Bug Summary

File:nnc/ccv_cnnp_model_addons.c
Warning:line 1189, column 2
The left operand of '%' is a garbage value

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/2026-09-23-222944-215595-1 -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;
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);
1
Assuming the condition is false
→
2
←
Taking false branch
→
3
←
Loop condition is false. Exiting loop
→
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__); }));
4
←
Assuming 'input_size' is equal to 1
→
5
←
Taking true branch
→
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__
); }));
6
←
Assuming 'output_size' is equal to 1
→
7
←
Taking true branch
→
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)
8
←
Assuming field 'format' is 0
→
9
←
Taking false branch
→
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;
10
←
'c' declared without an initial value
→
1174 switch (params.format)
11
←
'Default' branch taken. Execution continues on line 1189
→
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__
); }));
12
←
The left operand of '%' is a garbage value
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} ccv_cnnp_model_tanh_t;
1742
1743static 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)
1744{
1745 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);
1746 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"
, 1746, __extension__ __PRETTY_FUNCTION__); }));
1747 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", 1747, __extension__ __PRETTY_FUNCTION__
); }));
1748 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1749 ccv_nnc_tensor_param_t output_params;
1750 const ccv_nnc_cmd_t tanh = CMD_TANH_FORWARD()ccv_nnc_cmd(CCV_NNC_TANH_FORWARD, 0, ccv_nnc_cmd_auto, 0);
1751 ccv_nnc_hint_tensor_auto(tanh, (ccv_nnc_tensor_param_t []){
1752 params,
1753 }, 1, ccv_nnc_no_hint, &output_params, 1);
1754 const ccv_nnc_tensor_symbol_t tanh_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1755 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");
1756 outputs[0] = tanh_output;
1757}
1758
1759static ccv_cnnp_model_t* _ccv_cnnp_tanh_copy(const ccv_cnnp_model_t* const self, void* const context);
1760
1761static const ccv_cnnp_model_vtab_t ccv_cnnp_tanh_isa = {
1762 .build = _ccv_cnnp_tanh_build,
1763 .copy = _ccv_cnnp_tanh_copy,
1764};
1765
1766ccv_cnnp_model_t* ccv_cnnp_tanh(const char* const name)
1767{
1768 ccv_cnnp_model_tanh_t* const model_tanh = (ccv_cnnp_model_tanh_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_tanh_t));
1769 model_tanh->super.isa = &ccv_cnnp_tanh_isa;
1770 model_tanh->super.input_size = 1;
1771 model_tanh->super.outputs = &model_tanh->output;
1772 model_tanh->super.output_size = 1;
1773 ccv_cnnp_model_copy_name(&model_tanh->super, name);
1774 return (ccv_cnnp_model_t*)model_tanh;
1775}
1776
1777static ccv_cnnp_model_t* _ccv_cnnp_tanh_copy(const ccv_cnnp_model_t* const self, void* const context)
1778{
1779 return ccv_cnnp_tanh(self->name);
1780}
1781
1782// MARK - Exp Layer
1783
1784typedef struct {
1785 ccv_cnnp_model_t super;
1786 ccv_nnc_tensor_symbol_t output;
1787} ccv_cnnp_model_exp_t;
1788
1789static 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)
1790{
1791 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);
1792 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"
, 1792, __extension__ __PRETTY_FUNCTION__); }));
1793 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", 1793, __extension__ __PRETTY_FUNCTION__
); }));
1794 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1795 ccv_nnc_tensor_param_t output_params;
1796 const ccv_nnc_cmd_t exp = CMD_EWEXP_FORWARD()ccv_nnc_cmd(CCV_NNC_EWEXP_FORWARD, 0, ccv_nnc_cmd_auto, 0);
1797 ccv_nnc_hint_tensor_auto(exp, (ccv_nnc_tensor_param_t []){
1798 params,
1799 }, 1, ccv_nnc_no_hint, &output_params, 1);
1800 const ccv_nnc_tensor_symbol_t exp_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1801 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");
1802 outputs[0] = exp_output;
1803}
1804
1805static ccv_cnnp_model_t* _ccv_cnnp_exp_copy(const ccv_cnnp_model_t* const self, void* const context);
1806
1807static const ccv_cnnp_model_vtab_t ccv_cnnp_exp_isa = {
1808 .build = _ccv_cnnp_exp_build,
1809 .copy = _ccv_cnnp_exp_copy,
1810};
1811
1812ccv_cnnp_model_t* ccv_cnnp_exp(const char* const name)
1813{
1814 ccv_cnnp_model_exp_t* const model_exp = (ccv_cnnp_model_exp_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_exp_t));
1815 model_exp->super.isa = &ccv_cnnp_exp_isa;
1816 model_exp->super.input_size = 1;
1817 model_exp->super.outputs = &model_exp->output;
1818 model_exp->super.output_size = 1;
1819 ccv_cnnp_model_copy_name(&model_exp->super, name);
1820 return (ccv_cnnp_model_t*)model_exp;
1821}
1822
1823static ccv_cnnp_model_t* _ccv_cnnp_exp_copy(const ccv_cnnp_model_t* const self, void* const context)
1824{
1825 return ccv_cnnp_exp(self->name);
1826}
1827
1828// MARK - Softplus Layer
1829
1830typedef struct {
1831 ccv_cnnp_model_t super;
1832 ccv_nnc_tensor_symbol_t output;
1833} ccv_cnnp_model_softplus_t;
1834
1835static 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)
1836{
1837 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);
1838 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"
, 1838, __extension__ __PRETTY_FUNCTION__); }));
1839 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", 1839, __extension__ __PRETTY_FUNCTION__
); }));
1840 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1841 ccv_nnc_tensor_param_t output_params;
1842 const ccv_nnc_cmd_t softplus = CMD_EWSOFTPLUS_FORWARD()ccv_nnc_cmd(CCV_NNC_EWSOFTPLUS_FORWARD, 0, ccv_nnc_cmd_auto, 0
);
1843 ccv_nnc_hint_tensor_auto(softplus, (ccv_nnc_tensor_param_t []){
1844 params,
1845 }, 1, ccv_nnc_no_hint, &output_params, 1);
1846 const ccv_nnc_tensor_symbol_t softplus_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1847 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");
1848 outputs[0] = softplus_output;
1849}
1850
1851static ccv_cnnp_model_t* _ccv_cnnp_softplus_copy(const ccv_cnnp_model_t* const self, void* const context);
1852
1853static const ccv_cnnp_model_vtab_t ccv_cnnp_softplus_isa = {
1854 .build = _ccv_cnnp_softplus_build,
1855 .copy = _ccv_cnnp_softplus_copy,
1856};
1857
1858ccv_cnnp_model_t* ccv_cnnp_softplus(const char* const name)
1859{
1860 ccv_cnnp_model_softplus_t* const model_softplus = (ccv_cnnp_model_softplus_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_softplus_t));
1861 model_softplus->super.isa = &ccv_cnnp_softplus_isa;
1862 model_softplus->super.input_size = 1;
1863 model_softplus->super.outputs = &model_softplus->output;
1864 model_softplus->super.output_size = 1;
1865 ccv_cnnp_model_copy_name(&model_softplus->super, name);
1866 return (ccv_cnnp_model_t*)model_softplus;
1867}
1868
1869static ccv_cnnp_model_t* _ccv_cnnp_softplus_copy(const ccv_cnnp_model_t* const self, void* const context)
1870{
1871 return ccv_cnnp_softplus(self->name);
1872}
1873
1874// MARK - Swish Layer
1875
1876typedef struct {
1877 ccv_cnnp_model_t super;
1878 ccv_nnc_tensor_symbol_t output;
1879 float beta;
1880} ccv_cnnp_model_swish_t;
1881
1882static 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)
1883{
1884 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);
1885 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"
, 1885, __extension__ __PRETTY_FUNCTION__); }));
1886 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", 1886, __extension__ __PRETTY_FUNCTION__
); }));
1887 ccv_cnnp_model_swish_t* const self = (ccv_cnnp_model_swish_t*)super;
1888 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1889 ccv_nnc_tensor_param_t output_params;
1890 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);
1891 ccv_nnc_hint_tensor_auto(swish, (ccv_nnc_tensor_param_t []){
1892 params,
1893 }, 1, ccv_nnc_no_hint, &output_params, 1);
1894 const ccv_nnc_tensor_symbol_t swish_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1895 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");
1896 outputs[0] = swish_output;
1897}
1898
1899static ccv_cnnp_model_t* _ccv_cnnp_swish_copy(const ccv_cnnp_model_t* const self, void* const context);
1900
1901static const ccv_cnnp_model_vtab_t ccv_cnnp_swish_isa = {
1902 .build = _ccv_cnnp_swish_build,
1903 .copy = _ccv_cnnp_swish_copy,
1904};
1905
1906ccv_cnnp_model_t* ccv_cnnp_swish(const float beta, const char* const name)
1907{
1908 ccv_cnnp_model_swish_t* const model_swish = (ccv_cnnp_model_swish_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_swish_t));
1909 model_swish->super.isa = &ccv_cnnp_swish_isa;
1910 model_swish->super.input_size = 1;
1911 model_swish->super.outputs = &model_swish->output;
1912 model_swish->super.output_size = 1;
1913 model_swish->beta = beta;
1914 ccv_cnnp_model_copy_name(&model_swish->super, name);
1915 return (ccv_cnnp_model_t*)model_swish;
1916}
1917
1918static ccv_cnnp_model_t* _ccv_cnnp_swish_copy(const ccv_cnnp_model_t* const self, void* const context)
1919{
1920 const ccv_cnnp_model_swish_t* const swish = (const ccv_cnnp_model_swish_t*)self;
1921 return ccv_cnnp_swish(swish->beta, self->name);
1922}
1923
1924// MARK - Swish Mul Layer
1925
1926typedef struct {
1927 ccv_cnnp_model_t super;
1928 ccv_nnc_tensor_symbol_t output;
1929 float beta;
1930 float scale;
1931} ccv_cnnp_model_swish_mul_t;
1932
1933static 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)
1934{
1935 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);
1936 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"
, 1936, __extension__ __PRETTY_FUNCTION__); }));
1937 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", 1937, __extension__ __PRETTY_FUNCTION__
); }));
1938 const ccv_cnnp_model_swish_mul_t* const self = (const ccv_cnnp_model_swish_mul_t*)super;
1939 ccv_nnc_tensor_param_t input_params[2];
1940 int i;
1941 for (i = 0; i < 2; i++)
1942 input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
1943 ccv_nnc_tensor_param_t output_params;
1944 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);
1945 ccv_nnc_hint_tensor_auto(swish_mul, input_params, 2, ccv_nnc_no_hint, &output_params, 1);
1946 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1947 ccv_nnc_graph_exec_symbol_new(graph, swish_mul, inputs, input_size, outputs, output_size, "swish_mul");
1948}
1949
1950static ccv_cnnp_model_t* _ccv_cnnp_swish_mul_copy(const ccv_cnnp_model_t* const self, void* const context);
1951
1952static const ccv_cnnp_model_vtab_t ccv_cnnp_swish_mul_isa = {
1953 .build = _ccv_cnnp_swish_mul_build,
1954 .copy = _ccv_cnnp_swish_mul_copy,
1955};
1956
1957ccv_cnnp_model_t* ccv_cnnp_swish_mul(const float beta, const float scale, const char* const name)
1958{
1959 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));
1960 model_swish_mul->super.isa = &ccv_cnnp_swish_mul_isa;
1961 model_swish_mul->super.input_size = 2;
1962 model_swish_mul->super.outputs = &model_swish_mul->output;
1963 model_swish_mul->super.output_size = 1;
1964 model_swish_mul->beta = beta;
1965 model_swish_mul->scale = scale;
1966 ccv_cnnp_model_copy_name(&model_swish_mul->super, name);
1967 return (ccv_cnnp_model_t*)model_swish_mul;
1968}
1969
1970static ccv_cnnp_model_t* _ccv_cnnp_swish_mul_copy(const ccv_cnnp_model_t* const super, void* const context)
1971{
1972 const ccv_cnnp_model_swish_mul_t* const self = (const ccv_cnnp_model_swish_mul_t*)super;
1973 return ccv_cnnp_swish_mul(self->beta, self->scale, self->super.name);
1974}
1975
1976// MARK - GELU Layer
1977
1978typedef struct {
1979 ccv_cnnp_model_t super;
1980 ccv_nnc_tensor_symbol_t output;
1981 int tanh;
1982} ccv_cnnp_model_gelu_t;
1983
1984static 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)
1985{
1986 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);
1987 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"
, 1987, __extension__ __PRETTY_FUNCTION__); }));
1988 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", 1988, __extension__ __PRETTY_FUNCTION__
); }));
1989 ccv_cnnp_model_gelu_t* const self = (ccv_cnnp_model_gelu_t*)super;
1990 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1991 ccv_nnc_tensor_param_t output_params;
1992 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);
1993 ccv_nnc_hint_tensor_auto(gelu, (ccv_nnc_tensor_param_t []){
1994 params,
1995 }, 1, ccv_nnc_no_hint, &output_params, 1);
1996 const ccv_nnc_tensor_symbol_t gelu_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1997 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");
1998 outputs[0] = gelu_output;
1999}
2000
2001static ccv_cnnp_model_t* _ccv_cnnp_gelu_copy(const ccv_cnnp_model_t* const self, void* const context);
2002
2003static const ccv_cnnp_model_vtab_t ccv_cnnp_gelu_isa = {
2004 .build = _ccv_cnnp_gelu_build,
2005 .copy = _ccv_cnnp_gelu_copy,
2006};
2007
2008ccv_cnnp_model_t* ccv_cnnp_gelu(const int tanh, const char* const name)
2009{
2010 ccv_cnnp_model_gelu_t* const model_gelu = (ccv_cnnp_model_gelu_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_gelu_t));
2011 model_gelu->super.isa = &ccv_cnnp_gelu_isa;
2012 model_gelu->super.input_size = 1;
2013 model_gelu->super.outputs = &model_gelu->output;
2014 model_gelu->super.output_size = 1;
2015 model_gelu->tanh = tanh;
2016 ccv_cnnp_model_copy_name(&model_gelu->super, name);
2017 return (ccv_cnnp_model_t*)model_gelu;
2018}
2019
2020static ccv_cnnp_model_t* _ccv_cnnp_gelu_copy(const ccv_cnnp_model_t* const super, void* const context)
2021{
2022 ccv_cnnp_model_gelu_t* const self = (ccv_cnnp_model_gelu_t*)super;
2023 return ccv_cnnp_gelu(self->tanh, self->super.name);
2024}
2025
2026// MARK - Leaky ReLU Layer
2027
2028typedef struct {
2029 ccv_cnnp_model_t super;
2030 ccv_nnc_tensor_symbol_t output;
2031 float negative_slope;
2032} ccv_cnnp_model_leaky_relu_t;
2033
2034static 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)
2035{
2036 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);
2037 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"
, 2037, __extension__ __PRETTY_FUNCTION__); }));
2038 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", 2038, __extension__ __PRETTY_FUNCTION__
); }));
2039 ccv_cnnp_model_leaky_relu_t* const self = (ccv_cnnp_model_leaky_relu_t*)super;
2040 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2041 ccv_nnc_tensor_param_t output_params;
2042 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);
2043 ccv_nnc_hint_tensor_auto(leaky_relu, (ccv_nnc_tensor_param_t []){
2044 params,
2045 }, 1, ccv_nnc_no_hint, &output_params, 1);
2046 const ccv_nnc_tensor_symbol_t leaky_relu_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2047 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");
2048 outputs[0] = leaky_relu_output;
2049}
2050
2051static ccv_cnnp_model_t* _ccv_cnnp_leaky_relu_copy(const ccv_cnnp_model_t* const self, void* const context);
2052
2053static const ccv_cnnp_model_vtab_t ccv_cnnp_leaky_relu_isa = {
2054 .build = _ccv_cnnp_leaky_relu_build,
2055 .copy = _ccv_cnnp_leaky_relu_copy,
2056};
2057
2058ccv_cnnp_model_t* ccv_cnnp_leaky_relu(const float negative_slope, const char* const name)
2059{
2060 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));
2061 model_leaky_relu->super.isa = &ccv_cnnp_leaky_relu_isa;
2062 model_leaky_relu->super.input_size = 1;
2063 model_leaky_relu->super.outputs = &model_leaky_relu->output;
2064 model_leaky_relu->super.output_size = 1;
2065 model_leaky_relu->negative_slope = negative_slope;
2066 ccv_cnnp_model_copy_name(&model_leaky_relu->super, name);
2067 return (ccv_cnnp_model_t*)model_leaky_relu;
2068}
2069
2070static ccv_cnnp_model_t* _ccv_cnnp_leaky_relu_copy(const ccv_cnnp_model_t* const super, void* const context)
2071{
2072 ccv_cnnp_model_leaky_relu_t* const self = (ccv_cnnp_model_leaky_relu_t*)super;
2073 return ccv_cnnp_leaky_relu(self->negative_slope, self->super.name);
2074}
2075
2076// MARK - Softmax Layer
2077
2078typedef struct {
2079 ccv_cnnp_model_t super;
2080 ccv_nnc_tensor_symbol_t output;
2081} ccv_cnnp_model_softmax_t;
2082
2083static 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)
2084{
2085 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);
2086 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"
, 2086, __extension__ __PRETTY_FUNCTION__); }));
2087 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", 2087, __extension__ __PRETTY_FUNCTION__
); }));
2088 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2089 ccv_nnc_tensor_param_t output_params;
2090 const ccv_nnc_cmd_t softmax = CMD_SOFTMAX_FORWARD()ccv_nnc_cmd(CCV_NNC_SOFTMAX_FORWARD, 0, ccv_nnc_cmd_auto, 0);
2091 ccv_nnc_hint_tensor_auto(softmax, (ccv_nnc_tensor_param_t []){
2092 params,
2093 }, 1, ccv_nnc_no_hint, &output_params, 1);
2094 const ccv_nnc_tensor_symbol_t softmax_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2095 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");
2096 outputs[0] = softmax_output;
2097}
2098
2099static ccv_cnnp_model_t* _ccv_cnnp_softmax_copy(const ccv_cnnp_model_t* const self, void* const context);
2100
2101static const ccv_cnnp_model_vtab_t ccv_cnnp_softmax_isa = {
2102 .build = _ccv_cnnp_softmax_build,
2103 .copy = _ccv_cnnp_softmax_copy,
2104};
2105
2106ccv_cnnp_model_t* ccv_cnnp_softmax(const char* const name)
2107{
2108 ccv_cnnp_model_softmax_t* const model_softmax = (ccv_cnnp_model_softmax_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_softmax_t));
2109 model_softmax->super.isa = &ccv_cnnp_softmax_isa;
2110 model_softmax->super.input_size = 1;
2111 model_softmax->super.outputs = &model_softmax->output;
2112 model_softmax->super.output_size = 1;
2113 ccv_cnnp_model_copy_name(&model_softmax->super, name);
2114 return (ccv_cnnp_model_t*)model_softmax;
2115}
2116
2117static ccv_cnnp_model_t* _ccv_cnnp_softmax_copy(const ccv_cnnp_model_t* const self, void* const context)
2118{
2119 return ccv_cnnp_softmax(self->name);
2120}
2121
2122// MARK - Add Layer
2123
2124typedef struct {
2125 ccv_cnnp_model_t super;
2126 float p;
2127 float q;
2128 ccv_nnc_tensor_symbol_t output;
2129} ccv_cnnp_model_add_t;
2130
2131static 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)
2132{
2133 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);
2134 const ccv_cnnp_model_add_t* const self = (const ccv_cnnp_model_add_t*)super;
2135 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"
, 2135, __extension__ __PRETTY_FUNCTION__); }));
2136 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", 2136, __extension__ __PRETTY_FUNCTION__
); }));
2137 ccv_nnc_tensor_param_t input_params[2];
2138 int i;
2139 for (i = 0; i < 2; i++)
2140 input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
2141 ccv_nnc_tensor_param_t output_params;
2142 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);
2143 ccv_nnc_hint_tensor_auto(add, input_params, 2, ccv_nnc_no_hint, &output_params, 1);
2144 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2145 ccv_nnc_graph_exec_symbol_new(graph, add, inputs, input_size, outputs, output_size, "add");
2146}
2147
2148static ccv_cnnp_model_t* _ccv_cnnp_add_copy(const ccv_cnnp_model_t* const self, void* const context);
2149
2150static const ccv_cnnp_model_vtab_t ccv_cnnp_add_isa = {
2151 .build = _ccv_cnnp_add_build,
2152 .copy = _ccv_cnnp_add_copy,
2153};
2154
2155ccv_cnnp_model_t* ccv_cnnp_add(const float p, const float q, const char* const name)
2156{
2157 ccv_cnnp_model_add_t* const model_add = (ccv_cnnp_model_add_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_add_t));
2158 model_add->super.isa = &ccv_cnnp_add_isa;
2159 model_add->super.input_size = 2;
2160 model_add->super.outputs = &model_add->output;
2161 model_add->super.output_size = 1;
2162 model_add->p = p;
2163 model_add->q = q;
2164 ccv_cnnp_model_copy_name(&model_add->super, name);
2165 return (ccv_cnnp_model_t*)model_add;
2166}
2167
2168static ccv_cnnp_model_t* _ccv_cnnp_add_copy(const ccv_cnnp_model_t* const super, void* const context)
2169{
2170 const ccv_cnnp_model_add_t* const self = (const ccv_cnnp_model_add_t*)super;
2171 return ccv_cnnp_add(self->p, self->q, self->super.name);
2172}
2173
2174// MARK - Mul Layer
2175
2176typedef struct {
2177 ccv_cnnp_model_t super;
2178 ccv_nnc_tensor_symbol_t output;
2179 float p;
2180} ccv_cnnp_model_mul_t;
2181
2182static 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)
2183{
2184 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);
2185 const ccv_cnnp_model_mul_t* const self = (const ccv_cnnp_model_mul_t*)super;
2186 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"
, 2186, __extension__ __PRETTY_FUNCTION__); }));
2187 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", 2187, __extension__ __PRETTY_FUNCTION__
); }));
2188 ccv_nnc_tensor_param_t input_params[2];
2189 int i;
2190 for (i = 0; i < 2; i++)
2191 input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
2192 ccv_nnc_tensor_param_t output_params;
2193 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);
2194 ccv_nnc_hint_tensor_auto(mul, input_params, 2, ccv_nnc_no_hint, &output_params, 1);
2195 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2196 ccv_nnc_graph_exec_symbol_new(graph, mul, inputs, input_size, outputs, output_size, "mul");
2197}
2198
2199static ccv_cnnp_model_t* _ccv_cnnp_mul_copy(const ccv_cnnp_model_t* const self, void* const context);
2200
2201static const ccv_cnnp_model_vtab_t ccv_cnnp_mul_isa = {
2202 .build = _ccv_cnnp_mul_build,
2203 .copy = _ccv_cnnp_mul_copy,
2204};
2205
2206ccv_cnnp_model_t* ccv_cnnp_mul(const float p, const char* const name)
2207{
2208 ccv_cnnp_model_mul_t* const model_mul = (ccv_cnnp_model_mul_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_mul_t));
2209 model_mul->super.isa = &ccv_cnnp_mul_isa;
2210 model_mul->super.input_size = 2;
2211 model_mul->super.outputs = &model_mul->output;
2212 model_mul->super.output_size = 1;
2213 model_mul->p = p;
2214 ccv_cnnp_model_copy_name(&model_mul->super, name);
2215 return (ccv_cnnp_model_t*)model_mul;
2216}
2217
2218static ccv_cnnp_model_t* _ccv_cnnp_mul_copy(const ccv_cnnp_model_t* const super, void* const context)
2219{
2220 const ccv_cnnp_model_mul_t* const self = (const ccv_cnnp_model_mul_t*)super;
2221 return ccv_cnnp_mul(self->p, self->super.name);
2222}
2223
2224// MARK - Scalar Mul Layer
2225
2226typedef struct {
2227 ccv_cnnp_model_t super;
2228 ccv_nnc_tensor_symbol_t output;
2229 float a;
2230} ccv_cnnp_model_scalar_mul_t;
2231
2232static 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)
2233{
2234 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);
2235 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"
, 2235, __extension__ __PRETTY_FUNCTION__); }));
2236 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", 2236, __extension__ __PRETTY_FUNCTION__
); }));
2237 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2238 ccv_nnc_tensor_param_t output_params;
2239 ccv_cnnp_model_scalar_mul_t* const self = (ccv_cnnp_model_scalar_mul_t*)super;
2240 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);
2241 ccv_nnc_hint_tensor_auto(scalar_mul, (ccv_nnc_tensor_param_t []){
2242 params,
2243 }, 1, ccv_nnc_no_hint, &output_params, 1);
2244 const ccv_nnc_tensor_symbol_t scalar_mul_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2245 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");
2246 outputs[0] = scalar_mul_output;
2247}
2248
2249static ccv_cnnp_model_t* _ccv_cnnp_scalar_mul_copy(const ccv_cnnp_model_t* const super, void* const context);
2250
2251static const ccv_cnnp_model_vtab_t ccv_cnnp_scalar_mul_isa = {
2252 .build = _ccv_cnnp_scalar_mul_build,
2253 .copy = _ccv_cnnp_scalar_mul_copy,
2254};
2255
2256ccv_cnnp_model_t* ccv_cnnp_scalar_mul(const float a, const char* const name)
2257{
2258 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));
2259 model_scalar_mul->super.isa = &ccv_cnnp_scalar_mul_isa;
2260 model_scalar_mul->super.input_size = 1;
2261 model_scalar_mul->super.outputs = &model_scalar_mul->output;
2262 model_scalar_mul->super.output_size = 1;
2263 model_scalar_mul->a = a;
2264 ccv_cnnp_model_copy_name(&model_scalar_mul->super, name);
2265 return (ccv_cnnp_model_t*)model_scalar_mul;
2266}
2267
2268static ccv_cnnp_model_t* _ccv_cnnp_scalar_mul_copy(const ccv_cnnp_model_t* const super, void* const context)
2269{
2270 const ccv_cnnp_model_scalar_mul_t* const self = (const ccv_cnnp_model_scalar_mul_t*)super;
2271 return ccv_cnnp_scalar_mul(self->a, self->super.name);
2272}
2273
2274// MARK - Div Layer
2275
2276typedef struct {
2277 ccv_cnnp_model_t super;
2278 ccv_nnc_tensor_symbol_t output;
2279 int reciprocal;
2280} ccv_cnnp_model_div_t;
2281
2282static 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)
2283{
2284 const ccv_cnnp_model_div_t* const self = (const ccv_cnnp_model_div_t*)super;
2285 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);
2286 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", 2286, __extension__ __PRETTY_FUNCTION__
); }));
2287 ccv_nnc_tensor_param_t input_params[2];
2288 int i;
2289 ccv_nnc_tensor_param_t output_params;
2290 const ccv_nnc_cmd_t div = CMD_EWDIV_FORWARD()ccv_nnc_cmd(CCV_NNC_EWDIV_FORWARD, 0, ccv_nnc_cmd_auto, 0);
2291 if (self->reciprocal)
2292 {
2293 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"
, 2293, __extension__ __PRETTY_FUNCTION__); }));
2294 input_params[0] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2295 input_params[1] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2296 ccv_nnc_hint_tensor_auto(div, input_params, 2, ccv_nnc_no_hint, &output_params, 1);
2297 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2298 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");
2299 } else {
2300 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"
, 2300, __extension__ __PRETTY_FUNCTION__); }));
2301 for (i = 0; i < 2; i++)
2302 input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
2303 ccv_nnc_hint_tensor_auto(div, input_params, input_size, ccv_nnc_no_hint, &output_params, 1);
2304 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2305 ccv_nnc_graph_exec_symbol_new(graph, div, inputs, input_size, outputs, output_size, "div");
2306 }
2307}
2308
2309static ccv_cnnp_model_t* _ccv_cnnp_div_copy(const ccv_cnnp_model_t* const self, void* const context);
2310
2311static const ccv_cnnp_model_vtab_t ccv_cnnp_div_isa = {
2312 .build = _ccv_cnnp_div_build,
2313 .copy = _ccv_cnnp_div_copy,
2314};
2315
2316ccv_cnnp_model_t* ccv_cnnp_div(const int reciprocal, const char* const name)
2317{
2318 ccv_cnnp_model_div_t* const model_div = (ccv_cnnp_model_div_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_div_t));
2319 model_div->super.isa = &ccv_cnnp_div_isa;
2320 model_div->super.input_size = reciprocal ? 1 : 2;
2321 model_div->super.outputs = &model_div->output;
2322 model_div->super.output_size = 1;
2323 model_div->reciprocal = reciprocal;
2324 ccv_cnnp_model_copy_name(&model_div->super, name);
2325 return (ccv_cnnp_model_t*)model_div;
2326}
2327
2328static ccv_cnnp_model_t* _ccv_cnnp_div_copy(const ccv_cnnp_model_t* const super, void* const context)
2329{
2330 const ccv_cnnp_model_div_t* const self = (const ccv_cnnp_model_div_t*)super;
2331 return ccv_cnnp_div(self->reciprocal, self->super.name);
2332}
2333
2334// MARK - Sqrt Layer
2335
2336typedef struct {
2337 ccv_cnnp_model_t super;
2338 ccv_nnc_tensor_symbol_t output;
2339} ccv_cnnp_model_sqrt_t;
2340
2341static 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)
2342{
2343 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);
2344 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", 2344, __extension__ __PRETTY_FUNCTION__
); }));
2345 ccv_nnc_tensor_param_t input_params[1];
2346 ccv_nnc_tensor_param_t output_params;
2347 const ccv_nnc_cmd_t sqrt = CMD_EWSQRT_FORWARD()ccv_nnc_cmd(CCV_NNC_EWSQRT_FORWARD, 0, ccv_nnc_cmd_auto, 0);
2348 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"
, 2348, __extension__ __PRETTY_FUNCTION__); }));
2349 input_params[0] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2350 ccv_nnc_hint_tensor_auto(sqrt, input_params, 1, ccv_nnc_no_hint, &output_params, 1);
2351 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2352 ccv_nnc_graph_exec_symbol_new(graph, sqrt, inputs, 1, outputs, output_size, "sqrt");
2353}
2354
2355static ccv_cnnp_model_t* _ccv_cnnp_sqrt_copy(const ccv_cnnp_model_t* const self, void* const context);
2356
2357static const ccv_cnnp_model_vtab_t ccv_cnnp_sqrt_isa = {
2358 .build = _ccv_cnnp_sqrt_build,
2359 .copy = _ccv_cnnp_sqrt_copy,
2360};
2361
2362ccv_cnnp_model_t* ccv_cnnp_sqrt(const char* const name)
2363{
2364 ccv_cnnp_model_sqrt_t* const model_sqrt = (ccv_cnnp_model_sqrt_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_sqrt_t));
2365 model_sqrt->super.isa = &ccv_cnnp_sqrt_isa;
2366 model_sqrt->super.input_size = 1;
2367 model_sqrt->super.outputs = &model_sqrt->output;
2368 model_sqrt->super.output_size = 1;
2369 ccv_cnnp_model_copy_name(&model_sqrt->super, name);
2370 return (ccv_cnnp_model_t*)model_sqrt;
2371}
2372
2373static ccv_cnnp_model_t* _ccv_cnnp_sqrt_copy(const ccv_cnnp_model_t* const super, void* const context)
2374{
2375 const ccv_cnnp_model_sqrt_t* const self = (const ccv_cnnp_model_sqrt_t*)super;
2376 return ccv_cnnp_sqrt(self->super.name);
2377}
2378
2379// MARK - Signed Sqrt Layer
2380
2381typedef struct {
2382 ccv_cnnp_model_t super;
2383 ccv_nnc_tensor_symbol_t output;
2384 float minimum_magnitude;
2385} ccv_cnnp_model_signed_sqrt_t;
2386
2387static 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)
2388{
2389 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);
2390 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"
, 2390, __extension__ __PRETTY_FUNCTION__); }));
2391 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", 2391, __extension__ __PRETTY_FUNCTION__
); }));
2392 const ccv_cnnp_model_signed_sqrt_t* const self = (const ccv_cnnp_model_signed_sqrt_t*)super;
2393 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);
2394 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2395 ccv_nnc_tensor_param_t output_params;
2396 ccv_nnc_hint_tensor_auto(cmd, &params, 1, ccv_nnc_no_hint, &output_params, 1);
2397 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2398 ccv_nnc_graph_exec_symbol_new(graph, cmd, inputs, input_size, outputs, output_size, "signed_sqrt");
2399}
2400
2401static ccv_cnnp_model_t* _ccv_cnnp_signed_sqrt_copy(const ccv_cnnp_model_t* const self, void* const context);
2402
2403static const ccv_cnnp_model_vtab_t ccv_cnnp_signed_sqrt_isa = {
2404 .build = _ccv_cnnp_signed_sqrt_build,
2405 .copy = _ccv_cnnp_signed_sqrt_copy,
2406};
2407
2408ccv_cnnp_model_t* ccv_cnnp_signed_sqrt(const float minimum_magnitude, const char* const name)
2409{
2410 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", 2410, __extension__ __PRETTY_FUNCTION__
); }));
2411 ccv_cnnp_model_signed_sqrt_t* const model = (ccv_cnnp_model_signed_sqrt_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_signed_sqrt_t));
2412 model->super.isa = &ccv_cnnp_signed_sqrt_isa;
2413 model->super.input_size = 1;
2414 model->super.outputs = &model->output;
2415 model->super.output_size = 1;
2416 model->minimum_magnitude = minimum_magnitude;
2417 ccv_cnnp_model_copy_name(&model->super, name);
2418 return (ccv_cnnp_model_t*)model;
2419}
2420
2421static ccv_cnnp_model_t* _ccv_cnnp_signed_sqrt_copy(const ccv_cnnp_model_t* const super, void* const context)
2422{
2423 const ccv_cnnp_model_signed_sqrt_t* const self = (const ccv_cnnp_model_signed_sqrt_t*)super;
2424 return ccv_cnnp_signed_sqrt(self->minimum_magnitude, self->super.name);
2425}
2426
2427// MARK - Log Layer
2428
2429typedef struct {
2430 ccv_cnnp_model_t super;
2431 ccv_nnc_tensor_symbol_t output;
2432} ccv_cnnp_model_log_t;
2433
2434static 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)
2435{
2436 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);
2437 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", 2437, __extension__ __PRETTY_FUNCTION__
); }));
2438 ccv_nnc_tensor_param_t input_params[1];
2439 ccv_nnc_tensor_param_t output_params;
2440 const ccv_nnc_cmd_t log = CMD_EWLOG_FORWARD()ccv_nnc_cmd(CCV_NNC_EWLOG_FORWARD, 0, ccv_nnc_cmd_auto, 0);
2441 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"
, 2441, __extension__ __PRETTY_FUNCTION__); }));
2442 input_params[0] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2443 ccv_nnc_hint_tensor_auto(log, input_params, 1, ccv_nnc_no_hint, &output_params, 1);
2444 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2445 ccv_nnc_graph_exec_symbol_new(graph, log, inputs, 1, outputs, output_size, "log");
2446}
2447
2448static ccv_cnnp_model_t* _ccv_cnnp_log_copy(const ccv_cnnp_model_t* const self, void* const context);
2449
2450static const ccv_cnnp_model_vtab_t ccv_cnnp_log_isa = {
2451 .build = _ccv_cnnp_log_build,
2452 .copy = _ccv_cnnp_log_copy,
2453};
2454
2455ccv_cnnp_model_t* ccv_cnnp_log(const char* const name)
2456{
2457 ccv_cnnp_model_log_t* const model_log = (ccv_cnnp_model_log_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_log_t));
2458 model_log->super.isa = &ccv_cnnp_log_isa;
2459 model_log->super.input_size = 1;
2460 model_log->super.outputs = &model_log->output;
2461 model_log->super.output_size = 1;
2462 ccv_cnnp_model_copy_name(&model_log->super, name);
2463 return (ccv_cnnp_model_t*)model_log;
2464}
2465
2466static ccv_cnnp_model_t* _ccv_cnnp_log_copy(const ccv_cnnp_model_t* const super, void* const context)
2467{
2468 return ccv_cnnp_log(super->name);
2469}
2470
2471// MARK - Pow Layer
2472
2473typedef struct {
2474 ccv_cnnp_model_t super;
2475 ccv_nnc_tensor_symbol_t output;
2476 ccv_nnc_cmd_param_t params;
2477} ccv_cnnp_model_pow_t;
2478
2479static 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)
2480{
2481 ccv_cnnp_model_pow_t* const self = (ccv_cnnp_model_pow_t*)super;
2482 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);
2483 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"
, 2483, __extension__ __PRETTY_FUNCTION__); }));
2484 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", 2484, __extension__ __PRETTY_FUNCTION__
); }));
2485 ccv_nnc_tensor_param_t input_params[1];
2486 input_params[0] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2487 ccv_nnc_tensor_param_t output_params;
2488 const ccv_nnc_cmd_t pow = ccv_nnc_cmd(CCV_NNC_EWPOW_FORWARD, 0, self->params, 0);
2489 ccv_nnc_hint_tensor_auto(pow, input_params, 1, ccv_nnc_no_hint, &output_params, 1);
2490 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2491 ccv_nnc_graph_exec_symbol_new(graph, pow, inputs, input_size, outputs, output_size, "pow");
2492}
2493
2494static ccv_cnnp_model_t* _ccv_cnnp_pow_copy(const ccv_cnnp_model_t* const self, void* const context);
2495
2496static const ccv_cnnp_model_vtab_t ccv_cnnp_pow_isa = {
2497 .build = _ccv_cnnp_pow_build,
2498 .copy = _ccv_cnnp_pow_copy,
2499};
2500
2501ccv_cnnp_model_t* ccv_cnnp_pow(const float exponent, const char* const name)
2502{
2503 ccv_cnnp_model_pow_t* const model_pow = (ccv_cnnp_model_pow_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_pow_t));
2504 model_pow->super.isa = &ccv_cnnp_pow_isa;
2505 model_pow->super.input_size = 1;
2506 model_pow->super.outputs = &model_pow->output;
2507 model_pow->super.output_size = 1;
2508 model_pow->params = (ccv_nnc_cmd_param_t){
2509 .size = {
2510 .dim = { 1, 1, 1 }
2511 },
2512 .pow = {
2513 .exponent = exponent,
2514 },
2515 };
2516 ccv_cnnp_model_copy_name(&model_pow->super, name);
2517 return (ccv_cnnp_model_t*)model_pow;
2518}
2519
2520static ccv_cnnp_model_t* _ccv_cnnp_pow_copy(const ccv_cnnp_model_t* const super, void* const context)
2521{
2522 const ccv_cnnp_model_pow_t* const self = (const ccv_cnnp_model_pow_t*)super;
2523 return ccv_cnnp_pow(self->params.pow.exponent, super->name);
2524}
2525
2526// MARK - Sin Layer
2527
2528typedef struct {
2529 ccv_cnnp_model_t super;
2530 ccv_nnc_tensor_symbol_t output;
2531} ccv_cnnp_model_sin_t;
2532
2533static 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)
2534{
2535 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);
2536 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", 2536, __extension__ __PRETTY_FUNCTION__
); }));
2537 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"
, 2537, __extension__ __PRETTY_FUNCTION__); }));
2538 ccv_nnc_tensor_param_t input_params[1];
2539 ccv_nnc_tensor_param_t output_params;
2540 const ccv_nnc_cmd_t sin = CMD_EWSIN_FORWARD()ccv_nnc_cmd(CCV_NNC_EWSIN_FORWARD, 0, ccv_nnc_cmd_auto, 0);
2541 input_params[0] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2542 ccv_nnc_hint_tensor_auto(sin, input_params, 1, ccv_nnc_no_hint, &output_params, 1);
2543 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2544 ccv_nnc_graph_exec_symbol_new(graph, sin, inputs, 1, outputs, output_size, "sin");
2545}
2546
2547static ccv_cnnp_model_t* _ccv_cnnp_sin_copy(const ccv_cnnp_model_t* const self, void* const context);
2548
2549static const ccv_cnnp_model_vtab_t ccv_cnnp_sin_isa = {
2550 .build = _ccv_cnnp_sin_build,
2551 .copy = _ccv_cnnp_sin_copy,
2552};
2553
2554ccv_cnnp_model_t* ccv_cnnp_sin(const char* const name)
2555{
2556 ccv_cnnp_model_sin_t* const model_sin = (ccv_cnnp_model_sin_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_sin_t));
2557 model_sin->super.isa = &ccv_cnnp_sin_isa;
2558 model_sin->super.input_size = 1;
2559 model_sin->super.outputs = &model_sin->output;
2560 model_sin->super.output_size = 1;
2561 ccv_cnnp_model_copy_name(&model_sin->super, name);
2562 return (ccv_cnnp_model_t*)model_sin;
2563}
2564
2565static ccv_cnnp_model_t* _ccv_cnnp_sin_copy(const ccv_cnnp_model_t* const super, void* const context)
2566{
2567 return ccv_cnnp_sin(super->name);
2568}
2569
2570// MARK - Cos Layer
2571
2572typedef struct {
2573 ccv_cnnp_model_t super;
2574 ccv_nnc_tensor_symbol_t output;
2575} ccv_cnnp_model_cos_t;
2576
2577static 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)
2578{
2579 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);
2580 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", 2580, __extension__ __PRETTY_FUNCTION__
); }));
2581 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"
, 2581, __extension__ __PRETTY_FUNCTION__); }));
2582 ccv_nnc_tensor_param_t input_params[1];
2583 ccv_nnc_tensor_param_t output_params;
2584 const ccv_nnc_cmd_t cos = CMD_EWCOS_FORWARD()ccv_nnc_cmd(CCV_NNC_EWCOS_FORWARD, 0, ccv_nnc_cmd_auto, 0);
2585 input_params[0] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2586 ccv_nnc_hint_tensor_auto(cos, input_params, 1, ccv_nnc_no_hint, &output_params, 1);
2587 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2588 ccv_nnc_graph_exec_symbol_new(graph, cos, inputs, 1, outputs, output_size, "cos");
2589}
2590
2591static ccv_cnnp_model_t* _ccv_cnnp_cos_copy(const ccv_cnnp_model_t* const self, void* const context);
2592
2593static const ccv_cnnp_model_vtab_t ccv_cnnp_cos_isa = {
2594 .build = _ccv_cnnp_cos_build,
2595 .copy = _ccv_cnnp_cos_copy,
2596};
2597
2598ccv_cnnp_model_t* ccv_cnnp_cos(const char* const name)
2599{
2600 ccv_cnnp_model_cos_t* const model_cos = (ccv_cnnp_model_cos_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_cos_t));
2601 model_cos->super.isa = &ccv_cnnp_cos_isa;
2602 model_cos->super.input_size = 1;
2603 model_cos->super.outputs = &model_cos->output;
2604 model_cos->super.output_size = 1;
2605 ccv_cnnp_model_copy_name(&model_cos->super, name);
2606 return (ccv_cnnp_model_t*)model_cos;
2607}
2608
2609static ccv_cnnp_model_t* _ccv_cnnp_cos_copy(const ccv_cnnp_model_t* const super, void* const context)
2610{
2611 return ccv_cnnp_cos(super->name);
2612}
2613
2614// MARK - Rotate Half Layer
2615
2616typedef struct {
2617 ccv_cnnp_model_t super;
2618 ccv_nnc_tensor_symbol_t output;
2619} ccv_cnnp_model_rotate_half_t;
2620
2621static 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)
2622{
2623 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);
2624 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"
, 2624, __extension__ __PRETTY_FUNCTION__); }));
2625 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", 2625, __extension__ __PRETTY_FUNCTION__
); }));
2626 ccv_nnc_tensor_param_t input_params[1];
2627 ccv_nnc_tensor_param_t output_params;
2628 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);
2629 input_params[0] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2630 ccv_nnc_hint_tensor_auto(rotate_half, input_params, 1, ccv_nnc_no_hint, &output_params, 1);
2631 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2632 ccv_nnc_graph_exec_symbol_new(graph, rotate_half, inputs, 1, outputs, output_size, "rotate_half");
2633}
2634
2635static ccv_cnnp_model_t* _ccv_cnnp_rotate_half_copy(const ccv_cnnp_model_t* const self, void* const context);
2636
2637static const ccv_cnnp_model_vtab_t ccv_cnnp_rotate_half_isa = {
2638 .build = _ccv_cnnp_rotate_half_build,
2639 .copy = _ccv_cnnp_rotate_half_copy,
2640};
2641
2642ccv_cnnp_model_t* ccv_cnnp_rotate_half(const char* const name)
2643{
2644 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));
2645 model_rotate_half->super.isa = &ccv_cnnp_rotate_half_isa;
2646 model_rotate_half->super.input_size = 1;
2647 model_rotate_half->super.outputs = &model_rotate_half->output;
2648 model_rotate_half->super.output_size = 1;
2649 ccv_cnnp_model_copy_name(&model_rotate_half->super, name);
2650 return (ccv_cnnp_model_t*)model_rotate_half;
2651}
2652
2653static ccv_cnnp_model_t* _ccv_cnnp_rotate_half_copy(const ccv_cnnp_model_t* const super, void* const context)
2654{
2655 return ccv_cnnp_rotate_half(super->name);
2656}
2657
2658// MARK - Walsh-Hadamard Transform Layer
2659
2660typedef struct {
2661 ccv_cnnp_model_t super;
2662 ccv_nnc_tensor_symbol_t output;
2663 float scale;
2664} ccv_cnnp_model_walsh_hadamard_transform_t;
2665
2666static 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)
2667{
2668 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);
2669 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"
, 2669, __extension__ __PRETTY_FUNCTION__); }));
2670 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", 2670, __extension__ __PRETTY_FUNCTION__
); }));
2671 const ccv_cnnp_model_walsh_hadamard_transform_t* const self = (const ccv_cnnp_model_walsh_hadamard_transform_t*)super;
2672 ccv_nnc_tensor_param_t input_params[1];
2673 ccv_nnc_tensor_param_t output_params;
2674 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);
2675 input_params[0] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2676 ccv_nnc_hint_tensor_auto(walsh_hadamard_transform, input_params, 1, ccv_nnc_no_hint, &output_params, 1);
2677 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2678 ccv_nnc_graph_exec_symbol_new(graph, walsh_hadamard_transform, inputs, 1, outputs, output_size, "walsh_hadamard_transform");
2679}
2680
2681static ccv_cnnp_model_t* _ccv_cnnp_walsh_hadamard_transform_copy(const ccv_cnnp_model_t* const self, void* const context);
2682
2683static const ccv_cnnp_model_vtab_t ccv_cnnp_walsh_hadamard_transform_isa = {
2684 .build = _ccv_cnnp_walsh_hadamard_transform_build,
2685 .copy = _ccv_cnnp_walsh_hadamard_transform_copy,
2686};
2687
2688ccv_cnnp_model_t* ccv_cnnp_walsh_hadamard_transform(const float scale, const char* const name)
2689{
2690 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));
2691 model_walsh_hadamard_transform->super.isa = &ccv_cnnp_walsh_hadamard_transform_isa;
2692 model_walsh_hadamard_transform->super.input_size = 1;
2693 model_walsh_hadamard_transform->super.outputs = &model_walsh_hadamard_transform->output;
2694 model_walsh_hadamard_transform->super.output_size = 1;
2695 model_walsh_hadamard_transform->scale = scale;
2696 ccv_cnnp_model_copy_name(&model_walsh_hadamard_transform->super, name);
2697 return (ccv_cnnp_model_t*)model_walsh_hadamard_transform;
2698}
2699
2700static ccv_cnnp_model_t* _ccv_cnnp_walsh_hadamard_transform_copy(const ccv_cnnp_model_t* const super, void* const context)
2701{
2702 const ccv_cnnp_model_walsh_hadamard_transform_t* const self = (const ccv_cnnp_model_walsh_hadamard_transform_t*)super;
2703 return ccv_cnnp_walsh_hadamard_transform(self->scale, self->super.name);
2704}
2705
2706// MARK - Hyper Connection Layer
2707
2708typedef struct {
2709 ccv_cnnp_model_t super;
2710 ccv_nnc_tensor_symbol_t outputs[3];
2711 int count;
2712 int sinkhorn_iterations;
2713 float epsilon;
2714 int operation;
2715} ccv_cnnp_model_hyper_connection_t;
2716
2717static 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)
2718{
2719 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);
2720 const ccv_cnnp_model_hyper_connection_t* const self = (const ccv_cnnp_model_hyper_connection_t*)super;
2721 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", 2721, __extension__ __PRETTY_FUNCTION__
); }));
2722 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", 2722, __extension__ __PRETTY_FUNCTION__
); }));
2723 ccv_nnc_tensor_param_t input_params[4];
2724 int i;
2725 for (i = 0; i < input_size; i++)
2726 input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
2727 ccv_nnc_tensor_param_t output_params[3];
2728 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);
2729 ccv_nnc_hint_tensor_auto(hyper_connection, input_params, input_size, ccv_nnc_no_hint, output_params, output_size);
2730 for (i = 0; i < output_size; i++)
2731 outputs[i] = ccv_nnc_tensor_symbol_new(graph, output_params[i], 0);
2732 ccv_nnc_graph_exec_symbol_new(graph, hyper_connection, inputs, input_size, outputs, output_size, "hyper_connection");
2733}
2734
2735static ccv_cnnp_model_t* _ccv_cnnp_hyper_connection_copy(const ccv_cnnp_model_t* const self, void* const context);
2736
2737static const ccv_cnnp_model_vtab_t ccv_cnnp_hyper_connection_isa = {
2738 .build = _ccv_cnnp_hyper_connection_build,
2739 .copy = _ccv_cnnp_hyper_connection_copy,
2740};
2741
2742ccv_cnnp_model_t* ccv_cnnp_hyper_connection(const int count, const int sinkhorn_iterations, const float epsilon, const int operation, const char* const name)
2743{
2744 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", 2744, __extension__ __PRETTY_FUNCTION__
); }));
2745 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", 2745, __extension__ __PRETTY_FUNCTION__
); }));
2746 assert(epsilon >= 0)((void) sizeof ((epsilon >= 0) ? 1 : 0), __extension__ ({ if
(epsilon >= 0) ; else __assert_fail ("epsilon >= 0", "ccv_cnnp_model_addons.c"
, 2746, __extension__ __PRETTY_FUNCTION__); }));
2747 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", 2747, __extension__ __PRETTY_FUNCTION__
); }));
2748 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));
2749 model_hyper_connection->super.isa = &ccv_cnnp_hyper_connection_isa;
2750 model_hyper_connection->super.input_size = operation == 0 ? 3 : 4;
2751 model_hyper_connection->super.outputs = model_hyper_connection->outputs;
2752 model_hyper_connection->super.output_size = operation == 2 ? 1 : 3;
2753 model_hyper_connection->count = count;
2754 model_hyper_connection->sinkhorn_iterations = sinkhorn_iterations;
2755 model_hyper_connection->epsilon = epsilon;
2756 model_hyper_connection->operation = operation;
2757 ccv_cnnp_model_copy_name(&model_hyper_connection->super, name);
2758 return (ccv_cnnp_model_t*)model_hyper_connection;
2759}
2760
2761static ccv_cnnp_model_t* _ccv_cnnp_hyper_connection_copy(const ccv_cnnp_model_t* const super, void* const context)
2762{
2763 const ccv_cnnp_model_hyper_connection_t* const self = (const ccv_cnnp_model_hyper_connection_t*)super;
2764 return ccv_cnnp_hyper_connection(self->count, self->sinkhorn_iterations, self->epsilon, self->operation, self->super.name);
2765}
2766
2767// MARK - Gated Delta Layer
2768
2769typedef struct {
2770 ccv_cnnp_model_t super;
2771 ccv_nnc_tensor_symbol_t outputs[2];
2772 int log_decay;
2773 int state_checkpoint_count;
2774} ccv_cnnp_model_gated_delta_t;
2775
2776static 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)
2777{
2778 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);
2779 ccv_cnnp_model_gated_delta_t* const self = (ccv_cnnp_model_gated_delta_t*)super;
2780 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"
, 2780, __extension__ __PRETTY_FUNCTION__); }));
2781 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", 2781, __extension__ __PRETTY_FUNCTION__
); }));
2782 ccv_nnc_tensor_param_t input_params[6];
2783 int i;
2784 for (i = 0; i < 6; i++)
2785 input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
2786 ccv_nnc_tensor_param_t output_params[2];
2787 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);
2788 ccv_nnc_hint_tensor_auto(gated_delta, input_params, 6, ccv_nnc_no_hint, output_params, 2);
2789 for (i = 0; i < 2; i++)
2790 outputs[i] = ccv_nnc_tensor_symbol_new(graph, output_params[i], 0);
2791 ccv_nnc_graph_exec_symbol_new(graph, gated_delta, inputs, input_size, outputs, output_size, "gated_delta");
2792}
2793
2794static ccv_cnnp_model_t* _ccv_cnnp_gated_delta_copy(const ccv_cnnp_model_t* const self, void* const context);
2795
2796static const ccv_cnnp_model_vtab_t ccv_cnnp_gated_delta_isa = {
2797 .build = _ccv_cnnp_gated_delta_build,
2798 .copy = _ccv_cnnp_gated_delta_copy,
2799};
2800
2801ccv_cnnp_model_t* ccv_cnnp_gated_delta(const int log_decay, const int state_checkpoint_count, const char* const name)
2802{
2803 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));
2804 model_gated_delta->super.isa = &ccv_cnnp_gated_delta_isa;
2805 model_gated_delta->super.input_size = 6;
2806 model_gated_delta->super.outputs = model_gated_delta->outputs;
2807 model_gated_delta->super.output_size = 2;
2808 model_gated_delta->log_decay = log_decay;
2809 model_gated_delta->state_checkpoint_count = state_checkpoint_count;
2810 ccv_cnnp_model_copy_name(&model_gated_delta->super, name);
2811 return (ccv_cnnp_model_t*)model_gated_delta;
2812}
2813
2814static ccv_cnnp_model_t* _ccv_cnnp_gated_delta_copy(const ccv_cnnp_model_t* const super, void* const context)
2815{
2816 const ccv_cnnp_model_gated_delta_t* const self = (const ccv_cnnp_model_gated_delta_t*)super;
2817 return ccv_cnnp_gated_delta(self->log_decay, self->state_checkpoint_count, super->name);
2818}
2819
2820// MARK - Cmul Layer
2821
2822typedef struct {
2823 ccv_cnnp_model_t super;
2824 ccv_nnc_tensor_symbol_t output;
2825 int conjugate;
2826} ccv_cnnp_model_cmul_t;
2827
2828static 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)
2829{
2830 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);
2831 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"
, 2831, __extension__ __PRETTY_FUNCTION__); }));
2832 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", 2832, __extension__ __PRETTY_FUNCTION__
); }));
2833 ccv_nnc_tensor_param_t input_params[2];
2834 int i;
2835 for (i = 0; i < 2; i++)
2836 input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
2837 ccv_nnc_tensor_param_t output_params;
2838 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);
2839 mul.info.cmul.conjugate = ((ccv_cnnp_model_cmul_t*)super)->conjugate;
2840 ccv_nnc_hint_tensor_auto(mul, input_params, 2, ccv_nnc_no_hint, &output_params, 1);
2841 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2842 ccv_nnc_graph_exec_symbol_new(graph, mul, inputs, input_size, outputs, output_size, "cmul");
2843}
2844
2845static ccv_cnnp_model_t* _ccv_cnnp_cmul_copy(const ccv_cnnp_model_t* const self, void* const context);
2846
2847static const ccv_cnnp_model_vtab_t ccv_cnnp_cmul_isa = {
2848 .build = _ccv_cnnp_cmul_build,
2849 .copy = _ccv_cnnp_cmul_copy,
2850};
2851
2852ccv_cnnp_model_t* ccv_cnnp_cmul(const int conjugate, const char* const name)
2853{
2854 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"
, 2854, __extension__ __PRETTY_FUNCTION__); }));
2855 ccv_cnnp_model_cmul_t* const model_cmul = (ccv_cnnp_model_cmul_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_cmul_t));
2856 model_cmul->super.isa = &ccv_cnnp_cmul_isa;
2857 model_cmul->super.input_size = 2;
2858 model_cmul->super.outputs = &model_cmul->output;
2859 model_cmul->super.output_size = 1;
2860 model_cmul->conjugate = conjugate;
2861 ccv_cnnp_model_copy_name(&model_cmul->super, name);
2862 return (ccv_cnnp_model_t*)model_cmul;
2863}
2864
2865static ccv_cnnp_model_t* _ccv_cnnp_cmul_copy(const ccv_cnnp_model_t* const super, void* const context)
2866{
2867 const ccv_cnnp_model_cmul_t* const self = (const ccv_cnnp_model_cmul_t*)super;
2868 return ccv_cnnp_cmul(self->conjugate, super->name);
2869}
2870
2871// MARK - Transpose Layer
2872
2873typedef struct {
2874 ccv_cnnp_model_t super;
2875 ccv_nnc_tensor_symbol_t output;
2876 int transpose[2];
2877} ccv_cnnp_model_transpose_t;
2878
2879static 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)
2880{
2881 ccv_cnnp_model_transpose_t* const self = (ccv_cnnp_model_transpose_t*)super;
2882 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);
2883 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"
, 2883, __extension__ __PRETTY_FUNCTION__); }));
2884 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", 2884, __extension__ __PRETTY_FUNCTION__
); }));
2885 if (self->transpose[0] == self->transpose[1])
2886 {
2887 outputs[0] = inputs[0];
2888 return;
2889 }
2890 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2891 ccv_nnc_tensor_param_t output_params;
2892 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);
2893 ccv_nnc_hint_tensor_auto(transpose, (ccv_nnc_tensor_param_t []){
2894 params,
2895 }, 1, ccv_nnc_no_hint, &output_params, 1);
2896 const ccv_nnc_tensor_symbol_t transpose_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2897 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");
2898 outputs[0] = transpose_output;
2899}
2900
2901static ccv_cnnp_model_t* _ccv_cnnp_transpose_copy(const ccv_cnnp_model_t* const super, void* const context);
2902
2903static const ccv_cnnp_model_vtab_t ccv_cnnp_transpose_isa = {
2904 .build = _ccv_cnnp_transpose_build,
2905 .copy = _ccv_cnnp_transpose_copy,
2906};
2907
2908ccv_cnnp_model_t* ccv_cnnp_transpose(const int axis_a, const int axis_b, const char* const name)
2909{
2910 ccv_cnnp_model_transpose_t* const model_transpose = (ccv_cnnp_model_transpose_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_transpose_t));
2911 model_transpose->super.isa = &ccv_cnnp_transpose_isa;
2912 model_transpose->super.input_size = 1;
2913 model_transpose->super.outputs = &model_transpose->output;
2914 model_transpose->super.output_size = 1;
2915 model_transpose->transpose[0] = axis_a;
2916 model_transpose->transpose[1] = axis_b;
2917 ccv_cnnp_model_copy_name(&model_transpose->super, name);
2918 return (ccv_cnnp_model_t*)model_transpose;
2919}
2920
2921static ccv_cnnp_model_t* _ccv_cnnp_transpose_copy(const ccv_cnnp_model_t* const super, void* const context)
2922{
2923 const ccv_cnnp_model_transpose_t* const self = (const ccv_cnnp_model_transpose_t*)super;
2924 return ccv_cnnp_transpose(self->transpose[0], self->transpose[1], self->super.name);
2925}
2926
2927// MARK - Layer Norm Layer
2928
2929typedef struct {
2930 ccv_cnnp_model_t super;
2931 ccv_nnc_tensor_symbol_t output;
2932 ccv_nnc_tensor_symbol_t bias;
2933 ccv_nnc_tensor_symbol_t scale;
2934 ccv_nnc_cmd_param_t params;
2935} ccv_cnnp_model_layer_norm_t;
2936
2937static 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)
2938{
2939 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);
2940 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"
, 2940, __extension__ __PRETTY_FUNCTION__); }));
2941 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", 2941, __extension__ __PRETTY_FUNCTION__
); }));
2942 ccv_cnnp_model_layer_norm_t* const self = (ccv_cnnp_model_layer_norm_t*)super;
2943 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2944 ccv_nnc_tensor_param_t bias_params = params;
2945 const int nd = ccv_nnc_tensor_nd(params.dim);
2946 int i;
2947 for (i = 0; i < nd; i++)
2948 bias_params.dim[i] = 1;
2949 for (i = 0; i < self->params.lnorm.count; i++)
2950 bias_params.dim[self->params.lnorm.axis[i]] = params.dim[self->params.lnorm.axis[i]];
2951 if (self->params.lnorm.elementwise_affine)
2952 {
2953 // Both scale and bias are shared between if this model is reused.
2954 if (!self->scale.graph)
2955 self->scale = ccv_nnc_tensor_symbol_new(graph, bias_params, "scale");
2956 if (!self->bias.graph)
2957 self->bias = ccv_nnc_tensor_symbol_new(graph, bias_params, "bias");
2958 }
2959 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
};
2960 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
};
2961 const ccv_nnc_cmd_t layer_norm = ccv_nnc_cmd(CCV_NNC_LAYER_NORM_FORWARD, 0, self->params, 0);
2962 ccv_nnc_tensor_param_t output_params[3];
2963 if (self->params.lnorm.elementwise_affine)
2964 ccv_nnc_hint_tensor_auto(layer_norm, (ccv_nnc_tensor_param_t []){
2965 params,
2966 bias_params,
2967 bias_params,
2968 }, 3, ccv_nnc_no_hint, output_params, 3);
2969 else
2970 ccv_nnc_hint_tensor_auto(layer_norm, (ccv_nnc_tensor_param_t []){
2971 params,
2972 }, 1, ccv_nnc_no_hint, output_params, 3);
2973 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0);
2974 const ccv_nnc_tensor_symbol_t saved_mean = ccv_nnc_tensor_symbol_new(graph, output_params[1], "saved_mean");
2975 const ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(graph, output_params[2], "saved_inv_std");
2976 if (self->params.lnorm.elementwise_affine)
2977 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");
2978 else
2979 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");
2980 outputs[0] = output;
2981}
2982
2983static 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)
2984{
2985 ccv_cnnp_model_layer_norm_t* const self = (ccv_cnnp_model_layer_norm_t*)super;
2986 if (self->scale.graph)
2987 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);
2988 if (self->bias.graph)
2989 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);
2990}
2991
2992static 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)
2993{
2994 ccv_cnnp_model_layer_norm_t* const self = (ccv_cnnp_model_layer_norm_t*)super;
2995 if (self->scale.graph)
2996 add_to_array(parameters, self->scale, is_trainable);
2997 if (self->bias.graph)
2998 add_to_array(parameters, self->bias, is_trainable);
2999}
3000
3001static ccv_cnnp_model_t* _ccv_cnnp_layer_norm_copy(const ccv_cnnp_model_t* const super, void* const context);
3002
3003static const ccv_cnnp_model_vtab_t ccv_cnnp_layer_norm_isa = {
3004 .build = _ccv_cnnp_layer_norm_build,
3005 .init_states = _ccv_cnnp_layer_norm_init_states,
3006 .add_to_parameter = _ccv_cnnp_layer_norm_add_to_parameter,
3007 .copy = _ccv_cnnp_layer_norm_copy,
3008};
3009
3010ccv_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)
3011{
3012 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));
3013 model_layer_norm->super.isa = &ccv_cnnp_layer_norm_isa;
3014 model_layer_norm->super.input_size = 1;
3015 model_layer_norm->super.outputs = &model_layer_norm->output;
3016 model_layer_norm->super.output_size = 1;
3017 model_layer_norm->super.is_trainable = is_trainable;
3018 ccv_cnnp_model_copy_name(&model_layer_norm->super, name);
3019 model_layer_norm->scale.d = CCV_NNC_NO_TENSOR_SYMBOL;
3020 model_layer_norm->scale.graph = 0;
3021 model_layer_norm->bias.d = CCV_NNC_NO_TENSOR_SYMBOL;
3022 model_layer_norm->bias.graph = 0;
3023 model_layer_norm->params.lnorm.epsilon = epsilon;
3024 model_layer_norm->params.lnorm.scale = scale;
3025 model_layer_norm->params.lnorm.count = axis_count;
3026 model_layer_norm->params.lnorm.elementwise_affine = elementwise_affine;
3027 memcpy(model_layer_norm->params.lnorm.axis, axis, sizeof(int) * axis_count);
3028 return (ccv_cnnp_model_t*)model_layer_norm;
3029}
3030
3031static ccv_cnnp_model_t* _ccv_cnnp_layer_norm_copy(const ccv_cnnp_model_t* const super, void* const context)
3032{
3033 const ccv_cnnp_model_layer_norm_t* const self = (const ccv_cnnp_model_layer_norm_t*)super;
3034 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);
3035}
3036
3037// MARK - Group Norm Layer
3038
3039typedef struct {
3040 ccv_cnnp_model_t super;
3041 ccv_nnc_tensor_symbol_t output;
3042 ccv_nnc_tensor_symbol_t bias;
3043 ccv_nnc_tensor_symbol_t scale;
3044 ccv_nnc_cmd_param_t params;
3045} ccv_cnnp_model_group_norm_t;
3046
3047static 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)
3048{
3049 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);
3050 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"
, 3050, __extension__ __PRETTY_FUNCTION__); }));
3051 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", 3051, __extension__ __PRETTY_FUNCTION__
); }));
3052 ccv_cnnp_model_group_norm_t* const self = (ccv_cnnp_model_group_norm_t*)super;
3053 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3054 ccv_nnc_tensor_param_t bias_params = params;
3055 const int nd = ccv_nnc_tensor_nd(params.dim);
3056 int i;
3057 for (i = 0; i < nd; i++)
3058 bias_params.dim[i] = 1;
3059 bias_params.dim[self->params.gnorm.group_axis] = params.dim[self->params.gnorm.group_axis];
3060 if (self->params.gnorm.elementwise_affine)
3061 {
3062 // Both scale and bias are shared between if this model is reused.
3063 if (!self->scale.graph)
3064 self->scale = ccv_nnc_tensor_symbol_new(graph, bias_params, "scale");
3065 if (!self->bias.graph)
3066 self->bias = ccv_nnc_tensor_symbol_new(graph, bias_params, "bias");
3067 }
3068 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
};
3069 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
};
3070 const ccv_nnc_cmd_t group_norm = ccv_nnc_cmd(CCV_NNC_GROUP_NORM_FORWARD, 0, self->params, 0);
3071 ccv_nnc_tensor_param_t output_params[3];
3072 if (self->params.gnorm.elementwise_affine)
3073 ccv_nnc_hint_tensor_auto(group_norm, (ccv_nnc_tensor_param_t []){
3074 params,
3075 bias_params,
3076 bias_params,
3077 }, 3, ccv_nnc_no_hint, output_params, 3);
3078 else
3079 ccv_nnc_hint_tensor_auto(group_norm, (ccv_nnc_tensor_param_t []){
3080 params,
3081 }, 1, ccv_nnc_no_hint, output_params, 3);
3082 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0);
3083 const ccv_nnc_tensor_symbol_t saved_mean = ccv_nnc_tensor_symbol_new(graph, output_params[1], "saved_mean");
3084 const ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(graph, output_params[2], "saved_inv_std");
3085 if (self->params.gnorm.elementwise_affine)
3086 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");
3087 else
3088 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");
3089 outputs[0] = output;
3090}
3091
3092static 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)
3093{
3094 ccv_cnnp_model_group_norm_t* const self = (ccv_cnnp_model_group_norm_t*)super;
3095 if (self->scale.graph)
3096 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);
3097 if (self->bias.graph)
3098 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);
3099}
3100
3101static 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)
3102{
3103 ccv_cnnp_model_group_norm_t* const self = (ccv_cnnp_model_group_norm_t*)super;
3104 if (self->scale.graph)
3105 add_to_array(parameters, self->scale, is_trainable);
3106 if (self->bias.graph)
3107 add_to_array(parameters, self->bias, is_trainable);
3108}
3109
3110static ccv_cnnp_model_t* _ccv_cnnp_group_norm_copy(const ccv_cnnp_model_t* const super, void* const context);
3111
3112static const ccv_cnnp_model_vtab_t ccv_cnnp_group_norm_isa = {
3113 .build = _ccv_cnnp_group_norm_build,
3114 .init_states = _ccv_cnnp_group_norm_init_states,
3115 .add_to_parameter = _ccv_cnnp_group_norm_add_to_parameter,
3116 .copy = _ccv_cnnp_group_norm_copy,
3117};
3118
3119ccv_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)
3120{
3121 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));
3122 model_group_norm->super.isa = &ccv_cnnp_group_norm_isa;
3123 model_group_norm->super.input_size = 1;
3124 model_group_norm->super.outputs = &model_group_norm->output;
3125 model_group_norm->super.output_size = 1;
3126 model_group_norm->super.is_trainable = is_trainable;
3127 ccv_cnnp_model_copy_name(&model_group_norm->super, name);
3128 model_group_norm->scale.d = CCV_NNC_NO_TENSOR_SYMBOL;
3129 model_group_norm->scale.graph = 0;
3130 model_group_norm->bias.d = CCV_NNC_NO_TENSOR_SYMBOL;
3131 model_group_norm->bias.graph = 0;
3132 model_group_norm->params.gnorm.group_axis = group_axis;
3133 model_group_norm->params.gnorm.groups = groups;
3134 model_group_norm->params.gnorm.epsilon = epsilon;
3135 model_group_norm->params.gnorm.reduce_count = axis_count;
3136 model_group_norm->params.gnorm.elementwise_affine = elementwise_affine;
3137 memcpy(model_group_norm->params.gnorm.reduce_axis, reduce_axis, sizeof(int) * axis_count);
3138 return (ccv_cnnp_model_t*)model_group_norm;
3139}
3140
3141static ccv_cnnp_model_t* _ccv_cnnp_group_norm_copy(const ccv_cnnp_model_t* const super, void* const context)
3142{
3143 const ccv_cnnp_model_group_norm_t* const self = (const ccv_cnnp_model_group_norm_t*)super;
3144 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);
3145}
3146
3147// MARK - RMSNorm Layer
3148
3149typedef struct {
3150 ccv_cnnp_model_t super;
3151 ccv_nnc_tensor_symbol_t output;
3152 ccv_nnc_tensor_symbol_t scale;
3153 ccv_nnc_cmd_param_t params;
3154} ccv_cnnp_model_rmsnorm_t;
3155
3156static 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)
3157{
3158 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);
3159 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"
, 3159, __extension__ __PRETTY_FUNCTION__); }));
3160 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", 3160, __extension__ __PRETTY_FUNCTION__
); }));
3161 ccv_cnnp_model_rmsnorm_t* const self = (ccv_cnnp_model_rmsnorm_t*)super;
3162 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3163 ccv_nnc_tensor_param_t scale_params = params;
3164 const int nd = ccv_nnc_tensor_nd(params.dim);
3165 int i;
3166 for (i = 0; i < nd; i++)
3167 scale_params.dim[i] = 1;
3168 for (i = 0; i < self->params.rmsnorm.count; i++)
3169 scale_params.dim[self->params.rmsnorm.axis[i]] = params.dim[self->params.rmsnorm.axis[i]];
3170 // Both scale and bias are shared between if this model is reused.
3171 if (self->params.rmsnorm.elementwise_affine)
3172 {
3173 if (!self->scale.graph)
3174 self->scale = ccv_nnc_tensor_symbol_new(graph, scale_params, "scale");
3175 }
3176 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
};
3177 const ccv_nnc_cmd_t rmsnorm = ccv_nnc_cmd(CCV_NNC_RMSNORM_FORWARD, 0, self->params, 0);
3178 ccv_nnc_tensor_param_t output_params[2];
3179 if (self->params.rmsnorm.elementwise_affine)
3180 ccv_nnc_hint_tensor_auto(rmsnorm, (ccv_nnc_tensor_param_t []){
3181 params,
3182 scale_params,
3183 }, 2, ccv_nnc_no_hint, output_params, 2);
3184 else
3185 ccv_nnc_hint_tensor_auto(rmsnorm, (ccv_nnc_tensor_param_t []){
3186 params,
3187 }, 1, ccv_nnc_no_hint, output_params, 2);
3188 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0);
3189 const ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(graph, output_params[1], "saved_inv_std");
3190 if (self->params.rmsnorm.elementwise_affine)
3191 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");
3192 else
3193 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");
3194 outputs[0] = output;
3195}
3196
3197static 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)
3198{
3199 ccv_cnnp_model_rmsnorm_t* const self = (ccv_cnnp_model_rmsnorm_t*)super;
3200 if (self->scale.graph)
3201 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);
3202}
3203
3204static 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)
3205{
3206 ccv_cnnp_model_rmsnorm_t* const self = (ccv_cnnp_model_rmsnorm_t*)super;
3207 if (self->scale.graph)
3208 add_to_array(parameters, self->scale, is_trainable);
3209}
3210
3211static ccv_cnnp_model_t* _ccv_cnnp_rmsnorm_copy(const ccv_cnnp_model_t* const super, void* const context);
3212
3213static const ccv_cnnp_model_vtab_t ccv_cnnp_rmsnorm_isa = {
3214 .build = _ccv_cnnp_rmsnorm_build,
3215 .init_states = _ccv_cnnp_rmsnorm_init_states,
3216 .add_to_parameter = _ccv_cnnp_rmsnorm_add_to_parameter,
3217 .copy = _ccv_cnnp_rmsnorm_copy,
3218};
3219
3220ccv_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)
3221{
3222 ccv_cnnp_model_rmsnorm_t* const model_rmsnorm = (ccv_cnnp_model_rmsnorm_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_rmsnorm_t));
3223 model_rmsnorm->super.isa = &ccv_cnnp_rmsnorm_isa;
3224 model_rmsnorm->super.input_size = 1;
3225 model_rmsnorm->super.outputs = &model_rmsnorm->output;
3226 model_rmsnorm->super.output_size = 1;
3227 model_rmsnorm->super.is_trainable = is_trainable;
3228 ccv_cnnp_model_copy_name(&model_rmsnorm->super, name);
3229 model_rmsnorm->scale.d = CCV_NNC_NO_TENSOR_SYMBOL;
3230 model_rmsnorm->scale.graph = 0;
3231 model_rmsnorm->params.rmsnorm.epsilon = epsilon;
3232 model_rmsnorm->params.rmsnorm.scale = scale;
3233 model_rmsnorm->params.rmsnorm.count = axis_count;
3234 model_rmsnorm->params.rmsnorm.elementwise_affine = elementwise_affine;
3235 memcpy(model_rmsnorm->params.lnorm.axis, axis, sizeof(int) * axis_count);
3236 return (ccv_cnnp_model_t*)model_rmsnorm;
3237}
3238
3239static ccv_cnnp_model_t* _ccv_cnnp_rmsnorm_copy(const ccv_cnnp_model_t* const super, void* const context)
3240{
3241 const ccv_cnnp_model_rmsnorm_t* const self = (const ccv_cnnp_model_rmsnorm_t*)super;
3242 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);
3243}
3244
3245// MARK - RMSNorm CMul Layer
3246
3247typedef struct {
3248 ccv_cnnp_model_t super;
3249 ccv_nnc_tensor_symbol_t output;
3250 ccv_nnc_tensor_symbol_t scale;
3251 ccv_nnc_cmd_param_t params;
3252} ccv_cnnp_model_rmsnorm_cmul_t;
3253
3254static 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)
3255{
3256 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);
3257 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"
, 3257, __extension__ __PRETTY_FUNCTION__); }));
3258 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", 3258, __extension__ __PRETTY_FUNCTION__
); }));
3259 ccv_cnnp_model_rmsnorm_cmul_t* const self = (ccv_cnnp_model_rmsnorm_cmul_t*)super;
3260 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3261 const ccv_nnc_tensor_param_t rotation_params = ccv_nnc_tensor_symbol_params(graph, inputs[1]);
3262 ccv_nnc_tensor_param_t scale_params = params;
3263 const int nd = ccv_nnc_tensor_nd(params.dim);
3264 int i;
3265 for (i = 0; i < nd; i++)
3266 scale_params.dim[i] = 1;
3267 for (i = 0; i < self->params.rmsnorm_cmul.count; i++)
3268 scale_params.dim[self->params.rmsnorm_cmul.axis[i]] = params.dim[self->params.rmsnorm_cmul.axis[i]];
3269 if (self->params.rmsnorm_cmul.elementwise_affine)
3270 {
3271 if (!self->scale.graph)
3272 self->scale = ccv_nnc_tensor_symbol_new(graph, scale_params, "scale");
3273 }
3274 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
};
3275 const ccv_nnc_cmd_t rmsnorm_cmul = ccv_nnc_cmd(CCV_NNC_RMSNORM_CMUL_FORWARD, 0, self->params, 0);
3276 ccv_nnc_tensor_param_t output_params;
3277 if (self->params.rmsnorm_cmul.elementwise_affine)
3278 ccv_nnc_hint_tensor_auto(rmsnorm_cmul, (ccv_nnc_tensor_param_t []){
3279 params,
3280 rotation_params,
3281 scale_params,
3282 }, 3, ccv_nnc_no_hint, &output_params, 1);
3283 else
3284 ccv_nnc_hint_tensor_auto(rmsnorm_cmul, (ccv_nnc_tensor_param_t []){
3285 params,
3286 rotation_params,
3287 }, 2, ccv_nnc_no_hint, &output_params, 1);
3288 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3289 if (self->params.rmsnorm_cmul.elementwise_affine)
3290 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");
3291 else
3292 ccv_nnc_graph_exec_symbol_new(graph, rmsnorm_cmul, inputs, input_size, outputs, output_size, "rmsnorm_cmul");
3293}
3294
3295static 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)
3296{
3297 ccv_cnnp_model_rmsnorm_cmul_t* const self = (ccv_cnnp_model_rmsnorm_cmul_t*)super;
3298 if (self->scale.graph)
3299 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);
3300}
3301
3302static 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)
3303{
3304 ccv_cnnp_model_rmsnorm_cmul_t* const self = (ccv_cnnp_model_rmsnorm_cmul_t*)super;
3305 if (self->scale.graph)
3306 add_to_array(parameters, self->scale, is_trainable);
3307}
3308
3309static ccv_cnnp_model_t* _ccv_cnnp_rmsnorm_cmul_copy(const ccv_cnnp_model_t* const super, void* const context);
3310
3311static const ccv_cnnp_model_vtab_t ccv_cnnp_rmsnorm_cmul_isa = {
3312 .build = _ccv_cnnp_rmsnorm_cmul_build,
3313 .init_states = _ccv_cnnp_rmsnorm_cmul_init_states,
3314 .add_to_parameter = _ccv_cnnp_rmsnorm_cmul_add_to_parameter,
3315 .copy = _ccv_cnnp_rmsnorm_cmul_copy,
3316};
3317
3318ccv_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)
3319{
3320 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));
3321 model_rmsnorm_cmul->super.isa = &ccv_cnnp_rmsnorm_cmul_isa;
3322 model_rmsnorm_cmul->super.input_size = 2;
3323 model_rmsnorm_cmul->super.outputs = &model_rmsnorm_cmul->output;
3324 model_rmsnorm_cmul->super.output_size = 1;
3325 model_rmsnorm_cmul->super.is_trainable = is_trainable;
3326 ccv_cnnp_model_copy_name(&model_rmsnorm_cmul->super, name);
3327 model_rmsnorm_cmul->scale.d = CCV_NNC_NO_TENSOR_SYMBOL;
3328 model_rmsnorm_cmul->scale.graph = 0;
3329 model_rmsnorm_cmul->params.rmsnorm_cmul.epsilon = epsilon;
3330 model_rmsnorm_cmul->params.rmsnorm_cmul.count = axis_count;
3331 model_rmsnorm_cmul->params.rmsnorm_cmul.elementwise_affine = elementwise_affine;
3332 memcpy(model_rmsnorm_cmul->params.rmsnorm_cmul.axis, axis, sizeof(int) * axis_count);
3333 return (ccv_cnnp_model_t*)model_rmsnorm_cmul;
3334}
3335
3336static ccv_cnnp_model_t* _ccv_cnnp_rmsnorm_cmul_copy(const ccv_cnnp_model_t* const super, void* const context)
3337{
3338 const ccv_cnnp_model_rmsnorm_cmul_t* const self = (const ccv_cnnp_model_rmsnorm_cmul_t*)super;
3339 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);
3340}
3341
3342// MARK - RMSNorm Gated Layer
3343
3344typedef struct {
3345 ccv_cnnp_model_t super;
3346 ccv_nnc_tensor_symbol_t output;
3347 ccv_nnc_tensor_symbol_t scale;
3348 ccv_nnc_cmd_param_t params;
3349} ccv_cnnp_model_rmsnorm_gated_t;
3350
3351static 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)
3352{
3353 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);
3354 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"
, 3354, __extension__ __PRETTY_FUNCTION__); }));
3355 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", 3355, __extension__ __PRETTY_FUNCTION__
); }));
3356 ccv_cnnp_model_rmsnorm_gated_t* const self = (ccv_cnnp_model_rmsnorm_gated_t*)super;
3357 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3358 const ccv_nnc_tensor_param_t gate_params = ccv_nnc_tensor_symbol_params(graph, inputs[1]);
3359 ccv_nnc_tensor_param_t scale_params = params;
3360 const int nd = ccv_nnc_tensor_nd(params.dim);
3361 int i;
3362 for (i = 0; i < nd; i++)
3363 scale_params.dim[i] = 1;
3364 for (i = 0; i < self->params.rmsnorm_gated.count; i++)
3365 scale_params.dim[self->params.rmsnorm_gated.axis[i]] = params.dim[self->params.rmsnorm_gated.axis[i]];
3366 if (self->params.rmsnorm_gated.elementwise_affine)
3367 {
3368 if (!self->scale.graph)
3369 self->scale = ccv_nnc_tensor_symbol_new(graph, scale_params, "scale");
3370 }
3371 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
};
3372 const ccv_nnc_cmd_t rmsnorm_gated = ccv_nnc_cmd(CCV_NNC_RMSNORM_GATED_FORWARD, 0, self->params, 0);
3373 ccv_nnc_tensor_param_t output_params;
3374 if (self->params.rmsnorm_gated.elementwise_affine)
3375 ccv_nnc_hint_tensor_auto(rmsnorm_gated, (ccv_nnc_tensor_param_t []){
3376 params,
3377 gate_params,
3378 scale_params,
3379 }, 3, ccv_nnc_no_hint, &output_params, 1);
3380 else
3381 ccv_nnc_hint_tensor_auto(rmsnorm_gated, (ccv_nnc_tensor_param_t []){
3382 params,
3383 gate_params,
3384 }, 2, ccv_nnc_no_hint, &output_params, 1);
3385 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3386 if (self->params.rmsnorm_gated.elementwise_affine)
3387 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");
3388 else
3389 ccv_nnc_graph_exec_symbol_new(graph, rmsnorm_gated, inputs, input_size, outputs, output_size, "rmsnorm_gated");
3390}
3391
3392static 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)
3393{
3394 ccv_cnnp_model_rmsnorm_gated_t* const self = (ccv_cnnp_model_rmsnorm_gated_t*)super;
3395 if (self->scale.graph)
3396 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);
3397}
3398
3399static 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)
3400{
3401 ccv_cnnp_model_rmsnorm_gated_t* const self = (ccv_cnnp_model_rmsnorm_gated_t*)super;
3402 if (self->scale.graph)
3403 add_to_array(parameters, self->scale, is_trainable);
3404}
3405
3406static ccv_cnnp_model_t* _ccv_cnnp_rmsnorm_gated_copy(const ccv_cnnp_model_t* const super, void* const context);
3407
3408static const ccv_cnnp_model_vtab_t ccv_cnnp_rmsnorm_gated_isa = {
3409 .build = _ccv_cnnp_rmsnorm_gated_build,
3410 .init_states = _ccv_cnnp_rmsnorm_gated_init_states,
3411 .add_to_parameter = _ccv_cnnp_rmsnorm_gated_add_to_parameter,
3412 .copy = _ccv_cnnp_rmsnorm_gated_copy,
3413};
3414
3415ccv_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 is_trainable, const char* const name)
3416{
3417 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));
3418 model_rmsnorm_gated->super.isa = &ccv_cnnp_rmsnorm_gated_isa;
3419 model_rmsnorm_gated->super.input_size = 2;
3420 model_rmsnorm_gated->super.outputs = &model_rmsnorm_gated->output;
3421 model_rmsnorm_gated->super.output_size = 1;
3422 model_rmsnorm_gated->super.is_trainable = is_trainable;
3423 ccv_cnnp_model_copy_name(&model_rmsnorm_gated->super, name);
3424 model_rmsnorm_gated->scale.d = CCV_NNC_NO_TENSOR_SYMBOL;
3425 model_rmsnorm_gated->scale.graph = 0;
3426 model_rmsnorm_gated->params.rmsnorm_gated.epsilon = epsilon;
3427 model_rmsnorm_gated->params.rmsnorm_gated.count = axis_count;
3428 model_rmsnorm_gated->params.rmsnorm_gated.elementwise_affine = elementwise_affine;
3429 memcpy(model_rmsnorm_gated->params.rmsnorm_gated.axis, axis, sizeof(int) * axis_count);
3430 return (ccv_cnnp_model_t*)model_rmsnorm_gated;
3431}
3432
3433static ccv_cnnp_model_t* _ccv_cnnp_rmsnorm_gated_copy(const ccv_cnnp_model_t* const super, void* const context)
3434{
3435 const ccv_cnnp_model_rmsnorm_gated_t* const self = (const ccv_cnnp_model_rmsnorm_gated_t*)super;
3436 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->super.is_trainable, self->super.name);
3437}
3438
3439// MARK - Batched Matrix Mul Layer
3440
3441typedef struct {
3442 ccv_cnnp_model_t super;
3443 ccv_nnc_tensor_symbol_t output;
3444 int transpose_a[2];
3445 int transpose_b[2];
3446 int flags;
3447} ccv_cnnp_model_matmul_t;
3448
3449static 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)
3450{
3451 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);
3452 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"
, 3452, __extension__ __PRETTY_FUNCTION__); }));
3453 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", 3453, __extension__ __PRETTY_FUNCTION__
); }));
3454 ccv_cnnp_model_matmul_t* const self = (ccv_cnnp_model_matmul_t*)super;
3455 ccv_nnc_tensor_param_t a_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3456 ccv_nnc_tensor_param_t b_params = ccv_nnc_tensor_symbol_params(graph, inputs[1]);
3457 ccv_nnc_tensor_param_t output_params;
3458 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);
3459 matmul.info.blas.flags = self->flags;
3460 ccv_nnc_hint_tensor_auto(matmul, (ccv_nnc_tensor_param_t []){
3461 a_params,
3462 b_params,
3463 }, 2, ccv_nnc_no_hint, &output_params, 1);
3464 const ccv_nnc_tensor_symbol_t matmul_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3465 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");
3466 outputs[0] = matmul_output;
3467}
3468
3469static ccv_cnnp_model_t* _ccv_cnnp_matmul_copy(const ccv_cnnp_model_t* const super, void* const context);
3470
3471static const ccv_cnnp_model_vtab_t ccv_cnnp_matmul_isa = {
3472 .build = _ccv_cnnp_matmul_build,
3473 .copy = _ccv_cnnp_matmul_copy,
3474};
3475
3476ccv_cnnp_model_t* ccv_cnnp_matmul(const int transpose_a[2], const int transpose_b[2], const int flags, const char* const name)
3477{
3478 ccv_cnnp_model_matmul_t* const model_matmul = (ccv_cnnp_model_matmul_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_matmul_t));
3479 model_matmul->super.isa = &ccv_cnnp_matmul_isa;
3480 model_matmul->super.input_size = 2;
3481 model_matmul->super.outputs = &model_matmul->output;
3482 model_matmul->super.output_size = 1;
3483 model_matmul->transpose_a[0] = transpose_a[0];
3484 model_matmul->transpose_a[1] = transpose_a[1];
3485 model_matmul->transpose_b[0] = transpose_b[0];
3486 model_matmul->transpose_b[1] = transpose_b[1];
3487 model_matmul->flags = flags;
3488 ccv_cnnp_model_copy_name(&model_matmul->super, name);
3489 return (ccv_cnnp_model_t*)model_matmul;
3490}
3491
3492static ccv_cnnp_model_t* _ccv_cnnp_matmul_copy(const ccv_cnnp_model_t* const super, void* const context)
3493{
3494 const ccv_cnnp_model_matmul_t* const self = (const ccv_cnnp_model_matmul_t*)super;
3495 return ccv_cnnp_matmul(self->transpose_a, self->transpose_b, self->flags, self->super.name);
3496}
3497
3498// MARK - Dropout Layer
3499
3500typedef struct {
3501 ccv_cnnp_model_t super;
3502 ccv_nnc_tensor_symbol_t output;
3503 ccv_nnc_graph_exec_symbol_t dropout;
3504 float p;
3505 int entirety;
3506} ccv_cnnp_model_dropout_t;
3507
3508static 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)
3509{
3510 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);
3511 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"
, 3511, __extension__ __PRETTY_FUNCTION__); }));
3512 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", 3512, __extension__ __PRETTY_FUNCTION__
); }));
3513 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3514 ccv_nnc_tensor_param_t output_params[2];
3515 ccv_cnnp_model_dropout_t* const self = (ccv_cnnp_model_dropout_t*)super;
3516 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);
3517 ccv_nnc_hint_tensor_auto(dropout, (ccv_nnc_tensor_param_t []){
3518 params,
3519 }, 1, ccv_nnc_no_hint, output_params, 2);
3520 const ccv_nnc_tensor_symbol_t dropout_output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0);
3521 const ccv_nnc_tensor_symbol_t mask = ccv_nnc_tensor_symbol_new(graph, output_params[1], "mask");
3522 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");
3523 outputs[0] = dropout_output;
3524}
3525
3526static 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)
3527{
3528 ccv_cnnp_model_dropout_t* const self = (ccv_cnnp_model_dropout_t*)super;
3529 if (self->dropout.graph)
3530 {
3531 if (is_test)
3532 // During test, the dropout is not applied. Data transfer is perfect because if these are the same tensor, it will skip.
3533 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);
3534 else
3535 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);
3536 }
3537}
3538
3539static ccv_cnnp_model_t* _ccv_cnnp_dropout_copy(const ccv_cnnp_model_t* const super, void* const context);
3540
3541static const ccv_cnnp_model_vtab_t ccv_cnnp_dropout_isa = {
3542 .build = _ccv_cnnp_dropout_build,
3543 .set_is_test = _ccv_cnnp_dropout_set_is_test,
3544 .copy = _ccv_cnnp_dropout_copy,
3545};
3546
3547ccv_cnnp_model_t* ccv_cnnp_dropout(const float p, const int entirety, const char* const name)
3548{
3549 ccv_cnnp_model_dropout_t* const model_dropout = (ccv_cnnp_model_dropout_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_dropout_t));
3550 model_dropout->super.isa = &ccv_cnnp_dropout_isa;
3551 model_dropout->super.input_size = 1;
3552 model_dropout->super.outputs = &model_dropout->output;
3553 model_dropout->super.output_size = 1;
3554 model_dropout->p = p;
3555 model_dropout->entirety = entirety;
3556 ccv_cnnp_model_copy_name(&model_dropout->super, name);
3557 return (ccv_cnnp_model_t*)model_dropout;
3558}
3559
3560static ccv_cnnp_model_t* _ccv_cnnp_dropout_copy(const ccv_cnnp_model_t* const super, void* const context)
3561{
3562 const ccv_cnnp_model_dropout_t* const self = (const ccv_cnnp_model_dropout_t*)super;
3563 return ccv_cnnp_dropout(self->p, self->entirety, self->super.name);
3564}
3565
3566// MARK - Masked Fill Layer
3567
3568typedef struct {
3569 ccv_cnnp_model_t super;
3570 ccv_nnc_tensor_symbol_t output;
3571 float eq;
3572 float fill;
3573} ccv_cnnp_model_masked_fill_t;
3574
3575static 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)
3576{
3577 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);
3578 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"
, 3578, __extension__ __PRETTY_FUNCTION__); }));
3579 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", 3579, __extension__ __PRETTY_FUNCTION__
); }));
3580 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3581 ccv_cnnp_model_masked_fill_t* const self = (ccv_cnnp_model_masked_fill_t*)super;
3582 const ccv_nnc_tensor_symbol_t masked_fill_output = ccv_nnc_tensor_symbol_new(graph, params, 0);
3583 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");
3584 outputs[0] = masked_fill_output;
3585}
3586
3587static ccv_cnnp_model_t* _ccv_cnnp_masked_fill_copy(const ccv_cnnp_model_t* const super, void* const context);
3588
3589static const ccv_cnnp_model_vtab_t ccv_cnnp_masked_fill_isa = {
3590 .build = _ccv_cnnp_masked_fill_build,
3591 .copy = _ccv_cnnp_masked_fill_copy,
3592};
3593
3594ccv_cnnp_model_t* ccv_cnnp_masked_fill(const float eq, const float fill, const char* const name)
3595{
3596 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));
3597 model_masked_fill->super.isa = &ccv_cnnp_masked_fill_isa;
3598 model_masked_fill->super.input_size = 2;
3599 model_masked_fill->super.outputs = &model_masked_fill->output;
3600 model_masked_fill->super.output_size = 1;
3601 model_masked_fill->eq = eq;
3602 model_masked_fill->fill = fill;
3603 ccv_cnnp_model_copy_name(&model_masked_fill->super, name);
3604 return (ccv_cnnp_model_t*)model_masked_fill;
3605}
3606
3607static ccv_cnnp_model_t* _ccv_cnnp_masked_fill_copy(const ccv_cnnp_model_t* const super, void* const context)
3608{
3609 const ccv_cnnp_model_masked_fill_t* const self = (const ccv_cnnp_model_masked_fill_t*)super;
3610 return ccv_cnnp_masked_fill(self->eq, self->fill, self->super.name);
3611}
3612
3613// MARK - Index Select Layer
3614
3615typedef struct {
3616 ccv_cnnp_model_t super;
3617 ccv_nnc_tensor_symbol_t output;
3618} ccv_cnnp_model_index_select_t;
3619
3620static 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)
3621{
3622 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);
3623 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"
, 3623, __extension__ __PRETTY_FUNCTION__); }));
3624 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", 3624, __extension__ __PRETTY_FUNCTION__
); }));
3625 const ccv_nnc_tensor_param_t vocab_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3626 const ccv_nnc_tensor_param_t index_params = ccv_nnc_tensor_symbol_params(graph, inputs[1]);
3627 ccv_nnc_tensor_param_t output_params;
3628 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);
3629 ccv_nnc_hint_tensor_auto(index_select, (ccv_nnc_tensor_param_t []){
3630 vocab_params,
3631 index_params,
3632 }, 2, ccv_nnc_no_hint, &output_params, 1);
3633 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3634 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");
3635 outputs[0] = output;
3636}
3637
3638static ccv_cnnp_model_t* _ccv_cnnp_index_select_copy(const ccv_cnnp_model_t* const super, void* const context);
3639
3640static const ccv_cnnp_model_vtab_t ccv_cnnp_index_select_isa = {
3641 .build = _ccv_cnnp_index_select_build,
3642 .copy = _ccv_cnnp_index_select_copy,
3643};
3644
3645ccv_cnnp_model_t* ccv_cnnp_index_select(const char* const name)
3646{
3647 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));
3648 model_index_select->super.isa = &ccv_cnnp_index_select_isa;
3649 model_index_select->super.input_size = 2;
3650 model_index_select->super.outputs = &model_index_select->output;
3651 model_index_select->super.output_size = 1;
3652 ccv_cnnp_model_copy_name(&model_index_select->super, name);
3653 return (ccv_cnnp_model_t*)model_index_select;
3654}
3655
3656static ccv_cnnp_model_t* _ccv_cnnp_index_select_copy(const ccv_cnnp_model_t* const super, void* const context)
3657{
3658 ccv_cnnp_model_index_select_t* const self = (ccv_cnnp_model_index_select_t*)super;
3659 return ccv_cnnp_index_select(self->super.name);
3660}
3661
3662// MARK - Embedding Layer
3663
3664typedef struct {
3665 ccv_cnnp_model_t super;
3666 ccv_nnc_tensor_symbol_t output;
3667 ccv_nnc_tensor_symbol_t vocab;
3668 int datatype;
3669 int vocab_size;
3670 int embed_size;
3671} ccv_cnnp_model_embedding_t;
3672
3673static 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)
3674{
3675 ccv_cnnp_model_embedding_t* const self = (ccv_cnnp_model_embedding_t*)super;
3676 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);
3677 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"
, 3677, __extension__ __PRETTY_FUNCTION__); }));
3678 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", 3678, __extension__ __PRETTY_FUNCTION__
); }));
3679 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3680 ccv_nnc_tensor_param_t vocab_params = params;
3681 memset(vocab_params.dim, 0, sizeof(vocab_params.dim));
3682 vocab_params.datatype = self->datatype;
3683 vocab_params.dim[0] = self->vocab_size;
3684 vocab_params.dim[1] = self->embed_size;
3685 if (!self->vocab.graph)
3686 self->vocab = ccv_nnc_tensor_symbol_new(graph, vocab_params, "vocab");
3687 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", 3687
, __extension__ __PRETTY_FUNCTION__); }));
3688 const ccv_nnc_tensor_symbol_t vocab = ccv_cnnp_model_get_symbol(super, self->vocab);
3689 ccv_nnc_tensor_param_t output_params;
3690 const ccv_nnc_cmd_t embedding = CMD_INDEX_SELECT_FORWARD()ccv_nnc_cmd(CCV_NNC_INDEX_SELECT_FORWARD, 0, ccv_nnc_cmd_auto
, 0);
3691 ccv_nnc_hint_tensor_auto(embedding, (ccv_nnc_tensor_param_t []){
3692 vocab_params,
3693 params,
3694 }, 2, ccv_nnc_no_hint, &output_params, 1);
3695 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3696 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");
3697 outputs[0] = output;
3698}
3699
3700static 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)
3701{
3702 ccv_cnnp_model_embedding_t* const self = (ccv_cnnp_model_embedding_t*)super;
3703 const float std = sqrtf(2) / sqrtf(self->vocab_size + self->embed_size);
3704 const float bound = sqrtf(3) * std;
3705 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);
3706}
3707
3708static 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)
3709{
3710 ccv_cnnp_model_embedding_t* const self = (ccv_cnnp_model_embedding_t*)super;
3711 add_to_array(parameters, self->vocab, is_trainable);
3712}
3713
3714static ccv_cnnp_model_t* _ccv_cnnp_embedding_copy(const ccv_cnnp_model_t* const super, void* const context);
3715
3716static const ccv_cnnp_model_vtab_t ccv_cnnp_embedding_isa = {
3717 .build = _ccv_cnnp_embedding_build,
3718 .init_states = _ccv_cnnp_embedding_init_states,
3719 .add_to_parameter = _ccv_cnnp_embedding_add_to_parameter,
3720 .copy = _ccv_cnnp_embedding_copy,
3721};
3722
3723ccv_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)
3724{
3725 ccv_cnnp_model_embedding_t* const model_embedding = (ccv_cnnp_model_embedding_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_embedding_t));
3726 model_embedding->super.isa = &ccv_cnnp_embedding_isa;
3727 model_embedding->super.input_size = 1;
3728 model_embedding->super.outputs = &model_embedding->output;
3729 model_embedding->super.output_size = 1;
3730 model_embedding->super.is_trainable = is_trainable;
3731 ccv_cnnp_model_copy_name(&model_embedding->super, name);
3732 model_embedding->vocab.d = CCV_NNC_NO_TENSOR_SYMBOL;
3733 model_embedding->vocab.graph = 0;
3734 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", 3734, __extension__ __PRETTY_FUNCTION__
); }));
3735 model_embedding->datatype = datatype;
3736 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", 3736, __extension__ __PRETTY_FUNCTION__
); }));
3737 model_embedding->vocab_size = vocab_size;
3738 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", 3738, __extension__ __PRETTY_FUNCTION__
); }));
3739 model_embedding->embed_size = embed_size;
3740 return (ccv_cnnp_model_t*)model_embedding;
3741}
3742
3743static ccv_cnnp_model_t* _ccv_cnnp_embedding_copy(const ccv_cnnp_model_t* const super, void* const context)
3744{
3745 ccv_cnnp_model_embedding_t* const self = (ccv_cnnp_model_embedding_t*)super;
3746 return ccv_cnnp_embedding(self->datatype, self->vocab_size, self->embed_size, self->super.is_trainable, self->super.name);
3747}
3748
3749// MARK - Pool Layers
3750
3751typedef struct {
3752 ccv_cnnp_model_t super;
3753 ccv_nnc_tensor_symbol_t output;
3754 int type;
3755 float width_scale;
3756 float height_scale;
3757 int align_corners;
3758} ccv_cnnp_model_upsample_t;
3759
3760static 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)
3761{
3762 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);
3763 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"
, 3763, __extension__ __PRETTY_FUNCTION__); }));
3764 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", 3764, __extension__ __PRETTY_FUNCTION__
); }));
3765 ccv_cnnp_model_upsample_t* const self = (ccv_cnnp_model_upsample_t*)super;
3766 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3767 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);
3768 ccv_nnc_tensor_param_t output_params;
3769 ccv_nnc_hint_tensor_auto(cmd, &params, 1, ccv_nnc_no_hint, &output_params, 1);
3770 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3771 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");
3772 outputs[0] = output;
3773}
3774
3775static ccv_cnnp_model_t* _ccv_cnnp_upsample_copy(const ccv_cnnp_model_t* const super, void* const context);
3776
3777static const ccv_cnnp_model_vtab_t ccv_cnnp_upsample_isa = {
3778 .build = _ccv_cnnp_upsample_build,
3779 .copy = _ccv_cnnp_upsample_copy,
3780};
3781
3782ccv_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)
3783{
3784 ccv_cnnp_model_upsample_t* const model_upsample = (ccv_cnnp_model_upsample_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_upsample_t));
3785 model_upsample->super.isa = &ccv_cnnp_upsample_isa;
3786 model_upsample->super.input_size = 1;
3787 model_upsample->super.outputs = &model_upsample->output;
3788 model_upsample->super.output_size = 1;
3789 ccv_cnnp_model_copy_name(&model_upsample->super, name);
3790 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", 3790, __extension__ __PRETTY_FUNCTION__
); }));
3791 model_upsample->type = type;
3792 model_upsample->width_scale = width_scale;
3793 model_upsample->height_scale = height_scale;
3794 model_upsample->align_corners = align_corners;
3795 return (ccv_cnnp_model_t*)model_upsample;
3796}
3797
3798static ccv_cnnp_model_t* _ccv_cnnp_upsample_copy(const ccv_cnnp_model_t* const super, void* const context)
3799{
3800 const ccv_cnnp_model_upsample_t* const self = (const ccv_cnnp_model_upsample_t*)super;
3801 return ccv_cnnp_upsample(self->type, self->width_scale, self->height_scale, self->align_corners, self->super.name);
3802}
3803
3804// MARK - Reduce Sum Layer
3805
3806typedef struct {
3807 ccv_cnnp_model_t super;
3808 int axis[CCV_NNC_MAX_DIM_ALLOC(12)];
3809 int count;
3810 ccv_nnc_tensor_symbol_t output;
3811} ccv_cnnp_model_reduce_sum_t;
3812
3813static 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)
3814{
3815 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);
3816 const ccv_cnnp_model_reduce_sum_t* const self = (const ccv_cnnp_model_reduce_sum_t*)super;
3817 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"
, 3817, __extension__ __PRETTY_FUNCTION__); }));
3818 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", 3818, __extension__ __PRETTY_FUNCTION__
); }));
3819 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3820 ccv_nnc_tensor_param_t output_params;
3821 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);
3822 int i;
3823 for (i = 0; i < self->count; i++)
3824 reduce_sum.info.reduce.axis[i] = self->axis[i];
3825 reduce_sum.info.reduce.count = self->count;
3826 ccv_nnc_hint_tensor_auto(reduce_sum, &input_params, 1, ccv_nnc_no_hint, &output_params, 1);
3827 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3828 ccv_nnc_graph_exec_symbol_new(graph, reduce_sum, inputs, input_size, outputs, output_size, "reduce_sum");
3829}
3830
3831static ccv_cnnp_model_t* _ccv_cnnp_reduce_sum_copy(const ccv_cnnp_model_t* const self, void* const context);
3832
3833static const ccv_cnnp_model_vtab_t ccv_cnnp_reduce_sum_isa = {
3834 .build = _ccv_cnnp_reduce_sum_build,
3835 .copy = _ccv_cnnp_reduce_sum_copy,
3836};
3837
3838ccv_cnnp_model_t* ccv_cnnp_reduce_sum(const int* const axis, const int axis_count, const char* const name)
3839{
3840 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));
3841 model_reduce_sum->super.isa = &ccv_cnnp_reduce_sum_isa;
3842 model_reduce_sum->super.input_size = 1;
3843 model_reduce_sum->super.outputs = &model_reduce_sum->output;
3844 model_reduce_sum->super.output_size = 1;
3845 ccv_cnnp_model_copy_name(&model_reduce_sum->super, name);
3846 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", 3846, __extension__ __PRETTY_FUNCTION__
); }));
3847 int i;
3848 for (i = 0; i < axis_count; i++)
3849 model_reduce_sum->axis[i] = axis[i];
3850 model_reduce_sum->count = axis_count;
3851 return (ccv_cnnp_model_t*)model_reduce_sum;
3852}
3853
3854static ccv_cnnp_model_t* _ccv_cnnp_reduce_sum_copy(const ccv_cnnp_model_t* const super, void* const context)
3855{
3856 const ccv_cnnp_model_reduce_sum_t* const self = (const ccv_cnnp_model_reduce_sum_t*)super;
3857 return ccv_cnnp_reduce_sum(self->axis, self->count, self->super.name);
3858}
3859
3860// MARK - Reduce Mean Layer
3861
3862typedef struct {
3863 ccv_cnnp_model_t super;
3864 int axis[CCV_NNC_MAX_DIM_ALLOC(12)];
3865 int count;
3866 ccv_nnc_tensor_symbol_t output;
3867} ccv_cnnp_model_reduce_mean_t;
3868
3869static 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)
3870{
3871 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);
3872 const ccv_cnnp_model_reduce_mean_t* const self = (const ccv_cnnp_model_reduce_mean_t*)super;
3873 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"
, 3873, __extension__ __PRETTY_FUNCTION__); }));
3874 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", 3874, __extension__ __PRETTY_FUNCTION__
); }));
3875 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3876 ccv_nnc_tensor_param_t output_params;
3877 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);
3878 int i;
3879 for (i = 0; i < self->count; i++)
3880 reduce_mean.info.reduce.axis[i] = self->axis[i];
3881 reduce_mean.info.reduce.count = self->count;
3882 ccv_nnc_hint_tensor_auto(reduce_mean, &input_params, 1, ccv_nnc_no_hint, &output_params, 1);
3883 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3884 ccv_nnc_graph_exec_symbol_new(graph, reduce_mean, inputs, input_size, outputs, output_size, "reduce_mean");
3885}
3886
3887static ccv_cnnp_model_t* _ccv_cnnp_reduce_mean_copy(const ccv_cnnp_model_t* const self, void* const context);
3888
3889static const ccv_cnnp_model_vtab_t ccv_cnnp_reduce_mean_isa = {
3890 .build = _ccv_cnnp_reduce_mean_build,
3891 .copy = _ccv_cnnp_reduce_mean_copy,
3892};
3893
3894ccv_cnnp_model_t* ccv_cnnp_reduce_mean(const int* const axis, const int axis_count, const char* const name)
3895{
3896 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));
3897 model_reduce_mean->super.isa = &ccv_cnnp_reduce_mean_isa;
3898 model_reduce_mean->super.input_size = 1;
3899 model_reduce_mean->super.outputs = &model_reduce_mean->output;
3900 model_reduce_mean->super.output_size = 1;
3901 ccv_cnnp_model_copy_name(&model_reduce_mean->super, name);
3902 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", 3902, __extension__ __PRETTY_FUNCTION__
); }));
3903 int i;
3904 for (i = 0; i < axis_count; i++)
3905 model_reduce_mean->axis[i] = axis[i];
3906 model_reduce_mean->count = axis_count;
3907 return (ccv_cnnp_model_t*)model_reduce_mean;
3908}
3909
3910static ccv_cnnp_model_t* _ccv_cnnp_reduce_mean_copy(const ccv_cnnp_model_t* const super, void* const context)
3911{
3912 const ccv_cnnp_model_reduce_mean_t* const self = (const ccv_cnnp_model_reduce_mean_t*)super;
3913 return ccv_cnnp_reduce_mean(self->axis, self->count, self->super.name);
3914}
3915
3916// MARK - Reduce Max Layer
3917
3918typedef struct {
3919 ccv_cnnp_model_t super;
3920 int axis[CCV_NNC_MAX_DIM_ALLOC(12)];
3921 int count;
3922 ccv_nnc_tensor_symbol_t output;
3923} ccv_cnnp_model_reduce_max_t;
3924
3925static 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)
3926{
3927 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);
3928 const ccv_cnnp_model_reduce_max_t* const self = (const ccv_cnnp_model_reduce_max_t*)super;
3929 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"
, 3929, __extension__ __PRETTY_FUNCTION__); }));
3930 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", 3930, __extension__ __PRETTY_FUNCTION__
); }));
3931 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3932 ccv_nnc_tensor_param_t output_params;
3933 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);
3934 int i;
3935 for (i = 0; i < self->count; i++)
3936 reduce_max.info.reduce.axis[i] = self->axis[i];
3937 reduce_max.info.reduce.count = self->count;
3938 ccv_nnc_hint_tensor_auto(reduce_max, &input_params, 1, ccv_nnc_no_hint, &output_params, 1);
3939 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3940 ccv_nnc_graph_exec_symbol_new(graph, reduce_max, inputs, input_size, outputs, output_size, "reduce_max");
3941}
3942
3943static ccv_cnnp_model_t* _ccv_cnnp_reduce_max_copy(const ccv_cnnp_model_t* const self, void* const context);
3944
3945static const ccv_cnnp_model_vtab_t ccv_cnnp_reduce_max_isa = {
3946 .build = _ccv_cnnp_reduce_max_build,
3947 .copy = _ccv_cnnp_reduce_max_copy,
3948};
3949
3950ccv_cnnp_model_t* ccv_cnnp_reduce_max(const int* const axis, const int axis_count, const char* const name)
3951{
3952 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));
3953 model_reduce_max->super.isa = &ccv_cnnp_reduce_max_isa;
3954 model_reduce_max->super.input_size = 1;
3955 model_reduce_max->super.outputs = &model_reduce_max->output;
3956 model_reduce_max->super.output_size = 1;
3957 ccv_cnnp_model_copy_name(&model_reduce_max->super, name);
3958 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", 3958, __extension__ __PRETTY_FUNCTION__
); }));
3959 int i;
3960 for (i = 0; i < axis_count; i++)
3961 model_reduce_max->axis[i] = axis[i];
3962 model_reduce_max->count = axis_count;
3963 return (ccv_cnnp_model_t*)model_reduce_max;
3964}
3965
3966static ccv_cnnp_model_t* _ccv_cnnp_reduce_max_copy(const ccv_cnnp_model_t* const super, void* const context)
3967{
3968 const ccv_cnnp_model_reduce_max_t* const self = (const ccv_cnnp_model_reduce_max_t*)super;
3969 return ccv_cnnp_reduce_max(self->axis, self->count, self->super.name);
3970}
3971
3972// MARK - Reduce Min Layer
3973
3974typedef struct {
3975 ccv_cnnp_model_t super;
3976 int axis[CCV_NNC_MAX_DIM_ALLOC(12)];
3977 int count;
3978 ccv_nnc_tensor_symbol_t output;
3979} ccv_cnnp_model_reduce_min_t;
3980
3981static 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)
3982{
3983 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);
3984 const ccv_cnnp_model_reduce_min_t* const self = (const ccv_cnnp_model_reduce_min_t*)super;
3985 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"
, 3985, __extension__ __PRETTY_FUNCTION__); }));
3986 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", 3986, __extension__ __PRETTY_FUNCTION__
); }));
3987 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3988 ccv_nnc_tensor_param_t output_params;
3989 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);
3990 int i;
3991 for (i = 0; i < self->count; i++)
3992 reduce_min.info.reduce.axis[i] = self->axis[i];
3993 reduce_min.info.reduce.count = self->count;
3994 ccv_nnc_hint_tensor_auto(reduce_min, &input_params, 1, ccv_nnc_no_hint, &output_params, 1);
3995 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3996 ccv_nnc_graph_exec_symbol_new(graph, reduce_min, inputs, input_size, outputs, output_size, "reduce_min");
3997}
3998
3999static ccv_cnnp_model_t* _ccv_cnnp_reduce_min_copy(const ccv_cnnp_model_t* const self, void* const context);
4000
4001static const ccv_cnnp_model_vtab_t ccv_cnnp_reduce_min_isa = {
4002 .build = _ccv_cnnp_reduce_min_build,
4003 .copy = _ccv_cnnp_reduce_min_copy,
4004};
4005
4006ccv_cnnp_model_t* ccv_cnnp_reduce_min(const int* const axis, const int axis_count, const char* const name)
4007{
4008 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));
4009 model_reduce_min->super.isa = &ccv_cnnp_reduce_min_isa;
4010 model_reduce_min->super.input_size = 1;
4011 model_reduce_min->super.outputs = &model_reduce_min->output;
4012 model_reduce_min->super.output_size = 1;
4013 ccv_cnnp_model_copy_name(&model_reduce_min->super, name);
4014 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", 4014, __extension__ __PRETTY_FUNCTION__
); }));
4015 int i;
4016 for (i = 0; i < axis_count; i++)
4017 model_reduce_min->axis[i] = axis[i];
4018 model_reduce_min->count = axis_count;
4019 return (ccv_cnnp_model_t*)model_reduce_min;
4020}
4021
4022static ccv_cnnp_model_t* _ccv_cnnp_reduce_min_copy(const ccv_cnnp_model_t* const super, void* const context)
4023{
4024 const ccv_cnnp_model_reduce_min_t* const self = (const ccv_cnnp_model_reduce_min_t*)super;
4025 return ccv_cnnp_reduce_min(self->axis, self->count, self->super.name);
4026}
4027
4028// MARK - Reduce Norm2 Layer
4029
4030typedef struct {
4031 ccv_cnnp_model_t super;
4032 int axis[CCV_NNC_MAX_DIM_ALLOC(12)];
4033 int count;
4034 ccv_nnc_tensor_symbol_t output;
4035} ccv_cnnp_model_reduce_norm2_t;
4036
4037static 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)
4038{
4039 const ccv_cnnp_model_reduce_norm2_t* const self = (const ccv_cnnp_model_reduce_norm2_t*)super;
4040 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);
4041 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"
, 4041, __extension__ __PRETTY_FUNCTION__); }));
4042 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", 4042, __extension__ __PRETTY_FUNCTION__
); }));
4043 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4044 ccv_nnc_tensor_param_t output_params;
4045 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);
4046 int i;
4047 for (i = 0; i < self->count; i++)
4048 reduce_norm2.info.reduce.axis[i] = self->axis[i];
4049 reduce_norm2.info.reduce.count = self->count;
4050 ccv_nnc_hint_tensor_auto(reduce_norm2, &input_params, 1, ccv_nnc_no_hint, &output_params, 1);
4051 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
4052 ccv_nnc_graph_exec_symbol_new(graph, reduce_norm2, inputs, input_size, outputs, output_size, "reduce_norm2");
4053}
4054
4055static ccv_cnnp_model_t* _ccv_cnnp_reduce_norm2_copy(const ccv_cnnp_model_t* const self, void* const context);
4056
4057static const ccv_cnnp_model_vtab_t ccv_cnnp_reduce_norm2_isa = {
4058 .build = _ccv_cnnp_reduce_norm2_build,
4059 .copy = _ccv_cnnp_reduce_norm2_copy,
4060};
4061
4062ccv_cnnp_model_t* ccv_cnnp_reduce_norm2(const int* const axis, const int axis_count, const char* const name)
4063{
4064 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));
4065 model_reduce_norm2->super.isa = &ccv_cnnp_reduce_norm2_isa;
4066 model_reduce_norm2->super.input_size = 1;
4067 model_reduce_norm2->super.outputs = &model_reduce_norm2->output;
4068 model_reduce_norm2->super.output_size = 1;
4069 ccv_cnnp_model_copy_name(&model_reduce_norm2->super, name);
4070 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", 4070, __extension__ __PRETTY_FUNCTION__
); }));
4071 int i;
4072 for (i = 0; i < axis_count; i++)
4073 model_reduce_norm2->axis[i] = axis[i];
4074 model_reduce_norm2->count = axis_count;
4075 return (ccv_cnnp_model_t*)model_reduce_norm2;
4076}
4077
4078static ccv_cnnp_model_t* _ccv_cnnp_reduce_norm2_copy(const ccv_cnnp_model_t* const super, void* const context)
4079{
4080 const ccv_cnnp_model_reduce_norm2_t* const self = (const ccv_cnnp_model_reduce_norm2_t*)super;
4081 return ccv_cnnp_reduce_norm2(self->axis, self->count, self->super.name);
4082}
4083
4084// MARK - Argmax Layer
4085
4086typedef struct {
4087 ccv_cnnp_model_t super;
4088 int axis;
4089 ccv_nnc_tensor_symbol_t output;
4090} ccv_cnnp_model_argmax_t;
4091
4092static 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)
4093{
4094 const ccv_cnnp_model_argmax_t* const self = (const ccv_cnnp_model_argmax_t*)super;
4095 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);
4096 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"
, 4096, __extension__ __PRETTY_FUNCTION__); }));
4097 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", 4097, __extension__ __PRETTY_FUNCTION__
); }));
4098 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4099 ccv_nnc_tensor_param_t output_params;
4100 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);
4101 argmax.info.reduce.axis[0] = self->axis;
4102 argmax.info.reduce.count = 1;
4103 ccv_nnc_hint_tensor_auto(argmax, &input_params, 1, ccv_nnc_no_hint, &output_params, 1);
4104 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
4105 ccv_nnc_graph_exec_symbol_new(graph, argmax, inputs, input_size, outputs, output_size, "argmax");
4106}
4107
4108static ccv_cnnp_model_t* _ccv_cnnp_argmax_copy(const ccv_cnnp_model_t* const self, void* const context);
4109
4110static const ccv_cnnp_model_vtab_t ccv_cnnp_argmax_isa = {
4111 .build = _ccv_cnnp_argmax_build,
4112 .copy = _ccv_cnnp_argmax_copy,
4113};
4114
4115ccv_cnnp_model_t* ccv_cnnp_argmax(const int axis, const char* const name)
4116{
4117 ccv_cnnp_model_argmax_t* const model_argmax = (ccv_cnnp_model_argmax_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_argmax_t));
4118 model_argmax->super.isa = &ccv_cnnp_argmax_isa;
4119 model_argmax->super.input_size = 1;
4120 model_argmax->super.outputs = &model_argmax->output;
4121 model_argmax->super.output_size = 1;
4122 ccv_cnnp_model_copy_name(&model_argmax->super, name);
4123 model_argmax->axis = axis;
4124 return (ccv_cnnp_model_t*)model_argmax;
4125}
4126
4127static ccv_cnnp_model_t* _ccv_cnnp_argmax_copy(const ccv_cnnp_model_t* const super, void* const context)
4128{
4129 const ccv_cnnp_model_argmax_t* const self = (const ccv_cnnp_model_argmax_t*)super;
4130 return ccv_cnnp_argmax(self->axis, self->super.name);
4131}
4132
4133// MARK - Argmin Layer
4134
4135typedef struct {
4136 ccv_cnnp_model_t super;
4137 int axis;
4138 ccv_nnc_tensor_symbol_t output;
4139} ccv_cnnp_model_argmin_t;
4140
4141static 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)
4142{
4143 const ccv_cnnp_model_argmin_t* const self = (const ccv_cnnp_model_argmin_t*)super;
4144 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);
4145 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"
, 4145, __extension__ __PRETTY_FUNCTION__); }));
4146 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", 4146, __extension__ __PRETTY_FUNCTION__
); }));
4147 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4148 ccv_nnc_tensor_param_t output_params;
4149 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);
4150 argmin.info.reduce.axis[0] = self->axis;
4151 argmin.info.reduce.count = 1;
4152 ccv_nnc_hint_tensor_auto(argmin, &input_params, 1, ccv_nnc_no_hint, &output_params, 1);
4153 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
4154 ccv_nnc_graph_exec_symbol_new(graph, argmin, inputs, input_size, outputs, output_size, "argmin");
4155}
4156
4157static ccv_cnnp_model_t* _ccv_cnnp_argmin_copy(const ccv_cnnp_model_t* const self, void* const context);
4158
4159static const ccv_cnnp_model_vtab_t ccv_cnnp_argmin_isa = {
4160 .build = _ccv_cnnp_argmin_build,
4161 .copy = _ccv_cnnp_argmin_copy,
4162};
4163
4164ccv_cnnp_model_t* ccv_cnnp_argmin(const int axis, const char* const name)
4165{
4166 ccv_cnnp_model_argmin_t* const model_argmin = (ccv_cnnp_model_argmin_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_argmin_t));
4167 model_argmin->super.isa = &ccv_cnnp_argmin_isa;
4168 model_argmin->super.input_size = 1;
4169 model_argmin->super.outputs = &model_argmin->output;
4170 model_argmin->super.output_size = 1;
4171 ccv_cnnp_model_copy_name(&model_argmin->super, name);
4172 model_argmin->axis = axis;
4173 return (ccv_cnnp_model_t*)model_argmin;
4174}
4175
4176static ccv_cnnp_model_t* _ccv_cnnp_argmin_copy(const ccv_cnnp_model_t* const super, void* const context)
4177{
4178 const ccv_cnnp_model_argmin_t* const self = (const ccv_cnnp_model_argmin_t*)super;
4179 return ccv_cnnp_argmin(self->axis, self->super.name);
4180}
4181
4182// MARK - Min Layer
4183
4184typedef struct {
4185 ccv_cnnp_model_t super;
4186 ccv_nnc_tensor_symbol_t output;
4187} ccv_cnnp_model_min_t;
4188
4189static 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)
4190{
4191 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);
4192 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"
, 4192, __extension__ __PRETTY_FUNCTION__); }));
4193 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", 4193, __extension__ __PRETTY_FUNCTION__
); }));
4194 ccv_nnc_tensor_param_t input_params[2];
4195 int i;
4196 for (i = 0; i < 2; i++)
4197 input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
4198 ccv_nnc_tensor_param_t output_params;
4199 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);
4200 ccv_nnc_hint_tensor_auto(min, input_params, 2, ccv_nnc_no_hint, &output_params, 1);
4201 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
4202 ccv_nnc_graph_exec_symbol_new(graph, min, inputs, input_size, outputs, output_size, "min");
4203}
4204
4205static ccv_cnnp_model_t* _ccv_cnnp_min_copy(const ccv_cnnp_model_t* const self, void* const context);
4206
4207static const ccv_cnnp_model_vtab_t ccv_cnnp_min_isa = {
4208 .build = _ccv_cnnp_min_build,
4209 .copy = _ccv_cnnp_min_copy,
4210};
4211
4212ccv_cnnp_model_t* ccv_cnnp_min(const char* const name)
4213{
4214 ccv_cnnp_model_min_t* const model_min = (ccv_cnnp_model_min_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_min_t));
4215 model_min->super.isa = &ccv_cnnp_min_isa;
4216 model_min->super.input_size = 2;
4217 model_min->super.outputs = &model_min->output;
4218 model_min->super.output_size = 1;
4219 ccv_cnnp_model_copy_name(&model_min->super, name);
4220 return (ccv_cnnp_model_t*)model_min;
4221}
4222
4223static ccv_cnnp_model_t* _ccv_cnnp_min_copy(const ccv_cnnp_model_t* const super, void* const context)
4224{
4225 const ccv_cnnp_model_min_t* const self = (const ccv_cnnp_model_min_t*)super;
4226 return ccv_cnnp_min(self->super.name);
4227}
4228
4229// MARK - Max Layer
4230
4231typedef struct {
4232 ccv_cnnp_model_t super;
4233 ccv_nnc_tensor_symbol_t output;
4234} ccv_cnnp_model_max_t;
4235
4236static 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)
4237{
4238 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);
4239 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"
, 4239, __extension__ __PRETTY_FUNCTION__); }));
4240 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", 4240, __extension__ __PRETTY_FUNCTION__
); }));
4241 ccv_nnc_tensor_param_t input_params[2];
4242 int i;
4243 for (i = 0; i < 2; i++)
4244 input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
4245 ccv_nnc_tensor_param_t output_params;
4246 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);
4247 ccv_nnc_hint_tensor_auto(max, input_params, 2, ccv_nnc_no_hint, &output_params, 1);
4248 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
4249 ccv_nnc_graph_exec_symbol_new(graph, max, inputs, input_size, outputs, output_size, "max");
4250}
4251
4252static ccv_cnnp_model_t* _ccv_cnnp_max_copy(const ccv_cnnp_model_t* const self, void* const context);
4253
4254static const ccv_cnnp_model_vtab_t ccv_cnnp_max_isa = {
4255 .build = _ccv_cnnp_max_build,
4256 .copy = _ccv_cnnp_max_copy,
4257};
4258
4259ccv_cnnp_model_t* ccv_cnnp_max(const char* const name)
4260{
4261 ccv_cnnp_model_max_t* const model_max = (ccv_cnnp_model_max_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_max_t));
4262 model_max->super.isa = &ccv_cnnp_max_isa;
4263 model_max->super.input_size = 2;
4264 model_max->super.outputs = &model_max->output;
4265 model_max->super.output_size = 1;
4266 ccv_cnnp_model_copy_name(&model_max->super, name);
4267 return (ccv_cnnp_model_t*)model_max;
4268}
4269
4270static ccv_cnnp_model_t* _ccv_cnnp_max_copy(const ccv_cnnp_model_t* const super, void* const context)
4271{
4272 const ccv_cnnp_model_max_t* const self = (const ccv_cnnp_model_max_t*)super;
4273 return ccv_cnnp_max(self->super.name);
4274}
4275
4276// MARK - LSTM Layer
4277
4278typedef struct {
4279 ccv_cnnp_model_t super;
4280 int masked;
4281 ccv_nnc_tensor_symbol_t output;
4282 ccv_nnc_tensor_symbol_t weights;
4283 ccv_nnc_tensor_symbol_t reserves;
4284 ccv_nnc_cmd_param_t params;
4285 ccv_nnc_graph_exec_symbol_t lstm;
4286} ccv_cnnp_model_lstm_t;
4287
4288static int _ccv_cnnp_lstm_weight_dim(int bidirectional, int num_layers, int input_size, int hidden_size, int proj_size, int bias)
4289{
4290 const int D = !!bidirectional + 1;
4291 if (hidden_size == proj_size)
4292 return (num_layers * (bias ? 8 : 0) + (num_layers - 1) * (hidden_size * 4 * D + hidden_size * 4) + input_size * 4 + hidden_size * 4) * D;
4293 else
4294 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;
4295}
4296
4297static 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)
4298{
4299 ccv_cnnp_model_lstm_t* const self = (ccv_cnnp_model_lstm_t*)super;
4300 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);
4301 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", 4301, __extension__ __PRETTY_FUNCTION__
); }));
4302 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", 4302, __extension__ __PRETTY_FUNCTION__
); }));
4303 const int proj_size = self->params.rnn.proj_size == 0 ? self->params.rnn.hidden_size : self->params.rnn.proj_size;
4304 ccv_nnc_tensor_param_t input_params[5];
4305 input_params[0]= ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4306 if (input_size == 2)
4307 input_params[1] = ccv_nnc_tensor_symbol_params(graph, inputs[1]);
4308 input_params[4] = input_params[0];
4309 memset(input_params[4].dim, 0, sizeof(input_params[4].dim));
4310 const int x_nd = ccv_nnc_tensor_nd(input_params[0].dim);
4311 const int feature_count = input_params[0].dim[x_nd - 1];
4312 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);
4313 input_params[4].dim[1] = self->params.rnn.hidden_size;
4314 const ccv_nnc_cmd_t lstm = ccv_nnc_cmd(CCV_NNC_LSTM_FORWARD, 0, self->params, 0);
4315 ccv_nnc_tensor_param_t output_params[4];
4316 ccv_nnc_hint_tensor_auto(lstm, input_params, 5, ccv_nnc_no_hint, output_params, 4);
4317 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0);
4318 if (!self->weights.graph)
4319 self->weights = ccv_nnc_tensor_symbol_new(graph, input_params[4], "weights");
4320 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"
, 4320, __extension__ __PRETTY_FUNCTION__); }));
4321 const ccv_nnc_tensor_symbol_t weights = ccv_cnnp_model_get_symbol(super, self->weights);
4322 if (!self->reserves.graph)
4323 self->reserves = ccv_nnc_tensor_symbol_new(graph, output_params[3], "reserves");
4324 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"
, 4324, __extension__ __PRETTY_FUNCTION__); }));
4325 const ccv_nnc_tensor_symbol_t reserves = ccv_cnnp_model_get_symbol(super, self->reserves);
4326 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
};
4327 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");
4328}
4329
4330static 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)
4331{
4332 ccv_cnnp_model_lstm_t* const self = (ccv_cnnp_model_lstm_t*)super;
4333 if (self->weights.graph)
4334 {
4335 const float stdv = 1.0 / sqrt(self->params.rnn.hidden_size);
4336 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);
4337 }
4338}
4339
4340static 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)
4341{
4342 ccv_cnnp_model_lstm_t* const self = (ccv_cnnp_model_lstm_t*)super;
4343 if (self->weights.graph)
4344 add_to_array(parameters, self->weights, is_trainable);
4345}
4346
4347static 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)
4348{
4349 ccv_cnnp_model_lstm_t* const self = (ccv_cnnp_model_lstm_t*)super;
4350 if (self->lstm.graph)
4351 {
4352 self->params.rnn.is_test = is_test;
4353 updater(context, self->lstm, ccv_nnc_cmd(CCV_NNC_LSTM_FORWARD, 0, self->params, 0), ccv_nnc_no_hint);
4354 }
4355}
4356
4357static ccv_cnnp_model_t* _ccv_cnnp_lstm_copy(const ccv_cnnp_model_t* const self, void* const context);
4358
4359static const ccv_cnnp_model_vtab_t ccv_cnnp_lstm_isa = {
4360 .build = _ccv_cnnp_lstm_build,
4361 .init_states = _ccv_cnnp_lstm_init_states,
4362 .add_to_parameter = _ccv_cnnp_lstm_add_to_parameter,
4363 .copy = _ccv_cnnp_lstm_copy,
4364 .set_is_test = _ccv_cnnp_lstm_set_is_test,
4365};
4366
4367ccv_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)
4368{
4369 ccv_cnnp_model_lstm_t* const model_lstm = (ccv_cnnp_model_lstm_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_lstm_t));
4370 model_lstm->super.isa = &ccv_cnnp_lstm_isa;
4371 model_lstm->super.input_size = masked ? 2 : 1;
4372 model_lstm->super.outputs = &model_lstm->output;
4373 model_lstm->super.output_size = 1;
4374 model_lstm->super.is_trainable = is_trainable;
4375 ccv_cnnp_model_copy_name(&model_lstm->super, name);
4376 model_lstm->masked = masked;
4377 model_lstm->weights.d = CCV_NNC_NO_TENSOR_SYMBOL;
4378 model_lstm->weights.graph = 0;
4379 model_lstm->params.rnn.hidden_size = hidden_size;
4380 model_lstm->params.rnn.proj_size = proj_size;
4381 model_lstm->params.rnn.num_layers = num_layers;
4382 model_lstm->params.rnn.bias = bias;
4383 model_lstm->params.rnn.batch_first = batch_first;
4384 model_lstm->params.rnn.bidirectional = bidirectional;
4385 model_lstm->params.rnn.dropout = dropout;
4386 return (ccv_cnnp_model_t*)model_lstm;
4387}
4388
4389static ccv_cnnp_model_t* _ccv_cnnp_lstm_copy(const ccv_cnnp_model_t* const super, void* const context)
4390{
4391 const ccv_cnnp_model_lstm_t* const self = (const ccv_cnnp_model_lstm_t*)super;
4392 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);
4393}
4394
4395/// MARK - Datatype conversion layer.
4396
4397typedef struct {
4398 ccv_cnnp_model_t super;
4399 ccv_nnc_tensor_symbol_t output;
4400 int datatype;
4401 int ref_to_last;
4402} ccv_cnnp_model_datatype_conversion_t;
4403
4404static 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)
4405{
4406 ccv_cnnp_model_datatype_conversion_t* const self = (ccv_cnnp_model_datatype_conversion_t*)super;
4407 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);
4408 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4409 if (self->ref_to_last)
4410 {
4411 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", 4411, __extension__ __PRETTY_FUNCTION__
); }));
4412 const ccv_nnc_tensor_param_t last_params = ccv_nnc_tensor_symbol_params(graph, inputs[input_size - 1]);
4413 params.datatype = last_params.datatype;
4414 } else
4415 params.datatype = self->datatype;
4416 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", 4416, __extension__ __PRETTY_FUNCTION__
); }));
4417 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4418 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);
4419}
4420
4421static ccv_cnnp_model_t* _ccv_cnnp_datatype_conversion_copy(const ccv_cnnp_model_t* const self, void* const context);
4422
4423static const ccv_cnnp_model_vtab_t ccv_cnnp_datatype_conversion_isa = {
4424 .build = _ccv_cnnp_datatype_conversion_build,
4425 .copy = _ccv_cnnp_datatype_conversion_copy,
4426};
4427
4428ccv_cnnp_model_t* ccv_cnnp_datatype_conversion(const int datatype, const int ref_to_last, const char* const name)
4429{
4430 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));
4431 model_datatype_conversion->super.isa = &ccv_cnnp_datatype_conversion_isa;
4432 model_datatype_conversion->super.input_size = 0;
4433 model_datatype_conversion->super.outputs = &model_datatype_conversion->output;
4434 model_datatype_conversion->super.output_size = 1;
4435 model_datatype_conversion->datatype = datatype;
4436 model_datatype_conversion->ref_to_last = ref_to_last;
4437 ccv_cnnp_model_copy_name(&model_datatype_conversion->super, name);
4438 return (ccv_cnnp_model_t*)model_datatype_conversion;
4439}
4440
4441static ccv_cnnp_model_t* _ccv_cnnp_datatype_conversion_copy(const ccv_cnnp_model_t* const super, void* const context)
4442{
4443 ccv_cnnp_model_datatype_conversion_t* const self = (ccv_cnnp_model_datatype_conversion_t*)super;
4444 return ccv_cnnp_datatype_conversion(self->datatype, self->ref_to_last, self->super.name);
4445}
4446
4447/// MARK - Clamp layer.
4448
4449typedef struct {
4450 ccv_cnnp_model_t super;
4451 ccv_nnc_tensor_symbol_t output;
4452 float min;
4453 float max;
4454} ccv_cnnp_model_clamp_t;
4455
4456static 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)
4457{
4458 ccv_cnnp_model_clamp_t* const self = (ccv_cnnp_model_clamp_t*)super;
4459 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);
4460 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4461 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", 4461, __extension__ __PRETTY_FUNCTION__
); }));
4462 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4463 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);
4464}
4465
4466static ccv_cnnp_model_t* _ccv_cnnp_clamp_copy(const ccv_cnnp_model_t* const self, void* const context);
4467
4468static const ccv_cnnp_model_vtab_t ccv_cnnp_clamp_isa = {
4469 .build = _ccv_cnnp_clamp_build,
4470 .copy = _ccv_cnnp_clamp_copy,
4471};
4472
4473ccv_cnnp_model_t* ccv_cnnp_clamp(const float min, const float max, const char* const name)
4474{
4475 ccv_cnnp_model_clamp_t* const model_clamp = (ccv_cnnp_model_clamp_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_clamp_t));
4476 model_clamp->super.isa = &ccv_cnnp_clamp_isa;
4477 model_clamp->super.input_size = 0;
4478 model_clamp->super.outputs = &model_clamp->output;
4479 model_clamp->super.output_size = 1;
4480 model_clamp->min = min;
4481 model_clamp->max = max;
4482 ccv_cnnp_model_copy_name(&model_clamp->super, name);
4483 return (ccv_cnnp_model_t*)model_clamp;
4484}
4485
4486static ccv_cnnp_model_t* _ccv_cnnp_clamp_copy(const ccv_cnnp_model_t* const super, void* const context)
4487{
4488 ccv_cnnp_model_clamp_t* const self = (ccv_cnnp_model_clamp_t*)super;
4489 return ccv_cnnp_clamp(self->min, self->max, self->super.name);
4490}
4491
4492// MARK - Parameter Layer
4493
4494typedef struct {
4495 ccv_cnnp_model_t super;
4496 float init_bound;
4497 ccv_nnc_tensor_symbol_t weights;
4498 ccv_nnc_tensor_param_t weights_params;
4499 ccv_nnc_tensor_symbol_t output;
4500} ccv_cnnp_model_parameter_t;
4501
4502static 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)
4503{
4504 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);
4505 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", 4505, __extension__ __PRETTY_FUNCTION__
); }));
4506 ccv_cnnp_model_parameter_t* const self = (ccv_cnnp_model_parameter_t*)super;
4507 if (!self->weights.graph)
4508 self->weights = ccv_nnc_tensor_symbol_new(graph, self->weights_params, "weights");
4509 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"
, 4509, __extension__ __PRETTY_FUNCTION__); }));
4510 outputs[0] = ccv_cnnp_model_get_symbol(super, self->weights);
4511}
4512
4513static 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)
4514{
4515 ccv_cnnp_model_parameter_t* const self = (ccv_cnnp_model_parameter_t*)super;
4516 if (self->init_bound > 0)
4517 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);
4518 else
4519 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);
4520}
4521
4522static 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)
4523{
4524 ccv_cnnp_model_parameter_t* const self = (ccv_cnnp_model_parameter_t*)super;
4525 add_to_array(parameters, self->weights, is_trainable);
4526}
4527
4528static ccv_cnnp_model_t* _ccv_cnnp_parameter_copy(const ccv_cnnp_model_t* const super, void* const context);
4529
4530static const ccv_cnnp_model_vtab_t ccv_cnnp_parameter_isa = {
4531 .build = _ccv_cnnp_parameter_build,
4532 .init_states = _ccv_cnnp_parameter_init_states,
4533 .add_to_parameter = _ccv_cnnp_parameter_add_to_parameter,
4534 .copy = _ccv_cnnp_parameter_copy,
4535};
4536
4537ccv_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)
4538{
4539 ccv_cnnp_model_parameter_t* const model_parameter = (ccv_cnnp_model_parameter_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_parameter_t));
4540 model_parameter->super.isa = &ccv_cnnp_parameter_isa;
4541 model_parameter->super.input_size = 0;
4542 model_parameter->super.outputs = &model_parameter->output;
4543 model_parameter->super.output_size = 1;
4544 model_parameter->super.is_trainable = is_trainable;
4545 ccv_cnnp_model_copy_name(&model_parameter->super, name);
4546 model_parameter->weights.d = CCV_NNC_NO_TENSOR_SYMBOL;
4547 model_parameter->weights.graph = 0;
4548 model_parameter->weights_params = params;
4549 return (ccv_cnnp_model_t*)model_parameter;
4550}
4551
4552static ccv_cnnp_model_t* _ccv_cnnp_parameter_copy(const ccv_cnnp_model_t* const super, void* const context)
4553{
4554 const ccv_cnnp_model_parameter_t* const self = (const ccv_cnnp_model_parameter_t*)super;
4555 return ccv_cnnp_parameter(self->weights_params, self->init_bound, self->super.is_trainable, self->super.name);
4556}
4557
4558// MARK - Scalar Layer
4559
4560typedef struct {
4561 ccv_cnnp_model_t super;
4562 int type;
4563 int format;
4564 int datatype;
4565 float value;
4566 ccv_nnc_tensor_symbol_t output;
4567} ccv_cnnp_model_scalar_t;
4568
4569static 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)
4570{
4571 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);
4572 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", 4572, __extension__ __PRETTY_FUNCTION__
); }));
4573 ccv_cnnp_model_scalar_t* const self = (ccv_cnnp_model_scalar_t*)super;
4574 ccv_nnc_tensor_param_t params = {
4575 .type = self->type,
4576 .format = self->format,
4577 .datatype = self->datatype,
4578 .dim = {
4579 1
4580 }
4581 };
4582 if (input_size > 0)
4583 {
4584 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4585 params.type = input_params.type;
4586 params.format = input_params.format;
4587 params.datatype = input_params.datatype;
4588 }
4589 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4590 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);
4591}
4592
4593static ccv_cnnp_model_t* _ccv_cnnp_scalar_copy(const ccv_cnnp_model_t* const super, void* const context);
4594
4595static const ccv_cnnp_model_vtab_t ccv_cnnp_scalar_isa = {
4596 .build = _ccv_cnnp_scalar_build,
4597 .copy = _ccv_cnnp_scalar_copy,
4598};
4599
4600ccv_cnnp_model_t* ccv_cnnp_scalar(const int type, const int format, const int datatype, const float value, const char* const name)
4601{
4602 ccv_cnnp_model_scalar_t* const model_scalar = (ccv_cnnp_model_scalar_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_scalar_t));
4603 model_scalar->super.isa = &ccv_cnnp_scalar_isa;
4604 model_scalar->super.input_size = 0;
4605 model_scalar->super.outputs = &model_scalar->output;
4606 model_scalar->super.output_size = 1;
4607 ccv_cnnp_model_copy_name(&model_scalar->super, name);
4608 model_scalar->type = type;
4609 model_scalar->format = format;
4610 model_scalar->datatype = datatype;
4611 model_scalar->value = value;
4612 return (ccv_cnnp_model_t*)model_scalar;
4613}
4614
4615static ccv_cnnp_model_t* _ccv_cnnp_scalar_copy(const ccv_cnnp_model_t* const super, void* const context)
4616{
4617 const ccv_cnnp_model_scalar_t* const self = (const ccv_cnnp_model_scalar_t*)super;
4618 return ccv_cnnp_scalar(self->type, self->format, self->datatype, self->value, self->super.name);
4619}
4620
4621// MARK - Variable Layer
4622
4623typedef struct {
4624 ccv_cnnp_model_t super;
4625 ccv_nnc_tensor_param_t params;
4626 ccv_nnc_tensor_symbol_t output;
4627} ccv_cnnp_model_variable_t;
4628
4629static 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)
4630{
4631 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);
4632 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"
, 4632, __extension__ __PRETTY_FUNCTION__); }));
4633 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", 4633, __extension__ __PRETTY_FUNCTION__
); }));
4634 ccv_cnnp_model_variable_t* const self = (ccv_cnnp_model_variable_t*)super;
4635 outputs[0] = ccv_nnc_tensor_symbol_new(graph, self->params, 0);
4636}
4637
4638static ccv_cnnp_model_t* _ccv_cnnp_variable_copy(const ccv_cnnp_model_t* const super, void* const context);
4639
4640static const ccv_cnnp_model_vtab_t ccv_cnnp_variable_isa = {
4641 .build = _ccv_cnnp_variable_build,
4642 .copy = _ccv_cnnp_variable_copy,
4643};
4644
4645ccv_cnnp_model_t* ccv_cnnp_variable(const ccv_nnc_tensor_param_t params, const char* const name)
4646{
4647 ccv_cnnp_model_variable_t* const model_variable = (ccv_cnnp_model_variable_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_variable_t));
4648 model_variable->super.isa = &ccv_cnnp_variable_isa;
4649 model_variable->super.input_size = 0;
4650 model_variable->super.outputs = &model_variable->output;
4651 model_variable->super.output_size = 1;
4652 ccv_cnnp_model_copy_name(&model_variable->super, name);
4653 model_variable->params = params;
4654 return (ccv_cnnp_model_t*)model_variable;
4655}
4656
4657static ccv_cnnp_model_t* _ccv_cnnp_variable_copy(const ccv_cnnp_model_t* const super, void* const context)
4658{
4659 const ccv_cnnp_model_variable_t* const self = (const ccv_cnnp_model_variable_t*)super;
4660 return ccv_cnnp_variable(self->params, self->super.name);
4661}
4662
4663// MARK - Set Layer
4664
4665typedef struct {
4666 ccv_cnnp_model_t super;
4667 float value;
4668 ccv_nnc_tensor_symbol_t output;
4669} ccv_cnnp_model_set_t;
4670
4671static 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)
4672{
4673 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);
4674 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"
, 4674, __extension__ __PRETTY_FUNCTION__); }));
4675 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", 4675, __extension__ __PRETTY_FUNCTION__
); }));
4676 ccv_cnnp_model_set_t* const self = (ccv_cnnp_model_set_t*)super;
4677 outputs[0] = inputs[0];
4678 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");
4679}
4680
4681static ccv_cnnp_model_t* _ccv_cnnp_set_copy(const ccv_cnnp_model_t* const super, void* const context);
4682
4683static const ccv_cnnp_model_vtab_t ccv_cnnp_set_isa = {
4684 .build = _ccv_cnnp_set_build,
4685 .copy = _ccv_cnnp_set_copy,
4686};
4687
4688ccv_cnnp_model_t* ccv_cnnp_set(const float value, const char* const name)
4689{
4690 ccv_cnnp_model_set_t* const model_set = (ccv_cnnp_model_set_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_set_t));
4691 model_set->super.isa = &ccv_cnnp_set_isa;
4692 model_set->super.input_size = 1;
4693 model_set->super.outputs = &model_set->output;
4694 model_set->super.output_size = 1;
4695 ccv_cnnp_model_copy_name(&model_set->super, name);
4696 model_set->value = value;
4697 return (ccv_cnnp_model_t*)model_set;
4698}
4699
4700static ccv_cnnp_model_t* _ccv_cnnp_set_copy(const ccv_cnnp_model_t* const super, void* const context)
4701{
4702 const ccv_cnnp_model_set_t* const self = (const ccv_cnnp_model_set_t*)super;
4703 return ccv_cnnp_set(self->value, self->super.name);
4704}
4705
4706// MARK - Send Layer
4707
4708typedef struct {
4709 ccv_cnnp_model_t super;
4710 int device_id;
4711 ccv_nnc_tensor_symbol_t output;
4712} ccv_cnnp_model_send_t;
4713
4714static 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)
4715{
4716 ccv_cnnp_model_send_t* const self = (ccv_cnnp_model_send_t*)super;
4717 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);
4718 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"
, 4718, __extension__ __PRETTY_FUNCTION__); }));
4719 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", 4719, __extension__ __PRETTY_FUNCTION__
); }));
4720 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4721 params.type = (params.type & ~0x3) | CCV_TENSOR_GPU_MEMORY;
4722 CCV_TENSOR_SET_DEVICE_ID(params.type, self->device_id)(params.type) = (((params.type) & ~0xfff00) | (((self->
device_id) & 0xfff) << 8));
4723 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4724 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");
4725}
4726
4727static ccv_cnnp_model_t* _ccv_cnnp_send_copy(const ccv_cnnp_model_t* const super, void* const context);
4728
4729static const ccv_cnnp_model_vtab_t ccv_cnnp_send_isa = {
4730 .build = _ccv_cnnp_send_build,
4731 .copy = _ccv_cnnp_send_copy,
4732};
4733
4734ccv_cnnp_model_t* ccv_cnnp_send(const int device_id, const char* const name)
4735{
4736 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", 4736, __extension__ __PRETTY_FUNCTION__
); }));
4737 ccv_cnnp_model_send_t* const model_send = (ccv_cnnp_model_send_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_send_t));
4738 model_send->super.isa = &ccv_cnnp_send_isa;
4739 model_send->super.input_size = 1;
4740 model_send->super.outputs = &model_send->output;
4741 model_send->super.output_size = 1;
4742 ccv_cnnp_model_copy_name(&model_send->super, name);
4743 model_send->device_id = device_id;
4744 return (ccv_cnnp_model_t*)model_send;
4745}
4746
4747static ccv_cnnp_model_t* _ccv_cnnp_send_copy(const ccv_cnnp_model_t* const super, void* const context)
4748{
4749 const ccv_cnnp_model_send_t* const self = (const ccv_cnnp_model_send_t*)super;
4750 return ccv_cnnp_send(self->device_id, self->super.name);
4751}
4752
4753// MARK - Move Layer
4754
4755typedef struct {
4756 ccv_cnnp_model_t super;
4757 ccv_nnc_tensor_symbol_t output;
4758} ccv_cnnp_model_move_t;
4759
4760static 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)
4761{
4762 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);
4763 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"
, 4763, __extension__ __PRETTY_FUNCTION__); }));
4764 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", 4764, __extension__ __PRETTY_FUNCTION__
); }));
4765 outputs[0] = inputs[1];
4766 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");
4767}
4768
4769static ccv_cnnp_model_t* _ccv_cnnp_move_copy(const ccv_cnnp_model_t* const super, void* const context);
4770
4771static const ccv_cnnp_model_vtab_t ccv_cnnp_move_isa = {
4772 .build = _ccv_cnnp_move_build,
4773 .copy = _ccv_cnnp_move_copy,
4774};
4775
4776ccv_cnnp_model_t* ccv_cnnp_move(const char* const name)
4777{
4778 ccv_cnnp_model_move_t* const model_move = (ccv_cnnp_model_move_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_move_t));
4779 model_move->super.isa = &ccv_cnnp_move_isa;
4780 model_move->super.input_size = 2;
4781 model_move->super.outputs = &model_move->output;
4782 model_move->super.output_size = 1;
4783 ccv_cnnp_model_copy_name(&model_move->super, name);
4784 return (ccv_cnnp_model_t*)model_move;
4785}
4786
4787static ccv_cnnp_model_t* _ccv_cnnp_move_copy(const ccv_cnnp_model_t* const super, void* const context)
4788{
4789 const ccv_cnnp_model_move_t* const self = (const ccv_cnnp_model_move_t*)super;
4790 return ccv_cnnp_move(self->super.name);
4791}
4792
4793// MARK - "Making" Contiguous Layer
4794
4795typedef struct {
4796 ccv_cnnp_model_t super;
4797 ccv_nnc_tensor_symbol_t output;
4798} ccv_cnnp_model_contiguous_t;
4799
4800static 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)
4801{
4802 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);
4803 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"
, 4803, __extension__ __PRETTY_FUNCTION__); }));
4804 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", 4804, __extension__ __PRETTY_FUNCTION__
); }));
4805 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4806 ccv_nnc_tensor_symbol_t to = ccv_nnc_tensor_symbol_alias_to(graph, inputs[0]);
4807 if (to.d == CCV_NNC_NO_TENSOR_SYMBOL) // If we are not reshape an alias, it is straightforward.
4808 {
4809 outputs[0] = inputs[0];
4810 return;
4811 }
4812 // Otherwise, we need to check its stride to know if it is contiguous.
4813 int old_stride[CCV_NNC_MAX_DIM_ALLOC(12)];
4814 ccv_nnc_tensor_symbol_alias_params(graph, inputs[0], 0, old_stride);
4815 // We identify permute by checking if the stride is not in descending order.
4816 // This also covered "permute" through reshape, rather than using ccv_cnnp_permute directly.
4817 if (ccv_nnc_is_tensor_stride_packed(old_stride, params.dim))
4818 {
4819 outputs[0] = inputs[0];
4820 return;
4821 }
4822 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4823 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");
4824 ccv_nnc_graph_exec_symbol_set_flags(graph, make_contiguous, CCV_NNC_GRAPH_EXEC_DISABLE_OPT);
4825}
4826
4827static ccv_cnnp_model_t* _ccv_cnnp_contiguous_copy(const ccv_cnnp_model_t* const super, void* const context);
4828
4829static const ccv_cnnp_model_vtab_t ccv_cnnp_contiguous_isa = {
4830 .build = _ccv_cnnp_contiguous_build,
4831 .copy = _ccv_cnnp_contiguous_copy,
4832};
4833
4834ccv_cnnp_model_t* ccv_cnnp_contiguous(const char* const name)
4835{
4836 ccv_cnnp_model_contiguous_t* const model_contiguous = (ccv_cnnp_model_contiguous_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_contiguous_t));
4837 model_contiguous->super.isa = &ccv_cnnp_contiguous_isa;
4838 model_contiguous->super.input_size = 1;
4839 model_contiguous->super.outputs = &model_contiguous->output;
4840 model_contiguous->super.output_size = 1;
4841 ccv_cnnp_model_copy_name(&model_contiguous->super, name);
4842 return (ccv_cnnp_model_t*)model_contiguous;
4843}
4844
4845static ccv_cnnp_model_t* _ccv_cnnp_contiguous_copy(const ccv_cnnp_model_t* const super, void* const context)
4846{
4847 const ccv_cnnp_model_contiguous_t* const self = (const ccv_cnnp_model_contiguous_t*)super;
4848 return ccv_cnnp_contiguous(self->super.name);
4849}
4850
4851// MARK - "Making" Copy Layer
4852
4853typedef struct {
4854 ccv_cnnp_model_t super;
4855 ccv_nnc_tensor_symbol_t output;
4856} ccv_cnnp_model_copy_t;
4857
4858static 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)
4859{
4860 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);
4861 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"
, 4861, __extension__ __PRETTY_FUNCTION__); }));
4862 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", 4862, __extension__ __PRETTY_FUNCTION__
); }));
4863 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4864 ccv_nnc_tensor_symbol_t to = ccv_nnc_tensor_symbol_alias_to(graph, inputs[0]);
4865 if (to.d == CCV_NNC_NO_TENSOR_SYMBOL) // If we are not reshape an alias, it is straightforward.
4866 {
4867 outputs[0] = inputs[0];
4868 return;
4869 }
4870 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4871 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");
4872 ccv_nnc_graph_exec_symbol_set_flags(graph, make_contiguous, CCV_NNC_GRAPH_EXEC_DISABLE_OPT);
4873}
4874
4875static ccv_cnnp_model_t* _ccv_cnnp_copy_copy(const ccv_cnnp_model_t* const super, void* const context);
4876
4877static const ccv_cnnp_model_vtab_t ccv_cnnp_copy_isa = {
4878 .build = _ccv_cnnp_copy_build,
4879 .copy = _ccv_cnnp_copy_copy,
4880};
4881
4882ccv_cnnp_model_t* ccv_cnnp_copy(const char* const name)
4883{
4884 ccv_cnnp_model_copy_t* const model_copy = (ccv_cnnp_model_copy_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_copy_t));
4885 model_copy->super.isa = &ccv_cnnp_copy_isa;
4886 model_copy->super.input_size = 1;
4887 model_copy->super.outputs = &model_copy->output;
4888 model_copy->super.output_size = 1;
4889 ccv_cnnp_model_copy_name(&model_copy->super, name);
4890 return (ccv_cnnp_model_t*)model_copy;
4891}
4892
4893static ccv_cnnp_model_t* _ccv_cnnp_copy_copy(const ccv_cnnp_model_t* const super, void* const context)
4894{
4895 const ccv_cnnp_model_copy_t* const self = (const ccv_cnnp_model_copy_t*)super;
4896 return ccv_cnnp_copy(self->super.name);
4897}
4898
4899// MARK - All-To-All Layer
4900
4901typedef struct {
4902 ccv_cnnp_model_t super;
4903 int axis;
4904 ccv_nnc_tensor_symbol_t outputs[1];
4905} ccv_cnnp_model_all_to_all_t;
4906
4907static 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)
4908{
4909 ccv_cnnp_model_all_to_all_t* const self = (ccv_cnnp_model_all_to_all_t*)super;
4910 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);
4911 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", 4911, __extension__ __PRETTY_FUNCTION__
); }));
4912 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", 4912, __extension__ __PRETTY_FUNCTION__
); }));
4913 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", 4913, __extension__ __PRETTY_FUNCTION__
); }));
4914 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", 4914, __extension__ __PRETTY_FUNCTION__
); }));
4915 const ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4916 const int nd = ccv_nnc_tensor_nd(input_params.dim);
4917 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", 4917, __extension__ __PRETTY_FUNCTION__
); }));
4918 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", 4918, __extension__ __PRETTY_FUNCTION__
); }));
4919 int i;
4920 for (i = 0; i < input_size; i++)
4921 {
4922 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
4923 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", 4923, __extension__ __PRETTY_FUNCTION__
); }));
4924 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", 4924, __extension__ __PRETTY_FUNCTION__
); }));
4925 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", 4925, __extension__ __PRETTY_FUNCTION__
); }));
4926 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", 4926, __extension__ __PRETTY_FUNCTION__
); }));
4927 outputs[i] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4928 }
4929 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");
4930}
4931
4932static ccv_cnnp_model_t* _ccv_cnnp_all_to_all_copy(const ccv_cnnp_model_t* const super, void* const context);
4933
4934static const ccv_cnnp_model_vtab_t ccv_cnnp_all_to_all_isa = {
4935 .build = _ccv_cnnp_all_to_all_build,
4936 .copy = _ccv_cnnp_all_to_all_copy,
4937};
4938
4939ccv_cnnp_model_t* ccv_cnnp_all_to_all(const int count, const int axis, const char* const name)
4940{
4941 assert(count > 0)((void) sizeof ((count > 0) ? 1 : 0), __extension__ ({ if (
count > 0) ; else __assert_fail ("count > 0", "ccv_cnnp_model_addons.c"
, 4941, __extension__ __PRETTY_FUNCTION__); }));
4942 assert(axis >= 0)((void) sizeof ((axis >= 0) ? 1 : 0), __extension__ ({ if (
axis >= 0) ; else __assert_fail ("axis >= 0", "ccv_cnnp_model_addons.c"
, 4942, __extension__ __PRETTY_FUNCTION__); }));
4943 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));
4944 model_all_to_all->super.isa = &ccv_cnnp_all_to_all_isa;
4945 model_all_to_all->super.input_size = count;
4946 model_all_to_all->super.outputs = model_all_to_all->outputs;
4947 model_all_to_all->super.output_size = count;
4948 model_all_to_all->axis = axis;
4949 ccv_cnnp_model_copy_name(&model_all_to_all->super, name);
4950 return (ccv_cnnp_model_t*)model_all_to_all;
4951}
4952
4953static ccv_cnnp_model_t* _ccv_cnnp_all_to_all_copy(const ccv_cnnp_model_t* const super, void* const context)
4954{
4955 const ccv_cnnp_model_all_to_all_t* const self = (const ccv_cnnp_model_all_to_all_t*)super;
4956 return ccv_cnnp_all_to_all(self->super.output_size, self->axis, self->super.name);
4957}
4958
4959// MARK - Scaled-Dot Product Attention Layer
4960
4961typedef struct {
4962 ccv_cnnp_model_t super;
4963 ccv_nnc_tensor_symbol_t output;
4964 ccv_nnc_tensor_symbol_t weights;
4965 ccv_nnc_tensor_symbol_t bias;
4966 float scale;
4967 int is_causal;
4968 int has_attn_mask;
4969 int is_varlen;
4970 int max_seqlen_q;
4971 int max_seqlen_kv;
4972 int flags;
4973 int attention_sinks;
4974 int sliding_window;
4975 int fused_unify_head_weights;
4976 int no_bias;
4977} ccv_cnnp_model_scaled_dot_product_attention_t;
4978
4979static 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)
4980{
4981 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);
4982 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", 4982, __extension__ __PRETTY_FUNCTION__
); }));
4983 ccv_cnnp_model_scaled_dot_product_attention_t* const self = (ccv_cnnp_model_scaled_dot_product_attention_t*)super;
4984 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", 4984, __extension__ __PRETTY_FUNCTION__
); }));
4985 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", 4985, __extension__ __PRETTY_FUNCTION__
); }));
4986 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", 4986, __extension__ __PRETTY_FUNCTION__
); }));
4987 const ccv_nnc_tensor_param_t q_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4988 const ccv_nnc_tensor_param_t k_params = ccv_nnc_tensor_symbol_params(graph, inputs[1]);
4989 const ccv_nnc_tensor_param_t v_params = ccv_nnc_tensor_symbol_params(graph, inputs[2]);
4990 const int sink_idx = self->is_varlen ? 5 : (3 + self->has_attn_mask);
4991 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){};
4992 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){};
4993 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){};
4994 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){};
4995 const int v_nd = ccv_nnc_tensor_nd(v_params.dim);
4996 assert(v_nd == 3 || v_nd == 4)((void) sizeof ((v_nd == 3 || v_nd == 4) ? 1 : 0), __extension__
({ if (v_nd == 3 || v_nd == 4) ; else __assert_fail ("v_nd == 3 || v_nd == 4"
, "ccv_cnnp_model_addons.c", 4996, __extension__ __PRETTY_FUNCTION__
); }));
4997 const int hEv = (v_nd == 3 ? 1 : v_params.dim[2]) * v_params.dim[v_nd - 1];
4998 ccv_nnc_tensor_param_t weights_params = q_params;
4999 memset(weights_params.dim, 0, sizeof(weights_params.dim));
5000 weights_params.dim[0] = hEv;
5001 weights_params.dim[1] = hEv;
5002 ccv_nnc_tensor_param_t bias_params = q_params;
5003 memset(bias_params.dim, 0, sizeof(bias_params.dim));
5004 bias_params.dim[0] = hEv;
5005 ccv_nnc_cmd_t cmd = {0};
5006 cmd.cmd = CCV_NNC_SCALED_DOT_PRODUCT_ATTENTION_FORWARD;
5007 cmd.info.scaled_dot_product_attention.scale = self->scale;
5008 cmd.info.scaled_dot_product_attention.is_causal = self->is_causal;
5009 cmd.info.scaled_dot_product_attention.is_varlen = self->is_varlen;
5010 cmd.info.scaled_dot_product_attention.max_seqlen_q = self->max_seqlen_q;
5011 cmd.info.scaled_dot_product_attention.max_seqlen_kv = self->max_seqlen_kv;
5012 cmd.info.scaled_dot_product_attention.flags = self->flags;
5013 cmd.info.scaled_dot_product_attention.attention_sinks = self->attention_sinks;
5014 cmd.info.scaled_dot_product_attention.sliding_window = self->sliding_window;
5015 ccv_nnc_tensor_param_t output_params[3];
5016 ccv_nnc_tensor_symbol_t output;
5017 ccv_nnc_tensor_symbol_t saved_softmax_lse;
5018 ccv_nnc_tensor_symbol_t saved_v_proj = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
5019 ccv_nnc_tensor_symbol_t attn_mask = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
5020 ccv_nnc_tensor_symbol_t weights = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
5021 ccv_nnc_tensor_symbol_t bias = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
5022 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
};
5023 if (self->has_attn_mask)
5024 attn_mask = inputs[3];
5025 if (self->is_varlen)
5026 {
5027 ccv_nnc_tensor_param_t input_params[9] = {
5028 q_params,
5029 k_params,
5030 v_params,
5031 (ccv_nnc_tensor_param_t){},
5032 (ccv_nnc_tensor_param_t){},
5033 (ccv_nnc_tensor_param_t){},
5034 q_seq_offsets_params,
5035 kv_seq_offsets_params,
5036 sinks_params,
5037 };
5038 ccv_nnc_hint_tensor_auto(cmd, input_params, self->attention_sinks ? 9 : 8, ccv_nnc_no_hint, output_params, 2);
5039 output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0);
5040 saved_softmax_lse = ccv_nnc_tensor_symbol_new(graph, output_params[1], 0);
5041 if (self->attention_sinks)
5042 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");
5043 else
5044 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");
5045 } else if (self->fused_unify_head_weights)
5046 {
5047 if (!self->weights.graph)
5048 self->weights = ccv_nnc_tensor_symbol_new(graph, weights_params, "weights");
5049 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"
, 5049, __extension__ __PRETTY_FUNCTION__); }));
5050 weights = ccv_cnnp_model_get_symbol(super, self->weights);
5051 if (!self->no_bias)
5052 {
5053 if (!self->bias.graph)
5054 self->bias = ccv_nnc_tensor_symbol_new(graph, bias_params, "bias");
5055 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", 5055, __extension__ __PRETTY_FUNCTION__
); }));
5056 bias = ccv_cnnp_model_get_symbol(super, self->bias);
5057 }
5058 ccv_nnc_tensor_param_t input_params[9] = {
5059 q_params,
5060 k_params,
5061 v_params,
5062 attn_mask_params,
5063 weights_params,
5064 bias_params,
5065 (ccv_nnc_tensor_param_t){},
5066 (ccv_nnc_tensor_param_t){},
5067 sinks_params,
5068 };
5069 ccv_nnc_hint_tensor_auto(cmd, input_params, self->attention_sinks ? 9 : 6, ccv_nnc_no_hint, output_params, 3);
5070 output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0);
5071 saved_softmax_lse = ccv_nnc_tensor_symbol_new(graph, output_params[1], 0);
5072 saved_v_proj = ccv_nnc_tensor_symbol_new(graph, output_params[2], 0);
5073 if (self->attention_sinks)
5074 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");
5075 else
5076 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");
5077 } else {
5078 ccv_nnc_tensor_param_t input_params[9] = {
5079 q_params,
5080 k_params,
5081 v_params,
5082 attn_mask_params,
5083 (ccv_nnc_tensor_param_t){},
5084 (ccv_nnc_tensor_param_t){},
5085 (ccv_nnc_tensor_param_t){},
5086 (ccv_nnc_tensor_param_t){},
5087 sinks_params,
5088 };
5089 ccv_nnc_hint_tensor_auto(cmd, input_params, self->attention_sinks ? 9 : 3, ccv_nnc_no_hint, output_params, 2);
5090 output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0);
5091 saved_softmax_lse = ccv_nnc_tensor_symbol_new(graph, output_params[1], 0);
5092 if (self->attention_sinks)
5093 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");
5094 else
5095 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");
5096 }
5097 outputs[0] = output;
5098}
5099
5100static 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)
5101{
5102 ccv_cnnp_model_scaled_dot_product_attention_t* const self = (ccv_cnnp_model_scaled_dot_product_attention_t*)super;
5103 if (self->weights.graph)
5104 {
5105 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"
, 5105, __extension__ __PRETTY_FUNCTION__); }));
5106 const ccv_nnc_tensor_param_t weight_params = ccv_nnc_tensor_symbol_params(graph, self->weights);
5107 const int c = weight_params.dim[1];
5108 const float std = sqrtf(2) / sqrtf(c);
5109 const float bound = sqrtf(3) * std;
5110 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);
5111 if (self->bias.graph)
5112 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);
5113 }
5114}
5115
5116static 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)
5117{
5118 ccv_cnnp_model_scaled_dot_product_attention_t* const self = (ccv_cnnp_model_scaled_dot_product_attention_t*)super;
5119 if (self->weights.graph)
5120 {
5121 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"
, 5121, __extension__ __PRETTY_FUNCTION__); }));
5122 add_to_array(parameters, self->weights, is_trainable);
5123 if (self->bias.graph)
5124 add_to_array(parameters, self->bias, is_trainable);
5125 }
5126}
5127
5128static ccv_cnnp_model_t* _ccv_cnnp_scaled_dot_product_attention_copy(const ccv_cnnp_model_t* const super, void* const context);
5129
5130static const ccv_cnnp_model_vtab_t ccv_cnnp_scaled_dot_product_attention_isa = {
5131 .build = _ccv_cnnp_scaled_dot_product_attention_build,
5132 .copy = _ccv_cnnp_scaled_dot_product_attention_copy,
5133};
5134
5135static const ccv_cnnp_model_vtab_t ccv_cnnp_scaled_dot_product_attention_fused_isa = {
5136 .build = _ccv_cnnp_scaled_dot_product_attention_build,
5137 .init_states = _ccv_cnnp_scaled_dot_product_attention_init_states,
5138 .add_to_parameter = _ccv_cnnp_scaled_dot_product_attention_add_to_parameter,
5139 .copy = _ccv_cnnp_scaled_dot_product_attention_copy,
5140};
5141
5142ccv_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 attention_sinks, const int sliding_window, const int fused_unify_head_weights, const int no_bias, const int is_trainable, const char* const name)
5143{
5144 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", 5144, __extension__ __PRETTY_FUNCTION__
); }));
5145 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", 5145, __extension__ __PRETTY_FUNCTION__
); }));
5146 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"
, 5146, __extension__ __PRETTY_FUNCTION__); }));
5147 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"
, 5147, __extension__ __PRETTY_FUNCTION__); }));
5148 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));
5149 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;
5150 const int has_attention_sinks = attention_sinks ? 1 : 0;
5151 model_scaled_dot_product_attention->super.input_size = (is_varlen ? 5 : (has_attn_mask ? 4 : 3)) + has_attention_sinks;
5152 model_scaled_dot_product_attention->super.outputs = &model_scaled_dot_product_attention->output;
5153 model_scaled_dot_product_attention->super.output_size = 1;
5154 model_scaled_dot_product_attention->super.is_trainable = is_trainable;
5155 ccv_cnnp_model_copy_name(&model_scaled_dot_product_attention->super, name);
5156 model_scaled_dot_product_attention->weights.d = CCV_NNC_NO_TENSOR_SYMBOL;
5157 model_scaled_dot_product_attention->weights.graph = 0;
5158 model_scaled_dot_product_attention->bias.d = CCV_NNC_NO_TENSOR_SYMBOL;
5159 model_scaled_dot_product_attention->bias.graph = 0;
5160 model_scaled_dot_product_attention->scale = scale;
5161 model_scaled_dot_product_attention->is_causal = is_causal;
5162 model_scaled_dot_product_attention->has_attn_mask = has_attn_mask;
5163 model_scaled_dot_product_attention->is_varlen = is_varlen;
5164 model_scaled_dot_product_attention->max_seqlen_q = max_seqlen_q;
5165 model_scaled_dot_product_attention->max_seqlen_kv = max_seqlen_kv;
5166 model_scaled_dot_product_attention->flags = flags;
5167 model_scaled_dot_product_attention->attention_sinks = has_attention_sinks;
5168 model_scaled_dot_product_attention->sliding_window = sliding_window;
5169 model_scaled_dot_product_attention->fused_unify_head_weights = fused_unify_head_weights;
5170 model_scaled_dot_product_attention->no_bias = no_bias;
5171 return (ccv_cnnp_model_t*)model_scaled_dot_product_attention;
5172}
5173
5174static ccv_cnnp_model_t* _ccv_cnnp_scaled_dot_product_attention_copy(const ccv_cnnp_model_t* const super, void* const context)
5175{
5176 const ccv_cnnp_model_scaled_dot_product_attention_t* const self = (const ccv_cnnp_model_scaled_dot_product_attention_t*)super;
5177 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->attention_sinks, self->sliding_window, self->fused_unify_head_weights, self->no_bias, self->super.is_trainable, self->super.name);
5178}
5179
5180// MARK - Debug Layer
5181
5182typedef struct {
5183 ccv_cnnp_model_t super;
5184 ccv_nnc_tensor_symbol_t output;
5185 ccv_cnnp_model_debug_f debugger;
5186 ccv_cnnp_model_debug_context_deinit_f debug_deinit;
5187 ccv_cnnp_model_debug_context_copy_f debug_copy;
5188 void* debug_context;
5189} ccv_cnnp_model_debug_t;
5190
5191static 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)
5192{
5193 if (cmd.cmd == CCV_NNC_CUSTOM_BACKWARD)
5194 {
5195 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", 5195, __extension__ __PRETTY_FUNCTION__
); }));
5196 }
5197 ccv_cnnp_model_debug_t* const self = (ccv_cnnp_model_debug_t*)cmd.data;
5198 self->debugger(inputs, input_size, stream_context, self->debug_context);
5199 return CCV_NNC_EXEC_SUCCESS;
5200}
5201
5202static ccv_nnc_cmd_vtab_t ccv_cnnp_debug_exec_isa = {
5203 .exec = _ccv_cnnp_debug_exec
5204};
5205
5206static 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)
5207{
5208 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);
5209 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", 5209, __extension__ __PRETTY_FUNCTION__
); }));
5210 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", 5210, __extension__ __PRETTY_FUNCTION__
); }));
5211 ccv_nnc_tensor_symbol_t to = ccv_nnc_tensor_symbol_alias_to(graph, inputs[0]);
5212 ccv_nnc_tensor_param_t output_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
5213 if (to.d == CCV_NNC_NO_TENSOR_SYMBOL) // If we are not reshape an alias, it is straightforward.
5214 {
5215 int ofs[CCV_NNC_MAX_DIM_ALLOC(12)] = {0};
5216 int stride[CCV_NNC_MAX_DIM_ALLOC(12)];
5217 ccv_nnc_tensor_get_stride(output_params.dim, stride);
5218 outputs[0] = ccv_nnc_tensor_symbol_alias_new(graph, inputs[0], ofs, stride, output_params, 0);
5219 } else {
5220 int old_ofs[CCV_NNC_MAX_DIM_ALLOC(12)];
5221 int old_stride[CCV_NNC_MAX_DIM_ALLOC(12)];
5222 ccv_nnc_tensor_symbol_alias_params(graph, inputs[0], old_ofs, old_stride);
5223 outputs[0] = ccv_nnc_tensor_symbol_alias_new(graph, to, old_ofs, old_stride, output_params, 0);
5224 }
5225 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);
5226 cmd.data = self;
5227 ccv_nnc_graph_exec_symbol_t make_debug = ccv_nnc_graph_exec_symbol_new(graph, cmd, inputs, input_size, outputs, 1, "debug");
5228 // Disable any optimizations.
5229 ccv_nnc_graph_exec_symbol_set_flags(graph, make_debug, CCV_NNC_GRAPH_EXEC_DISABLE_OPT);
5230}
5231
5232static void _ccv_cnnp_debug_deinit(ccv_cnnp_model_t* const super)
5233{
5234 const ccv_cnnp_model_debug_t* const self = (const ccv_cnnp_model_debug_t*)super;
5235 if (self->debug_deinit && self->debug_context)
5236 self->debug_deinit(self->debug_context);
5237}
5238
5239static ccv_cnnp_model_t* _ccv_cnnp_debug_copy(const ccv_cnnp_model_t* const super, void* const context);
5240
5241static const ccv_cnnp_model_vtab_t ccv_cnnp_debug_isa = {
5242 .build = _ccv_cnnp_debug_build,
5243 .deinit = _ccv_cnnp_debug_deinit,
5244 .copy = _ccv_cnnp_debug_copy,
5245};
5246
5247ccv_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)
5248{
5249 ccv_cnnp_model_debug_t* const model_debug = (ccv_cnnp_model_debug_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_debug_t));
5250 model_debug->super.isa = &ccv_cnnp_debug_isa;
5251 model_debug->super.input_size = 0;
5252 model_debug->super.outputs = &model_debug->output;
5253 model_debug->super.output_size = 1;
5254 model_debug->debugger = func;
5255 model_debug->debug_context = context;
5256 model_debug->debug_deinit = deinit;
5257 model_debug->debug_copy = copy;
5258 ccv_cnnp_model_copy_name(&model_debug->super, name);
5259 return (ccv_cnnp_model_t*)model_debug;
5260}
5261
5262static ccv_cnnp_model_t* _ccv_cnnp_debug_copy(const ccv_cnnp_model_t* const super, void* const context)
5263{
5264 const ccv_cnnp_model_debug_t* const self = (const ccv_cnnp_model_debug_t*)super;
5265 void* debug_context = self->debug_context;
5266 if (self->debug_copy && self->debug_context)
5267 debug_context = self->debug_copy(self->debug_context);
5268 return ccv_cnnp_debug(self->debugger, debug_context, self->debug_deinit, self->debug_copy, self->super.name);
5269}
5270
5271/// MARK - Sort layer.
5272
5273typedef struct {
5274 ccv_cnnp_model_t super;
5275 ccv_nnc_tensor_symbol_t outputs[2];
5276 int along_axis;
5277 int descending;
5278} ccv_cnnp_model_sort_t;
5279
5280static 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)
5281{
5282 ccv_cnnp_model_sort_t* const self = (ccv_cnnp_model_sort_t*)super;
5283 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);
5284 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
5285 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", 5285, __extension__ __PRETTY_FUNCTION__
); }));
5286 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
5287 params.datatype = CCV_32S;
5288 outputs[1] = ccv_nnc_tensor_symbol_new(graph, params, 0);
5289 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");
5290}
5291
5292static ccv_cnnp_model_t* _ccv_cnnp_sort_copy(const ccv_cnnp_model_t* const self, void* const context);
5293
5294static const ccv_cnnp_model_vtab_t ccv_cnnp_sort_isa = {
5295 .build = _ccv_cnnp_sort_build,
5296 .copy = _ccv_cnnp_sort_copy,
5297};
5298
5299ccv_cnnp_model_t* ccv_cnnp_sort(const int along_axis, const int descending, const char* const name)
5300{
5301 ccv_cnnp_model_sort_t* const model_sort = (ccv_cnnp_model_sort_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_sort_t));
5302 model_sort->super.isa = &ccv_cnnp_sort_isa;
5303 model_sort->super.input_size = 0;
5304 model_sort->super.outputs = model_sort->outputs;
5305 model_sort->super.output_size = 2;
5306 model_sort->along_axis = along_axis;
5307 model_sort->descending = descending;
5308 ccv_cnnp_model_copy_name(&model_sort->super, name);
5309 return (ccv_cnnp_model_t*)model_sort;
5310}
5311
5312static ccv_cnnp_model_t* _ccv_cnnp_sort_copy(const ccv_cnnp_model_t* const super, void* const context)
5313{
5314 ccv_cnnp_model_sort_t* const self = (ccv_cnnp_model_sort_t*)super;
5315 return ccv_cnnp_sort(self->along_axis, self->descending, self->super.name);
5316}
5317
5318/// MARK - Partition layer.
5319
5320typedef struct {
5321 ccv_cnnp_model_t super;
5322 ccv_nnc_tensor_symbol_t outputs[2];
5323 int kth;
5324 int along_axis;
5325 int descending;
5326} ccv_cnnp_model_partition_t;
5327
5328static 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)
5329{
5330 ccv_cnnp_model_partition_t* const self = (ccv_cnnp_model_partition_t*)super;
5331 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);
5332 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
5333 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", 5333, __extension__ __PRETTY_FUNCTION__
); }));
5334 if (self->kth > 0)
5335 params.dim[self->along_axis] = self->kth;
5336 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
5337 params.datatype = CCV_32S;
5338 outputs[1] = ccv_nnc_tensor_symbol_new(graph, params, 0);
5339 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");
5340}
5341
5342static ccv_cnnp_model_t* _ccv_cnnp_partition_copy(const ccv_cnnp_model_t* const self, void* const context);
5343
5344static const ccv_cnnp_model_vtab_t ccv_cnnp_partition_isa = {
5345 .build = _ccv_cnnp_partition_build,
5346 .copy = _ccv_cnnp_partition_copy,
5347};
5348
5349ccv_cnnp_model_t* ccv_cnnp_partition(const int kth, const int along_axis, const int descending, const char* const name)
5350{
5351 ccv_cnnp_model_partition_t* const model_partition = (ccv_cnnp_model_partition_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_partition_t));
5352 model_partition->super.isa = &ccv_cnnp_partition_isa;
5353 model_partition->super.input_size = 0;
5354 model_partition->super.outputs = model_partition->outputs;
5355 model_partition->super.output_size = 2;
5356 model_partition->kth = kth;
5357 model_partition->along_axis = along_axis;
5358 model_partition->descending = descending;
5359 ccv_cnnp_model_copy_name(&model_partition->super, name);
5360 return (ccv_cnnp_model_t*)model_partition;
5361}
5362
5363static ccv_cnnp_model_t* _ccv_cnnp_partition_copy(const ccv_cnnp_model_t* const super, void* const context)
5364{
5365 ccv_cnnp_model_partition_t* const self = (ccv_cnnp_model_partition_t*)super;
5366 return ccv_cnnp_partition(self->kth, self->along_axis, self->descending, self->super.name);
5367}
5368
5369/// MARK - Unique consecutive layer.
5370
5371typedef struct {
5372 ccv_cnnp_model_t super;
5373 ccv_nnc_tensor_symbol_t outputs[2];
5374 int bincount;
5375} ccv_cnnp_model_unique_consecutive_t;
5376
5377static 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)
5378{
5379 ccv_cnnp_model_unique_consecutive_t* const self = (ccv_cnnp_model_unique_consecutive_t*)super;
5380 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);
5381 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
5382 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", 5382, __extension__ __PRETTY_FUNCTION__
); }));
5383 if (self->bincount > 0)
5384 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; }
);
5385 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
5386 params.datatype = CCV_32S;
5387 outputs[1] = ccv_nnc_tensor_symbol_new(graph, params, 0);
5388 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");
5389}
5390
5391static ccv_cnnp_model_t* _ccv_cnnp_unique_consecutive_copy(const ccv_cnnp_model_t* const self, void* const context);
5392
5393static const ccv_cnnp_model_vtab_t ccv_cnnp_unique_consecutive_isa = {
5394 .build = _ccv_cnnp_unique_consecutive_build,
5395 .copy = _ccv_cnnp_unique_consecutive_copy,
5396};
5397
5398ccv_cnnp_model_t* ccv_cnnp_unique_consecutive(const int bincount, const char* const name)
5399{
5400 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));
5401 model_unique_consecutive->super.isa = &ccv_cnnp_unique_consecutive_isa;
5402 model_unique_consecutive->super.input_size = 0;
5403 model_unique_consecutive->super.outputs = model_unique_consecutive->outputs;
5404 model_unique_consecutive->super.output_size = 2;
5405 model_unique_consecutive->bincount = bincount;
5406 ccv_cnnp_model_copy_name(&model_unique_consecutive->super, name);
5407 return (ccv_cnnp_model_t*)model_unique_consecutive;
5408}
5409
5410static ccv_cnnp_model_t* _ccv_cnnp_unique_consecutive_copy(const ccv_cnnp_model_t* const super, void* const context)
5411{
5412 ccv_cnnp_model_unique_consecutive_t* const self = (ccv_cnnp_model_unique_consecutive_t*)super;
5413 return ccv_cnnp_unique_consecutive(self->bincount, self->super.name);
5414}
5415
5416/// MARK - Scatter add layer.
5417
5418typedef struct {
5419 ccv_cnnp_model_t super;
5420 ccv_nnc_tensor_symbol_t output;
5421 int bincount;
5422 int count_per_output;
5423} ccv_cnnp_model_scatter_add_t;
5424
5425static 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)
5426{
5427 ccv_cnnp_model_scatter_add_t* const self = (ccv_cnnp_model_scatter_add_t*)super;
5428 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);
5429 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
5430 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", 5430, __extension__ __PRETTY_FUNCTION__
); }));
5431 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", 5431, __extension__ __PRETTY_FUNCTION__
); }));
5432 if (params.dim[0] != 0)
5433 params.dim[0] = self->bincount;
5434 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
5435 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");
5436}
5437
5438static ccv_cnnp_model_t* _ccv_cnnp_scatter_add_copy(const ccv_cnnp_model_t* const self, void* const context);
5439
5440static const ccv_cnnp_model_vtab_t ccv_cnnp_scatter_add_isa = {
5441 .build = _ccv_cnnp_scatter_add_build,
5442 .copy = _ccv_cnnp_scatter_add_copy,
5443};
5444
5445ccv_cnnp_model_t* ccv_cnnp_scatter_add(const int bincount, const int count_per_output, const char* const name)
5446{
5447 assert(bincount >= 0)((void) sizeof ((bincount >= 0) ? 1 : 0), __extension__ ({
if (bincount >= 0) ; else __assert_fail ("bincount >= 0"
, "ccv_cnnp_model_addons.c", 5447, __extension__ __PRETTY_FUNCTION__
); }));
5448 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", 5448, __extension__ __PRETTY_FUNCTION__
); }));
5449 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));
5450 model_scatter_add->super.isa = &ccv_cnnp_scatter_add_isa;
5451 model_scatter_add->super.input_size = 0;
5452 model_scatter_add->super.outputs = &model_scatter_add->output;
5453 model_scatter_add->super.output_size = 1;
5454 model_scatter_add->bincount = bincount;
5455 model_scatter_add->count_per_output = count_per_output;
5456 ccv_cnnp_model_copy_name(&model_scatter_add->super, name);
5457 return (ccv_cnnp_model_t*)model_scatter_add;
5458}
5459
5460static ccv_cnnp_model_t* _ccv_cnnp_scatter_add_copy(const ccv_cnnp_model_t* const super, void* const context)
5461{
5462 ccv_cnnp_model_scatter_add_t* const self = (ccv_cnnp_model_scatter_add_t*)super;
5463 return ccv_cnnp_scatter_add(self->bincount, self->count_per_output, self->super.name);
5464}
5465
5466// MARK - Segmented Dense Layer
5467
5468typedef struct {
5469 ccv_cnnp_model_t super;
5470 ccv_nnc_tensor_symbol_t output;
5471 ccv_nnc_tensor_symbol_t weights;
5472 ccv_nnc_tensor_symbol_t bias;
5473 int segments;
5474 int count;
5475 int no_bias;
5476 int flags;
5477 int functional;
5478} ccv_cnnp_model_segmented_dense_t;
5479
5480static 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)
5481{
5482 ccv_cnnp_model_segmented_dense_t* const self = (ccv_cnnp_model_segmented_dense_t*)super;
5483 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);
5484 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", 5484, __extension__ __PRETTY_FUNCTION__
); }));
5485 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", 5485, __extension__ __PRETTY_FUNCTION__
); }));
5486 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
5487 const ccv_nnc_tensor_param_t indices_params = ccv_nnc_tensor_symbol_params(graph, inputs[1]);
5488 const ccv_nnc_tensor_param_t counts_params = ccv_nnc_tensor_symbol_params(graph, inputs[2]);
5489 ccv_nnc_tensor_param_t weights_params;
5490 ccv_nnc_tensor_symbol_t weights;
5491 if (self->functional)
5492 {
5493 weights = inputs[3];
5494 weights_params = ccv_nnc_tensor_symbol_params(graph, weights);
5495 } else {
5496 weights_params = params;
5497 memset(weights_params.dim, 0, sizeof(weights_params.dim));
5498 if (params.dim[0] != 0)
5499 {
5500 weights_params.dim[0] = self->segments;
5501 weights_params.dim[1] = self->count;
5502 weights_params.dim[2] = params.dim[ccv_nnc_tensor_nd(params.dim) - 1];
5503 }
5504 if (!self->weights.graph)
5505 self->weights = ccv_nnc_tensor_symbol_new(graph, weights_params, "weights");
5506 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"
, 5506, __extension__ __PRETTY_FUNCTION__); }));
5507 weights = ccv_cnnp_model_get_symbol(super, self->weights);
5508 }
5509 ccv_nnc_tensor_param_t bias_params = {};
5510 ccv_nnc_tensor_symbol_t bias = {};
5511 if (!self->no_bias)
5512 {
5513 if (self->functional)
5514 {
5515 bias = inputs[4];
5516 bias_params = ccv_nnc_tensor_symbol_params(graph, bias);
5517 } else {
5518 bias_params = params;
5519 memset(bias_params.dim, 0, sizeof(bias_params.dim));
5520 bias_params.dim[0] = self->segments;
5521 bias_params.dim[1] = self->count;
5522 if (!self->bias.graph)
5523 self->bias = ccv_nnc_tensor_symbol_new(graph, bias_params, "bias");
5524 bias = ccv_cnnp_model_get_symbol(super, self->bias);
5525 }
5526 }
5527 ccv_nnc_cmd_t cmd = {0};
5528 cmd.cmd = CCV_NNC_SEGMENTED_GEMM_FORWARD;
5529 cmd.info.blas.a[0] = 1;
5530 cmd.info.blas.a[1] = 1;
5531 cmd.info.blas.transpose_b[0] = 1;
5532 cmd.info.blas.transpose_b[1] = 2;
5533 cmd.info.blas.flags = self->flags;
5534 ccv_nnc_tensor_param_t output_params;
5535 ccv_nnc_hint_tensor_auto(cmd, (ccv_nnc_tensor_param_t []){
5536 params, indices_params, counts_params,
5537 weights_params,
5538 bias_params,
5539 }, 4 + !self->no_bias, ccv_nnc_no_hint, &output_params, 1);
5540 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
5541 if (self->no_bias)
5542 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");
5543 else
5544 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");
5545 outputs[0] = output;
5546}
5547
5548static 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)
5549{
5550 ccv_cnnp_model_segmented_dense_t* const self = (ccv_cnnp_model_segmented_dense_t*)super;
5551 if (self->functional)
5552 return;
5553 const ccv_nnc_tensor_param_t weight_params = ccv_nnc_tensor_symbol_params(graph, self->weights);
5554 const int c = weight_params.dim[1];
5555 const float std = sqrtf(2) / sqrtf(c);
5556 const float bound = sqrtf(3) * std;
5557 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);
5558 if (self->bias.graph)
5559 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);
5560}
5561
5562static 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)
5563{
5564 ccv_cnnp_model_segmented_dense_t* const self = (ccv_cnnp_model_segmented_dense_t*)super;
5565 if (self->functional)
5566 return;
5567 add_to_array(parameters, self->weights, is_trainable);
5568 if (self->bias.graph)
5569 add_to_array(parameters, self->bias, is_trainable);
5570}
5571
5572static ccv_cnnp_model_t* _ccv_cnnp_segmented_dense_copy(const ccv_cnnp_model_t* const super, void* const context);
5573
5574static const ccv_cnnp_model_vtab_t ccv_cnnp_segmented_dense_isa = {
5575 .build = _ccv_cnnp_segmented_dense_build,
5576 .init_states = _ccv_cnnp_segmented_dense_init_states,
5577 .add_to_parameter = _ccv_cnnp_segmented_dense_add_to_parameter,
5578 .copy = _ccv_cnnp_segmented_dense_copy,
5579};
5580
5581ccv_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)
5582{
5583 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));
5584 model_segmented_dense->super.isa = &ccv_cnnp_segmented_dense_isa;
5585 model_segmented_dense->super.input_size = functional ? 4 + !no_bias : 3;
5586 model_segmented_dense->super.outputs = &model_segmented_dense->output;
5587 model_segmented_dense->super.output_size = 1;
5588 model_segmented_dense->super.is_trainable = is_trainable;
5589 ccv_cnnp_model_copy_name(&model_segmented_dense->super, name);
5590 model_segmented_dense->weights.d = CCV_NNC_NO_TENSOR_SYMBOL;
5591 model_segmented_dense->weights.graph = 0;
5592 model_segmented_dense->bias.d = CCV_NNC_NO_TENSOR_SYMBOL;
5593 model_segmented_dense->bias.graph = 0;
5594 model_segmented_dense->segments = segments;
5595 model_segmented_dense->count = count;
5596 model_segmented_dense->no_bias = no_bias;
5597 model_segmented_dense->flags = flags;
5598 model_segmented_dense->functional = functional;
5599 return (ccv_cnnp_model_t*)model_segmented_dense;
5600}
5601
5602static ccv_cnnp_model_t* _ccv_cnnp_segmented_dense_copy(const ccv_cnnp_model_t* const super, void* const context)
5603{
5604 const ccv_cnnp_model_segmented_dense_t* const self = (const ccv_cnnp_model_segmented_dense_t*)super;
5605 return ccv_cnnp_segmented_dense(self->segments, self->count, self->no_bias, self->flags, self->functional, self->super.is_trainable, self->super.name);
5606}
5607
5608// MARK - SwiGLU Layer
5609
5610typedef struct {
5611 ccv_cnnp_model_t super;
5612 ccv_nnc_tensor_symbol_t output;
5613 ccv_nnc_tensor_symbol_t gate_weights;
5614 ccv_nnc_tensor_symbol_t up_weights;
5615 int count;
5616 float clamp;
5617} ccv_cnnp_model_swiglu_t;
5618
5619static 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)
5620{
5621 ccv_cnnp_model_swiglu_t* const self = (ccv_cnnp_model_swiglu_t*)super;
5622 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);
5623 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"
, 5623, __extension__ __PRETTY_FUNCTION__); }));
5624 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", 5624, __extension__ __PRETTY_FUNCTION__
); }));
5625 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
5626 const int params_nd = ccv_nnc_tensor_nd(params.dim);
5627 ccv_nnc_tensor_param_t weights_params = params;
5628 memset(weights_params.dim, 0, sizeof(weights_params.dim));
5629 if (params_nd > 0)
5630 {
5631 weights_params.dim[0] = self->count;
5632 weights_params.dim[1] = params.dim[params_nd - 1];
5633 }
5634 if (!self->gate_weights.graph)
5635 self->gate_weights = ccv_nnc_tensor_symbol_new(graph, weights_params, "gate.weight");
5636 if (!self->up_weights.graph)
5637 self->up_weights = ccv_nnc_tensor_symbol_new(graph, weights_params, "up.weight");
5638 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", 5638, __extension__ __PRETTY_FUNCTION__
); }));
5639 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"
, 5639, __extension__ __PRETTY_FUNCTION__); }));
5640 const ccv_nnc_tensor_symbol_t gate_weights = ccv_cnnp_model_get_symbol(super, self->gate_weights);
5641 const ccv_nnc_tensor_symbol_t up_weights = ccv_cnnp_model_get_symbol(super, self->up_weights);
5642 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);
5643 const ccv_nnc_tensor_symbol_t command_inputs[] = {
5644 inputs[0], gate_weights, up_weights,
5645 };
5646 ccv_nnc_tensor_param_t command_input_params[3];
5647 int i;
5648 for (i = 0; i < 3; i++)
5649 command_input_params[i] = ccv_nnc_tensor_symbol_params(graph, command_inputs[i]);
5650 ccv_nnc_tensor_param_t output_params;
5651 ccv_nnc_hint_tensor_auto(cmd, command_input_params, 3, ccv_nnc_no_hint, &output_params, 1);
5652 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
5653 ccv_nnc_graph_exec_symbol_new(graph, cmd, command_inputs, 3, outputs, 1, "swiglu");
5654}
5655
5656static 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)
5657{
5658 ccv_cnnp_model_swiglu_t* const self = (ccv_cnnp_model_swiglu_t*)super;
5659 const ccv_nnc_tensor_param_t weight_params = ccv_nnc_tensor_symbol_params(graph, self->gate_weights);
5660 const float bound = sqrtf(6.f / weight_params.dim[1]);
5661 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);
5662 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);
5663}
5664
5665static 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)
5666{
5667 ccv_cnnp_model_swiglu_t* const self = (ccv_cnnp_model_swiglu_t*)super;
5668 add_to_array(parameters, self->gate_weights, is_trainable);
5669 add_to_array(parameters, self->up_weights, is_trainable);
5670}
5671
5672static ccv_cnnp_model_t* _ccv_cnnp_swiglu_copy(const ccv_cnnp_model_t* const super, void* const context);
5673
5674static const ccv_cnnp_model_vtab_t ccv_cnnp_swiglu_isa = {
5675 .build = _ccv_cnnp_swiglu_build,
5676 .init_states = _ccv_cnnp_swiglu_init_states,
5677 .add_to_parameter = _ccv_cnnp_swiglu_add_to_parameter,
5678 .copy = _ccv_cnnp_swiglu_copy,
5679};
5680
5681ccv_cnnp_model_t* ccv_cnnp_swiglu(const int count, const float clamp, const int is_trainable, const char* const name)
5682{
5683 ccv_cnnp_model_swiglu_t* const model = (ccv_cnnp_model_swiglu_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_swiglu_t));
5684 model->super.isa = &ccv_cnnp_swiglu_isa;
5685 model->super.input_size = 1;
5686 model->super.outputs = &model->output;
5687 model->super.output_size = 1;
5688 model->super.is_trainable = is_trainable;
5689 ccv_cnnp_model_copy_name(&model->super, name);
5690 model->gate_weights.d = CCV_NNC_NO_TENSOR_SYMBOL;
5691 model->up_weights.d = CCV_NNC_NO_TENSOR_SYMBOL;
5692 model->count = count;
5693 model->clamp = clamp;
5694 return (ccv_cnnp_model_t*)model;
5695}
5696
5697static ccv_cnnp_model_t* _ccv_cnnp_swiglu_copy(const ccv_cnnp_model_t* const super, void* const context)
5698{
5699 const ccv_cnnp_model_swiglu_t* const self = (const ccv_cnnp_model_swiglu_t*)super;
5700 return ccv_cnnp_swiglu(self->count, self->clamp, self->super.is_trainable, self->super.name);
5701}
5702
5703// MARK - Segmented SwiGLU Layer
5704
5705typedef struct {
5706 ccv_cnnp_model_t super;
5707 ccv_nnc_tensor_symbol_t output;
5708 ccv_nnc_tensor_symbol_t gate_weights;
5709 ccv_nnc_tensor_symbol_t up_weights;
5710 int segments;
5711 int count;
5712 float clamp;
5713 int functional;
5714} ccv_cnnp_model_segmented_swiglu_t;
5715
5716static 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)
5717{
5718 ccv_cnnp_model_segmented_swiglu_t* const self = (ccv_cnnp_model_segmented_swiglu_t*)super;
5719 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);
5720 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", 5720, __extension__ __PRETTY_FUNCTION__
); }));
5721 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", 5721, __extension__ __PRETTY_FUNCTION__
); }));
5722 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
5723 ccv_nnc_tensor_symbol_t gate_weights;
5724 ccv_nnc_tensor_symbol_t up_weights;
5725 if (self->functional)
5726 {
5727 gate_weights = inputs[3];
5728 up_weights = inputs[4];
5729 } else {
5730 const int params_nd = ccv_nnc_tensor_nd(params.dim);
5731 ccv_nnc_tensor_param_t weights_params = params;
5732 memset(weights_params.dim, 0, sizeof(weights_params.dim));
5733 if (params_nd > 0)
5734 {
5735 weights_params.dim[0] = self->segments;
5736 weights_params.dim[1] = self->count;
5737 weights_params.dim[2] = params.dim[params_nd - 1];
5738 }
5739 if (!self->gate_weights.graph)
5740 self->gate_weights = ccv_nnc_tensor_symbol_new(graph, weights_params, "gate.weight");
5741 if (!self->up_weights.graph)
5742 self->up_weights = ccv_nnc_tensor_symbol_new(graph, weights_params, "up.weight");
5743 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", 5743, __extension__ __PRETTY_FUNCTION__
); }));
5744 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"
, 5744, __extension__ __PRETTY_FUNCTION__); }));
5745 gate_weights = ccv_cnnp_model_get_symbol(super, self->gate_weights);
5746 up_weights = ccv_cnnp_model_get_symbol(super, self->up_weights);
5747 }
5748 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);
5749 const ccv_nnc_tensor_symbol_t command_inputs[] = {
5750 inputs[0], inputs[1], inputs[2], gate_weights, up_weights, inputs[self->functional ? 5 : 3],
5751 };
5752 ccv_nnc_tensor_param_t command_input_params[6];
5753 int i;
5754 for (i = 0; i < 6; i++)
5755 command_input_params[i] = ccv_nnc_tensor_symbol_params(graph, command_inputs[i]);
5756 ccv_nnc_tensor_param_t output_params;
5757 ccv_nnc_hint_tensor_auto(cmd, command_input_params, 6, ccv_nnc_no_hint, &output_params, 1);
5758 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
5759 ccv_nnc_graph_exec_symbol_new(graph, cmd, command_inputs, 6, outputs, 1, "segmented_swiglu");
5760}
5761
5762static 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)
5763{
5764 ccv_cnnp_model_segmented_swiglu_t* const self = (ccv_cnnp_model_segmented_swiglu_t*)super;
5765 if (self->functional)
5766 return;
5767 const ccv_nnc_tensor_param_t weight_params = ccv_nnc_tensor_symbol_params(graph, self->gate_weights);
5768 const float bound = sqrtf(6.f / weight_params.dim[2]);
5769 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);
5770 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);
5771}
5772
5773static 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)
5774{
5775 ccv_cnnp_model_segmented_swiglu_t* const self = (ccv_cnnp_model_segmented_swiglu_t*)super;
5776 if (self->functional)
5777 return;
5778 add_to_array(parameters, self->gate_weights, is_trainable);
5779 add_to_array(parameters, self->up_weights, is_trainable);
5780}
5781
5782static ccv_cnnp_model_t* _ccv_cnnp_segmented_swiglu_copy(const ccv_cnnp_model_t* const super, void* const context);
5783
5784static const ccv_cnnp_model_vtab_t ccv_cnnp_segmented_swiglu_isa = {
5785 .build = _ccv_cnnp_segmented_swiglu_build,
5786 .init_states = _ccv_cnnp_segmented_swiglu_init_states,
5787 .add_to_parameter = _ccv_cnnp_segmented_swiglu_add_to_parameter,
5788 .copy = _ccv_cnnp_segmented_swiglu_copy,
5789};
5790
5791ccv_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)
5792{
5793 ccv_cnnp_model_segmented_swiglu_t* const model = (ccv_cnnp_model_segmented_swiglu_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_segmented_swiglu_t));
5794 model->super.isa = &ccv_cnnp_segmented_swiglu_isa;
5795 model->super.input_size = functional ? 6 : 4;
5796 model->super.outputs = &model->output;
5797 model->super.output_size = 1;
5798 model->super.is_trainable = is_trainable;
5799 ccv_cnnp_model_copy_name(&model->super, name);
5800 model->gate_weights.d = CCV_NNC_NO_TENSOR_SYMBOL;
5801 model->up_weights.d = CCV_NNC_NO_TENSOR_SYMBOL;
5802 model->segments = segments;
5803 model->count = count;
5804 model->clamp = clamp;
5805 model->functional = functional;
5806 return (ccv_cnnp_model_t*)model;
5807}
5808
5809static ccv_cnnp_model_t* _ccv_cnnp_segmented_swiglu_copy(const ccv_cnnp_model_t* const super, void* const context)
5810{
5811 const ccv_cnnp_model_segmented_swiglu_t* const self = (const ccv_cnnp_model_segmented_swiglu_t*)super;
5812 return ccv_cnnp_segmented_swiglu(self->segments, self->count, self->clamp, self->functional, self->super.is_trainable, self->super.name);
5813}
5814
5815// MARK - MoE Weights Streaming
5816
5817typedef struct {
5818 ccv_cnnp_model_t super;
5819 ccv_nnc_tensor_symbol_t outputs[6];
5820 int resident_slots;
5821 int routing_width;
5822} ccv_cnnp_model_moe_weights_streaming_t;
5823
5824static 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)
5825{
5826 ccv_cnnp_model_moe_weights_streaming_t* const self = (ccv_cnnp_model_moe_weights_streaming_t*)super;
5827 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);
5828 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"
, 5828, __extension__ __PRETTY_FUNCTION__); }));
5829 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", 5829, __extension__ __PRETTY_FUNCTION__
); }));
5830 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);
5831 ccv_nnc_tensor_param_t input_params[6];
5832 ccv_nnc_tensor_param_t output_params[6];
5833 int i;
5834 for (i = 0; i < 6; i++)
5835 input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
5836 ccv_nnc_hint_tensor_auto(cmd, input_params, 6, ccv_nnc_no_hint, output_params, 6);
5837 for (i = 0; i < 6; i++)
5838 outputs[i] = ccv_nnc_tensor_symbol_new(graph, output_params[i], 0);
5839 ccv_nnc_graph_exec_symbol_new(graph, cmd, inputs, 6, outputs, 6, "moe_weights_streaming");
5840}
5841
5842static ccv_cnnp_model_t* _ccv_cnnp_moe_weights_streaming_copy(const ccv_cnnp_model_t* const super, void* const context);
5843
5844static const ccv_cnnp_model_vtab_t ccv_cnnp_moe_weights_streaming_isa = {
5845 .build = _ccv_cnnp_moe_weights_streaming_build,
5846 .copy = _ccv_cnnp_moe_weights_streaming_copy,
5847};
5848
5849ccv_cnnp_model_t* ccv_cnnp_moe_weights_streaming(const int resident_slots, const int routing_width, const char* const name)
5850{
5851 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", 5851, __extension__ __PRETTY_FUNCTION__
); }));
5852 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", 5852, __extension__ __PRETTY_FUNCTION__
); }));
5853 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));
5854 model->super.isa = &ccv_cnnp_moe_weights_streaming_isa;
5855 model->super.input_size = 6;
5856 model->super.outputs = model->outputs;
5857 model->super.output_size = 6;
5858 model->resident_slots = resident_slots;
5859 model->routing_width = routing_width;
5860 ccv_cnnp_model_copy_name(&model->super, name);
5861 return (ccv_cnnp_model_t*)model;
5862}
5863
5864static ccv_cnnp_model_t* _ccv_cnnp_moe_weights_streaming_copy(const ccv_cnnp_model_t* const super, void* const context)
5865{
5866 const ccv_cnnp_model_moe_weights_streaming_t* const self = (const ccv_cnnp_model_moe_weights_streaming_t*)super;
5867 return ccv_cnnp_moe_weights_streaming(self->resident_slots, self->routing_width, self->super.name);
5868}