTY_FUNCTION__); }));
4769 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 4769, __extension__ __PRETTY_FUNCTION__
); }));
4770 outputs[0] = inputs[1];
4771 ccv_nnc_graph_exec_symbol_new(graph, CMD_FORMAT_TRANSFORM_FORWARD()ccv_nnc_cmd(CCV_NNC_FORMAT_TRANSFORM_FORWARD, 0, ccv_nnc_cmd_auto
, 0), inputs, 1, outputs, 1, "move");
4772}
4773
4774static ccv_cnnp_model_t* _ccv_cnnp_move_copy(const ccv_cnnp_model_t* const super, void* const context);
4775
4776static const ccv_cnnp_model_vtab_t ccv_cnnp_move_isa = {
4777 .build = _ccv_cnnp_move_build,
4778 .copy = _ccv_cnnp_move_copy,
4779};
4780
4781ccv_cnnp_model_t* ccv_cnnp_move(const char* const name)
4782{
4783 ccv_cnnp_model_move_t* const model_move = (ccv_cnnp_model_move_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_move_t));
4784 model_move->super.isa = &ccv_cnnp_move_isa;
4785 model_move->super.input_size = 2;
4786 model_move->super.outputs = &model_move->output;
4787 model_move->super.output_size = 1;
4788 ccv_cnnp_model_copy_name(&model_move->super, name);
4789 return (ccv_cnnp_model_t*)model_move;
4790}
4791
4792static ccv_cnnp_model_t* _ccv_cnnp_move_copy(const ccv_cnnp_model_t* const super, void* const context)
4793{
4794 const ccv_cnnp_model_move_t* const self = (const ccv_cnnp_model_move_t*)super;
4795 return ccv_cnnp_move(self->super.name);
4796}
4797
4798// MARK - "Making" Contiguous Layer
4799
4800typedef struct {
4801 ccv_cnnp_model_t super;
4802 ccv_nnc_tensor_symbol_t output;
4803} ccv_cnnp_model_contiguous_t;
4804
4805static void _ccv_cnnp_contiguous_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
4806{
4807 PRINT(CCV_CLI_VERBOSE, "[cnnp_contiguous_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_contiguous_build] -\n"); fflush(stdout); } }
while (0);
4808 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 4808, __extension__ __PRETTY_FUNCTION__); }));
4809 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 4809, __extension__ __PRETTY_FUNCTION__
); }));
4810 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4811 ccv_nnc_tensor_symbol_t to = ccv_nnc_tensor_symbol_alias_to(graph, inputs[0]);
4812 if (to.d == CCV_NNC_NO_TENSOR_SYMBOL) // If we are not reshape an alias, it is straightforward.
4813 {
4814 outputs[0] = inputs[0];
4815 return;
4816 }
4817 // Otherwise, we need to check its stride to know if it is contiguous.
4818 int old_stride[CCV_NNC_MAX_DIM_ALLOC(12)];
4819 ccv_nnc_tensor_symbol_alias_params(graph, inputs[0], 0, old_stride);
4820 // We identify permute by checking if the stride is not in descending order.
4821 // This also covered "permute" through reshape, rather than using ccv_cnnp_permute directly.
4822 if (ccv_nnc_is_tensor_stride_packed(old_stride, params.dim))
4823 {
4824 outputs[0] = inputs[0];
4825 return;
4826 }
4827 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4828 ccv_nnc_graph_exec_symbol_t make_contiguous = ccv_nnc_graph_exec_symbol_new(graph, CMD_FORMAT_TRANSFORM_FORWARD()ccv_nnc_cmd(CCV_NNC_FORMAT_TRANSFORM_FORWARD, 0, ccv_nnc_cmd_auto
, 0), inputs, 1, outputs, 1, "contiguous");
4829 ccv_nnc_graph_exec_symbol_set_flags(graph, make_contiguous, CCV_NNC_GRAPH_EXEC_DISABLE_OPT);
4830}
4831
4832static ccv_cnnp_model_t* _ccv_cnnp_contiguous_copy(const ccv_cnnp_model_t* const super, void* const context);
4833
4834static const ccv_cnnp_model_vtab_t ccv_cnnp_contiguous_isa = {
4835 .build = _ccv_cnnp_contiguous_build,
4836 .copy = _ccv_cnnp_contiguous_copy,
4837};
4838
4839ccv_cnnp_model_t* ccv_cnnp_contiguous(const char* const name)
4840{
4841 ccv_cnnp_model_contiguous_t* const model_contiguous = (ccv_cnnp_model_contiguous_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_contiguous_t));
4842 model_contiguous->super.isa = &ccv_cnnp_contiguous_isa;
4843 model_contiguous->super.input_size = 1;
4844 model_contiguous->super.outputs = &model_contiguous->output;
4845 model_contiguous->super.output_size = 1;
4846 ccv_cnnp_model_copy_name(&model_contiguous->super, name);
4847 return (ccv_cnnp_model_t*)model_contiguous;
4848}
4849
4850static ccv_cnnp_model_t* _ccv_cnnp_contiguous_copy(const ccv_cnnp_model_t* const super, void* const context)
4851{
4852 const ccv_cnnp_model_contiguous_t* const self = (const ccv_cnnp_model_contiguous_t*)super;
4853 return ccv_cnnp_contiguous(self->super.name);
4854}
4855
4856// MARK - "Making" Copy Layer
4857
4858typedef struct {
4859 ccv_cnnp_model_t super;
4860 ccv_nnc_tensor_symbol_t output;
4861} ccv_cnnp_model_copy_t;
4862
4863static void _ccv_cnnp_copy_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
4864{
4865 PRINT(CCV_CLI_VERBOSE, "[cnnp_copy_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_copy_build] -\n"); fflush(stdout); } } while
(0);
4866 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 4866, __extension__ __PRETTY_FUNCTION__); }));
4867 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 4867, __extension__ __PRETTY_FUNCTION__
); }));
4868 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4869 ccv_nnc_tensor_symbol_t to = ccv_nnc_tensor_symbol_alias_to(graph, inputs[0]);
4870 if (to.d == CCV_NNC_NO_TENSOR_SYMBOL) // If we are not reshape an alias, it is straightforward.
4871 {
4872 outputs[0] = inputs[0];
4873 return;
4874 }
4875 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4876 ccv_nnc_graph_exec_symbol_t make_contiguous = ccv_nnc_graph_exec_symbol_new(graph, CMD_FORMAT_TRANSFORM_FORWARD()ccv_nnc_cmd(CCV_NNC_FORMAT_TRANSFORM_FORWARD, 0, ccv_nnc_cmd_auto
, 0), inputs, 1, outputs, 1, "contiguous");
4877 ccv_nnc_graph_exec_symbol_set_flags(graph, make_contiguous, CCV_NNC_GRAPH_EXEC_DISABLE_OPT);
4878}
4879
4880static ccv_cnnp_model_t* _ccv_cnnp_copy_copy(const ccv_cnnp_model_t* const super, void* const context);
4881
4882static const ccv_cnnp_model_vtab_t ccv_cnnp_copy_isa = {
4883 .build = _ccv_cnnp_copy_build,
4884 .copy = _ccv_cnnp_copy_copy,
4885};
4886
4887ccv_cnnp_model_t* ccv_cnnp_copy(const char* const name)
4888{
4889 ccv_cnnp_model_copy_t* const model_copy = (ccv_cnnp_model_copy_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_copy_t));
4890 model_copy->super.isa = &ccv_cnnp_copy_isa;
4891 model_copy->super.input_size = 1;
4892 model_copy->super.outputs = &model_copy->output;
4893 model_copy->super.output_size = 1;
4894 ccv_cnnp_model_copy_name(&model_copy->super, name);
4895 return (ccv_cnnp_model_t*)model_copy;
4896}
4897
4898static ccv_cnnp_model_t* _ccv_cnnp_copy_copy(const ccv_cnnp_model_t* const super, void* const context)
4899{
4900 const ccv_cnnp_model_copy_t* const self = (const ccv_cnnp_model_copy_t*)super;
4901 return ccv_cnnp_copy(self->super.name);
4902}
4903
4904// MARK - All-To-All Layer
4905
4906typedef struct {
4907 ccv_cnnp_model_t super;
4908 int axis;
4909 ccv_nnc_tensor_symbol_t outputs[1];
4910} ccv_cnnp_model_all_to_all_t;
4911
4912static void _ccv_cnnp_all_to_all_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
4913{
4914 ccv_cnnp_model_all_to_all_t* const self = (ccv_cnnp_model_all_to_all_t*)super;
4915 PRINT(CCV_CLI_VERBOSE, "[cnnp_all_to_all_build] - axis: %d\n", self->axis)do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_all_to_all_build] - axis: %d\n", self->axis
); fflush(stdout); } } while (0);
4916 assert(input_size == self->super.input_size)((void) sizeof ((input_size == self->super.input_size) ? 1
: 0), __extension__ ({ if (input_size == self->super.input_size
) ; else __assert_fail ("input_size == self->super.input_size"
, "ccv_cnnp_model_addons.c", 4916, __extension__ __PRETTY_FUNCTION__
); }));
4917 assert(output_size == self->super.output_size)((void) sizeof ((output_size == self->super.output_size) ?
1 : 0), __extension__ ({ if (output_size == self->super.output_size
) ; else __assert_fail ("output_size == self->super.output_size"
, "ccv_cnnp_model_addons.c", 4917, __extension__ __PRETTY_FUNCTION__
); }));
4918 assert(input_size == output_size)((void) sizeof ((input_size == output_size) ? 1 : 0), __extension__
({ if (input_size == output_size) ; else __assert_fail ("input_size == output_size"
, "ccv_cnnp_model_addons.c", 4918, __extension__ __PRETTY_FUNCTION__
); }));
4919 assert(input_size > 0)((void) sizeof ((input_size > 0) ? 1 : 0), __extension__ (
{ if (input_size > 0) ; else __assert_fail ("input_size > 0"
, "ccv_cnnp_model_addons.c", 4919, __extension__ __PRETTY_FUNCTION__
); }));
4920 const ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4921 const int nd = ccv_nnc_tensor_nd(input_params.dim);
4922 assert(self->axis >= 0 && self->axis < nd)((void) sizeof ((self->axis >= 0 && self->axis
< nd) ? 1 : 0), __extension__ ({ if (self->axis >= 0
&& self->axis < nd) ; else __assert_fail ("self->axis >= 0 && self->axis < nd"
, "ccv_cnnp_model_addons.c", 4922, __extension__ __PRETTY_FUNCTION__
); }));
4923 assert(input_params.dim[self->axis] % input_size == 0)((void) sizeof ((input_params.dim[self->axis] % input_size
== 0) ? 1 : 0), __extension__ ({ if (input_params.dim[self->
axis] % input_size == 0) ; else __assert_fail ("input_params.dim[self->axis] % input_size == 0"
, "ccv_cnnp_model_addons.c", 4923, __extension__ __PRETTY_FUNCTION__
); }));
4924 int i;
4925 for (i = 0; i < input_size; i++)
4926 {
4927 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
4928 assert(params.format == input_params.format)((void) sizeof ((params.format == input_params.format) ? 1 : 0
), __extension__ ({ if (params.format == input_params.format)
; else __assert_fail ("params.format == input_params.format"
, "ccv_cnnp_model_addons.c", 4928, __extension__ __PRETTY_FUNCTION__
); }));
4929 assert(params.datatype == input_params.datatype)((void) sizeof ((params.datatype == input_params.datatype) ? 1
: 0), __extension__ ({ if (params.datatype == input_params.datatype
) ; else __assert_fail ("params.datatype == input_params.datatype"
, "ccv_cnnp_model_addons.c", 4929, __extension__ __PRETTY_FUNCTION__
); }));
4930 assert(CCV_TENSOR_GET_MEMORY(params.type) == CCV_TENSOR_GET_MEMORY(input_params.type))((void) sizeof ((((params.type) & 0x3) == ((input_params.
type) & 0x3)) ? 1 : 0), __extension__ ({ if (((params.type
) & 0x3) == ((input_params.type) & 0x3)) ; else __assert_fail
("CCV_TENSOR_GET_MEMORY(params.type) == CCV_TENSOR_GET_MEMORY(input_params.type)"
, "ccv_cnnp_model_addons.c", 4930, __extension__ __PRETTY_FUNCTION__
); }));
4931 assert(memcmp(params.dim, input_params.dim, sizeof(input_params.dim)) == 0)((void) sizeof ((memcmp(params.dim, input_params.dim, sizeof(
input_params.dim)) == 0) ? 1 : 0), __extension__ ({ if (memcmp
(params.dim, input_params.dim, sizeof(input_params.dim)) == 0
) ; else __assert_fail ("memcmp(params.dim, input_params.dim, sizeof(input_params.dim)) == 0"
, "ccv_cnnp_model_addons.c", 4931, __extension__ __PRETTY_FUNCTION__
); }));
4932 outputs[i] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4933 }
4934 ccv_nnc_graph_exec_symbol_new(graph, CMD_COMM_ALL_TO_ALL_FORWARD(self->axis)ccv_nnc_cmd(CCV_NNC_COMM_ALL_TO_ALL_FORWARD, 0, ((ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.all_to_all={.axis=self->axis}}), 0
), inputs, input_size, outputs, output_size, "all_to_all");
4935}
4936
4937static ccv_cnnp_model_t* _ccv_cnnp_all_to_all_copy(const ccv_cnnp_model_t* const super, void* const context);
4938
4939static const ccv_cnnp_model_vtab_t ccv_cnnp_all_to_all_isa = {
4940 .build = _ccv_cnnp_all_to_all_build,
4941 .copy = _ccv_cnnp_all_to_all_copy,
4942};
4943
4944ccv_cnnp_model_t* ccv_cnnp_all_to_all(const int count, const int axis, const char* const name)
4945{
4946 assert(count > 0)((void) sizeof ((count > 0) ? 1 : 0), __extension__ ({ if (
count > 0) ; else __assert_fail ("count > 0", "ccv_cnnp_model_addons.c"
, 4946, __extension__ __PRETTY_FUNCTION__); }));
4947 assert(axis >= 0)((void) sizeof ((axis >= 0) ? 1 : 0), __extension__ ({ if (
axis >= 0) ; else __assert_fail ("axis >= 0", "ccv_cnnp_model_addons.c"
, 4947, __extension__ __PRETTY_FUNCTION__); }));
4948 ccv_cnnp_model_all_to_all_t* const model_all_to_all = (ccv_cnnp_model_all_to_all_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_all_to_all_t) + sizeof(ccv_nnc_tensor_symbol_t) * (count - 1));
4949 model_all_to_all->super.isa = &ccv_cnnp_all_to_all_isa;
4950 model_all_to_all->super.input_size = count;
4951 model_all_to_all->super.outputs = model_all_to_all->outputs;
4952 model_all_to_all->super.output_size = count;
4953 model_all_to_all->axis = axis;
4954 ccv_cnnp_model_copy_name(&model_all_to_all->super, name);
4955 return (ccv_cnnp_model_t*)model_all_to_all;
4956}
4957
4958static ccv_cnnp_model_t* _ccv_cnnp_all_to_all_copy(const ccv_cnnp_model_t* const super, void* const context)
4959{
4960 const ccv_cnnp_model_all_to_all_t* const self = (const ccv_cnnp_model_all_to_all_t*)super;
4961 return ccv_cnnp_all_to_all(self->super.output_size, self->axis, self->super.name);
4962}
4963
4964// MARK - Scaled-Dot Product Attention Layer
4965
4966typedef struct {
4967 ccv_cnnp_model_t super;
4968 ccv_nnc_tensor_symbol_t output;
4969 ccv_nnc_tensor_symbol_t weights;
4970 ccv_nnc_tensor_symbol_t bias;
4971 float scale;
4972 int is_causal;
4973 int has_attn_mask;
4974 int is_varlen;
4975 int max_seqlen_q;
4976 int max_seqlen_kv;
4977 int flags;
4978 int use_hadamard;
4979 int attention_sinks;
4980 int sliding_window;
4981 int fused_unify_head_weights;
4982 int no_bias;
4983} ccv_cnnp_model_scaled_dot_product_attention_t;
4984
4985static void _ccv_cnnp_scaled_dot_product_attention_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
4986{
4987 PRINT(CCV_CLI_VERBOSE, "[cnnp_scaled_dot_product_attention_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_scaled_dot_product_attention_build] -\n"); fflush
(stdout); } } while (0);
4988 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 4988, __extension__ __PRETTY_FUNCTION__
); }));
4989 ccv_cnnp_model_scaled_dot_product_attention_t* const self = (ccv_cnnp_model_scaled_dot_product_attention_t*)super;
4990 assert(input_size == (self->is_varlen ? 5 : (self->has_attn_mask ? 4 : 3)) + self->attention_sinks)((void) sizeof ((input_size == (self->is_varlen ? 5 : (self
->has_attn_mask ? 4 : 3)) + self->attention_sinks) ? 1 :
0), __extension__ ({ if (input_size == (self->is_varlen ?
5 : (self->has_attn_mask ? 4 : 3)) + self->attention_sinks
) ; else __assert_fail ("input_size == (self->is_varlen ? 5 : (self->has_attn_mask ? 4 : 3)) + self->attention_sinks"
, "ccv_cnnp_model_addons.c", 4990, __extension__ __PRETTY_FUNCTION__
); }));
4991 assert(!self->is_varlen || !self->has_attn_mask)((void) sizeof ((!self->is_varlen || !self->has_attn_mask
) ? 1 : 0), __extension__ ({ if (!self->is_varlen || !self
->has_attn_mask) ; else __assert_fail ("!self->is_varlen || !self->has_attn_mask"
, "ccv_cnnp_model_addons.c", 4991, __extension__ __PRETTY_FUNCTION__
); }));
4992 assert(!self->is_varlen || !self->fused_unify_head_weights)((void) sizeof ((!self->is_varlen || !self->fused_unify_head_weights
) ? 1 : 0), __extension__ ({ if (!self->is_varlen || !self
->fused_unify_head_weights) ; else __assert_fail ("!self->is_varlen || !self->fused_unify_head_weights"
, "ccv_cnnp_model_addons.c", 4992, __extension__ __PRETTY_FUNCTION__
); }));
4993 const ccv_nnc_tensor_param_t q_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4994 const ccv_nnc_tensor_param_t k_params = ccv_nnc_tensor_symbol_params(graph, inputs[1]);
4995 const ccv_nnc_tensor_param_t v_params = ccv_nnc_tensor_symbol_params(graph, inputs[2]);
4996 const int sink_idx = self->is_varlen ? 5 : (3 + self->has_attn_mask);
4997 const ccv_nnc_tensor_param_t sinks_params = self->attention_sinks ? ccv_nnc_tensor_symbol_params(graph, inputs[sink_idx]) : (ccv_nnc_tensor_param_t){};
4998 const ccv_nnc_tensor_param_t attn_mask_params = self->has_attn_mask ? ccv_nnc_tensor_symbol_params(graph, inputs[3]) : (ccv_nnc_tensor_param_t){};
4999 const ccv_nnc_tensor_param_t q_seq_offsets_params = self->is_varlen ? ccv_nnc_tensor_symbol_params(graph, inputs[3]) : (ccv_nnc_tensor_param_t){};
5000 const ccv_nnc_tensor_param_t kv_seq_offsets_params = self->is_varlen ? ccv_nnc_tensor_symbol_params(graph, inputs[4]) : (ccv_nnc_tensor_param_t){};
5001 const int v_nd = ccv_nnc_tensor_nd(v_params.dim);
5002 // 0-length values have no rank, the command infers 0-length outputs from them.
5003 assert(v_nd == 3 || v_nd == 4 || v_params.dim[0] == 0)((void) sizeof ((v_nd == 3 || v_nd == 4 || v_params.dim[0] ==
0) ? 1 : 0), __extension__ ({ if (v_nd == 3 || v_nd == 4 || v_params
.dim[0] == 0) ; else __assert_fail ("v_nd == 3 || v_nd == 4 || v_params.dim[0] == 0"
, "ccv_cnnp_model_addons.c", 5003, __extension__ __PRETTY_FUNCTION__
); }));
5004 const int hEv = v_nd == 0 ? 0 : (v_nd == 3 ? 1 : v_params.dim[2]) * v_params.dim[v_nd - 1];
5005 ccv_nnc_tensor_param_t weights_params = q_params;
5006 memset(weights_params.dim, 0, sizeof(weights_params.dim));
5007 weights_params.dim[0] = hEv;
5008 weights_params.dim[1] = hEv;
5009 ccv_nnc_tensor_param_t bias_params = q_params;
5010 memset(bias_params.dim, 0, sizeof(bias_params.dim));
5011 bias_params.dim[0] = hEv;
5012 ccv_nnc_cmd_t cmd = {0};
5013 cmd.cmd = CCV_NNC_SCALED_DOT_PRODUCT_ATTENTION_FORWARD;
5014 cmd.info.scaled_dot_product_attention.scale = self->scale;
5015 cmd.info.scaled_dot_product_attention.is_causal = self->is_causal;
5016 cmd.info.scaled_dot_product_attention.is_varlen = self->is_varlen;
5017 cmd.info.scaled_dot_product_attention.max_seqlen_q = self->max_seqlen_q;
5018 cmd.info.scaled_dot_product_attention.max_seqlen_kv = self->max_seqlen_kv;
5019 cmd.info.scaled_dot_product_attention.flags = self->flags;
5020 cmd.info.scaled_dot_product_attention.use_hadamard = self->use_hadamard;
5021 cmd.info.scaled_dot_product_attention.attention_sinks = self->attention_sinks;
5022 cmd.info.scaled_dot_product_attention.sliding_window = self->sliding_window;
5023 ccv_nnc_tensor_param_t output_params[3];
5024 ccv_nnc_tensor_symbol_t output;
5025 ccv_nnc_tensor_symbol_t saved_softmax_lse;
5026 ccv_nnc_tensor_symbol_t saved_v_proj = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
5027 ccv_nnc_tensor_symbol_t attn_mask = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
5028 ccv_nnc_tensor_symbol_t weights = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
5029 ccv_nnc_tensor_symbol_t bias = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
5030 ccv_nnc_tensor_symbol_t sinks = self->attention_sinks ? inputs[sink_idx] : NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
5031 if (self->has_attn_mask)
5032 attn_mask = inputs[3];
5033 if (self->is_varlen)
5034 {
5035 ccv_nnc_tensor_param_t input_params[9] = {
5036 q_params,
5037 k_params,
5038 v_params,
5039 (ccv_nnc_tensor_param_t){},
5040 (ccv_nnc_tensor_param_t){},
5041 (ccv_nnc_tensor_param_t){},
5042 q_seq_offsets_params,
5043 kv_seq_offsets_params,
5044 sinks_params,
5045 };
5046 ccv_nnc_hint_tensor_auto(cmd, input_params, self->attention_sinks ? 9 : 8, ccv_nnc_no_hint, output_params, 2);
5047 output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0);
5048 saved_softmax_lse = ccv_nnc_tensor_symbol_new(graph, output_params[1], 0);
5049 if (self->attention_sinks)
5050 ccv_nnc_graph_exec_symbol_new(graph, cmd, TENSOR_SYMBOL_LIST(inputs[0], inputs[1], inputs[2], NO_TENSOR_SYMBOL, NO_TENSOR_SYMBOL, NO_TENSOR_SYMBOL, inputs[3], inputs[4], sinks)(const ccv_nnc_tensor_symbol_t []){inputs[0], inputs[1], inputs
[2], (const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
}, (const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
}, (const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
}, inputs[3], inputs[4], sinks}, (1 +1 +1 +1 +1 +1 +1 +1 +1 +
1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(output, saved_softmax_lse, saved_v_proj)(const ccv_nnc_tensor_symbol_t []){output, saved_softmax_lse,
saved_v_proj}, (1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 -1), "scaled_dot_product_attention");
5051 else
5052 ccv_nnc_graph_exec_symbol_new(graph, cmd, TENSOR_SYMBOL_LIST(inputs[0], inputs[1], inputs[2], NO_TENSOR_SYMBOL, NO_TENSOR_SYMBOL, NO_TENSOR_SYMBOL, inputs[3], inputs[4])(const ccv_nnc_tensor_symbol_t []){inputs[0], inputs[1], inputs
[2], (const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
}, (const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
}, (const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
}, inputs[3], inputs[4]}, (1 +1 +1 +1 +1 +1 +1 +1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(output, saved_softmax_lse, saved_v_proj)(const ccv_nnc_tensor_symbol_t []){output, saved_softmax_lse,
saved_v_proj}, (1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 -1), "scaled_dot_product_attention");
5053 } else if (self->fused_unify_head_weights)
5054 {
5055 if (!self->weights.graph)
5056 self->weights = ccv_nnc_tensor_symbol_new(graph, weights_params, "weights");
5057 assert(self->weights.graph == graph)((void) sizeof ((self->weights.graph == graph) ? 1 : 0), __extension__
({ if (self->weights.graph == graph) ; else __assert_fail
("self->weights.graph == graph", "ccv_cnnp_model_addons.c"
, 5057, __extension__ __PRETTY_FUNCTION__); }));
5058 weights = ccv_cnnp_model_get_symbol(super, self->weights);
5059 if (!self->no_bias)
5060 {
5061 if (!self->bias.graph)
5062 self->bias = ccv_nnc_tensor_symbol_new(graph, bias_params, "bias");
5063 assert(self->bias.graph == graph)((void) sizeof ((self->bias.graph == graph) ? 1 : 0), __extension__
({ if (self->bias.graph == graph) ; else __assert_fail ("self->bias.graph == graph"
, "ccv_cnnp_model_addons.c", 5063, __extension__ __PRETTY_FUNCTION__
); }));
5064 bias = ccv_cnnp_model_get_symbol(super, self->bias);
5065 }
5066 ccv_nnc_tensor_param_t input_params[9] = {
5067 q_params,
5068 k_params,
5069 v_params,
5070 attn_mask_params,
5071 weights_params,
5072 bias_params,
5073 (ccv_nnc_tensor_param_t){},
5074 (ccv_nnc_tensor_param_t){},
5075 sinks_params,
5076 };
5077 ccv_nnc_hint_tensor_auto(cmd, input_params, self->attention_sinks ? 9 : 6, ccv_nnc_no_hint, output_params, 3);
5078 output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0);
5079 saved_softmax_lse = ccv_nnc_tensor_symbol_new(graph, output_params[1], 0);
5080 saved_v_proj = ccv_nnc_tensor_symbol_new(graph, output_params[2], 0);
5081 if (self->attention_sinks)
5082 ccv_nnc_graph_exec_symbol_new(graph, cmd, TENSOR_SYMBOL_LIST(inputs[0], inputs[1], inputs[2], attn_mask, weights, bias, NO_TENSOR_SYMBOL, NO_TENSOR_SYMBOL, sinks)(const ccv_nnc_tensor_symbol_t []){inputs[0], inputs[1], inputs
[2], attn_mask, weights, bias, (const ccv_nnc_tensor_symbol_t
){.d = CCV_NNC_NO_TENSOR_SYMBOL}, (const ccv_nnc_tensor_symbol_t
){.d = CCV_NNC_NO_TENSOR_SYMBOL}, sinks}, (1 +1 +1 +1 +1 +1 +
1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(output, saved_softmax_lse, saved_v_proj)(const ccv_nnc_tensor_symbol_t []){output, saved_softmax_lse,
saved_v_proj}, (1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 -1), "scaled_dot_product_attention");
5083 else
5084 ccv_nnc_graph_exec_symbol_new(graph, cmd, TENSOR_SYMBOL_LIST(inputs[0], inputs[1], inputs[2], attn_mask, weights, bias)(const ccv_nnc_tensor_symbol_t []){inputs[0], inputs[1], inputs
[2], attn_mask, weights, bias}, (1 +1 +1 +1 +1 +1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(output, saved_softmax_lse, saved_v_proj)(const ccv_nnc_tensor_symbol_t []){output, saved_softmax_lse,
saved_v_proj}, (1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 -1), "scaled_dot_product_attention");
5085 } else {
5086 ccv_nnc_tensor_param_t input_params[9] = {
5087 q_params,
5088 k_params,
5089 v_params,
5090 attn_mask_params,
5091 (ccv_nnc_tensor_param_t){},
5092 (ccv_nnc_tensor_param_t){},
5093 (ccv_nnc_tensor_param_t){},
5094 (ccv_nnc_tensor_param_t){},
5095 sinks_params,
5096 };
5097 ccv_nnc_hint_tensor_auto(cmd, input_params, self->attention_sinks ? 9 : 3, ccv_nnc_no_hint, output_params, 2);
5098 output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0);
5099 saved_softmax_lse = ccv_nnc_tensor_symbol_new(graph, output_params[1], 0);
5100 if (self->attention_sinks)
5101 ccv_nnc_graph_exec_symbol_new(graph, cmd, TENSOR_SYMBOL_LIST(inputs[0], inputs[1], inputs[2], attn_mask, weights, bias, NO_TENSOR_SYMBOL, NO_TENSOR_SYMBOL, sinks)(const ccv_nnc_tensor_symbol_t []){inputs[0], inputs[1], inputs
[2], attn_mask, weights, bias, (const ccv_nnc_tensor_symbol_t
){.d = CCV_NNC_NO_TENSOR_SYMBOL}, (const ccv_nnc_tensor_symbol_t
){.d = CCV_NNC_NO_TENSOR_SYMBOL}, sinks}, (1 +1 +1 +1 +1 +1 +
1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(output, saved_softmax_lse, saved_v_proj)(const ccv_nnc_tensor_symbol_t []){output, saved_softmax_lse,
saved_v_proj}, (1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 -1), "scaled_dot_product_attention");
5102 else
5103 ccv_nnc_graph_exec_symbol_new(graph, cmd, TENSOR_SYMBOL_LIST(inputs[0], inputs[1], inputs[2], attn_mask, weights, bias)(const ccv_nnc_tensor_symbol_t []){inputs[0], inputs[1], inputs
[2], attn_mask, weights, bias}, (1 +1 +1 +1 +1 +1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(output, saved_softmax_lse, saved_v_proj)(const ccv_nnc_tensor_symbol_t []){output, saved_softmax_lse,
saved_v_proj}, (1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 -1), "scaled_dot_product_attention");
5104 }
5105 outputs[0] = output;
5106}
5107
5108static void _ccv_cnnp_scaled_dot_product_attention_init_states(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_cnnp_state_initializer_f initializer, void* const context)
5109{
5110 ccv_cnnp_model_scaled_dot_product_attention_t* const self = (ccv_cnnp_model_scaled_dot_product_attention_t*)super;
5111 if (self->weights.graph)
5112 {
5113 assert(self->fused_unify_head_weights)((void) sizeof ((self->fused_unify_head_weights) ? 1 : 0),
__extension__ ({ if (self->fused_unify_head_weights) ; else
__assert_fail ("self->fused_unify_head_weights", "ccv_cnnp_model_addons.c"
, 5113, __extension__ __PRETTY_FUNCTION__); }));
5114 const ccv_nnc_tensor_param_t weight_params = ccv_nnc_tensor_symbol_params(graph, self->weights);
5115 const int c = weight_params.dim[1];
5116 const float std = sqrtf(2) / sqrtf(c);
5117 const float bound = sqrtf(3) * std;
5118 initializer(context, CMD_RANDOM_UNIFORM_FORWARD(-bound, bound)ccv_nnc_cmd(CCV_NNC_RANDOM_UNIFORM_FORWARD, 0, (ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.blas={.a={-bound, bound}}}, 0), ccv_nnc_no_hint, 0, 0, self->weights);
5119 if (self->bias.graph)
5120 initializer(context, CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, self->bias);
5121 }
5122}
5123
5124static void _ccv_cnnp_scaled_dot_product_attention_add_to_parameter(ccv_cnnp_model_t* const super, const ccv_cnnp_add_to_array_f add_to_array, void* const parameters, const int is_trainable)
5125{
5126 ccv_cnnp_model_scaled_dot_product_attention_t* const self = (ccv_cnnp_model_scaled_dot_product_attention_t*)super;
5127 if (self->weights.graph)
5128 {
5129 assert(self->fused_unify_head_weights)((void) sizeof ((self->fused_unify_head_weights) ? 1 : 0),
__extension__ ({ if (self->fused_unify_head_weights) ; else
__assert_fail ("self->fused_unify_head_weights", "ccv_cnnp_model_addons.c"
, 5129, __extension__ __PRETTY_FUNCTION__); }));
5130 add_to_array(parameters, self->weights, is_trainable);
5131 if (self->bias.graph)
5132 add_to_array(parameters, self->bias, is_trainable);
5133 }
5134}
5135
5136static ccv_cnnp_model_t* _ccv_cnnp_scaled_dot_product_attention_copy(const ccv_cnnp_model_t* const super, void* const context);
5137
5138static const ccv_cnnp_model_vtab_t ccv_cnnp_scaled_dot_product_attention_isa = {
5139 .build = _ccv_cnnp_scaled_dot_product_attention_build,
5140 .copy = _ccv_cnnp_scaled_dot_product_attention_copy,
5141};
5142
5143static const ccv_cnnp_model_vtab_t ccv_cnnp_scaled_dot_product_attention_fused_isa = {
5144 .build = _ccv_cnnp_scaled_dot_product_attention_build,
5145 .init_states = _ccv_cnnp_scaled_dot_product_attention_init_states,
5146 .add_to_parameter = _ccv_cnnp_scaled_dot_product_attention_add_to_parameter,
5147 .copy = _ccv_cnnp_scaled_dot_product_attention_copy,
5148};
5149
5150ccv_cnnp_model_t* ccv_cnnp_scaled_dot_product_attention(const float scale, const int is_causal, const int has_attn_mask, const int is_varlen, const int max_seqlen_q, const int max_seqlen_kv, const int flags, const int use_hadamard, const int attention_sinks, const int sliding_window, const int fused_unify_head_weights, const int no_bias, const int is_trainable, const char* const name)
5151{
5152 assert(!is_varlen || !has_attn_mask)((void) sizeof ((!is_varlen || !has_attn_mask) ? 1 : 0), __extension__
({ if (!is_varlen || !has_attn_mask) ; else __assert_fail ("!is_varlen || !has_attn_mask"
, "ccv_cnnp_model_addons.c", 5152, __extension__ __PRETTY_FUNCTION__
); }));
5153 assert(!is_varlen || !fused_unify_head_weights)((void) sizeof ((!is_varlen || !fused_unify_head_weights) ? 1
: 0), __extension__ ({ if (!is_varlen || !fused_unify_head_weights
) ; else __assert_fail ("!is_varlen || !fused_unify_head_weights"
, "ccv_cnnp_model_addons.c", 5153, __extension__ __PRETTY_FUNCTION__
); }));
5154 assert(sliding_window == 0 || is_causal)((void) sizeof ((sliding_window == 0 || is_causal) ? 1 : 0), __extension__
({ if (sliding_window == 0 || is_causal) ; else __assert_fail
("sliding_window == 0 || is_causal", "ccv_cnnp_model_addons.c"
, 5154, __extension__ __PRETTY_FUNCTION__); }));
5155 assert(sliding_window == 0 || !is_varlen)((void) sizeof ((sliding_window == 0 || !is_varlen) ? 1 : 0),
__extension__ ({ if (sliding_window == 0 || !is_varlen) ; else
__assert_fail ("sliding_window == 0 || !is_varlen", "ccv_cnnp_model_addons.c"
, 5155, __extension__ __PRETTY_FUNCTION__); }));
5156 ccv_cnnp_model_scaled_dot_product_attention_t* const model_scaled_dot_product_attention = (ccv_cnnp_model_scaled_dot_product_attention_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_scaled_dot_product_attention_t));
5157 model_scaled_dot_product_attention->super.isa = fused_unify_head_weights ? &ccv_cnnp_scaled_dot_product_attention_fused_isa : &ccv_cnnp_scaled_dot_product_attention_isa;
5158 const int has_attention_sinks = attention_sinks ? 1 : 0;
5159 model_scaled_dot_product_attention->super.input_size = (is_varlen ? 5 : (has_attn_mask ? 4 : 3)) + has_attention_sinks;
5160 model_scaled_dot_product_attention->super.outputs = &model_scaled_dot_product_attention->output;
5161 model_scaled_dot_product_attention->super.output_size = 1;
5162 model_scaled_dot_product_attention->super.is_trainable = is_trainable;
5163 ccv_cnnp_model_copy_name(&model_scaled_dot_product_attention->super, name);
5164 model_scaled_dot_product_attention->weights.d = CCV_NNC_NO_TENSOR_SYMBOL;
5165 model_scaled_dot_product_attention->weights.graph = 0;
5166 model_scaled_dot_product_attention->bias.d = CCV_NNC_NO_TENSOR_SYMBOL;
5167 model_scaled_dot_product_attention->bias.graph = 0;
5168 model_scaled_dot_product_attention->scale = scale;
5169 model_scaled_dot_product_attention->is_causal = is_causal;
5170 model_scaled_dot_product_attention->has_attn_mask = has_attn_mask;
5171 model_scaled_dot_product_attention->is_varlen = is_varlen;
5172 model_scaled_dot_product_attention->max_seqlen_q = max_seqlen_q;
5173 model_scaled_dot_product_attention->max_seqlen_kv = max_seqlen_kv;
5174 model_scaled_dot_product_attention->flags = flags;
5175 model_scaled_dot_product_attention->use_hadamard = use_hadamard;
5176 model_scaled_dot_product_attention->attention_sinks = has_attention_sinks;
5177 model_scaled_dot_product_attention->sliding_window = sliding_window;
5178 model_scaled_dot_product_attention->fused_unify_head_weights = fused_unify_head_weights;
5179 model_scaled_dot_product_attention->no_bias = no_bias;
5180 return (ccv_cnnp_model_t*)model_scaled_dot_product_attention;
5181}
5182
5183static ccv_cnnp_model_t* _ccv_cnnp_scaled_dot_product_attention_copy(const ccv_cnnp_model_t* const super, void* const context)
5184{
5185 const ccv_cnnp_model_scaled_dot_product_attention_t* const self = (const ccv_cnnp_model_scaled_dot_product_attention_t*)super;
5186 return ccv_cnnp_scaled_dot_product_attention(self->scale, self->is_causal, self->has_attn_mask, self->is_varlen, self->max_seqlen_q, self->max_seqlen_kv, self->flags, self->use_hadamard, self->attention_sinks, self->sliding_window, self->fused_unify_head_weights, self->no_bias, self->super.is_trainable, self->super.name);
5187}
5188
5189// MARK - Debug Layer
5190
5191typedef struct {
5192 ccv_cnnp_model_t super;
5193 ccv_nnc_tensor_symbol_t output;
5194 ccv_cnnp_model_debug_f debugger;
5195 ccv_cnnp_model_debug_context_deinit_f debug_deinit;
5196 ccv_cnnp_model_debug_context_copy_f debug_copy;
5197 void* debug_context;
5198} ccv_cnnp_model_debug_t;
5199
5200static int _ccv_cnnp_debug_exec(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint, const int flags, ccv_nnc_tensor_t* const* const inputs, const int input_size, ccv_nnc_tensor_t* const* const outputs, const int output_size, ccv_nnc_stream_context_t* const stream_context)
5201{
5202 if (cmd.cmd == CCV_NNC_CUSTOM_BACKWARD)
5203 {
5204 assert(0 && "don't support debug backward pass yet")((void) sizeof ((0 && "don't support debug backward pass yet"
) ? 1 : 0), __extension__ ({ if (0 && "don't support debug backward pass yet"
) ; else __assert_fail ("0 && \"don't support debug backward pass yet\""
, "ccv_cnnp_model_addons.c", 5204, __extension__ __PRETTY_FUNCTION__
); }));
5205 }
5206 ccv_cnnp_model_debug_t* const self = (ccv_cnnp_model_debug_t*)cmd.data;
5207 self->debugger(inputs, input_size, stream_context, self->debug_context);
5208 return CCV_NNC_EXEC_SUCCESS;
5209}
5210
5211static ccv_nnc_cmd_vtab_t ccv_cnnp_debug_exec_isa = {
5212 .exec = _ccv_cnnp_debug_exec
5213};
5214
5215static void _ccv_cnnp_debug_build(ccv_cnnp_model_t* const self, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
5216{
5217 PRINT(CCV_CLI_VERBOSE, "[cnnp_debug_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_debug_build] -\n"); fflush(stdout); } } while
(0);
5218 assert(input_size >= 1)((void) sizeof ((input_size >= 1) ? 1 : 0), __extension__ (
{ if (input_size >= 1) ; else __assert_fail ("input_size >= 1"
, "ccv_cnnp_model_addons.c", 5218, __extension__ __PRETTY_FUNCTION__
); }));
5219 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 5219, __extension__ __PRETTY_FUNCTION__
); }));
5220 ccv_nnc_tensor_symbol_t to = ccv_nnc_tensor_symbol_alias_to(graph, inputs[0]);
5221 ccv_nnc_tensor_param_t output_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
5222 if (to.d == CCV_NNC_NO_TENSOR_SYMBOL) // If we are not reshape an alias, it is straightforward.
5223 {
5224 int ofs[CCV_NNC_MAX_DIM_ALLOC(12)] = {0};
5225 int stride[CCV_NNC_MAX_DIM_ALLOC(12)];
5226 ccv_nnc_tensor_get_stride(output_params.dim, stride);
5227 outputs[0] = ccv_nnc_tensor_symbol_alias_new(graph, inputs[0], ofs, stride, output_params, 0);
5228 } else {
5229 int old_ofs[CCV_NNC_MAX_DIM_ALLOC(12)];
5230 int old_stride[CCV_NNC_MAX_DIM_ALLOC(12)];
5231 ccv_nnc_tensor_symbol_alias_params(graph, inputs[0], old_ofs, old_stride);
5232 outputs[0] = ccv_nnc_tensor_symbol_alias_new(graph, to, old_ofs, old_stride, output_params, 0);
5233 }
5234 ccv_nnc_cmd_t cmd = ccv_nnc_cmd(CCV_NNC_CUSTOM_FORWARD, (ccv_nnc_cmd_vtab_t*)&ccv_cnnp_debug_exec_isa, (ccv_nnc_cmd_param_t){}, 0);
5235 cmd.data = self;
5236 ccv_nnc_graph_exec_symbol_t make_debug = ccv_nnc_graph_exec_symbol_new(graph, cmd, inputs, input_size, outputs, 1, "debug");
5237 // Disable any optimizations.
5238 ccv_nnc_graph_exec_symbol_set_flags(graph, make_debug, CCV_NNC_GRAPH_EXEC_DISABLE_OPT);
5239}
5240
5241static void _ccv_cnnp_debug_deinit(ccv_cnnp_model_t* const super)
5242{
5243 const ccv_cnnp_model_debug_t* const self = (const ccv_cnnp_model_debug_t*)super;
5244 if (self->debug_deinit && self->debug_context)
5245 self->debug_deinit(self->debug_context);
5246}
5247
5248static ccv_cnnp_model_t* _ccv_cnnp_debug_copy(const ccv_cnnp_model_t* const super, void* const context);
5249
5250static const ccv_cnnp_model_vtab_t ccv_cnnp_debug_isa = {
5251 .build = _ccv_cnnp_debug_build,
5252 .deinit = _ccv_cnnp_debug_deinit,
5253 .copy = _ccv_cnnp_debug_copy,
5254};
5255
5256ccv_cnnp_model_t* ccv_cnnp_debug(ccv_cnnp_model_debug_f func, void* const context, ccv_cnnp_model_debug_context_deinit_f deinit, ccv_cnnp_model_debug_context_copy_f copy, const char* const name)
5257{
5258 ccv_cnnp_model_debug_t* const model_debug = (ccv_cnnp_model_debug_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_debug_t));
5259 model_debug->super.isa = &ccv_cnnp_debug_isa;
5260 model_debug->super.input_size = 0;
5261 model_debug->super.outputs = &model_debug->output;
5262 model_debug->super.output_size = 1;
5263 model_debug->debugger = func;
5264 model_debug->debug_context = context;
5265 model_debug->debug_deinit = deinit;
5266 model_debug->debug_copy = copy;
5267 ccv_cnnp_model_copy_name(&model_debug->super, name);
5268 return (ccv_cnnp_model_t*)model_debug;
5269}
5270
5271static ccv_cnnp_model_t* _ccv_cnnp_debug_copy(const ccv_cnnp_model_t* const super, void* const context)
5272{
5273 const ccv_cnnp_model_debug_t* const self = (const ccv_cnnp_model_debug_t*)super;
5274 void* debug_context = self->debug_context;
5275 if (self->debug_copy && self->debug_context)
5276 debug_context = self->debug_copy(self->debug_context);
5277 return ccv_cnnp_debug(self->debugger, debug_context, self->debug_deinit, self->debug_copy, self->super.name);
5278}
5279
5280/// MARK - Sort layer.
5281
5282typedef struct {
5283 ccv_cnnp_model_t super;
5284 ccv_nnc_tensor_symbol_t outputs[2];
5285 int along_axis;
5286 int descending;
5287} ccv_cnnp_model_sort_t;
5288
5289static void _ccv_cnnp_sort_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
5290{
5291 ccv_cnnp_model_sort_t* const self = (ccv_cnnp_model_sort_t*)super;
5292 PRINT(CCV_CLI_VERBOSE, "[cnnp_sort_build] - along_axis: %d, descending: %d\n", self->along_axis, self->descending)do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_sort_build] - along_axis: %d, descending: %d\n"
, self->along_axis, self->descending); fflush(stdout); }
} while (0);
5293 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
5294 assert(output_size == 2)((void) sizeof ((output_size == 2) ? 1 : 0), __extension__ ({
if (output_size == 2) ; else __assert_fail ("output_size == 2"
, "ccv_cnnp_model_addons.c", 5294, __extension__ __PRETTY_FUNCTION__
); }));
5295 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
5296 params.datatype = CCV_32S;
5297 outputs[1] = ccv_nnc_tensor_symbol_new(graph, params, 0);
5298 ccv_nnc_graph_exec_symbol_new(graph, CMD_SORT_FORWARD(self->along_axis, self->descending)ccv_nnc_cmd(CCV_NNC_SORT_FORWARD, 0, ((ccv_nnc_cmd_param_t){.
size={.dim={1,1,1}},.sort={.along_axis=self->along_axis,.descending
=self->descending}}), 0), inputs, input_size, outputs, output_size, "sort");
5299}
5300
5301static ccv_cnnp_model_t* _ccv_cnnp_sort_copy(const ccv_cnnp_model_t* const self, void* const context);
5302
5303static const ccv_cnnp_model_vtab_t ccv_cnnp_sort_isa = {
5304 .build = _ccv_cnnp_sort_build,
5305 .copy = _ccv_cnnp_sort_copy,
5306};
5307
5308ccv_cnnp_model_t* ccv_cnnp_sort(const int along_axis, const int descending, const char* const name)
5309{
5310 ccv_cnnp_model_sort_t* const model_sort = (ccv_cnnp_model_sort_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_sort_t));
5311 model_sort->super.isa = &ccv_cnnp_sort_isa;
5312 model_sort->super.input_size = 0;
5313 model_sort->super.outputs = model_sort->outputs;
5314 model_sort->super.output_size = 2;
5315 model_sort->along_axis = along_axis;
5316 model_sort->descending = descending;
5317 ccv_cnnp_model_copy_name(&model_sort->super, name);
5318 return (ccv_cnnp_model_t*)model_sort;
5319}
5320
5321static ccv_cnnp_model_t* _ccv_cnnp_sort_copy(const ccv_cnnp_model_t* const super, void* const context)
5322{
5323 ccv_cnnp_model_sort_t* const self = (ccv_cnnp_model_sort_t*)super;
5324 return ccv_cnnp_sort(self->along_axis, self->descending, self->super.name);
5325}
5326
5327/// MARK - Partition layer.
5328
5329typedef struct {
5330 ccv_cnnp_model_t super;
5331 ccv_nnc_tensor_symbol_t outputs[2];
5332 int kth;
5333 int along_axis;
5334 int descending;
5335} ccv_cnnp_model_partition_t;
5336
5337static void _ccv_cnnp_partition_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
5338{
5339 ccv_cnnp_model_partition_t* const self = (ccv_cnnp_model_partition_t*)super;
5340 PRINT(CCV_CLI_VERBOSE, "[cnnp_partition_build] - kth: %d, along_axis: %d, descending: %d\n", self->kth, self->along_axis, self->descending)do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_partition_build] - kth: %d, along_axis: %d, descending: %d\n"
, self->kth, self->along_axis, self->descending); fflush
(stdout); } } while (0);
5341 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
5342 assert(output_size == 2)((void) sizeof ((output_size == 2) ? 1 : 0), __extension__ ({
if (output_size == 2) ; else __assert_fail ("output_size == 2"
, "ccv_cnnp_model_addons.c", 5342, __extension__ __PRETTY_FUNCTION__
); }));
5343 if (self->kth > 0)
5344 params.dim[self->along_axis] = self->kth;
5345 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
5346 params.datatype = CCV_32S;
5347 outputs[1] = ccv_nnc_tensor_symbol_new(graph, params, 0);
5348 ccv_nnc_graph_exec_symbol_new(graph, CMD_PARTITION_FORWARD(self->kth, self->along_axis, self->descending)ccv_nnc_cmd(CCV_NNC_PARTITION_FORWARD, 0, ((ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.partition={.kth=self->kth,.along_axis
=self->along_axis,.descending=self->descending}}), 0), inputs, input_size, outputs, output_size, "partition");
5349}
5350
5351static ccv_cnnp_model_t* _ccv_cnnp_partition_copy(const ccv_cnnp_model_t* const self, void* const context);
5352
5353static const ccv_cnnp_model_vtab_t ccv_cnnp_partition_isa = {
5354 .build = _ccv_cnnp_partition_build,
5355 .copy = _ccv_cnnp_partition_copy,
5356};
5357
5358ccv_cnnp_model_t* ccv_cnnp_partition(const int kth, const int along_axis, const int descending, const char* const name)
5359{
5360 ccv_cnnp_model_partition_t* const model_partition = (ccv_cnnp_model_partition_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_partition_t));
5361 model_partition->super.isa = &ccv_cnnp_partition_isa;
5362 model_partition->super.input_size = 0;
5363 model_partition->super.outputs = model_partition->outputs;
5364 model_partition->super.output_size = 2;
5365 model_partition->kth = kth;
5366 model_partition->along_axis = along_axis;
5367 model_partition->descending = descending;
5368 ccv_cnnp_model_copy_name(&model_partition->super, name);
5369 return (ccv_cnnp_model_t*)model_partition;
5370}
5371
5372static ccv_cnnp_model_t* _ccv_cnnp_partition_copy(const ccv_cnnp_model_t* const super, void* const context)
5373{
5374 ccv_cnnp_model_partition_t* const self = (ccv_cnnp_model_partition_t*)super;
5375 return ccv_cnnp_partition(self->kth, self->along_axis, self->descending, self->super.name);
5376}
5377
5378/// MARK - Unique consecutive layer.
5379
5380typedef struct {
5381 ccv_cnnp_model_t super;
5382 ccv_nnc_tensor_symbol_t outputs[2];
5383 int bincount;
5384} ccv_cnnp_model_unique_consecutive_t;
5385
5386static void _ccv_cnnp_unique_consecutive_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
5387{
5388 ccv_cnnp_model_unique_consecutive_t* const self = (ccv_cnnp_model_unique_consecutive_t*)super;
5389 PRINT(CCV_CLI_VERBOSE, "[cnnp_unique_consecutive_build] - bincount: %d\n", self->bincount)do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_unique_consecutive_build] - bincount: %d\n",
self->bincount); fflush(stdout); } } while (0);
5390 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
5391 assert(output_size == 2)((void) sizeof ((output_size == 2) ? 1 : 0), __extension__ ({
if (output_size == 2) ; else __assert_fail ("output_size == 2"
, "ccv_cnnp_model_addons.c", 5391, __extension__ __PRETTY_FUNCTION__
); }));
5392 if (self->bincount > 0)
5393 params.dim[0] = ccv_min(params.dim[0], self->bincount)({ typeof (params.dim[0]) _a = (params.dim[0]); typeof (self->
bincount) _b = (self->bincount); (_a < _b) ? _a : _b; }
);
5394 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
5395 params.datatype = CCV_32S;
5396 outputs[1] = ccv_nnc_tensor_symbol_new(graph, params, 0);
5397 ccv_nnc_graph_exec_symbol_new(graph, CMD_UNIQUE_CONSECUTIVE_FORWARD(self->bincount)ccv_nnc_cmd(CCV_NNC_UNIQUE_CONSECUTIVE_FORWARD, 0, ((ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.unique_consecutive={.bincount=self->
bincount}}), 0), inputs, input_size, outputs, output_size, "unique_consecutive");
5398}
5399
5400static ccv_cnnp_model_t* _ccv_cnnp_unique_consecutive_copy(const ccv_cnnp_model_t* const self, void* const context);
5401
5402static const ccv_cnnp_model_vtab_t ccv_cnnp_unique_consecutive_isa = {
5403 .build = _ccv_cnnp_unique_consecutive_build,
5404 .copy = _ccv_cnnp_unique_consecutive_copy,
5405};
5406
5407ccv_cnnp_model_t* ccv_cnnp_unique_consecutive(const int bincount, const char* const name)
5408{
5409 ccv_cnnp_model_unique_consecutive_t* const model_unique_consecutive = (ccv_cnnp_model_unique_consecutive_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_unique_consecutive_t));
5410 model_unique_consecutive->super.isa = &ccv_cnnp_unique_consecutive_isa;
5411 model_unique_consecutive->super.input_size = 0;
5412 model_unique_consecutive->super.outputs = model_unique_consecutive->outputs;
5413 model_unique_consecutive->super.output_size = 2;
5414 model_unique_consecutive->bincount = bincount;
5415 ccv_cnnp_model_copy_name(&model_unique_consecutive->super, name);
5416 return (ccv_cnnp_model_t*)model_unique_consecutive;
5417}
5418
5419static ccv_cnnp_model_t* _ccv_cnnp_unique_consecutive_copy(const ccv_cnnp_model_t* const super, void* const context)
5420{
5421 ccv_cnnp_model_unique_consecutive_t* const self = (ccv_cnnp_model_unique_consecutive_t*)super;
5422 return ccv_cnnp_unique_consecutive(self->bincount, self->super.name);
5423}
5424
5425/// MARK - Scatter add layer.
5426
5427typedef struct {
5428 ccv_cnnp_model_t super;
5429 ccv_nnc_tensor_symbol_t output;
5430 int bincount;
5431 int count_per_output;
5432} ccv_cnnp_model_scatter_add_t;
5433
5434static void _ccv_cnnp_scatter_add_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
5435{
5436 ccv_cnnp_model_scatter_add_t* const self = (ccv_cnnp_model_scatter_add_t*)super;
5437 PRINT(CCV_CLI_VERBOSE, "[cnnp_scatter_add_build] - bincount: %d, count_per_output: %d\n", self->bincount, self->count_per_output)do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_scatter_add_build] - bincount: %d, count_per_output: %d\n"
, self->bincount, self->count_per_output); fflush(stdout
); } } while (0);
5438 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
5439 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 5439, __extension__ __PRETTY_FUNCTION__
); }));
5440 assert(self->bincount >= 0 && (self->bincount > 0 || params.dim[0] == 0))((void) sizeof ((self->bincount >= 0 && (self->
bincount > 0 || params.dim[0] == 0)) ? 1 : 0), __extension__
({ if (self->bincount >= 0 && (self->bincount
> 0 || params.dim[0] == 0)) ; else __assert_fail ("self->bincount >= 0 && (self->bincount > 0 || params.dim[0] == 0)"
, "ccv_cnnp_model_addons.c", 5440, __extension__ __PRETTY_FUNCTION__
); }));
5441 if (params.dim[0] != 0)
5442 params.dim[0] = self->bincount;
5443 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
5444 ccv_nnc_graph_exec_symbol_new(graph, CMD_SCATTER_ADD_FORWARD(self->bincount, self->count_per_output)ccv_nnc_cmd(CCV_NNC_SCATTER_ADD_FORWARD, 0, ((ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.scatter_add={.bincount=self->bincount
,.count_per_output=self->count_per_output}}), 0), inputs, input_size, outputs, output_size, "scatter_add");
5445}
5446
5447static ccv_cnnp_model_t* _ccv_cnnp_scatter_add_copy(const ccv_cnnp_model_t* const self, void* const context);
5448
5449static const ccv_cnnp_model_vtab_t ccv_cnnp_scatter_add_isa = {
5450 .build = _ccv_cnnp_scatter_add_build,
5451 .copy = _ccv_cnnp_scatter_add_copy,
5452};
5453
5454ccv_cnnp_model_t* ccv_cnnp_scatter_add(const int bincount, const int count_per_output, const char* const name)
5455{
5456 assert(bincount >= 0)((void) sizeof ((bincount >= 0) ? 1 : 0), __extension__ ({
if (bincount >= 0) ; else __assert_fail ("bincount >= 0"
, "ccv_cnnp_model_addons.c", 5456, __extension__ __PRETTY_FUNCTION__
); }));
5457 assert(count_per_output >= 0)((void) sizeof ((count_per_output >= 0) ? 1 : 0), __extension__
({ if (count_per_output >= 0) ; else __assert_fail ("count_per_output >= 0"
, "ccv_cnnp_model_addons.c", 5457, __extension__ __PRETTY_FUNCTION__
); }));
5458 ccv_cnnp_model_scatter_add_t* const model_scatter_add = (ccv_cnnp_model_scatter_add_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_scatter_add_t));
5459 model_scatter_add->super.isa = &ccv_cnnp_scatter_add_isa;
5460 model_scatter_add->super.input_size = 0;
5461 model_scatter_add->super.outputs = &model_scatter_add->output;
5462 model_scatter_add->super.output_size = 1;
5463 model_scatter_add->bincount = bincount;
5464 model_scatter_add->count_per_output = count_per_output;
5465 ccv_cnnp_model_copy_name(&model_scatter_add->super, name);
5466 return (ccv_cnnp_model_t*)model_scatter_add;
5467}
5468
5469static ccv_cnnp_model_t* _ccv_cnnp_scatter_add_copy(const ccv_cnnp_model_t* const super, void* const context)
5470{
5471 ccv_cnnp_model_scatter_add_t* const self = (ccv_cnnp_model_scatter_add_t*)super;
5472 return ccv_cnnp_scatter_add(self->bincount, self->count_per_output, self->super.name);
5473}
5474
5475// MARK - Segmented Dense Layer
5476
5477typedef struct {
5478 ccv_cnnp_model_t super;
5479 ccv_nnc_tensor_symbol_t output;
5480 ccv_nnc_tensor_symbol_t weights;
5481 ccv_nnc_tensor_symbol_t bias;
5482 int segments;
5483 int count;
5484 int no_bias;
5485 int flags;
5486 int functional;
5487} ccv_cnnp_model_segmented_dense_t;
5488
5489static void _ccv_cnnp_segmented_dense_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
5490{
5491 ccv_cnnp_model_segmented_dense_t* const self = (ccv_cnnp_model_segmented_dense_t*)super;
5492 PRINT(CCV_CLI_VERBOSE, "[cnnp_segmented_dense_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_segmented_dense_build] -\n"); fflush(stdout)
; } } while (0);
5493 assert(input_size == (self->functional ? 4 + !self->no_bias : 3))((void) sizeof ((input_size == (self->functional ? 4 + !self
->no_bias : 3)) ? 1 : 0), __extension__ ({ if (input_size ==
(self->functional ? 4 + !self->no_bias : 3)) ; else __assert_fail
("input_size == (self->functional ? 4 + !self->no_bias : 3)"
, "ccv_cnnp_model_addons.c", 5493, __extension__ __PRETTY_FUNCTION__
); }));
5494 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 5494, __extension__ __PRETTY_FUNCTION__
); }));
5495 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
5496 const ccv_nnc_tensor_param_t indices_params = ccv_nnc_tensor_symbol_params(graph, inputs[1]);
5497 const ccv_nnc_tensor_param_t counts_params = ccv_nnc_tensor_symbol_params(graph, inputs[2]);
5498 ccv_nnc_tensor_param_t weights_params;
5499 ccv_nnc_tensor_symbol_t weights;
5500 if (self->functional)
5501 {
5502 weights = inputs[3];
5503 weights_params = ccv_nnc_tensor_symbol_params(graph, weights);
5504 } else {
5505 weights_params = params;
5506 memset(weights_params.dim, 0, sizeof(weights_params.dim));
5507 if (params.dim[0] != 0)
5508 {
5509 weights_params.dim[0] = self->segments;
5510 weights_params.dim[1] = self->count;
5511 weights_params.dim[2] = params.dim[ccv_nnc_tensor_nd(params.dim) - 1];
5512 }
5513 if (!self->weights.graph)
5514 self->weights = ccv_nnc_tensor_symbol_new(graph, weights_params, "weights");
5515 assert(self->weights.graph == graph)((void) sizeof ((self->weights.graph == graph) ? 1 : 0), __extension__
({ if (self->weights.graph == graph) ; else __assert_fail
("self->weights.graph == graph", "ccv_cnnp_model_addons.c"
, 5515, __extension__ __PRETTY_FUNCTION__); }));
5516 weights = ccv_cnnp_model_get_symbol(super, self->weights);
5517 }
5518 ccv_nnc_tensor_param_t bias_params = {};
5519 ccv_nnc_tensor_symbol_t bias = {};
5520 if (!self->no_bias)
5521 {
5522 if (self->functional)
5523 {
5524 bias = inputs[4];
5525 bias_params = ccv_nnc_tensor_symbol_params(graph, bias);
5526 } else {
5527 bias_params = params;
5528 memset(bias_params.dim, 0, sizeof(bias_params.dim));
5529 bias_params.dim[0] = self->segments;
5530 bias_params.dim[1] = self->count;
5531 if (!self->bias.graph)
5532 self->bias = ccv_nnc_tensor_symbol_new(graph, bias_params, "bias");
5533 bias = ccv_cnnp_model_get_symbol(super, self->bias);
5534 }
5535 }
5536 ccv_nnc_cmd_t cmd = {0};
5537 cmd.cmd = CCV_NNC_SEGMENTED_GEMM_FORWARD;
5538 cmd.info.blas.a[0] = 1;
5539 cmd.info.blas.a[1] = 1;
5540 cmd.info.blas.transpose_b[0] = 1;
5541 cmd.info.blas.transpose_b[1] = 2;
5542 cmd.info.blas.flags = self->flags;
5543 ccv_nnc_tensor_param_t output_params;
5544 ccv_nnc_hint_tensor_auto(cmd, (ccv_nnc_tensor_param_t []){
5545 params, indices_params, counts_params,
5546 weights_params,
5547 bias_params,
5548 }, 4 + !self->no_bias, ccv_nnc_no_hint, &output_params, 1);
5549 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
5550 if (self->no_bias)
5551 ccv_nnc_graph_exec_symbol_new(graph, cmd, TENSOR_SYMBOL_LIST(inputs[0], inputs[1], inputs[2], weights)(const ccv_nnc_tensor_symbol_t []){inputs[0], inputs[1], inputs
[2], weights}, (1 +1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(output)(const ccv_nnc_tensor_symbol_t []){output}, (1 +1 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "segmented_dense");
5552 else
5553 ccv_nnc_graph_exec_symbol_new(graph, cmd, TENSOR_SYMBOL_LIST(inputs[0], inputs[1], inputs[2], weights, bias)(const ccv_nnc_tensor_symbol_t []){inputs[0], inputs[1], inputs
[2], weights, bias}, (1 +1 +1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(output)(const ccv_nnc_tensor_symbol_t []){output}, (1 +1 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "segmented_dense");
5554 outputs[0] = output;
5555}
5556
5557static void _ccv_cnnp_segmented_dense_init_states(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_cnnp_state_initializer_f initializer, void* const context)
5558{
5559 ccv_cnnp_model_segmented_dense_t* const self = (ccv_cnnp_model_segmented_dense_t*)super;
5560 if (self->functional)
5561 return;
5562 const ccv_nnc_tensor_param_t weight_params = ccv_nnc_tensor_symbol_params(graph, self->weights);
5563 const int c = weight_params.dim[1];
5564 const float std = sqrtf(2) / sqrtf(c);
5565 const float bound = sqrtf(3) * std;
5566 initializer(context, CMD_RANDOM_UNIFORM_FORWARD(-bound, bound)ccv_nnc_cmd(CCV_NNC_RANDOM_UNIFORM_FORWARD, 0, (ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.blas={.a={-bound, bound}}}, 0), ccv_nnc_no_hint, 0, 0, self->weights);
5567 if (self->bias.graph)
5568 initializer(context, CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, self->bias);
5569}
5570
5571static void _ccv_cnnp_segmented_dense_add_to_parameter(ccv_cnnp_model_t* const super, const ccv_cnnp_add_to_array_f add_to_array, void* const parameters, const int is_trainable)
5572{
5573 ccv_cnnp_model_segmented_dense_t* const self = (ccv_cnnp_model_segmented_dense_t*)super;
5574 if (self->functional)
5575 return;
5576 add_to_array(parameters, self->weights, is_trainable);
5577 if (self->bias.graph)
5578 add_to_array(parameters, self->bias, is_trainable);
5579}
5580
5581static ccv_cnnp_model_t* _ccv_cnnp_segmented_dense_copy(const ccv_cnnp_model_t* const super, void* const context);
5582
5583static const ccv_cnnp_model_vtab_t ccv_cnnp_segmented_dense_isa = {
5584 .build = _ccv_cnnp_segmented_dense_build,
5585 .init_states = _ccv_cnnp_segmented_dense_init_states,
5586 .add_to_parameter = _ccv_cnnp_segmented_dense_add_to_parameter,
5587 .copy = _ccv_cnnp_segmented_dense_copy,
5588};
5589
5590ccv_cnnp_model_t* ccv_cnnp_segmented_dense(const int segments, const int count, const int no_bias, const int flags, const int functional, const int is_trainable, const char* const name)
5591{
5592 ccv_cnnp_model_segmented_dense_t* const model_segmented_dense = (ccv_cnnp_model_segmented_dense_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_segmented_dense_t));
5593 model_segmented_dense->super.isa = &ccv_cnnp_segmented_dense_isa;
5594 model_segmented_dense->super.input_size = functional ? 4 + !no_bias : 3;
5595 model_segmented_dense->super.outputs = &model_segmented_dense->output;
5596 model_segmented_dense->super.output_size = 1;
5597 model_segmented_dense->super.is_trainable = is_trainable;
5598 ccv_cnnp_model_copy_name(&model_segmented_dense->super, name);
5599 model_segmented_dense->weights.d = CCV_NNC_NO_TENSOR_SYMBOL;
5600 model_segmented_dense->weights.graph = 0;
5601 model_segmented_dense->bias.d = CCV_NNC_NO_TENSOR_SYMBOL;
5602 model_segmented_dense->bias.graph = 0;
5603 model_segmented_dense->segments = segments;
5604 model_segmented_dense->count = count;
5605 model_segmented_dense->no_bias = no_bias;
5606 model_segmented_dense->flags = flags;
5607 model_segmented_dense->functional = functional;
5608 return (ccv_cnnp_model_t*)model_segmented_dense;
5609}
5610
5611static ccv_cnnp_model_t* _ccv_cnnp_segmented_dense_copy(const ccv_cnnp_model_t* const super, void* const context)
5612{
5613 const ccv_cnnp_model_segmented_dense_t* const self = (const ccv_cnnp_model_segmented_dense_t*)super;
5614 return ccv_cnnp_segmented_dense(self->segments, self->count, self->no_bias, self->flags, self->functional, self->super.is_trainable, self->super.name);
5615}
5616
5617// MARK - SwiGLU Layer
5618
5619typedef struct {
5620 ccv_cnnp_model_t super;
5621 ccv_nnc_tensor_symbol_t output;
5622 ccv_nnc_tensor_symbol_t gate_weights;
5623 ccv_nnc_tensor_symbol_t up_weights;
5624 int count;
5625 float clamp;
5626} ccv_cnnp_model_swiglu_t;
5627
5628static void _ccv_cnnp_swiglu_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
5629{
5630 ccv_cnnp_model_swiglu_t* const self = (ccv_cnnp_model_swiglu_t*)super;
5631 PRINT(CCV_CLI_VERBOSE, "[cnnp_swiglu_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_swiglu_build] -\n"); fflush(stdout); } } while
(0);
5632 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 5632, __extension__ __PRETTY_FUNCTION__); }));
5633 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 5633, __extension__ __PRETTY_FUNCTION__
); }));
5634 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
5635 const int params_nd = ccv_nnc_tensor_nd(params.dim);
5636 ccv_nnc_tensor_param_t weights_params = params;
5637 memset(weights_params.dim, 0, sizeof(weights_params.dim));
5638 if (params_nd > 0)
5639 {
5640 weights_params.dim[0] = self->count;
5641 weights_params.dim[1] = params.dim[params_nd - 1];
5642 }
5643 if (!self->gate_weights.graph)
5644 self->gate_weights = ccv_nnc_tensor_symbol_new(graph, weights_params, "gate.weight");
5645 if (!self->up_weights.graph)
5646 self->up_weights = ccv_nnc_tensor_symbol_new(graph, weights_params, "up.weight");
5647 assert(self->gate_weights.graph == graph)((void) sizeof ((self->gate_weights.graph == graph) ? 1 : 0
), __extension__ ({ if (self->gate_weights.graph == graph)
; else __assert_fail ("self->gate_weights.graph == graph"
, "ccv_cnnp_model_addons.c", 5647, __extension__ __PRETTY_FUNCTION__
); }));
5648 assert(self->up_weights.graph == graph)((void) sizeof ((self->up_weights.graph == graph) ? 1 : 0)
, __extension__ ({ if (self->up_weights.graph == graph) ; else
__assert_fail ("self->up_weights.graph == graph", "ccv_cnnp_model_addons.c"
, 5648, __extension__ __PRETTY_FUNCTION__); }));
5649 const ccv_nnc_tensor_symbol_t gate_weights = ccv_cnnp_model_get_symbol(super, self->gate_weights);
5650 const ccv_nnc_tensor_symbol_t up_weights = ccv_cnnp_model_get_symbol(super, self->up_weights);
5651 const ccv_nnc_cmd_t cmd = CMD_SWIGLU_FORWARD(self->clamp)ccv_nnc_cmd(CCV_NNC_SWIGLU_FORWARD, 0, ((ccv_nnc_cmd_param_t)
{.size={.dim={1,1,1}},.swiglu={.clamp=self->clamp}}), 0);
5652 const ccv_nnc_tensor_symbol_t command_inputs[] = {
5653 inputs[0], gate_weights, up_weights,
5654 };
5655 ccv_nnc_tensor_param_t command_input_params[3];
5656 int i;
5657 for (i = 0; i < 3; i++)
5658 command_input_params[i] = ccv_nnc_tensor_symbol_params(graph, command_inputs[i]);
5659 ccv_nnc_tensor_param_t output_params;
5660 ccv_nnc_hint_tensor_auto(cmd, command_input_params, 3, ccv_nnc_no_hint, &output_params, 1);
5661 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
5662 ccv_nnc_graph_exec_symbol_new(graph, cmd, command_inputs, 3, outputs, 1, "swiglu");
5663}
5664
5665static void _ccv_cnnp_swiglu_init_states(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_cnnp_state_initializer_f initializer, void* const context)
5666{
5667 ccv_cnnp_model_swiglu_t* const self = (ccv_cnnp_model_swiglu_t*)super;
5668 const ccv_nnc_tensor_param_t weight_params = ccv_nnc_tensor_symbol_params(graph, self->gate_weights);
5669 const float bound = sqrtf(6.f / weight_params.dim[1]);
5670 initializer(context, CMD_RANDOM_UNIFORM_FORWARD(-bound, bound)ccv_nnc_cmd(CCV_NNC_RANDOM_UNIFORM_FORWARD, 0, (ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.blas={.a={-bound, bound}}}, 0), ccv_nnc_no_hint, 0, 0, self->gate_weights);
5671 initializer(context, CMD_RANDOM_UNIFORM_FORWARD(-bound, bound)ccv_nnc_cmd(CCV_NNC_RANDOM_UNIFORM_FORWARD, 0, (ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.blas={.a={-bound, bound}}}, 0), ccv_nnc_no_hint, 0, 0, self->up_weights);
5672}
5673
5674static void _ccv_cnnp_swiglu_add_to_parameter(ccv_cnnp_model_t* const super, const ccv_cnnp_add_to_array_f add_to_array, void* const parameters, const int is_trainable)
5675{
5676 ccv_cnnp_model_swiglu_t* const self = (ccv_cnnp_model_swiglu_t*)super;
5677 add_to_array(parameters, self->gate_weights, is_trainable);
5678 add_to_array(parameters, self->up_weights, is_trainable);
5679}
5680
5681static ccv_cnnp_model_t* _ccv_cnnp_swiglu_copy(const ccv_cnnp_model_t* const super, void* const context);
5682
5683static const ccv_cnnp_model_vtab_t ccv_cnnp_swiglu_isa = {
5684 .build = _ccv_cnnp_swiglu_build,
5685 .init_states = _ccv_cnnp_swiglu_init_states,
5686 .add_to_parameter = _ccv_cnnp_swiglu_add_to_parameter,
5687 .copy = _ccv_cnnp_swiglu_copy,
5688};
5689
5690ccv_cnnp_model_t* ccv_cnnp_swiglu(const int count, const float clamp, const int is_trainable, const char* const name)
5691{
5692 ccv_cnnp_model_swiglu_t* const model = (ccv_cnnp_model_swiglu_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_swiglu_t));
5693 model->super.isa = &ccv_cnnp_swiglu_isa;
5694 model->super.input_size = 1;
5695 model->super.outputs = &model->output;
5696 model->super.output_size = 1;
5697 model->super.is_trainable = is_trainable;
5698 ccv_cnnp_model_copy_name(&model->super, name);
5699 model->gate_weights.d = CCV_NNC_NO_TENSOR_SYMBOL;
5700 model->up_weights.d = CCV_NNC_NO_TENSOR_SYMBOL;
5701 model->count = count;
5702 model->clamp = clamp;
5703 return (ccv_cnnp_model_t*)model;
5704}
5705
5706static ccv_cnnp_model_t* _ccv_cnnp_swiglu_copy(const ccv_cnnp_model_t* const super, void* const context)
5707{
5708 const ccv_cnnp_model_swiglu_t* const self = (const ccv_cnnp_model_swiglu_t*)super;
5709 return ccv_cnnp_swiglu(self->count, self->clamp, self->super.is_trainable, self->super.name);
5710}
5711
5712// MARK - Segmented SwiGLU Layer
5713
5714typedef struct {
5715 ccv_cnnp_model_t super;
5716 ccv_nnc_tensor_symbol_t output;
5717 ccv_nnc_tensor_symbol_t gate_weights;
5718 ccv_nnc_tensor_symbol_t up_weights;
5719 int segments;
5720 int count;
5721 float clamp;
5722 int functional;
5723} ccv_cnnp_model_segmented_swiglu_t;
5724
5725static void _ccv_cnnp_segmented_swiglu_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
5726{
5727 ccv_cnnp_model_segmented_swiglu_t* const self = (ccv_cnnp_model_segmented_swiglu_t*)super;
5728 PRINT(CCV_CLI_VERBOSE, "[cnnp_segmented_swiglu_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_segmented_swiglu_build] -\n"); fflush(stdout
); } } while (0);
5729 assert(input_size == (self->functional ? 6 : 4))((void) sizeof ((input_size == (self->functional ? 6 : 4))
? 1 : 0), __extension__ ({ if (input_size == (self->functional
? 6 : 4)) ; else __assert_fail ("input_size == (self->functional ? 6 : 4)"
, "ccv_cnnp_model_addons.c", 5729, __extension__ __PRETTY_FUNCTION__
); }));
5730 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 5730, __extension__ __PRETTY_FUNCTION__
); }));
5731 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
5732 ccv_nnc_tensor_symbol_t gate_weights;
5733 ccv_nnc_tensor_symbol_t up_weights;
5734 if (self->functional)
5735 {
5736 gate_weights = inputs[3];
5737 up_weights = inputs[4];
5738 } else {
5739 const int params_nd = ccv_nnc_tensor_nd(params.dim);
5740 ccv_nnc_tensor_param_t weights_params = params;
5741 memset(weights_params.dim, 0, sizeof(weights_params.dim));
5742 if (params_nd > 0)
5743 {
5744 weights_params.dim[0] = self->segments;
5745 weights_params.dim[1] = self->count;
5746 weights_params.dim[2] = params.dim[params_nd - 1];
5747 }
5748 if (!self->gate_weights.graph)
5749 self->gate_weights = ccv_nnc_tensor_symbol_new(graph, weights_params, "gate.weight");
5750 if (!self->up_weights.graph)
5751 self->up_weights = ccv_nnc_tensor_symbol_new(graph, weights_params, "up.weight");
5752 assert(self->gate_weights.graph == graph)((void) sizeof ((self->gate_weights.graph == graph) ? 1 : 0
), __extension__ ({ if (self->gate_weights.graph == graph)
; else __assert_fail ("self->gate_weights.graph == graph"
, "ccv_cnnp_model_addons.c", 5752, __extension__ __PRETTY_FUNCTION__
); }));
5753 assert(self->up_weights.graph == graph)((void) sizeof ((self->up_weights.graph == graph) ? 1 : 0)
, __extension__ ({ if (self->up_weights.graph == graph) ; else
__assert_fail ("self->up_weights.graph == graph", "ccv_cnnp_model_addons.c"
, 5753, __extension__ __PRETTY_FUNCTION__); }));
5754 gate_weights = ccv_cnnp_model_get_symbol(super, self->gate_weights);
5755 up_weights = ccv_cnnp_model_get_symbol(super, self->up_weights);
5756 }
5757 const ccv_nnc_cmd_t cmd = CMD_SEGMENTED_SWIGLU_FORWARD(self->clamp)ccv_nnc_cmd(CCV_NNC_SEGMENTED_SWIGLU_FORWARD, 0, ((ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.segmented_swiglu={.clamp=self->clamp
}}), 0);
5758 const ccv_nnc_tensor_symbol_t command_inputs[] = {
5759 inputs[0], inputs[1], inputs[2], gate_weights, up_weights, inputs[self->functional ? 5 : 3],
5760 };
5761 ccv_nnc_tensor_param_t command_input_params[6];
5762 int i;
5763 for (i = 0; i < 6; i++)
5764 command_input_params[i] = ccv_nnc_tensor_symbol_params(graph, command_inputs[i]);
5765 ccv_nnc_tensor_param_t output_params;
5766 ccv_nnc_hint_tensor_auto(cmd, command_input_params, 6, ccv_nnc_no_hint, &output_params, 1);
5767 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
5768 ccv_nnc_graph_exec_symbol_new(graph, cmd, command_inputs, 6, outputs, 1, "segmented_swiglu");
5769}
5770
5771static void _ccv_cnnp_segmented_swiglu_init_states(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_cnnp_state_initializer_f initializer, void* const context)
5772{
5773 ccv_cnnp_model_segmented_swiglu_t* const self = (ccv_cnnp_model_segmented_swiglu_t*)super;
5774 if (self->functional)
5775 return;
5776 const ccv_nnc_tensor_param_t weight_params = ccv_nnc_tensor_symbol_params(graph, self->gate_weights);
5777 const float bound = sqrtf(6.f / weight_params.dim[2]);
5778 initializer(context, CMD_RANDOM_UNIFORM_FORWARD(-bound, bound)ccv_nnc_cmd(CCV_NNC_RANDOM_UNIFORM_FORWARD, 0, (ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.blas={.a={-bound, bound}}}, 0), ccv_nnc_no_hint, 0, 0, self->gate_weights);
5779 initializer(context, CMD_RANDOM_UNIFORM_FORWARD(-bound, bound)ccv_nnc_cmd(CCV_NNC_RANDOM_UNIFORM_FORWARD, 0, (ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.blas={.a={-bound, bound}}}, 0), ccv_nnc_no_hint, 0, 0, self->up_weights);
5780}
5781
5782static void _ccv_cnnp_segmented_swiglu_add_to_parameter(ccv_cnnp_model_t* const super, const ccv_cnnp_add_to_array_f add_to_array, void* const parameters, const int is_trainable)
5783{
5784 ccv_cnnp_model_segmented_swiglu_t* const self = (ccv_cnnp_model_segmented_swiglu_t*)super;
5785 if (self->functional)
5786 return;
5787 add_to_array(parameters, self->gate_weights, is_trainable);
5788 add_to_array(parameters, self->up_weights, is_trainable);
5789}
5790
5791static ccv_cnnp_model_t* _ccv_cnnp_segmented_swiglu_copy(const ccv_cnnp_model_t* const super, void* const context);
5792
5793static const ccv_cnnp_model_vtab_t ccv_cnnp_segmented_swiglu_isa = {
5794 .build = _ccv_cnnp_segmented_swiglu_build,
5795 .init_states = _ccv_cnnp_segmented_swiglu_init_states,
5796 .add_to_parameter = _ccv_cnnp_segmented_swiglu_add_to_parameter,
5797 .copy = _ccv_cnnp_segmented_swiglu_copy,
5798};
5799
5800ccv_cnnp_model_t* ccv_cnnp_segmented_swiglu(const int segments, const int count, const float clamp, const int functional, const int is_trainable, const char* const name)
5801{
5802 ccv_cnnp_model_segmented_swiglu_t* const model = (ccv_cnnp_model_segmented_swiglu_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_segmented_swiglu_t));
5803 model->super.isa = &ccv_cnnp_segmented_swiglu_isa;
5804 model->super.input_size = functional ? 6 : 4;
5805 model->super.outputs = &model->output;
5806 model->super.output_size = 1;
5807 model->super.is_trainable = is_trainable;
5808 ccv_cnnp_model_copy_name(&model->super, name);
5809 model->gate_weights.d = CCV_NNC_NO_TENSOR_SYMBOL;
5810 model->up_weights.d = CCV_NNC_NO_TENSOR_SYMBOL;
5811 model->segments = segments;
5812 model->count = count;
5813 model->clamp = clamp;
5814 model->functional = functional;
5815 return (ccv_cnnp_model_t*)model;
5816}
5817
5818static ccv_cnnp_model_t* _ccv_cnnp_segmented_swiglu_copy(const ccv_cnnp_model_t* const super, void* const context)
5819{
5820 const ccv_cnnp_model_segmented_swiglu_t* const self = (const ccv_cnnp_model_segmented_swiglu_t*)super;
5821 return ccv_cnnp_segmented_swiglu(self->segments, self->count, self->clamp, self->functional, self->super.is_trainable, self->super.name);
5822}
5823
5824// MARK - MoE Weights Streaming
5825
5826typedef struct {
5827 ccv_cnnp_model_t super;
5828 ccv_nnc_tensor_symbol_t outputs[6];
5829 int resident_slots;
5830 int routing_width;
5831} ccv_cnnp_model_moe_weights_streaming_t;
5832
5833static void _ccv_cnnp_moe_weights_streaming_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
5834{
5835 ccv_cnnp_model_moe_weights_streaming_t* const self = (ccv_cnnp_model_moe_weights_streaming_t*)super;
5836 PRINT(CCV_CLI_VERBOSE, "[cnnp_moe_weights_streaming_build] - resident_slots: %d, routing_width: %d\n", self->resident_slots, self->routing_width)do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_moe_weights_streaming_build] - resident_slots: %d, routing_width: %d\n"
, self->resident_slots, self->routing_width); fflush(stdout
); } } while (0);
5837 assert(input_size == 6)((void) sizeof ((input_size == 6) ? 1 : 0), __extension__ ({ if
(input_size == 6) ; else __assert_fail ("input_size == 6", "ccv_cnnp_model_addons.c"
, 5837, __extension__ __PRETTY_FUNCTION__); }));
5838 assert(output_size == 6)((void) sizeof ((output_size == 6) ? 1 : 0), __extension__ ({
if (output_size == 6) ; else __assert_fail ("output_size == 6"
, "ccv_cnnp_model_addons.c", 5838, __extension__ __PRETTY_FUNCTION__
); }));
5839 const ccv_nnc_cmd_t cmd = CMD_MOE_WEIGHTS_STREAMING_FORWARD(self->resident_slots, self->routing_width)ccv_nnc_cmd(CCV_NNC_MOE_WEIGHTS_STREAMING_FORWARD, 0, ((ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.moe_weights_streaming={.resident_slots
=(self->resident_slots),.routing_width=(self->routing_width
)}}), 0);
5840 ccv_nnc_tensor_param_t input_params[6];
5841 ccv_nnc_tensor_param_t output_params[6];
5842 int i;
5843 for (i = 0; i < 6; i++)
5844 input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
5845 ccv_nnc_hint_tensor_auto(cmd, input_params, 6, ccv_nnc_no_hint, output_params, 6);
5846 for (i = 0; i < 6; i++)
5847 outputs[i] = ccv_nnc_tensor_symbol_new(graph, output_params[i], 0);
5848 ccv_nnc_graph_exec_symbol_new(graph, cmd, inputs, 6, outputs, 6, "moe_weights_streaming");
5849}
5850
5851static ccv_cnnp_model_t* _ccv_cnnp_moe_weights_streaming_copy(const ccv_cnnp_model_t* const super, void* const context);
5852
5853static const ccv_cnnp_model_vtab_t ccv_cnnp_moe_weights_streaming_isa = {
5854 .build = _ccv_cnnp_moe_weights_streaming_build,
5855 .copy = _ccv_cnnp_moe_weights_streaming_copy,
5856};
5857
5858ccv_cnnp_model_t* ccv_cnnp_moe_weights_streaming(const int resident_slots, const int routing_width, const char* const name)
5859{
5860 assert(resident_slots > 0)((void) sizeof ((resident_slots > 0) ? 1 : 0), __extension__
({ if (resident_slots > 0) ; else __assert_fail ("resident_slots > 0"
, "ccv_cnnp_model_addons.c", 5860, __extension__ __PRETTY_FUNCTION__
); }));
5861 assert(routing_width > 0)((void) sizeof ((routing_width > 0) ? 1 : 0), __extension__
({ if (routing_width > 0) ; else __assert_fail ("routing_width > 0"
, "ccv_cnnp_model_addons.c", 5861, __extension__ __PRETTY_FUNCTION__
); }));
5862 ccv_cnnp_model_moe_weights_streaming_t* const model = (ccv_cnnp_model_moe_weights_streaming_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_moe_weights_streaming_t));
5863 model->super.isa = &ccv_cnnp_moe_weights_streaming_isa;
5864 model->super.input_size = 6;
5865 model->super.outputs = model->outputs;
5866 model->super.output_size = 6;
5867 model->resident_slots = resident_slots;
5868 model->routing_width = routing_width;
5869 ccv_cnnp_model_copy_name(&model->super, name);
5870 return (ccv_cnnp_model_t*)model;
5871}
5872
5873static ccv_cnnp_model_t* _ccv_cnnp_moe_weights_streaming_copy(const ccv_cnnp_model_t* const super, void* const context)
5874{
5875 const ccv_cnnp_model_moe_weights_streaming_t* const self = (const ccv_cnnp_model_moe_weights_streaming_t*)super;
5876 return ccv_cnnp_moe_weights_streaming(self->resident_slots, self->routing_width, self->super.name);
5877}