Bug Summary

File:nnc/ccv_cnnp_model_addons.c
Warning:line 400, column 3
Declared variable-length array (VLA) has negative size

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/2026-05-01-233500-1493103-1 -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);
1
Assuming the condition is false
→
2
←
Taking false branch
→
3
←
Loop condition is false. Exiting loop
→
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__
); }));
4
←
Assuming 'output_size' is equal to 1
→
5
←
Taking true branch
→
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)
6
←
Assuming the condition is false
→
7
←
Taking false branch
→
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 assert(axis < nd)((void) sizeof ((axis < nd) ? 1 : 0), __extension__ ({ if (
axis < nd) ; else __assert_fail ("axis < nd", "ccv_cnnp_model_addons.c"
, 367, __extension__ __PRETTY_FUNCTION__); }));
8
←
Assuming 'axis' is < 'nd'
→
9
←
Taking true branch
→
368 output_params.dim[axis] = 0;
369 int input_is_contiguous = 1;
370 for (i = 0; i < input_size; i++)
10
←
Assuming 'i' is >= 'input_size'
→
11
←
Loop condition is false. Execution continues on line 394
→
371 {
372 const ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
373 const int input_nd = ccv_nnc_tensor_nd(input_params.dim);
374 if (input_nd == 0)
375 {
376 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);
377 input_is_contiguous = 0;
378 continue;
379 }
380 if (CCV_CLI_OUTPUT_LEVEL_IS(CCV_CLI_VERBOSE)(CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
381 {
382 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);
383 int i;
384 for (i = 1; i < CCV_NNC_MAX_DIM_ALLOC(12) && input_params.dim[i] > 0; i++)
385 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);
386 PRINT(CCV_CLI_VERBOSE, ")\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf(")\n"); fflush(stdout); } } while (0);
387 }
388 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"
, 388, __extension__ __PRETTY_FUNCTION__); }));
389 for (j = 0; j < nd; j++)
390 if (j != axis)
391 { 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", 391, __extension__ __PRETTY_FUNCTION__
); })); }
392 output_params.dim[axis] += input_params.dim[axis];
393 }
394 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
395 int ofs[CCV_NNC_MAX_DIM_ALLOC(12)] = {};
396 int stride[CCV_NNC_MAX_DIM_ALLOC(12)] = {};
397 ccv_nnc_tensor_get_stride(output_params.dim, stride);
398 if (input_is_contiguous
11.1
'input_is_contiguous' is 1
)
12
←
Taking true branch
→
399 {
400 ccv_nnc_tensor_symbol_t aliases[input_size];
13
←
Declared variable-length array (VLA) has negative size
401 for (i = 0; i < input_size; i++)
402 {
403 const ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
404 aliases[i] = ccv_nnc_tensor_symbol_alias_new(graph, outputs[0], ofs, stride, input_params, 0);
405 ofs[axis] += input_params.dim[axis];
406 }
407 // Format transform is more flexible.
408 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");
409 } else {
410 ccv_nnc_tensor_symbol_t aliases[input_size];
411 for (i = 0; i < input_size; i++)
412 {
413 const ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
414 if (input_params.dim[0] == 0)
415 {
416 // Create a new alias anyway, but not going to use it, in this way, the alias count will match during absorb.
417 aliases[i] = ccv_nnc_tensor_symbol_alias_new(graph, outputs[0], ofs, stride, input_params, 0);
418 continue;
419 }
420 aliases[i] = ccv_nnc_tensor_symbol_alias_new(graph, outputs[0], ofs, stride, input_params, 0);
421 ofs[axis] += input_params.dim[axis];
422 }
423 // Format transform is more flexible.
424 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");
425 }
426}
427
428static ccv_cnnp_model_t* _ccv_cnnp_concat_copy(const ccv_cnnp_model_t* const self, void* const context);
429
430static const ccv_cnnp_model_vtab_t ccv_cnnp_concat_isa = {
431 .build = _ccv_cnnp_concat_build,
432 .copy = _ccv_cnnp_concat_copy,
433};
434
435ccv_cnnp_model_t* ccv_cnnp_concat(const int axis, const char* const name)
436{
437 ccv_cnnp_model_concat_t* const model_concat = (ccv_cnnp_model_concat_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_concat_t));
438 model_concat->super.isa = &ccv_cnnp_concat_isa;
439 model_concat->super.input_size = 0;
440 model_concat->super.outputs = &model_concat->output;
441 model_concat->super.output_size = 1;
442 model_concat->axis = axis;
443 ccv_cnnp_model_copy_name(&model_concat->super, name);
444 return (ccv_cnnp_model_t*)model_concat;
445}
446
447static ccv_cnnp_model_t* _ccv_cnnp_concat_copy(const ccv_cnnp_model_t* const super, void* const context)
448{
449 const ccv_cnnp_model_concat_t* const self = (const ccv_cnnp_model_concat_t*)super;
450 return ccv_cnnp_concat(self->axis, self->super.name);
451}
452
453typedef struct {
454 ccv_cnnp_model_t super;
455 int axis;
456 ccv_nnc_tensor_symbol_t outputs[1];
457} ccv_cnnp_model_chunk_t;
458
459static 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)
460{
461 const ccv_cnnp_model_concat_t* const self = (const ccv_cnnp_model_concat_t*)super;
462 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);
463 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"
, 463, __extension__ __PRETTY_FUNCTION__); }));
464 const ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
465 if (CCV_CLI_OUTPUT_LEVEL_IS(CCV_CLI_VERBOSE)(CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
466 {
467 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);
468 int i;
469 for (i = 1; i < CCV_NNC_MAX_DIM_ALLOC(12) && input_params.dim[i] > 0; i++)
470 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);
471 PRINT(CCV_CLI_VERBOSE, ")\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf(")\n"); fflush(stdout); } } while (0);
472 }
473 ccv_nnc_tensor_param_t output_params = input_params;
474 int i;
475 const int nd = ccv_nnc_tensor_nd(output_params.dim);
476 const int axis = self->axis;
477 assert(axis < nd)((void) sizeof ((axis < nd) ? 1 : 0), __extension__ ({ if (
axis < nd) ; else __assert_fail ("axis < nd", "ccv_cnnp_model_addons.c"
, 477, __extension__ __PRETTY_FUNCTION__); }));
478 const int n = self->super.output_size;
479 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", 479, __extension__ __PRETTY_FUNCTION__
); }));
480 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"
, 480, __extension__ __PRETTY_FUNCTION__); }));
481 output_params.dim[axis] = output_params.dim[axis] / n;
482 int ofs[CCV_NNC_MAX_DIM_ALLOC(12)] = {};
483 int stride[CCV_NNC_MAX_DIM_ALLOC(12)] = {};
484 ccv_nnc_tensor_get_stride(input_params.dim, stride);
485 ccv_nnc_tensor_symbol_t to = ccv_nnc_tensor_symbol_alias_to(graph, inputs[0]);
486 if (to.d == CCV_NNC_NO_TENSOR_SYMBOL) // If we are not reshape an alias, it is straightforward.
487 {
488 for (i = 0; i < output_size; i++)
489 {
490 outputs[i] = ccv_nnc_tensor_symbol_alias_new(graph, inputs[0], ofs, stride, output_params, 0);
491 ofs[axis] += output_params.dim[axis];
492 }
493 } else {
494 // Otherwise, we need to check if it is permute. For permute, we cannot do alias directly.
495 // We need to first materialize the permute and then run reshape on top of it, otherwise it will be wrong.
496 int old_stride[CCV_NNC_MAX_DIM_ALLOC(12)];
497 ccv_nnc_tensor_symbol_alias_params(graph, inputs[0], 0, old_stride);
498 // We identify permute by checking if the stride is not in descending order.
499 // This also covered "permute" through reshape, rather than using ccv_cnnp_permute directly.
500 int i, no_permute = 1;
501 for (i = 1; no_permute && i < nd; i++)
502 if (old_stride[i - 1] < old_stride[i])
503 no_permute = 0;
504 if (no_permute)
505 { // Just straightforward reshape if there is no no permute.
506 for (i = 0; i < output_size; i++)
507 {
508 outputs[i] = ccv_nnc_tensor_symbol_alias_new(graph, inputs[0], ofs, old_stride, output_params, 0);
509 ofs[axis] += output_params.dim[axis];
510 }
511 } else {
512 // Otherwise, we first do format transform to plain tensor and then do reshape.
513 ccv_nnc_tensor_symbol_t permuted = ccv_nnc_tensor_symbol_new(graph, input_params, 0);
514 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");
515 for (i = 0; i < output_size; i++)
516 {
517 outputs[i] = ccv_nnc_tensor_symbol_alias_new(graph, permuted, ofs, stride, output_params, 0);
518 ofs[axis] += output_params.dim[axis];
519 }
520 }
521 }
522}
523
524static ccv_cnnp_model_t* _ccv_cnnp_chunk_copy(const ccv_cnnp_model_t* const self, void* const context);
525
526static const ccv_cnnp_model_vtab_t ccv_cnnp_chunk_isa = {
527 .build = _ccv_cnnp_chunk_build,
528 .copy = _ccv_cnnp_chunk_copy,
529};
530
531ccv_cnnp_model_t* ccv_cnnp_chunk(const int n, const int axis, const char* const name)
532{
533 assert(n >= 1)((void) sizeof ((n >= 1) ? 1 : 0), __extension__ ({ if (n >=
1) ; else __assert_fail ("n >= 1", "ccv_cnnp_model_addons.c"
, 533, __extension__ __PRETTY_FUNCTION__); }));
534 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));
535 model_chunk->super.isa = &ccv_cnnp_chunk_isa;
536 model_chunk->super.input_size = 1;
537 model_chunk->super.outputs = model_chunk->outputs;
538 model_chunk->super.output_size = n;
539 model_chunk->axis = axis;
540 ccv_cnnp_model_copy_name(&model_chunk->super, name);
541 return (ccv_cnnp_model_t*)model_chunk;
542}
543
544static ccv_cnnp_model_t* _ccv_cnnp_chunk_copy(const ccv_cnnp_model_t* const super, void* const context)
545{
546 const ccv_cnnp_model_chunk_t* const self = (const ccv_cnnp_model_chunk_t*)super;
547 return ccv_cnnp_chunk(self->super.output_size, self->axis, self->super.name);
548}
549
550typedef struct {
551 ccv_cnnp_model_t super;
552 ccv_nnc_tensor_symbol_t output;
553 int format;
554 int dim[CCV_NNC_MAX_DIM_ALLOC(12)];
555 int ofs[CCV_NNC_MAX_DIM_ALLOC(12)];
556 int stride[CCV_NNC_MAX_DIM_ALLOC(12)];
557} ccv_cnnp_model_reshape_t;
558
559static 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)
560{
561 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"
, 561, __extension__ __PRETTY_FUNCTION__); }));
562 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", 562, __extension__ __PRETTY_FUNCTION__
); }));
563 ccv_cnnp_model_reshape_t* const self = (ccv_cnnp_model_reshape_t*)super;
564 if (CCV_CLI_OUTPUT_LEVEL_IS(CCV_CLI_VERBOSE)(CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
565 {
566 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);
567 int i;
568 for (i = 1; i < CCV_NNC_MAX_DIM_ALLOC(12) && self->dim[i] > 0; i++)
569 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);
570 const int count = i;
571 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);
572 for (i = 1; i < count; i++)
573 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);
574 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);
575 for (i = 1; i < count; i++)
576 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);
577 PRINT(CCV_CLI_VERBOSE, ")\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf(")\n"); fflush(stdout); } } while (0);
578 }
579 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
580 int dim[CCV_NNC_MAX_DIM_ALLOC(12)];
581 memcpy(dim, self->dim, sizeof(dim));
582 int i, auto_idx = -1;
583 size_t known = 1;
584 const size_t tensor_count = ccv_nnc_tensor_count(params);
585 for (i = 0; i < CCV_NNC_MAX_DIM_ALLOC(12) && dim[i]; i++)
586 if (dim[i] == -1)
587 auto_idx = i;
588 else
589 known *= dim[i];
590 if (auto_idx >= 0)
591 {
592 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", 592, __extension__ __PRETTY_FUNCTION__
); }));
593 dim[auto_idx] = tensor_count / known;
594 }
595 if (CCV_CLI_OUTPUT_LEVEL_IS(CCV_CLI_VERBOSE)(CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
596 {
597 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);
598 int i;
599 for (i = 1; i < CCV_NNC_MAX_DIM_ALLOC(12) && params.dim[i] > 0; i++)
600 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);
601 PRINT(CCV_CLI_VERBOSE, ")\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf(")\n"); fflush(stdout); } } while (0);
602 }
603 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", 603, __extension__ __PRETTY_FUNCTION__
); }));
604 ccv_nnc_tensor_symbol_t to = ccv_nnc_tensor_symbol_alias_to(graph, inputs[0]);
605 int stride_from_dim[CCV_NNC_MAX_DIM_ALLOC(12)];
606 if (to.d == CCV_NNC_NO_TENSOR_SYMBOL) // If we are not reshape an alias, it is straightforward.
607 {
608 memcpy(params.dim, dim, sizeof(params.dim));
609 int* stride;
610 if (self->stride[0] == 0)
611 {
612 ccv_nnc_tensor_get_stride(dim, stride_from_dim);
613 stride = stride_from_dim;
614 } else
615 stride = self->stride;
616 if (self->format > 0)
617 params.format = self->format;
618 outputs[0] = ccv_nnc_tensor_symbol_alias_new(graph, inputs[0], self->ofs, stride, params, 0);
619 } else {
620 // Otherwise, we need to check if it is permute. For permute, we cannot do alias directly.
621 // We need to first materialize the permute and then run reshape on top of it, otherwise it will be wrong.
622 int old_stride[CCV_NNC_MAX_DIM_ALLOC(12)];
623 ccv_nnc_tensor_symbol_alias_params(graph, inputs[0], 0, old_stride);
624 // We identify permute by checking if the stride is not in descending order.
625 // This also covered "permute" through reshape, rather than using ccv_cnnp_permute directly.
626 const int nd = ccv_nnc_tensor_nd(params.dim);
627 const int new_nd = ccv_nnc_tensor_nd(dim);
628 int i, no_permute = 1;
629 // If the new dim has different nd, or we actually have a stride, we need to check if it is no permute or not.
630 if (new_nd != nd || (self->stride[0] != 0 && memcmp(self->stride, old_stride, sizeof(self->stride))))
631 for (i = 1; no_permute && i < nd; i++)
632 if (old_stride[i - 1] < old_stride[i])
633 no_permute = 0;
634 if (no_permute)
635 { // Just straightforward reshape if there is no no permute.
636 memcpy(params.dim, dim, sizeof(params.dim));
637 int* stride;
638 if (self->stride[0] == 0)
639 {
640 if (new_nd != nd) // Cannot use old stride.
641 {
642 ccv_nnc_tensor_get_stride(dim, stride_from_dim);
643 stride = stride_from_dim;
644 } else
645 stride = old_stride;
646 } else
647 stride = self->stride;
648 if (self->format > 0)
649 params.format = self->format;
650 outputs[0] = ccv_nnc_tensor_symbol_alias_new(graph, inputs[0], self->ofs, stride, params, 0);
651 } else {
652 // Otherwise, we first do format transform to plain tensor and then do reshape.
653 ccv_nnc_tensor_symbol_t permuted = ccv_nnc_tensor_symbol_new(graph, params, 0);
654 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");
655 memcpy(params.dim, dim, sizeof(params.dim));
656 int* stride;
657 if (self->stride[0] == 0)
658 {
659 ccv_nnc_tensor_get_stride(dim, stride_from_dim);
660 stride = stride_from_dim;
661 } else
662 stride = self->stride;
663 if (self->format > 0)
664 params.format = self->format;
665 // And then we create alias against the permuted one.
666 outputs[0] = ccv_nnc_tensor_symbol_alias_new(graph, permuted, self->ofs, stride, params, 0);
667 }
668 }
669}
670
671static ccv_cnnp_model_t* _ccv_cnnp_reshape_copy(const ccv_cnnp_model_t* const super, void* const context);
672
673static const ccv_cnnp_model_vtab_t ccv_cnnp_reshape_isa = {
674 .build = _ccv_cnnp_reshape_build,
675 .copy = _ccv_cnnp_reshape_copy,
676};
677
678ccv_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)
679{
680 ccv_cnnp_model_reshape_t* const model_reshape = (ccv_cnnp_model_reshape_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_reshape_t));
681 model_reshape->super.isa = &ccv_cnnp_reshape_isa;
682 model_reshape->super.input_size = 1;
683 model_reshape->super.outputs = &model_reshape->output;
684 model_reshape->super.output_size = 1;
685 ccv_cnnp_model_copy_name(&model_reshape->super, name);
686 model_reshape->format = format;
687 memcpy(model_reshape->dim, dim, sizeof(model_reshape->dim));
688 memcpy(model_reshape->ofs, ofs, sizeof(model_reshape->ofs));
689 if (stride[0] != 0)
690 memcpy(model_reshape->stride, stride, sizeof(model_reshape->stride));
691 return (ccv_cnnp_model_t*)model_reshape;
692}
693
694static ccv_cnnp_model_t* _ccv_cnnp_reshape_copy(const ccv_cnnp_model_t* const super, void* const context)
695{
696 const ccv_cnnp_model_reshape_t* const self = (const ccv_cnnp_model_reshape_t*)super;
697 return ccv_cnnp_reshape(self->format, self->dim, self->ofs, self->stride, self->super.name);
698}
699
700typedef struct {
701 ccv_cnnp_model_t super;
702 ccv_nnc_tensor_symbol_t output;
703 int type;
704 int begin[CCV_NNC_MAX_DIM_ALLOC(12)];
705 int end[CCV_NNC_MAX_DIM_ALLOC(12)];
706} ccv_cnnp_model_pad_t;
707
708static 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)
709{
710 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"
, 710, __extension__ __PRETTY_FUNCTION__); }));
711 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", 711, __extension__ __PRETTY_FUNCTION__
); }));
712 ccv_cnnp_model_pad_t* const self = (ccv_cnnp_model_pad_t*)super;
713 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);
714 const ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
715 const int nd = ccv_nnc_tensor_nd(input_params.dim);
716 ccv_nnc_tensor_param_t params = input_params;
717 int i;
718 for (i = 0 ; i < nd; i++)
719 params.dim[i] += self->begin[i] + self->end[i];
720 const ccv_nnc_tensor_symbol_t padded = ccv_nnc_tensor_symbol_new(graph, params, 0);
721 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);
722 memcpy(pad.info.size.dim, self->begin, sizeof(pad.info.size.dim));
723 memcpy(pad.info.pad.end, self->end, sizeof(pad.info.pad.end));
724 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");
725 outputs[0] = padded;
726}
727
728static ccv_cnnp_model_t* _ccv_cnnp_pad_copy(const ccv_cnnp_model_t* const super, void* const context);
729
730static const ccv_cnnp_model_vtab_t ccv_cnnp_pad_isa = {
731 .build = _ccv_cnnp_pad_build,
732 .copy = _ccv_cnnp_pad_copy,
733};
734
735ccv_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)
736{
737 ccv_cnnp_model_pad_t* const model_pad = (ccv_cnnp_model_pad_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_pad_t));
738 model_pad->super.isa = &ccv_cnnp_pad_isa;
739 model_pad->super.input_size = 1;
740 model_pad->super.outputs = &model_pad->output;
741 model_pad->super.output_size = 1;
742 ccv_cnnp_model_copy_name(&model_pad->super, name);
743 model_pad->type = type;
744 memcpy(model_pad->begin, begin, sizeof(model_pad->begin));
745 memcpy(model_pad->end, end, sizeof(model_pad->end));
746 return (ccv_cnnp_model_t*)model_pad;
747}
748
749static ccv_cnnp_model_t* _ccv_cnnp_pad_copy(const ccv_cnnp_model_t* const super, void* const context)
750{
751 const ccv_cnnp_model_pad_t* const self = (const ccv_cnnp_model_pad_t*)super;
752 return ccv_cnnp_pad(self->type, self->begin, self->end, self->super.name);
753}
754
755typedef struct {
756 ccv_cnnp_model_t super;
757 ccv_nnc_tensor_symbol_t output;
758} ccv_cnnp_model_identity_t;
759
760static 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)
761{
762 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"
, 762, __extension__ __PRETTY_FUNCTION__); }));
763 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", 763, __extension__ __PRETTY_FUNCTION__
); }));
764 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);
765 outputs[0] = inputs[0];
766}
767
768static ccv_cnnp_model_t* _ccv_cnnp_identity_copy(const ccv_cnnp_model_t* const super, void* const context);
769
770static const ccv_cnnp_model_vtab_t ccv_cnnp_identity_isa = {
771 .build = _ccv_cnnp_identity_build,
772 .copy = _ccv_cnnp_identity_copy,
773};
774
775ccv_cnnp_model_t* ccv_cnnp_identity(const char* const name)
776{
777 ccv_cnnp_model_identity_t* const model_identity = (ccv_cnnp_model_identity_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_identity_t));
778 model_identity->super.isa = &ccv_cnnp_identity_isa;
779 model_identity->super.input_size = 1;
780 model_identity->super.outputs = &model_identity->output;
781 model_identity->super.output_size = 1;
782 ccv_cnnp_model_copy_name(&model_identity->super, name);
783 return (ccv_cnnp_model_t*)model_identity;
784}
785
786static ccv_cnnp_model_t* _ccv_cnnp_identity_copy(const ccv_cnnp_model_t* const super, void* const context)
787{
788 const ccv_cnnp_model_identity_t* const self = (const ccv_cnnp_model_identity_t*)super;
789 return ccv_cnnp_identity(self->super.name);
790}
791
792typedef struct {
793 ccv_cnnp_model_t super;
794 ccv_nnc_tensor_symbol_t output;
795 int index[CCV_NNC_MAX_DIM_ALLOC(12)];
796} ccv_cnnp_model_permute_t;
797
798static 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)
799{
800 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"
, 800, __extension__ __PRETTY_FUNCTION__); }));
801 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", 801, __extension__ __PRETTY_FUNCTION__
); }));
802 ccv_cnnp_model_permute_t* const self = (ccv_cnnp_model_permute_t*)super;
803 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);
804 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
805 ccv_nnc_tensor_symbol_t to = ccv_nnc_tensor_symbol_alias_to(graph, inputs[0]);
806 const int nd = ccv_nnc_tensor_nd(params.dim);
807 int input_dim[CCV_NNC_MAX_DIM_ALLOC(12)];
808 memcpy(input_dim, params.dim, sizeof(params.dim));
809 int input_stride[CCV_NNC_MAX_DIM_ALLOC(12)] = {};
810 int output_stride[CCV_NNC_MAX_DIM_ALLOC(12)] = {};
811 if (to.d == CCV_NNC_NO_TENSOR_SYMBOL) // If it is not an alias. Find stride and permute.
812 {
813 ccv_nnc_tensor_get_stride(input_dim, input_stride);
814 int i;
815 for (i = 0; i < nd; i++)
816 {
817 const int idx = self->index[i];
818 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"
, 818, __extension__ __PRETTY_FUNCTION__); }));
819 params.dim[i] = input_dim[idx];
820 output_stride[i] = input_stride[idx];
821 }
822 outputs[0] = ccv_nnc_tensor_symbol_alias_new(graph, inputs[0], ccv_nnc_no_ofs, output_stride, params, 0);
823 } else {
824 // if it is an alias, we can get the stride from it and use that.
825 int input_ofs[CCV_NNC_MAX_DIM_ALLOC(12)];
826 ccv_nnc_tensor_symbol_alias_params(graph, inputs[0], input_ofs, input_stride);
827 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", 827, __extension__ __PRETTY_FUNCTION__
); }));
828 int output_ofs[CCV_NNC_MAX_DIM_ALLOC(12)] = {};
829 int i;
830 for (i = 0; i < nd; i++)
831 {
832 const int idx = self->index[i];
833 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"
, 833, __extension__ __PRETTY_FUNCTION__); }));
834 params.dim[i] = input_dim[idx];
835 output_stride[i] = input_stride[idx];
836 output_ofs[i] = input_ofs[idx];
837 }
838 outputs[0] = ccv_nnc_tensor_symbol_alias_new(graph, inputs[0], output_ofs, output_stride, params, 0);
839 }
840}
841
842static ccv_cnnp_model_t* _ccv_cnnp_permute_copy(const ccv_cnnp_model_t* const super, void* const context);
843
844static const ccv_cnnp_model_vtab_t ccv_cnnp_permute_isa = {
845 .build = _ccv_cnnp_permute_build,
846 .copy = _ccv_cnnp_permute_copy,
847};
848
849ccv_cnnp_model_t* ccv_cnnp_permute(const int index[CCV_NNC_MAX_DIM_ALLOC(12)], const char* const name)
850{
851 ccv_cnnp_model_permute_t* const model_permute = (ccv_cnnp_model_permute_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_permute_t));
852 model_permute->super.isa = &ccv_cnnp_permute_isa;
853 model_permute->super.input_size = 1;
854 model_permute->super.outputs = &model_permute->output;
855 model_permute->super.output_size = 1;
856 ccv_cnnp_model_copy_name(&model_permute->super, name);
857 memcpy(model_permute->index, index, sizeof(model_permute->index));
858 return (ccv_cnnp_model_t*)model_permute;
859}
860
861static ccv_cnnp_model_t* _ccv_cnnp_permute_copy(const ccv_cnnp_model_t* const super, void* const context)
862{
863 const ccv_cnnp_model_permute_t* const self = (const ccv_cnnp_model_permute_t*)super;
864 return ccv_cnnp_permute(self->index, self->super.name);
865}
866
867typedef struct {
868 ccv_cnnp_model_t super;
869 int index;
870 ccv_nnc_tensor_symbol_t output;
871} ccv_cnnp_model_extract_t;
872
873static 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)
874{
875 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", 875, __extension__ __PRETTY_FUNCTION__
); }));
876 ccv_cnnp_model_extract_t* const self = (ccv_cnnp_model_extract_t*)super;
877 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);
878 outputs[0] = inputs[self->index];
879}
880
881static ccv_cnnp_model_t* _ccv_cnnp_extract_copy(const ccv_cnnp_model_t* const self, void* const context);
882
883static const ccv_cnnp_model_vtab_t ccv_cnnp_extract_isa = {
884 .build = _ccv_cnnp_extract_build,
885 .copy = _ccv_cnnp_extract_copy,
886};
887
888ccv_cnnp_model_t* ccv_cnnp_extract(const int index, const char* const name)
889{
890 ccv_cnnp_model_extract_t* const model_extract = (ccv_cnnp_model_extract_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_extract_t));
891 model_extract->index = index;
892 model_extract->super.isa = &ccv_cnnp_extract_isa;
893 model_extract->super.input_size = 0;
894 model_extract->super.outputs = &model_extract->output;
895 model_extract->super.output_size = 1;
896 ccv_cnnp_model_copy_name(&model_extract->super, name);
897 return (ccv_cnnp_model_t*)model_extract;
898}
899
900static ccv_cnnp_model_t* _ccv_cnnp_extract_copy(const ccv_cnnp_model_t* const super, void* const context)
901{
902 ccv_cnnp_model_extract_t* const self = (ccv_cnnp_model_extract_t*)super;
903 return ccv_cnnp_extract(self->index, self->super.name);
904}
905
906typedef struct {
907 ccv_cnnp_model_t super;
908 ccv_nnc_tensor_symbol_t output;
909} ccv_cnnp_model_flatten_t;
910
911static 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)
912{
913 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);
914 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"
, 914, __extension__ __PRETTY_FUNCTION__); }));
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 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
917 ccv_nnc_tensor_param_t output_params = params;
918 memset(output_params.dim, 0, sizeof(output_params.dim));
919 output_params.dim[0] = ccv_nnc_tensor_get_n(params);
920 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", 920, __extension__ __PRETTY_FUNCTION__
); }));
921 output_params.dim[1] = ccv_nnc_tensor_count(params) / output_params.dim[0];
922 int stride[CCV_NNC_MAX_DIM_ALLOC(12)] = {};
923 ccv_nnc_tensor_get_stride(output_params.dim, stride);
924 outputs[0] = ccv_nnc_tensor_symbol_alias_new(graph, inputs[0], DIM_ALLOC()(int [(12)]){}, stride, output_params, 0);
925}
926
927static ccv_cnnp_model_t* _ccv_cnnp_flatten_copy(const ccv_cnnp_model_t* const self, void* const context);
928
929static const ccv_cnnp_model_vtab_t ccv_cnnp_flatten_isa = {
930 .build = _ccv_cnnp_flatten_build,
931 .copy = _ccv_cnnp_flatten_copy,
932};
933
934ccv_cnnp_model_t* ccv_cnnp_flatten(const char* const name)
935{
936 ccv_cnnp_model_flatten_t* const model_flatten = (ccv_cnnp_model_flatten_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_flatten_t));
937 model_flatten->super.isa = &ccv_cnnp_flatten_isa;
938 model_flatten->super.input_size = 1;
939 model_flatten->super.outputs = &model_flatten->output;
940 model_flatten->super.output_size = 1;
941 ccv_cnnp_model_copy_name(&model_flatten->super, name);
942 return (ccv_cnnp_model_t*)model_flatten;
943}
944
945static ccv_cnnp_model_t* _ccv_cnnp_flatten_copy(const ccv_cnnp_model_t* const self, void* const context)
946{
947 return ccv_cnnp_flatten(self->name);
948}
949
950// MARK - Batch Norm Layer
951
952typedef struct {
953 ccv_cnnp_model_t super;
954 ccv_nnc_tensor_symbol_t output;
955 ccv_nnc_tensor_symbol_t bias;
956 ccv_nnc_tensor_symbol_t scale;
957 ccv_nnc_graph_exec_symbol_t batch_norm;
958 ccv_nnc_cmd_param_t params;
959 ccv_array_t* zero_inits;
960 ccv_array_t* retainables;
961} ccv_cnnp_model_batch_norm_t;
962
963static 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)
964{
965 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"
, 965, __extension__ __PRETTY_FUNCTION__); }));
966 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", 966, __extension__ __PRETTY_FUNCTION__
); }));
967 ccv_cnnp_model_batch_norm_t* const self = (ccv_cnnp_model_batch_norm_t*)super;
968 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);
969 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
970 const int nd = ccv_nnc_tensor_nd(params.dim);
971 ccv_nnc_tensor_param_t bias_params = params;
972 memset(bias_params.dim, 0, sizeof(bias_params.dim));
973 // If the accuracy is not enough, bump it to 32-bit floating point.
974 if (bias_params.datatype != CCV_32F && bias_params.datatype != CCV_64F)
975 bias_params.datatype = CCV_32F;
976 bias_params.dim[0] = nd > 1 ? ccv_nnc_tensor_get_c(params) : params.dim[0];
977 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, params, 0);
978 // Both scale and bias are shared between if this model is reused.
979 if (!self->scale.graph)
980 self->scale = ccv_nnc_tensor_symbol_new(graph, bias_params, "scale");
981 if (!self->bias.graph)
982 self->bias = ccv_nnc_tensor_symbol_new(graph, bias_params, "bias");
983 const ccv_nnc_tensor_symbol_t scale = ccv_cnnp_model_get_symbol(super, self->scale);
984 const ccv_nnc_tensor_symbol_t bias = ccv_cnnp_model_get_symbol(super, self->bias);
985 const ccv_nnc_tensor_symbol_t mean = ccv_nnc_tensor_symbol_new(graph, bias_params, "mean");
986 const ccv_nnc_tensor_symbol_t var = ccv_nnc_tensor_symbol_new(graph, bias_params, "var");
987 // Otherwise, notice mean, var, saved_mean, saved_inv_std are not reused.
988 if (!self->zero_inits)
989 self->zero_inits = ccv_array_new(sizeof(ccv_nnc_tensor_symbol_t), 0, 0);
990 ccv_array_push(self->zero_inits, &mean);
991 ccv_array_push(self->zero_inits, &var);
992 const ccv_nnc_tensor_symbol_t out_mean = ccv_nnc_tensor_symbol_new(graph, bias_params, "out_mean");
993 const ccv_nnc_tensor_symbol_t out_var = ccv_nnc_tensor_symbol_new(graph, bias_params, "out_var");
994 if (!self->retainables)
995 self->retainables = ccv_array_new(sizeof(ccv_nnc_tensor_symbol_t), 0, 0);
996 ccv_array_push(self->retainables, &out_mean);
997 ccv_array_push(self->retainables, &out_var);
998 const ccv_nnc_tensor_symbol_t saved_mean = ccv_nnc_tensor_symbol_new(graph, bias_params, "saved_mean");
999 const ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(graph, bias_params, "saved_inv_std");
1000 const int hw = ccv_nnc_tensor_hw(params, ccv_nnc_tensor_nd(params.dim), CCV_NNC_MAX_DIM(2));
1001 ccv_nnc_cmd_param_t batch_norm = self->params;
1002 batch_norm.bnorm.count = hw >= 0 ? CCV_NNC_MAX_DIM(2) + 1 : 1;
1003 int i;
1004 batch_norm.bnorm.axis[0] = (params.format == CCV_TENSOR_FORMAT_CHWN) ? 3 : 0;
1005 if (hw >= 0)
1006 for (i = 0; i < CCV_NNC_MAX_DIM(2); i++)
1007 batch_norm.bnorm.axis[i + 1] = i + hw;
1008 self->params = batch_norm;
1009 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");
1010 outputs[0] = output;
1011}
1012
1013static 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)
1014{
1015 ccv_cnnp_model_batch_norm_t* const self = (ccv_cnnp_model_batch_norm_t*)super;
1016 if (self->scale.graph)
1017 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);
1018 if (self->bias.graph)
1019 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);
1020 int i;
1021 if (self->zero_inits)
1022 for (i = 0; i < self->zero_inits->rnum; i++)
1023 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))));
1024}
1025
1026static 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)
1027{
1028 ccv_cnnp_model_batch_norm_t* const self = (ccv_cnnp_model_batch_norm_t*)super;
1029 if (self->scale.graph)
1030 add_to_array(parameters, self->scale, is_trainable);
1031 if (self->bias.graph)
1032 add_to_array(parameters, self->bias, is_trainable);
1033}
1034
1035static 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)
1036{
1037 ccv_cnnp_model_batch_norm_t* const self = (ccv_cnnp_model_batch_norm_t*)super;
1038 int i;
1039 if (self->retainables)
1040 for (i = 0; i < self->retainables->rnum; i++)
1041 {
1042 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)));
1043 add_to_array(outputs, symbol, 0);
1044 }
1045}
1046
1047static 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)
1048{
1049 ccv_cnnp_model_batch_norm_t* const self = (ccv_cnnp_model_batch_norm_t*)super;
1050 if (self->batch_norm.graph)
1051 {
1052 self->params.bnorm.is_test = is_test;
1053 updater(context, self->batch_norm, ccv_nnc_cmd(CCV_NNC_BATCH_NORM_FORWARD, 0, self->params, 0), ccv_nnc_no_hint);
1054 }
1055}
1056
1057static void _ccv_cnnp_batch_norm_deinit(ccv_cnnp_model_t* const super)
1058{
1059 ccv_cnnp_model_batch_norm_t* const self = (ccv_cnnp_model_batch_norm_t*)super;
1060 if (self->zero_inits)
1061 ccv_array_free(self->zero_inits);
1062 if (self->retainables)
1063 ccv_array_free(self->retainables);
1064}
1065
1066static ccv_cnnp_model_t* _ccv_cnnp_batch_norm_copy(const ccv_cnnp_model_t* const super, void* const context);
1067
1068static const ccv_cnnp_model_vtab_t ccv_cnnp_batch_norm_isa = {
1069 .build = _ccv_cnnp_batch_norm_build,
1070 .init_states = _ccv_cnnp_batch_norm_init_states,
1071 .add_to_parameter = _ccv_cnnp_batch_norm_add_to_parameter,
1072 .add_to_output = _ccv_cnnp_batch_norm_add_to_output,
1073 .copy = _ccv_cnnp_batch_norm_copy,
1074 .set_is_test = _ccv_cnnp_batch_norm_set_is_test,
1075 .deinit = _ccv_cnnp_batch_norm_deinit,
1076};
1077
1078ccv_cnnp_model_t* ccv_cnnp_batch_norm(const float momentum, const float epsilon, const int is_trainable, const char* const name)
1079{
1080 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));
1081 model_batch_norm->super.isa = &ccv_cnnp_batch_norm_isa;
1082 model_batch_norm->super.input_size = 1;
1083 model_batch_norm->super.outputs = &model_batch_norm->output;
1084 model_batch_norm->super.output_size = 1;
1085 model_batch_norm->super.is_trainable = is_trainable;
1086 ccv_cnnp_model_copy_name(&model_batch_norm->super, name);
1087 model_batch_norm->scale.d = CCV_NNC_NO_TENSOR_SYMBOL;
1088 model_batch_norm->scale.graph = 0;
1089 model_batch_norm->bias.d = CCV_NNC_NO_TENSOR_SYMBOL;
1090 model_batch_norm->bias.graph = 0;
1091 model_batch_norm->params.bnorm.momentum = momentum;
1092 model_batch_norm->params.bnorm.epsilon = epsilon;
1093 return (ccv_cnnp_model_t*)model_batch_norm;
1094}
1095
1096static ccv_cnnp_model_t* _ccv_cnnp_batch_norm_copy(const ccv_cnnp_model_t* const super, void* const context)
1097{
1098 const ccv_cnnp_model_batch_norm_t* const self = (const ccv_cnnp_model_batch_norm_t*)super;
1099 return ccv_cnnp_batch_norm(self->params.bnorm.momentum, self->params.bnorm.epsilon, self->super.is_trainable, self->super.name);
1100}
1101
1102// MARK - Convolution Layer
1103
1104typedef struct {
1105 ccv_cnnp_model_t super;
1106 ccv_nnc_tensor_symbol_t output;
1107 ccv_nnc_tensor_symbol_t weights;
1108 ccv_nnc_tensor_symbol_t bias;
1109 int groups;
1110 int filters;
1111 int kdim[CCV_NNC_MAX_DIM_ALLOC(12)];
1112 int dilation[CCV_NNC_MAX_DIM_ALLOC(12)];
1113 int no_bias;
1114 int format;
1115 ccv_nnc_hint_t hint;
1116} ccv_cnnp_model_convolution_t;
1117
1118static 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)
1119{
1120 ccv_cnnp_model_convolution_t* const self = (ccv_cnnp_model_convolution_t*)super;
1121 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);
1122 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"
, 1122, __extension__ __PRETTY_FUNCTION__); }));
1123 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", 1123, __extension__ __PRETTY_FUNCTION__
); }));
1124 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1125 int i;
1126 const int k_nd = ccv_nnc_tensor_nd(self->kdim);
1127 const int nd = k_nd + 2;
1128 ccv_nnc_tensor_param_t weights_params = params;
1129 if (self->format)
1130 weights_params.format = self->format;
1131 ccv_nnc_tensor_set_n(&weights_params, self->filters);
1132 const int a_nd = ccv_nnc_tensor_nd(params.dim);
1133 int c;
1134 switch (params.format)
1135 {
1136 case CCV_TENSOR_FORMAT_NHWC:
1137 c = params.dim[a_nd - 1];
1138 break;
1139 case CCV_TENSOR_FORMAT_NCHW:
1140 if (a_nd == k_nd + 1)
1141 c = params.dim[0];
1142 else
1143 c = params.dim[a_nd <= 1 ? 0 : 1];
1144 break;
1145 case CCV_TENSOR_FORMAT_CHWN:
1146 c = params.dim[0];
1147 break;
1148 }
1149 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", 1149, __extension__ __PRETTY_FUNCTION__
); }));
1150 ccv_nnc_tensor_set_c(&weights_params, nd, c / self->groups);
1151 int hw = -1;
1152 if (weights_params.format == CCV_TENSOR_FORMAT_NHWC || weights_params.format == CCV_TENSOR_FORMAT_CHWN)
1153 hw = 1;
1154 else if (weights_params.format == CCV_TENSOR_FORMAT_NCHW)
1155 hw = 2;
1156 assert(hw >= 0)((void) sizeof ((hw >= 0) ? 1 : 0), __extension__ ({ if (hw
>= 0) ; else __assert_fail ("hw >= 0", "ccv_cnnp_model_addons.c"
, 1156, __extension__ __PRETTY_FUNCTION__); }));
1157 for (i = 0; i < k_nd; i++)
1158 weights_params.dim[i + hw] = self->kdim[i];
1159 if (!self->weights.graph)
1160 self->weights = ccv_nnc_tensor_symbol_new(graph, weights_params, "weights");
1161 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"
, 1161, __extension__ __PRETTY_FUNCTION__); }));
1162 const ccv_nnc_tensor_symbol_t weights = ccv_cnnp_model_get_symbol(super, self->weights);
1163 ccv_nnc_tensor_param_t bias_params = params;
1164 if (self->format)
1165 bias_params.format = self->format;
1166 memset(bias_params.dim, 0, sizeof(bias_params.dim));
1167 bias_params.dim[0] = self->filters;
1168 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);
1169 for (i = 0; i < k_nd; i++)
1170 cmd.info.size.dim[i] = self->kdim[i];
1171 cmd.info.size.dim[k_nd] = c;
1172 memcpy(cmd.info.convolution.dilation, self->dilation, sizeof(self->dilation));
1173 ccv_nnc_tensor_param_t output_params;
1174 // Dilate weight size based on the dilation factor.
1175 for (i = 0; i < k_nd; i++)
1176 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;
1177 ccv_nnc_hint_tensor_auto(cmd, (ccv_nnc_tensor_param_t []){
1178 params,
1179 weights_params,
1180 bias_params,
1181 }, 3, self->hint, &output_params, 1);
1182 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1183 ccv_nnc_graph_exec_symbol_t convolution;
1184 if (self->no_bias)
1185 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");
1186 else {
1187 if (!self->bias.graph)
1188 self->bias = ccv_nnc_tensor_symbol_new(graph, bias_params, "bias");
1189 const ccv_nnc_tensor_symbol_t bias = ccv_cnnp_model_get_symbol(super, self->bias);
1190 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");
1191 }
1192 ccv_nnc_graph_exec_symbol_set_hint(graph, convolution, self->hint);
1193 outputs[0] = output;
1194}
1195
1196static 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)
1197{
1198 ccv_cnnp_model_convolution_t* const self = (ccv_cnnp_model_convolution_t*)super;
1199 const ccv_nnc_tensor_param_t weight_params = ccv_nnc_tensor_symbol_params(graph, self->weights);
1200 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
; });
1201 const int count = ccv_nnc_tensor_count(weight_params);
1202 const float std = sqrtf(2) / sqrtf(count / n);
1203 const float bound = sqrtf(3) * std;
1204 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);
1205 if (self->bias.graph)
1206 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);
1207}
1208
1209static 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)
1210{
1211 ccv_cnnp_model_convolution_t* const self = (ccv_cnnp_model_convolution_t*)super;
1212 add_to_array(parameters, self->weights, is_trainable);
1213 if (self->bias.graph)
1214 add_to_array(parameters, self->bias, is_trainable);
1215}
1216
1217static ccv_cnnp_model_t* _ccv_cnnp_convolution_copy(const ccv_cnnp_model_t* const super, void* const context);
1218
1219static const ccv_cnnp_model_vtab_t ccv_cnnp_convolution_isa = {
1220 .build = _ccv_cnnp_convolution_build,
1221 .init_states = _ccv_cnnp_convolution_init_states,
1222 .add_to_parameter = _ccv_cnnp_convolution_add_to_parameter,
1223 .copy = _ccv_cnnp_convolution_copy,
1224};
1225
1226ccv_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)
1227{
1228 ccv_cnnp_model_convolution_t* const model_convolution = (ccv_cnnp_model_convolution_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_convolution_t));
1229 model_convolution->super.isa = &ccv_cnnp_convolution_isa;
1230 model_convolution->super.input_size = 1;
1231 model_convolution->super.outputs = &model_convolution->output;
1232 model_convolution->super.output_size = 1;
1233 model_convolution->super.is_trainable = is_trainable;
1234 ccv_cnnp_model_copy_name(&model_convolution->super, name);
1235 model_convolution->weights.d = CCV_NNC_NO_TENSOR_SYMBOL;
1236 model_convolution->weights.graph = 0;
1237 model_convolution->bias.d = CCV_NNC_NO_TENSOR_SYMBOL;
1238 model_convolution->bias.graph = 0;
1239 model_convolution->groups = groups;
1240 model_convolution->filters = filters;
1241 memcpy(model_convolution->kdim, kdim, sizeof(model_convolution->kdim));
1242 memcpy(model_convolution->dilation, dilation, sizeof(model_convolution->dilation));
1243 model_convolution->no_bias = no_bias;
1244 model_convolution->hint = hint;
1245 model_convolution->format = format;
1246 return (ccv_cnnp_model_t*)model_convolution;
1247}
1248
1249static ccv_cnnp_model_t* _ccv_cnnp_convolution_copy(const ccv_cnnp_model_t* const super, void* const context)
1250{
1251 ccv_cnnp_model_convolution_t* const self = (ccv_cnnp_model_convolution_t*)super;
1252 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);
1253}
1254
1255// MARK - Convolution Transpose Layer
1256
1257typedef struct {
1258 ccv_cnnp_model_t super;
1259 ccv_nnc_tensor_symbol_t output;
1260 ccv_nnc_tensor_symbol_t weights;
1261 ccv_nnc_tensor_symbol_t bias;
1262 int groups;
1263 int filters;
1264 int kdim[CCV_NNC_MAX_DIM_ALLOC(12)];
1265 int dilation[CCV_NNC_MAX_DIM_ALLOC(12)];
1266 int output_padding;
1267 int no_bias;
1268 int format;
1269 ccv_nnc_hint_t hint;
1270} ccv_cnnp_model_convolution_transpose_t;
1271
1272static 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)
1273{
1274 ccv_cnnp_model_convolution_transpose_t* const self = (ccv_cnnp_model_convolution_transpose_t*)super;
1275 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);
1276 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"
, 1276, __extension__ __PRETTY_FUNCTION__); }));
1277 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", 1277, __extension__ __PRETTY_FUNCTION__
); }));
1278 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1279 int i;
1280 const int nd = CCV_NNC_MAX_DIM(2) + 2;
1281 ccv_nnc_tensor_param_t weights_params = params;
1282 if (self->format)
1283 weights_params.format = self->format;
1284 const int c = ccv_nnc_tensor_get_c(params);
1285 ccv_nnc_tensor_set_n(&weights_params, c);
1286 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", 1286, __extension__ __PRETTY_FUNCTION__
); }));
1287 ccv_nnc_tensor_set_c(&weights_params, nd, self->filters / self->groups);
1288 const int hw = ccv_nnc_tensor_hw(weights_params, nd, CCV_NNC_MAX_DIM(2));
1289 assert(hw >= 0)((void) sizeof ((hw >= 0) ? 1 : 0), __extension__ ({ if (hw
>= 0) ; else __assert_fail ("hw >= 0", "ccv_cnnp_model_addons.c"
, 1289, __extension__ __PRETTY_FUNCTION__); }));
1290 for (i = 0; i < CCV_NNC_MAX_DIM(2); i++)
1291 weights_params.dim[i + hw] = self->kdim[i];
1292 if (!self->weights.graph)
1293 self->weights = ccv_nnc_tensor_symbol_new(graph, weights_params, "weights");
1294 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"
, 1294, __extension__ __PRETTY_FUNCTION__); }));
1295 const ccv_nnc_tensor_symbol_t weights = ccv_cnnp_model_get_symbol(super, self->weights);
1296 ccv_nnc_tensor_param_t bias_params = params;
1297 if (self->format)
1298 bias_params.format = self->format;
1299 memset(bias_params.dim, 0, sizeof(bias_params.dim));
1300 bias_params.dim[0] = self->filters;
1301 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);
1302 for (i = 0; i < CCV_NNC_MAX_DIM(2); i++)
1303 cmd.info.size.dim[i] = self->kdim[i];
1304 cmd.info.size.dim[CCV_NNC_MAX_DIM(2)] = c;
1305 memcpy(cmd.info.convolution_transpose.dilation, self->dilation, sizeof(self->dilation));
1306 ccv_nnc_tensor_param_t output_params;
1307 // Dilate weight size based on the dilation factor.
1308 for (i = 0; i < CCV_NNC_MAX_DIM(2); i++)
1309 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;
1310 ccv_nnc_hint_tensor_auto(cmd, (ccv_nnc_tensor_param_t []){
1311 params,
1312 weights_params,
1313 bias_params,
1314 }, 3, self->hint, &output_params, 1);
1315 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1316 ccv_nnc_graph_exec_symbol_t convolution_transpose;
1317 if (self->no_bias)
1318 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");
1319 else {
1320 if (!self->bias.graph)
1321 self->bias = ccv_nnc_tensor_symbol_new(graph, bias_params, "bias");
1322 const ccv_nnc_tensor_symbol_t bias = ccv_cnnp_model_get_symbol(super, self->bias);
1323 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");
1324 }
1325 ccv_nnc_graph_exec_symbol_set_hint(graph, convolution_transpose, self->hint);
1326 outputs[0] = output;
1327}
1328
1329static 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)
1330{
1331 ccv_cnnp_model_convolution_transpose_t* const self = (ccv_cnnp_model_convolution_transpose_t*)super;
1332 const ccv_nnc_tensor_param_t weight_params = ccv_nnc_tensor_symbol_params(graph, self->weights);
1333 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
; });
1334 const int count = ccv_nnc_tensor_count(weight_params);
1335 const float std = sqrtf(2) / sqrtf(count / n);
1336 const float bound = sqrtf(3) * std;
1337 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);
1338 if (self->bias.graph)
1339 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);
1340}
1341
1342static 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)
1343{
1344 ccv_cnnp_model_convolution_transpose_t* const self = (ccv_cnnp_model_convolution_transpose_t*)super;
1345 add_to_array(parameters, self->weights, is_trainable);
1346 if (self->bias.graph)
1347 add_to_array(parameters, self->bias, is_trainable);
1348}
1349
1350static ccv_cnnp_model_t* _ccv_cnnp_convolution_transpose_copy(const ccv_cnnp_model_t* const super, void* const context);
1351
1352static const ccv_cnnp_model_vtab_t ccv_cnnp_convolution_transpose_isa = {
1353 .build = _ccv_cnnp_convolution_transpose_build,
1354 .init_states = _ccv_cnnp_convolution_transpose_init_states,
1355 .add_to_parameter = _ccv_cnnp_convolution_transpose_add_to_parameter,
1356 .copy = _ccv_cnnp_convolution_transpose_copy,
1357};
1358
1359ccv_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)
1360{
1361 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));
1362 model_convolution_transpose->super.isa = &ccv_cnnp_convolution_transpose_isa;
1363 model_convolution_transpose->super.input_size = 1;
1364 model_convolution_transpose->super.outputs = &model_convolution_transpose->output;
1365 model_convolution_transpose->super.output_size = 1;
1366 model_convolution_transpose->super.is_trainable = is_trainable;
1367 ccv_cnnp_model_copy_name(&model_convolution_transpose->super, name);
1368 model_convolution_transpose->weights.d = CCV_NNC_NO_TENSOR_SYMBOL;
1369 model_convolution_transpose->weights.graph = 0;
1370 model_convolution_transpose->bias.d = CCV_NNC_NO_TENSOR_SYMBOL;
1371 model_convolution_transpose->bias.graph = 0;
1372 model_convolution_transpose->groups = groups;
1373 model_convolution_transpose->filters = filters;
1374 memcpy(model_convolution_transpose->kdim, kdim, sizeof(model_convolution_transpose->kdim));
1375 memcpy(model_convolution_transpose->dilation, dilation, sizeof(model_convolution_transpose->dilation));
1376 model_convolution_transpose->output_padding = output_padding;
1377 model_convolution_transpose->no_bias = no_bias;
1378 model_convolution_transpose->hint = hint;
1379 model_convolution_transpose->format = format;
1380 return (ccv_cnnp_model_t*)model_convolution_transpose;
1381}
1382
1383static ccv_cnnp_model_t* _ccv_cnnp_convolution_transpose_copy(const ccv_cnnp_model_t* const super, void* const context)
1384{
1385 ccv_cnnp_model_convolution_transpose_t* const self = (ccv_cnnp_model_convolution_transpose_t*)super;
1386 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);
1387}
1388
1389// MARK - Dense Layer
1390
1391typedef struct {
1392 ccv_cnnp_model_t super;
1393 ccv_nnc_tensor_symbol_t output;
1394 ccv_nnc_tensor_symbol_t weights;
1395 ccv_nnc_tensor_symbol_t bias;
1396 int count;
1397 int no_bias;
1398 int flags;
1399} ccv_cnnp_model_dense_t;
1400
1401static 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)
1402{
1403 ccv_cnnp_model_dense_t* const self = (ccv_cnnp_model_dense_t*)super;
1404 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);
1405 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"
, 1405, __extension__ __PRETTY_FUNCTION__); }));
1406 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", 1406, __extension__ __PRETTY_FUNCTION__
); }));
1407 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1408 ccv_nnc_tensor_param_t weights_params = params;
1409 memset(weights_params.dim, 0, sizeof(weights_params.dim));
1410 weights_params.dim[0] = self->count;
1411 weights_params.dim[1] = params.dim[ccv_nnc_tensor_nd(params.dim) - 1];
1412 if (!self->weights.graph)
1413 self->weights = ccv_nnc_tensor_symbol_new(graph, weights_params, "weights");
1414 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"
, 1414, __extension__ __PRETTY_FUNCTION__); }));
1415 const ccv_nnc_tensor_symbol_t weights = ccv_cnnp_model_get_symbol(super, self->weights);
1416 ccv_nnc_tensor_param_t bias_params = params;
1417 memset(bias_params.dim, 0, sizeof(bias_params.dim));
1418 bias_params.dim[0] = self->count;
1419 ccv_nnc_cmd_t cmd = {0};
1420 cmd.cmd = CCV_NNC_GEMM_FORWARD;
1421 cmd.info.blas.a[0] = 1;
1422 cmd.info.blas.a[1] = 1;
1423 cmd.info.blas.transpose_b[0] = 0;
1424 cmd.info.blas.transpose_b[1] = 1;
1425 cmd.info.blas.flags = self->flags;
1426 ccv_nnc_tensor_param_t output_params;
1427 ccv_nnc_hint_tensor_auto(cmd, (ccv_nnc_tensor_param_t []){
1428 params,
1429 weights_params,
1430 bias_params,
1431 }, 3, ccv_nnc_no_hint, &output_params, 1);
1432 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1433 if (self->no_bias)
1434 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");
1435 else {
1436 if (!self->bias.graph)
1437 self->bias = ccv_nnc_tensor_symbol_new(graph, bias_params, "bias");
1438 const ccv_nnc_tensor_symbol_t bias = ccv_cnnp_model_get_symbol(super, self->bias);
1439 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");
1440 }
1441 outputs[0] = output;
1442}
1443
1444static 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)
1445{
1446 ccv_cnnp_model_dense_t* const self = (ccv_cnnp_model_dense_t*)super;
1447 const ccv_nnc_tensor_param_t weight_params = ccv_nnc_tensor_symbol_params(graph, self->weights);
1448 const int c = weight_params.dim[1];
1449 const float std = sqrtf(2) / sqrtf(c);
1450 const float bound = sqrtf(3) * std;
1451 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);
1452 if (self->bias.graph)
1453 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);
1454}
1455
1456static 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)
1457{
1458 ccv_cnnp_model_dense_t* const self = (ccv_cnnp_model_dense_t*)super;
1459 add_to_array(parameters, self->weights, is_trainable);
1460 if (self->bias.graph)
1461 add_to_array(parameters, self->bias, is_trainable);
1462}
1463
1464static ccv_cnnp_model_t* _ccv_cnnp_dense_copy(const ccv_cnnp_model_t* const super, void* const context);
1465
1466static const ccv_cnnp_model_vtab_t ccv_cnnp_dense_isa = {
1467 .build = _ccv_cnnp_dense_build,
1468 .init_states = _ccv_cnnp_dense_init_states,
1469 .add_to_parameter = _ccv_cnnp_dense_add_to_parameter,
1470 .copy = _ccv_cnnp_dense_copy,
1471};
1472
1473ccv_cnnp_model_t* ccv_cnnp_dense(const int count, const int no_bias, const int flags, const int is_trainable, const char* const name)
1474{
1475 ccv_cnnp_model_dense_t* const model_dense = (ccv_cnnp_model_dense_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_dense_t));
1476 model_dense->super.isa = &ccv_cnnp_dense_isa;
1477 model_dense->super.input_size = 1;
1478 model_dense->super.outputs = &model_dense->output;
1479 model_dense->super.output_size = 1;
1480 model_dense->super.is_trainable = is_trainable;
1481 ccv_cnnp_model_copy_name(&model_dense->super, name);
1482 model_dense->weights.d = CCV_NNC_NO_TENSOR_SYMBOL;
1483 model_dense->weights.graph = 0;
1484 model_dense->bias.d = CCV_NNC_NO_TENSOR_SYMBOL;
1485 model_dense->bias.graph = 0;
1486 model_dense->count = count;
1487 model_dense->no_bias = no_bias;
1488 model_dense->flags = flags;
1489 return (ccv_cnnp_model_t*)model_dense;
1490}
1491
1492static ccv_cnnp_model_t* _ccv_cnnp_dense_copy(const ccv_cnnp_model_t* const super, void* const context)
1493{
1494 const ccv_cnnp_model_dense_t* const self = (const ccv_cnnp_model_dense_t*)super;
1495 return ccv_cnnp_dense(self->count, self->no_bias, self->flags, self->super.is_trainable, self->super.name);
1496}
1497
1498// MARK - Pool Layers
1499
1500typedef struct {
1501 ccv_cnnp_model_t super;
1502 ccv_nnc_tensor_symbol_t output;
1503 int kdim[CCV_NNC_MAX_DIM_ALLOC(12)];
1504 ccv_nnc_hint_t hint;
1505} ccv_cnnp_model_pool_t;
1506
1507static 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)
1508{
1509 ccv_cnnp_model_pool_t* const self = (ccv_cnnp_model_pool_t*)super;
1510 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);
1511 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"
, 1511, __extension__ __PRETTY_FUNCTION__); }));
1512 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", 1512, __extension__ __PRETTY_FUNCTION__
); }));
1513 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1514 const int hw = ccv_nnc_tensor_hw(params, ccv_nnc_tensor_nd(params.dim), CCV_NNC_MAX_DIM(2));
1515 ccv_nnc_cmd_t cmd;
1516 if (hw >= 0 && self->kdim[0] == 0 && self->kdim[1] == 0)
1517 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);
1518 else
1519 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);
1520 ccv_nnc_tensor_param_t output_params;
1521 ccv_nnc_hint_tensor_auto(cmd, &params, 1, self->hint, &output_params, 1);
1522 const ccv_nnc_tensor_symbol_t pool_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1523 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");
1524 ccv_nnc_graph_exec_symbol_set_hint(graph, exec, self->hint);
1525 outputs[0] = pool_output;
1526}
1527
1528static ccv_cnnp_model_t* _ccv_cnnp_max_pool_copy(const ccv_cnnp_model_t* const super, void* const context);
1529
1530static const ccv_cnnp_model_vtab_t ccv_cnnp_max_pool_isa = {
1531 .build = _ccv_cnnp_max_pool_build,
1532 .copy = _ccv_cnnp_max_pool_copy,
1533};
1534
1535ccv_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)
1536{
1537 ccv_cnnp_model_pool_t* const model_pool = (ccv_cnnp_model_pool_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_pool_t));
1538 model_pool->super.isa = &ccv_cnnp_max_pool_isa;
1539 model_pool->super.input_size = 1;
1540 model_pool->super.outputs = &model_pool->output;
1541 model_pool->super.output_size = 1;
1542 ccv_cnnp_model_copy_name(&model_pool->super, name);
1543 memcpy(model_pool->kdim, kdim, sizeof(model_pool->kdim));
1544 model_pool->hint = hint;
1545 return (ccv_cnnp_model_t*)model_pool;
1546}
1547
1548static ccv_cnnp_model_t* _ccv_cnnp_max_pool_copy(const ccv_cnnp_model_t* const super, void* const context)
1549{
1550 const ccv_cnnp_model_pool_t* const self = (const ccv_cnnp_model_pool_t*)super;
1551 return ccv_cnnp_max_pool(self->kdim, self->hint, self->super.name);
1552}
1553
1554static 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)
1555{
1556 ccv_cnnp_model_pool_t* const self = (ccv_cnnp_model_pool_t*)super;
1557 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);
1558 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"
, 1558, __extension__ __PRETTY_FUNCTION__); }));
1559 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", 1559, __extension__ __PRETTY_FUNCTION__
); }));
1560 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1561 const int hw = ccv_nnc_tensor_hw(params, ccv_nnc_tensor_nd(params.dim), CCV_NNC_MAX_DIM(2));
1562 ccv_nnc_cmd_t cmd;
1563 if (hw >= 0 && self->kdim[0] == 0 && self->kdim[1] == 0)
1564 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);
1565 else
1566 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);
1567 ccv_nnc_tensor_param_t output_params;
1568 ccv_nnc_hint_tensor_auto(cmd, &params, 1, self->hint, &output_params, 1);
1569 const ccv_nnc_tensor_symbol_t pool_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1570 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");
1571 ccv_nnc_graph_exec_symbol_set_hint(graph, exec, self->hint);
1572 outputs[0] = pool_output;
1573}
1574
1575static ccv_cnnp_model_t* _ccv_cnnp_average_pool_copy(const ccv_cnnp_model_t* const super, void* const context);
1576
1577static const ccv_cnnp_model_vtab_t ccv_cnnp_average_pool_isa = {
1578 .build = _ccv_cnnp_average_pool_build,
1579 .copy = _ccv_cnnp_average_pool_copy,
1580};
1581
1582ccv_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)
1583{
1584 ccv_cnnp_model_pool_t* const model_pool = (ccv_cnnp_model_pool_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_pool_t));
1585 model_pool->super.isa = &ccv_cnnp_average_pool_isa;
1586 model_pool->super.input_size = 1;
1587 model_pool->super.outputs = &model_pool->output;
1588 model_pool->super.output_size = 1;
1589 ccv_cnnp_model_copy_name(&model_pool->super, name);
1590 memcpy(model_pool->kdim, kdim, sizeof(model_pool->kdim));
1591 model_pool->hint = hint;
1592 return (ccv_cnnp_model_t*)model_pool;
1593}
1594
1595static ccv_cnnp_model_t* _ccv_cnnp_average_pool_copy(const ccv_cnnp_model_t* const super, void* const context)
1596{
1597 const ccv_cnnp_model_pool_t* const self = (const ccv_cnnp_model_pool_t*)super;
1598 return ccv_cnnp_average_pool(self->kdim, self->hint, self->super.name);
1599}
1600
1601// MARK - RELU Layer
1602
1603typedef struct {
1604 ccv_cnnp_model_t super;
1605 ccv_nnc_tensor_symbol_t output;
1606} ccv_cnnp_model_relu_t;
1607
1608static 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)
1609{
1610 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);
1611 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"
, 1611, __extension__ __PRETTY_FUNCTION__); }));
1612 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", 1612, __extension__ __PRETTY_FUNCTION__
); }));
1613 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1614 ccv_nnc_tensor_param_t output_params;
1615 const ccv_nnc_cmd_t relu = CMD_RELU_FORWARD()ccv_nnc_cmd(CCV_NNC_RELU_FORWARD, 0, ccv_nnc_cmd_auto, 0);
1616 ccv_nnc_hint_tensor_auto(relu, (ccv_nnc_tensor_param_t []){
1617 params,
1618 }, 1, ccv_nnc_no_hint, &output_params, 1);
1619 const ccv_nnc_tensor_symbol_t relu_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1620 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");
1621 outputs[0] = relu_output;
1622}
1623
1624static ccv_cnnp_model_t* _ccv_cnnp_relu_copy(const ccv_cnnp_model_t* const self, void* const context);
1625
1626static const ccv_cnnp_model_vtab_t ccv_cnnp_relu_isa = {
1627 .build = _ccv_cnnp_relu_build,
1628 .copy = _ccv_cnnp_relu_copy,
1629};
1630
1631ccv_cnnp_model_t* ccv_cnnp_relu(const char* const name)
1632{
1633 ccv_cnnp_model_relu_t* const model_relu = (ccv_cnnp_model_relu_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_relu_t));
1634 model_relu->super.isa = &ccv_cnnp_relu_isa;
1635 model_relu->super.input_size = 1;
1636 model_relu->super.outputs = &model_relu->output;
1637 model_relu->super.output_size = 1;
1638 ccv_cnnp_model_copy_name(&model_relu->super, name);
1639 return (ccv_cnnp_model_t*)model_relu;
1640}
1641
1642static ccv_cnnp_model_t* _ccv_cnnp_relu_copy(const ccv_cnnp_model_t* const self, void* const context)
1643{
1644 return ccv_cnnp_relu(self->name);
1645}
1646
1647// MARK - Sigmoid Layer
1648
1649typedef struct {
1650 ccv_cnnp_model_t super;
1651 ccv_nnc_tensor_symbol_t output;
1652} ccv_cnnp_model_sigmoid_t;
1653
1654static 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)
1655{
1656 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);
1657 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"
, 1657, __extension__ __PRETTY_FUNCTION__); }));
1658 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", 1658, __extension__ __PRETTY_FUNCTION__
); }));
1659 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1660 ccv_nnc_tensor_param_t output_params;
1661 const ccv_nnc_cmd_t sigmoid = CMD_SIGMOID_FORWARD()ccv_nnc_cmd(CCV_NNC_SIGMOID_FORWARD, 0, ccv_nnc_cmd_auto, 0);
1662 ccv_nnc_hint_tensor_auto(sigmoid, (ccv_nnc_tensor_param_t []){
1663 params,
1664 }, 1, ccv_nnc_no_hint, &output_params, 1);
1665 const ccv_nnc_tensor_symbol_t sigmoid_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1666 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");
1667 outputs[0] = sigmoid_output;
1668}
1669
1670static ccv_cnnp_model_t* _ccv_cnnp_sigmoid_copy(const ccv_cnnp_model_t* const self, void* const context);
1671
1672static const ccv_cnnp_model_vtab_t ccv_cnnp_sigmoid_isa = {
1673 .build = _ccv_cnnp_sigmoid_build,
1674 .copy = _ccv_cnnp_sigmoid_copy,
1675};
1676
1677ccv_cnnp_model_t* ccv_cnnp_sigmoid(const char* const name)
1678{
1679 ccv_cnnp_model_sigmoid_t* const model_sigmoid = (ccv_cnnp_model_sigmoid_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_sigmoid_t));
1680 model_sigmoid->super.isa = &ccv_cnnp_sigmoid_isa;
1681 model_sigmoid->super.input_size = 1;
1682 model_sigmoid->super.outputs = &model_sigmoid->output;
1683 model_sigmoid->super.output_size = 1;
1684 ccv_cnnp_model_copy_name(&model_sigmoid->super, name);
1685 return (ccv_cnnp_model_t*)model_sigmoid;
1686}
1687
1688static ccv_cnnp_model_t* _ccv_cnnp_sigmoid_copy(const ccv_cnnp_model_t* const self, void* const context)
1689{
1690 return ccv_cnnp_sigmoid(self->name);
1691}
1692
1693// MARK - Tanh Layer
1694
1695typedef struct {
1696 ccv_cnnp_model_t super;
1697 ccv_nnc_tensor_symbol_t output;
1698} ccv_cnnp_model_tanh_t;
1699
1700static 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)
1701{
1702 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);
1703 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"
, 1703, __extension__ __PRETTY_FUNCTION__); }));
1704 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", 1704, __extension__ __PRETTY_FUNCTION__
); }));
1705 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1706 ccv_nnc_tensor_param_t output_params;
1707 const ccv_nnc_cmd_t tanh = CMD_TANH_FORWARD()ccv_nnc_cmd(CCV_NNC_TANH_FORWARD, 0, ccv_nnc_cmd_auto, 0);
1708 ccv_nnc_hint_tensor_auto(tanh, (ccv_nnc_tensor_param_t []){
1709 params,
1710 }, 1, ccv_nnc_no_hint, &output_params, 1);
1711 const ccv_nnc_tensor_symbol_t tanh_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1712 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");
1713 outputs[0] = tanh_output;
1714}
1715
1716static ccv_cnnp_model_t* _ccv_cnnp_tanh_copy(const ccv_cnnp_model_t* const self, void* const context);
1717
1718static const ccv_cnnp_model_vtab_t ccv_cnnp_tanh_isa = {
1719 .build = _ccv_cnnp_tanh_build,
1720 .copy = _ccv_cnnp_tanh_copy,
1721};
1722
1723ccv_cnnp_model_t* ccv_cnnp_tanh(const char* const name)
1724{
1725 ccv_cnnp_model_tanh_t* const model_tanh = (ccv_cnnp_model_tanh_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_tanh_t));
1726 model_tanh->super.isa = &ccv_cnnp_tanh_isa;
1727 model_tanh->super.input_size = 1;
1728 model_tanh->super.outputs = &model_tanh->output;
1729 model_tanh->super.output_size = 1;
1730 ccv_cnnp_model_copy_name(&model_tanh->super, name);
1731 return (ccv_cnnp_model_t*)model_tanh;
1732}
1733
1734static ccv_cnnp_model_t* _ccv_cnnp_tanh_copy(const ccv_cnnp_model_t* const self, void* const context)
1735{
1736 return ccv_cnnp_tanh(self->name);
1737}
1738
1739// MARK - Exp Layer
1740
1741typedef struct {
1742 ccv_cnnp_model_t super;
1743 ccv_nnc_tensor_symbol_t output;
1744} ccv_cnnp_model_exp_t;
1745
1746static 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)
1747{
1748 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);
1749 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"
, 1749, __extension__ __PRETTY_FUNCTION__); }));
1750 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", 1750, __extension__ __PRETTY_FUNCTION__
); }));
1751 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1752 ccv_nnc_tensor_param_t output_params;
1753 const ccv_nnc_cmd_t exp = CMD_EWEXP_FORWARD()ccv_nnc_cmd(CCV_NNC_EWEXP_FORWARD, 0, ccv_nnc_cmd_auto, 0);
1754 ccv_nnc_hint_tensor_auto(exp, (ccv_nnc_tensor_param_t []){
1755 params,
1756 }, 1, ccv_nnc_no_hint, &output_params, 1);
1757 const ccv_nnc_tensor_symbol_t exp_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1758 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");
1759 outputs[0] = exp_output;
1760}
1761
1762static ccv_cnnp_model_t* _ccv_cnnp_exp_copy(const ccv_cnnp_model_t* const self, void* const context);
1763
1764static const ccv_cnnp_model_vtab_t ccv_cnnp_exp_isa = {
1765 .build = _ccv_cnnp_exp_build,
1766 .copy = _ccv_cnnp_exp_copy,
1767};
1768
1769ccv_cnnp_model_t* ccv_cnnp_exp(const char* const name)
1770{
1771 ccv_cnnp_model_exp_t* const model_exp = (ccv_cnnp_model_exp_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_exp_t));
1772 model_exp->super.isa = &ccv_cnnp_exp_isa;
1773 model_exp->super.input_size = 1;
1774 model_exp->super.outputs = &model_exp->output;
1775 model_exp->super.output_size = 1;
1776 ccv_cnnp_model_copy_name(&model_exp->super, name);
1777 return (ccv_cnnp_model_t*)model_exp;
1778}
1779
1780static ccv_cnnp_model_t* _ccv_cnnp_exp_copy(const ccv_cnnp_model_t* const self, void* const context)
1781{
1782 return ccv_cnnp_exp(self->name);
1783}
1784
1785// MARK - Softplus Layer
1786
1787typedef struct {
1788 ccv_cnnp_model_t super;
1789 ccv_nnc_tensor_symbol_t output;
1790} ccv_cnnp_model_softplus_t;
1791
1792static 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)
1793{
1794 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);
1795 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"
, 1795, __extension__ __PRETTY_FUNCTION__); }));
1796 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", 1796, __extension__ __PRETTY_FUNCTION__
); }));
1797 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1798 ccv_nnc_tensor_param_t output_params;
1799 const ccv_nnc_cmd_t softplus = CMD_EWSOFTPLUS_FORWARD()ccv_nnc_cmd(CCV_NNC_EWSOFTPLUS_FORWARD, 0, ccv_nnc_cmd_auto, 0
);
1800 ccv_nnc_hint_tensor_auto(softplus, (ccv_nnc_tensor_param_t []){
1801 params,
1802 }, 1, ccv_nnc_no_hint, &output_params, 1);
1803 const ccv_nnc_tensor_symbol_t softplus_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1804 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");
1805 outputs[0] = softplus_output;
1806}
1807
1808static ccv_cnnp_model_t* _ccv_cnnp_softplus_copy(const ccv_cnnp_model_t* const self, void* const context);
1809
1810static const ccv_cnnp_model_vtab_t ccv_cnnp_softplus_isa = {
1811 .build = _ccv_cnnp_softplus_build,
1812 .copy = _ccv_cnnp_softplus_copy,
1813};
1814
1815ccv_cnnp_model_t* ccv_cnnp_softplus(const char* const name)
1816{
1817 ccv_cnnp_model_softplus_t* const model_softplus = (ccv_cnnp_model_softplus_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_softplus_t));
1818 model_softplus->super.isa = &ccv_cnnp_softplus_isa;
1819 model_softplus->super.input_size = 1;
1820 model_softplus->super.outputs = &model_softplus->output;
1821 model_softplus->super.output_size = 1;
1822 ccv_cnnp_model_copy_name(&model_softplus->super, name);
1823 return (ccv_cnnp_model_t*)model_softplus;
1824}
1825
1826static ccv_cnnp_model_t* _ccv_cnnp_softplus_copy(const ccv_cnnp_model_t* const self, void* const context)
1827{
1828 return ccv_cnnp_softplus(self->name);
1829}
1830
1831// MARK - Swish Layer
1832
1833typedef struct {
1834 ccv_cnnp_model_t super;
1835 ccv_nnc_tensor_symbol_t output;
1836 float beta;
1837} ccv_cnnp_model_swish_t;
1838
1839static 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)
1840{
1841 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);
1842 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 1842, __extension__ __PRETTY_FUNCTION__); }));
1843 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 1843, __extension__ __PRETTY_FUNCTION__
); }));
1844 ccv_cnnp_model_swish_t* const self = (ccv_cnnp_model_swish_t*)super;
1845 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1846 ccv_nnc_tensor_param_t output_params;
1847 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);
1848 ccv_nnc_hint_tensor_auto(swish, (ccv_nnc_tensor_param_t []){
1849 params,
1850 }, 1, ccv_nnc_no_hint, &output_params, 1);
1851 const ccv_nnc_tensor_symbol_t swish_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1852 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");
1853 outputs[0] = swish_output;
1854}
1855
1856static ccv_cnnp_model_t* _ccv_cnnp_swish_copy(const ccv_cnnp_model_t* const self, void* const context);
1857
1858static const ccv_cnnp_model_vtab_t ccv_cnnp_swish_isa = {
1859 .build = _ccv_cnnp_swish_build,
1860 .copy = _ccv_cnnp_swish_copy,
1861};
1862
1863ccv_cnnp_model_t* ccv_cnnp_swish(const float beta, const char* const name)
1864{
1865 ccv_cnnp_model_swish_t* const model_swish = (ccv_cnnp_model_swish_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_swish_t));
1866 model_swish->super.isa = &ccv_cnnp_swish_isa;
1867 model_swish->super.input_size = 1;
1868 model_swish->super.outputs = &model_swish->output;
1869 model_swish->super.output_size = 1;
1870 model_swish->beta = beta;
1871 ccv_cnnp_model_copy_name(&model_swish->super, name);
1872 return (ccv_cnnp_model_t*)model_swish;
1873}
1874
1875static ccv_cnnp_model_t* _ccv_cnnp_swish_copy(const ccv_cnnp_model_t* const self, void* const context)
1876{
1877 const ccv_cnnp_model_swish_t* const swish = (const ccv_cnnp_model_swish_t*)self;
1878 return ccv_cnnp_swish(swish->beta, self->name);
1879}
1880
1881// MARK - GELU Layer
1882
1883typedef struct {
1884 ccv_cnnp_model_t super;
1885 ccv_nnc_tensor_symbol_t output;
1886 int tanh;
1887} ccv_cnnp_model_gelu_t;
1888
1889static 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)
1890{
1891 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);
1892 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"
, 1892, __extension__ __PRETTY_FUNCTION__); }));
1893 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", 1893, __extension__ __PRETTY_FUNCTION__
); }));
1894 ccv_cnnp_model_gelu_t* const self = (ccv_cnnp_model_gelu_t*)super;
1895 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1896 ccv_nnc_tensor_param_t output_params;
1897 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);
1898 ccv_nnc_hint_tensor_auto(gelu, (ccv_nnc_tensor_param_t []){
1899 params,
1900 }, 1, ccv_nnc_no_hint, &output_params, 1);
1901 const ccv_nnc_tensor_symbol_t gelu_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1902 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");
1903 outputs[0] = gelu_output;
1904}
1905
1906static ccv_cnnp_model_t* _ccv_cnnp_gelu_copy(const ccv_cnnp_model_t* const self, void* const context);
1907
1908static const ccv_cnnp_model_vtab_t ccv_cnnp_gelu_isa = {
1909 .build = _ccv_cnnp_gelu_build,
1910 .copy = _ccv_cnnp_gelu_copy,
1911};
1912
1913ccv_cnnp_model_t* ccv_cnnp_gelu(const int tanh, const char* const name)
1914{
1915 ccv_cnnp_model_gelu_t* const model_gelu = (ccv_cnnp_model_gelu_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_gelu_t));
1916 model_gelu->super.isa = &ccv_cnnp_gelu_isa;
1917 model_gelu->super.input_size = 1;
1918 model_gelu->super.outputs = &model_gelu->output;
1919 model_gelu->super.output_size = 1;
1920 model_gelu->tanh = tanh;
1921 ccv_cnnp_model_copy_name(&model_gelu->super, name);
1922 return (ccv_cnnp_model_t*)model_gelu;
1923}
1924
1925static ccv_cnnp_model_t* _ccv_cnnp_gelu_copy(const ccv_cnnp_model_t* const super, void* const context)
1926{
1927 ccv_cnnp_model_gelu_t* const self = (ccv_cnnp_model_gelu_t*)super;
1928 return ccv_cnnp_gelu(self->tanh, self->super.name);
1929}
1930
1931// MARK - Leaky ReLU Layer
1932
1933typedef struct {
1934 ccv_cnnp_model_t super;
1935 ccv_nnc_tensor_symbol_t output;
1936 float negative_slope;
1937} ccv_cnnp_model_leaky_relu_t;
1938
1939static 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)
1940{
1941 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);
1942 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"
, 1942, __extension__ __PRETTY_FUNCTION__); }));
1943 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", 1943, __extension__ __PRETTY_FUNCTION__
); }));
1944 ccv_cnnp_model_leaky_relu_t* const self = (ccv_cnnp_model_leaky_relu_t*)super;
1945 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1946 ccv_nnc_tensor_param_t output_params;
1947 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);
1948 ccv_nnc_hint_tensor_auto(leaky_relu, (ccv_nnc_tensor_param_t []){
1949 params,
1950 }, 1, ccv_nnc_no_hint, &output_params, 1);
1951 const ccv_nnc_tensor_symbol_t leaky_relu_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1952 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");
1953 outputs[0] = leaky_relu_output;
1954}
1955
1956static ccv_cnnp_model_t* _ccv_cnnp_leaky_relu_copy(const ccv_cnnp_model_t* const self, void* const context);
1957
1958static const ccv_cnnp_model_vtab_t ccv_cnnp_leaky_relu_isa = {
1959 .build = _ccv_cnnp_leaky_relu_build,
1960 .copy = _ccv_cnnp_leaky_relu_copy,
1961};
1962
1963ccv_cnnp_model_t* ccv_cnnp_leaky_relu(const float negative_slope, const char* const name)
1964{
1965 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));
1966 model_leaky_relu->super.isa = &ccv_cnnp_leaky_relu_isa;
1967 model_leaky_relu->super.input_size = 1;
1968 model_leaky_relu->super.outputs = &model_leaky_relu->output;
1969 model_leaky_relu->super.output_size = 1;
1970 model_leaky_relu->negative_slope = negative_slope;
1971 ccv_cnnp_model_copy_name(&model_leaky_relu->super, name);
1972 return (ccv_cnnp_model_t*)model_leaky_relu;
1973}
1974
1975static ccv_cnnp_model_t* _ccv_cnnp_leaky_relu_copy(const ccv_cnnp_model_t* const super, void* const context)
1976{
1977 ccv_cnnp_model_leaky_relu_t* const self = (ccv_cnnp_model_leaky_relu_t*)super;
1978 return ccv_cnnp_leaky_relu(self->negative_slope, self->super.name);
1979}
1980
1981// MARK - Softmax Layer
1982
1983typedef struct {
1984 ccv_cnnp_model_t super;
1985 ccv_nnc_tensor_symbol_t output;
1986} ccv_cnnp_model_softmax_t;
1987
1988static 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)
1989{
1990 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);
1991 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 1991, __extension__ __PRETTY_FUNCTION__); }));
1992 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 1992, __extension__ __PRETTY_FUNCTION__
); }));
1993 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1994 ccv_nnc_tensor_param_t output_params;
1995 const ccv_nnc_cmd_t softmax = CMD_SOFTMAX_FORWARD()ccv_nnc_cmd(CCV_NNC_SOFTMAX_FORWARD, 0, ccv_nnc_cmd_auto, 0);
1996 ccv_nnc_hint_tensor_auto(softmax, (ccv_nnc_tensor_param_t []){
1997 params,
1998 }, 1, ccv_nnc_no_hint, &output_params, 1);
1999 const ccv_nnc_tensor_symbol_t softmax_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2000 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");
2001 outputs[0] = softmax_output;
2002}
2003
2004static ccv_cnnp_model_t* _ccv_cnnp_softmax_copy(const ccv_cnnp_model_t* const self, void* const context);
2005
2006static const ccv_cnnp_model_vtab_t ccv_cnnp_softmax_isa = {
2007 .build = _ccv_cnnp_softmax_build,
2008 .copy = _ccv_cnnp_softmax_copy,
2009};
2010
2011ccv_cnnp_model_t* ccv_cnnp_softmax(const char* const name)
2012{
2013 ccv_cnnp_model_softmax_t* const model_softmax = (ccv_cnnp_model_softmax_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_softmax_t));
2014 model_softmax->super.isa = &ccv_cnnp_softmax_isa;
2015 model_softmax->super.input_size = 1;
2016 model_softmax->super.outputs = &model_softmax->output;
2017 model_softmax->super.output_size = 1;
2018 ccv_cnnp_model_copy_name(&model_softmax->super, name);
2019 return (ccv_cnnp_model_t*)model_softmax;
2020}
2021
2022static ccv_cnnp_model_t* _ccv_cnnp_softmax_copy(const ccv_cnnp_model_t* const self, void* const context)
2023{
2024 return ccv_cnnp_softmax(self->name);
2025}
2026
2027// MARK - Add Layer
2028
2029typedef struct {
2030 ccv_cnnp_model_t super;
2031 float p;
2032 float q;
2033 ccv_nnc_tensor_symbol_t output;
2034} ccv_cnnp_model_add_t;
2035
2036static 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)
2037{
2038 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);
2039 const ccv_cnnp_model_add_t* const self = (const ccv_cnnp_model_add_t*)super;
2040 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"
, 2040, __extension__ __PRETTY_FUNCTION__); }));
2041 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", 2041, __extension__ __PRETTY_FUNCTION__
); }));
2042 ccv_nnc_tensor_param_t input_params[2];
2043 int i;
2044 for (i = 0; i < 2; i++)
2045 input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
2046 ccv_nnc_tensor_param_t output_params;
2047 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);
2048 ccv_nnc_hint_tensor_auto(add, input_params, 2, ccv_nnc_no_hint, &output_params, 1);
2049 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2050 ccv_nnc_graph_exec_symbol_new(graph, add, inputs, input_size, outputs, output_size, "add");
2051}
2052
2053static ccv_cnnp_model_t* _ccv_cnnp_add_copy(const ccv_cnnp_model_t* const self, void* const context);
2054
2055static const ccv_cnnp_model_vtab_t ccv_cnnp_add_isa = {
2056 .build = _ccv_cnnp_add_build,
2057 .copy = _ccv_cnnp_add_copy,
2058};
2059
2060ccv_cnnp_model_t* ccv_cnnp_add(const float p, const float q, const char* const name)
2061{
2062 ccv_cnnp_model_add_t* const model_add = (ccv_cnnp_model_add_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_add_t));
2063 model_add->super.isa = &ccv_cnnp_add_isa;
2064 model_add->super.input_size = 2;
2065 model_add->super.outputs = &model_add->output;
2066 model_add->super.output_size = 1;
2067 model_add->p = p;
2068 model_add->q = q;
2069 ccv_cnnp_model_copy_name(&model_add->super, name);
2070 return (ccv_cnnp_model_t*)model_add;
2071}
2072
2073static ccv_cnnp_model_t* _ccv_cnnp_add_copy(const ccv_cnnp_model_t* const super, void* const context)
2074{
2075 const ccv_cnnp_model_add_t* const self = (const ccv_cnnp_model_add_t*)super;
2076 return ccv_cnnp_add(self->p, self->q, self->super.name);
2077}
2078
2079// MARK - Mul Layer
2080
2081typedef struct {
2082 ccv_cnnp_model_t super;
2083 ccv_nnc_tensor_symbol_t output;
2084 float p;
2085} ccv_cnnp_model_mul_t;
2086
2087static 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)
2088{
2089 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);
2090 const ccv_cnnp_model_mul_t* const self = (const ccv_cnnp_model_mul_t*)super;
2091 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"
, 2091, __extension__ __PRETTY_FUNCTION__); }));
2092 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", 2092, __extension__ __PRETTY_FUNCTION__
); }));
2093 ccv_nnc_tensor_param_t input_params[2];
2094 int i;
2095 for (i = 0; i < 2; i++)
2096 input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
2097 ccv_nnc_tensor_param_t output_params;
2098 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);
2099 ccv_nnc_hint_tensor_auto(mul, input_params, 2, ccv_nnc_no_hint, &output_params, 1);
2100 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2101 ccv_nnc_graph_exec_symbol_new(graph, mul, inputs, input_size, outputs, output_size, "mul");
2102}
2103
2104static ccv_cnnp_model_t* _ccv_cnnp_mul_copy(const ccv_cnnp_model_t* const self, void* const context);
2105
2106static const ccv_cnnp_model_vtab_t ccv_cnnp_mul_isa = {
2107 .build = _ccv_cnnp_mul_build,
2108 .copy = _ccv_cnnp_mul_copy,
2109};
2110
2111ccv_cnnp_model_t* ccv_cnnp_mul(const float p, const char* const name)
2112{
2113 ccv_cnnp_model_mul_t* const model_mul = (ccv_cnnp_model_mul_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_mul_t));
2114 model_mul->super.isa = &ccv_cnnp_mul_isa;
2115 model_mul->super.input_size = 2;
2116 model_mul->super.outputs = &model_mul->output;
2117 model_mul->super.output_size = 1;
2118 model_mul->p = p;
2119 ccv_cnnp_model_copy_name(&model_mul->super, name);
2120 return (ccv_cnnp_model_t*)model_mul;
2121}
2122
2123static ccv_cnnp_model_t* _ccv_cnnp_mul_copy(const ccv_cnnp_model_t* const super, void* const context)
2124{
2125 const ccv_cnnp_model_mul_t* const self = (const ccv_cnnp_model_mul_t*)super;
2126 return ccv_cnnp_mul(self->p, self->super.name);
2127}
2128
2129// MARK - Scalar Mul Layer
2130
2131typedef struct {
2132 ccv_cnnp_model_t super;
2133 ccv_nnc_tensor_symbol_t output;
2134 float a;
2135} ccv_cnnp_model_scalar_mul_t;
2136
2137static 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)
2138{
2139 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);
2140 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"
, 2140, __extension__ __PRETTY_FUNCTION__); }));
2141 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", 2141, __extension__ __PRETTY_FUNCTION__
); }));
2142 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2143 ccv_nnc_tensor_param_t output_params;
2144 ccv_cnnp_model_scalar_mul_t* const self = (ccv_cnnp_model_scalar_mul_t*)super;
2145 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);
2146 ccv_nnc_hint_tensor_auto(scalar_mul, (ccv_nnc_tensor_param_t []){
2147 params,
2148 }, 1, ccv_nnc_no_hint, &output_params, 1);
2149 const ccv_nnc_tensor_symbol_t scalar_mul_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2150 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");
2151 outputs[0] = scalar_mul_output;
2152}
2153
2154static ccv_cnnp_model_t* _ccv_cnnp_scalar_mul_copy(const ccv_cnnp_model_t* const super, void* const context);
2155
2156static const ccv_cnnp_model_vtab_t ccv_cnnp_scalar_mul_isa = {
2157 .build = _ccv_cnnp_scalar_mul_build,
2158 .copy = _ccv_cnnp_scalar_mul_copy,
2159};
2160
2161ccv_cnnp_model_t* ccv_cnnp_scalar_mul(const float a, const char* const name)
2162{
2163 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));
2164 model_scalar_mul->super.isa = &ccv_cnnp_scalar_mul_isa;
2165 model_scalar_mul->super.input_size = 1;
2166 model_scalar_mul->super.outputs = &model_scalar_mul->output;
2167 model_scalar_mul->super.output_size = 1;
2168 model_scalar_mul->a = a;
2169 ccv_cnnp_model_copy_name(&model_scalar_mul->super, name);
2170 return (ccv_cnnp_model_t*)model_scalar_mul;
2171}
2172
2173static ccv_cnnp_model_t* _ccv_cnnp_scalar_mul_copy(const ccv_cnnp_model_t* const super, void* const context)
2174{
2175 const ccv_cnnp_model_scalar_mul_t* const self = (const ccv_cnnp_model_scalar_mul_t*)super;
2176 return ccv_cnnp_scalar_mul(self->a, self->super.name);
2177}
2178
2179// MARK - Div Layer
2180
2181typedef struct {
2182 ccv_cnnp_model_t super;
2183 ccv_nnc_tensor_symbol_t output;
2184 int reciprocal;
2185} ccv_cnnp_model_div_t;
2186
2187static 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)
2188{
2189 const ccv_cnnp_model_div_t* const self = (const ccv_cnnp_model_div_t*)super;
2190 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);
2191 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 2191, __extension__ __PRETTY_FUNCTION__
); }));
2192 ccv_nnc_tensor_param_t input_params[2];
2193 int i;
2194 ccv_nnc_tensor_param_t output_params;
2195 const ccv_nnc_cmd_t div = CMD_EWDIV_FORWARD()ccv_nnc_cmd(CCV_NNC_EWDIV_FORWARD, 0, ccv_nnc_cmd_auto, 0);
2196 if (self->reciprocal)
2197 {
2198 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"
, 2198, __extension__ __PRETTY_FUNCTION__); }));
2199 input_params[0] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2200 input_params[1] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2201 ccv_nnc_hint_tensor_auto(div, input_params, 2, ccv_nnc_no_hint, &output_params, 1);
2202 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2203 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");
2204 } else {
2205 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"
, 2205, __extension__ __PRETTY_FUNCTION__); }));
2206 for (i = 0; i < 2; i++)
2207 input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
2208 ccv_nnc_hint_tensor_auto(div, input_params, input_size, ccv_nnc_no_hint, &output_params, 1);
2209 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2210 ccv_nnc_graph_exec_symbol_new(graph, div, inputs, input_size, outputs, output_size, "div");
2211 }
2212}
2213
2214static ccv_cnnp_model_t* _ccv_cnnp_div_copy(const ccv_cnnp_model_t* const self, void* const context);
2215
2216static const ccv_cnnp_model_vtab_t ccv_cnnp_div_isa = {
2217 .build = _ccv_cnnp_div_build,
2218 .copy = _ccv_cnnp_div_copy,
2219};
2220
2221ccv_cnnp_model_t* ccv_cnnp_div(const int reciprocal, const char* const name)
2222{
2223 ccv_cnnp_model_div_t* const model_div = (ccv_cnnp_model_div_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_div_t));
2224 model_div->super.isa = &ccv_cnnp_div_isa;
2225 model_div->super.input_size = reciprocal ? 1 : 2;
2226 model_div->super.outputs = &model_div->output;
2227 model_div->super.output_size = 1;
2228 model_div->reciprocal = reciprocal;
2229 ccv_cnnp_model_copy_name(&model_div->super, name);
2230 return (ccv_cnnp_model_t*)model_div;
2231}
2232
2233static ccv_cnnp_model_t* _ccv_cnnp_div_copy(const ccv_cnnp_model_t* const super, void* const context)
2234{
2235 const ccv_cnnp_model_div_t* const self = (const ccv_cnnp_model_div_t*)super;
2236 return ccv_cnnp_div(self->reciprocal, self->super.name);
2237}
2238
2239// MARK - Sqrt Layer
2240
2241typedef struct {
2242 ccv_cnnp_model_t super;
2243 ccv_nnc_tensor_symbol_t output;
2244} ccv_cnnp_model_sqrt_t;
2245
2246static 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)
2247{
2248 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);
2249 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", 2249, __extension__ __PRETTY_FUNCTION__
); }));
2250 ccv_nnc_tensor_param_t input_params[1];
2251 ccv_nnc_tensor_param_t output_params;
2252 const ccv_nnc_cmd_t sqrt = CMD_EWSQRT_FORWARD()ccv_nnc_cmd(CCV_NNC_EWSQRT_FORWARD, 0, ccv_nnc_cmd_auto, 0);
2253 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"
, 2253, __extension__ __PRETTY_FUNCTION__); }));
2254 input_params[0] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2255 ccv_nnc_hint_tensor_auto(sqrt, input_params, 1, ccv_nnc_no_hint, &output_params, 1);
2256 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2257 ccv_nnc_graph_exec_symbol_new(graph, sqrt, inputs, 1, outputs, output_size, "sqrt");
2258}
2259
2260static ccv_cnnp_model_t* _ccv_cnnp_sqrt_copy(const ccv_cnnp_model_t* const self, void* const context);
2261
2262static const ccv_cnnp_model_vtab_t ccv_cnnp_sqrt_isa = {
2263 .build = _ccv_cnnp_sqrt_build,
2264 .copy = _ccv_cnnp_sqrt_copy,
2265};
2266
2267ccv_cnnp_model_t* ccv_cnnp_sqrt(const char* const name)
2268{
2269 ccv_cnnp_model_sqrt_t* const model_sqrt = (ccv_cnnp_model_sqrt_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_sqrt_t));
2270 model_sqrt->super.isa = &ccv_cnnp_sqrt_isa;
2271 model_sqrt->super.input_size = 1;
2272 model_sqrt->super.outputs = &model_sqrt->output;
2273 model_sqrt->super.output_size = 1;
2274 ccv_cnnp_model_copy_name(&model_sqrt->super, name);
2275 return (ccv_cnnp_model_t*)model_sqrt;
2276}
2277
2278static ccv_cnnp_model_t* _ccv_cnnp_sqrt_copy(const ccv_cnnp_model_t* const super, void* const context)
2279{
2280 const ccv_cnnp_model_sqrt_t* const self = (const ccv_cnnp_model_sqrt_t*)super;
2281 return ccv_cnnp_sqrt(self->super.name);
2282}
2283
2284// MARK - Log Layer
2285
2286typedef struct {
2287 ccv_cnnp_model_t super;
2288 ccv_nnc_tensor_symbol_t output;
2289} ccv_cnnp_model_log_t;
2290
2291static 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)
2292{
2293 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);
2294 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", 2294, __extension__ __PRETTY_FUNCTION__
); }));
2295 ccv_nnc_tensor_param_t input_params[1];
2296 ccv_nnc_tensor_param_t output_params;
2297 const ccv_nnc_cmd_t log = CMD_EWLOG_FORWARD()ccv_nnc_cmd(CCV_NNC_EWLOG_FORWARD, 0, ccv_nnc_cmd_auto, 0);
2298 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"
, 2298, __extension__ __PRETTY_FUNCTION__); }));
2299 input_params[0] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2300 ccv_nnc_hint_tensor_auto(log, input_params, 1, ccv_nnc_no_hint, &output_params, 1);
2301 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2302 ccv_nnc_graph_exec_symbol_new(graph, log, inputs, 1, outputs, output_size, "log");
2303}
2304
2305static ccv_cnnp_model_t* _ccv_cnnp_log_copy(const ccv_cnnp_model_t* const self, void* const context);
2306
2307static const ccv_cnnp_model_vtab_t ccv_cnnp_log_isa = {
2308 .build = _ccv_cnnp_log_build,
2309 .copy = _ccv_cnnp_log_copy,
2310};
2311
2312ccv_cnnp_model_t* ccv_cnnp_log(const char* const name)
2313{
2314 ccv_cnnp_model_log_t* const model_log = (ccv_cnnp_model_log_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_log_t));
2315 model_log->super.isa = &ccv_cnnp_log_isa;
2316 model_log->super.input_size = 1;
2317 model_log->super.outputs = &model_log->output;
2318 model_log->super.output_size = 1;
2319 ccv_cnnp_model_copy_name(&model_log->super, name);
2320 return (ccv_cnnp_model_t*)model_log;
2321}
2322
2323static ccv_cnnp_model_t* _ccv_cnnp_log_copy(const ccv_cnnp_model_t* const super, void* const context)
2324{
2325 return ccv_cnnp_log(super->name);
2326}
2327
2328// MARK - Pow Layer
2329
2330typedef struct {
2331 ccv_cnnp_model_t super;
2332 ccv_nnc_tensor_symbol_t output;
2333 ccv_nnc_cmd_param_t params;
2334} ccv_cnnp_model_pow_t;
2335
2336static 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)
2337{
2338 ccv_cnnp_model_pow_t* const self = (ccv_cnnp_model_pow_t*)super;
2339 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);
2340 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"
, 2340, __extension__ __PRETTY_FUNCTION__); }));
2341 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", 2341, __extension__ __PRETTY_FUNCTION__
); }));
2342 ccv_nnc_tensor_param_t input_params[1];
2343 input_params[0] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2344 ccv_nnc_tensor_param_t output_params;
2345 const ccv_nnc_cmd_t pow = ccv_nnc_cmd(CCV_NNC_EWPOW_FORWARD, 0, self->params, 0);
2346 ccv_nnc_hint_tensor_auto(pow, input_params, 1, ccv_nnc_no_hint, &output_params, 1);
2347 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2348 ccv_nnc_graph_exec_symbol_new(graph, pow, inputs, input_size, outputs, output_size, "pow");
2349}
2350
2351static ccv_cnnp_model_t* _ccv_cnnp_pow_copy(const ccv_cnnp_model_t* const self, void* const context);
2352
2353static const ccv_cnnp_model_vtab_t ccv_cnnp_pow_isa = {
2354 .build = _ccv_cnnp_pow_build,
2355 .copy = _ccv_cnnp_pow_copy,
2356};
2357
2358ccv_cnnp_model_t* ccv_cnnp_pow(const float exponent, const char* const name)
2359{
2360 ccv_cnnp_model_pow_t* const model_pow = (ccv_cnnp_model_pow_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_pow_t));
2361 model_pow->super.isa = &ccv_cnnp_pow_isa;
2362 model_pow->super.input_size = 1;
2363 model_pow->super.outputs = &model_pow->output;
2364 model_pow->super.output_size = 1;
2365 model_pow->params = (ccv_nnc_cmd_param_t){
2366 .size = {
2367 .dim = { 1, 1, 1 }
2368 },
2369 .pow = {
2370 .exponent = exponent,
2371 },
2372 };
2373 ccv_cnnp_model_copy_name(&model_pow->super, name);
2374 return (ccv_cnnp_model_t*)model_pow;
2375}
2376
2377static ccv_cnnp_model_t* _ccv_cnnp_pow_copy(const ccv_cnnp_model_t* const super, void* const context)
2378{
2379 const ccv_cnnp_model_pow_t* const self = (const ccv_cnnp_model_pow_t*)super;
2380 return ccv_cnnp_pow(self->params.pow.exponent, super->name);
2381}
2382
2383// MARK - Sin Layer
2384
2385typedef struct {
2386 ccv_cnnp_model_t super;
2387 ccv_nnc_tensor_symbol_t output;
2388} ccv_cnnp_model_sin_t;
2389
2390static 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)
2391{
2392 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);
2393 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", 2393, __extension__ __PRETTY_FUNCTION__
); }));
2394 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 2394, __extension__ __PRETTY_FUNCTION__); }));
2395 ccv_nnc_tensor_param_t input_params[1];
2396 ccv_nnc_tensor_param_t output_params;
2397 const ccv_nnc_cmd_t sin = CMD_EWSIN_FORWARD()ccv_nnc_cmd(CCV_NNC_EWSIN_FORWARD, 0, ccv_nnc_cmd_auto, 0);
2398 input_params[0] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2399 ccv_nnc_hint_tensor_auto(sin, input_params, 1, ccv_nnc_no_hint, &output_params, 1);
2400 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2401 ccv_nnc_graph_exec_symbol_new(graph, sin, inputs, 1, outputs, output_size, "sin");
2402}
2403
2404static ccv_cnnp_model_t* _ccv_cnnp_sin_copy(const ccv_cnnp_model_t* const self, void* const context);
2405
2406static const ccv_cnnp_model_vtab_t ccv_cnnp_sin_isa = {
2407 .build = _ccv_cnnp_sin_build,
2408 .copy = _ccv_cnnp_sin_copy,
2409};
2410
2411ccv_cnnp_model_t* ccv_cnnp_sin(const char* const name)
2412{
2413 ccv_cnnp_model_sin_t* const model_sin = (ccv_cnnp_model_sin_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_sin_t));
2414 model_sin->super.isa = &ccv_cnnp_sin_isa;
2415 model_sin->super.input_size = 1;
2416 model_sin->super.outputs = &model_sin->output;
2417 model_sin->super.output_size = 1;
2418 ccv_cnnp_model_copy_name(&model_sin->super, name);
2419 return (ccv_cnnp_model_t*)model_sin;
2420}
2421
2422static ccv_cnnp_model_t* _ccv_cnnp_sin_copy(const ccv_cnnp_model_t* const super, void* const context)
2423{
2424 return ccv_cnnp_sin(super->name);
2425}
2426
2427// MARK - Cos Layer
2428
2429typedef struct {
2430 ccv_cnnp_model_t super;
2431 ccv_nnc_tensor_symbol_t output;
2432} ccv_cnnp_model_cos_t;
2433
2434static 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)
2435{
2436 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);
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 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"
, 2438, __extension__ __PRETTY_FUNCTION__); }));
2439 ccv_nnc_tensor_param_t input_params[1];
2440 ccv_nnc_tensor_param_t output_params;
2441 const ccv_nnc_cmd_t cos = CMD_EWCOS_FORWARD()ccv_nnc_cmd(CCV_NNC_EWCOS_FORWARD, 0, ccv_nnc_cmd_auto, 0);
2442 input_params[0] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2443 ccv_nnc_hint_tensor_auto(cos, 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, cos, inputs, 1, outputs, output_size, "cos");
2446}
2447
2448static ccv_cnnp_model_t* _ccv_cnnp_cos_copy(const ccv_cnnp_model_t* const self, void* const context);
2449
2450static const ccv_cnnp_model_vtab_t ccv_cnnp_cos_isa = {
2451 .build = _ccv_cnnp_cos_build,
2452 .copy = _ccv_cnnp_cos_copy,
2453};
2454
2455ccv_cnnp_model_t* ccv_cnnp_cos(const char* const name)
2456{
2457 ccv_cnnp_model_cos_t* const model_cos = (ccv_cnnp_model_cos_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_cos_t));
2458 model_cos->super.isa = &ccv_cnnp_cos_isa;
2459 model_cos->super.input_size = 1;
2460 model_cos->super.outputs = &model_cos->output;
2461 model_cos->super.output_size = 1;
2462 ccv_cnnp_model_copy_name(&model_cos->super, name);
2463 return (ccv_cnnp_model_t*)model_cos;
2464}
2465
2466static ccv_cnnp_model_t* _ccv_cnnp_cos_copy(const ccv_cnnp_model_t* const super, void* const context)
2467{
2468 return ccv_cnnp_cos(super->name);
2469}
2470
2471// MARK - Rotate Half Layer
2472
2473typedef struct {
2474 ccv_cnnp_model_t super;
2475 ccv_nnc_tensor_symbol_t output;
2476} ccv_cnnp_model_rotate_half_t;
2477
2478static 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)
2479{
2480 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);
2481 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"
, 2481, __extension__ __PRETTY_FUNCTION__); }));
2482 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", 2482, __extension__ __PRETTY_FUNCTION__
); }));
2483 ccv_nnc_tensor_param_t input_params[1];
2484 ccv_nnc_tensor_param_t output_params;
2485 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);
2486 input_params[0] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2487 ccv_nnc_hint_tensor_auto(rotate_half, input_params, 1, ccv_nnc_no_hint, &output_params, 1);
2488 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2489 ccv_nnc_graph_exec_symbol_new(graph, rotate_half, inputs, 1, outputs, output_size, "rotate_half");
2490}
2491
2492static ccv_cnnp_model_t* _ccv_cnnp_rotate_half_copy(const ccv_cnnp_model_t* const self, void* const context);
2493
2494static const ccv_cnnp_model_vtab_t ccv_cnnp_rotate_half_isa = {
2495 .build = _ccv_cnnp_rotate_half_build,
2496 .copy = _ccv_cnnp_rotate_half_copy,
2497};
2498
2499ccv_cnnp_model_t* ccv_cnnp_rotate_half(const char* const name)
2500{
2501 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));
2502 model_rotate_half->super.isa = &ccv_cnnp_rotate_half_isa;
2503 model_rotate_half->super.input_size = 1;
2504 model_rotate_half->super.outputs = &model_rotate_half->output;
2505 model_rotate_half->super.output_size = 1;
2506 ccv_cnnp_model_copy_name(&model_rotate_half->super, name);
2507 return (ccv_cnnp_model_t*)model_rotate_half;
2508}
2509
2510static ccv_cnnp_model_t* _ccv_cnnp_rotate_half_copy(const ccv_cnnp_model_t* const super, void* const context)
2511{
2512 return ccv_cnnp_rotate_half(super->name);
2513}
2514
2515// MARK - Cmul Layer
2516
2517typedef struct {
2518 ccv_cnnp_model_t super;
2519 ccv_nnc_tensor_symbol_t output;
2520} ccv_cnnp_model_cmul_t;
2521
2522static 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)
2523{
2524 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);
2525 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"
, 2525, __extension__ __PRETTY_FUNCTION__); }));
2526 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", 2526, __extension__ __PRETTY_FUNCTION__
); }));
2527 ccv_nnc_tensor_param_t input_params[2];
2528 int i;
2529 for (i = 0; i < 2; i++)
2530 input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
2531 ccv_nnc_tensor_param_t output_params;
2532 const 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);
2533 ccv_nnc_hint_tensor_auto(mul, input_params, 2, ccv_nnc_no_hint, &output_params, 1);
2534 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2535 ccv_nnc_graph_exec_symbol_new(graph, mul, inputs, input_size, outputs, output_size, "cmul");
2536}
2537
2538static ccv_cnnp_model_t* _ccv_cnnp_cmul_copy(const ccv_cnnp_model_t* const self, void* const context);
2539
2540static const ccv_cnnp_model_vtab_t ccv_cnnp_cmul_isa = {
2541 .build = _ccv_cnnp_cmul_build,
2542 .copy = _ccv_cnnp_cmul_copy,
2543};
2544
2545ccv_cnnp_model_t* ccv_cnnp_cmul(const char* const name)
2546{
2547 ccv_cnnp_model_cmul_t* const model_cmul = (ccv_cnnp_model_cmul_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_cmul_t));
2548 model_cmul->super.isa = &ccv_cnnp_cmul_isa;
2549 model_cmul->super.input_size = 2;
2550 model_cmul->super.outputs = &model_cmul->output;
2551 model_cmul->super.output_size = 1;
2552 ccv_cnnp_model_copy_name(&model_cmul->super, name);
2553 return (ccv_cnnp_model_t*)model_cmul;
2554}
2555
2556static ccv_cnnp_model_t* _ccv_cnnp_cmul_copy(const ccv_cnnp_model_t* const super, void* const context)
2557{
2558 return ccv_cnnp_cmul(super->name);
2559}
2560
2561// MARK - Transpose Layer
2562
2563typedef struct {
2564 ccv_cnnp_model_t super;
2565 ccv_nnc_tensor_symbol_t output;
2566 int transpose[2];
2567} ccv_cnnp_model_transpose_t;
2568
2569static 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)
2570{
2571 ccv_cnnp_model_transpose_t* const self = (ccv_cnnp_model_transpose_t*)super;
2572 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);
2573 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"
, 2573, __extension__ __PRETTY_FUNCTION__); }));
2574 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", 2574, __extension__ __PRETTY_FUNCTION__
); }));
2575 if (self->transpose[0] == self->transpose[1])
2576 {
2577 outputs[0] = inputs[0];
2578 return;
2579 }
2580 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2581 ccv_nnc_tensor_param_t output_params;
2582 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);
2583 ccv_nnc_hint_tensor_auto(transpose, (ccv_nnc_tensor_param_t []){
2584 params,
2585 }, 1, ccv_nnc_no_hint, &output_params, 1);
2586 const ccv_nnc_tensor_symbol_t transpose_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2587 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");
2588 outputs[0] = transpose_output;
2589}
2590
2591static ccv_cnnp_model_t* _ccv_cnnp_transpose_copy(const ccv_cnnp_model_t* const super, void* const context);
2592
2593static const ccv_cnnp_model_vtab_t ccv_cnnp_transpose_isa = {
2594 .build = _ccv_cnnp_transpose_build,
2595 .copy = _ccv_cnnp_transpose_copy,
2596};
2597
2598ccv_cnnp_model_t* ccv_cnnp_transpose(const int axis_a, const int axis_b, const char* const name)
2599{
2600 ccv_cnnp_model_transpose_t* const model_transpose = (ccv_cnnp_model_transpose_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_transpose_t));
2601 model_transpose->super.isa = &ccv_cnnp_transpose_isa;
2602 model_transpose->super.input_size = 1;
2603 model_transpose->super.outputs = &model_transpose->output;
2604 model_transpose->super.output_size = 1;
2605 model_transpose->transpose[0] = axis_a;
2606 model_transpose->transpose[1] = axis_b;
2607 ccv_cnnp_model_copy_name(&model_transpose->super, name);
2608 return (ccv_cnnp_model_t*)model_transpose;
2609}
2610
2611static ccv_cnnp_model_t* _ccv_cnnp_transpose_copy(const ccv_cnnp_model_t* const super, void* const context)
2612{
2613 const ccv_cnnp_model_transpose_t* const self = (const ccv_cnnp_model_transpose_t*)super;
2614 return ccv_cnnp_transpose(self->transpose[0], self->transpose[1], self->super.name);
2615}
2616
2617// MARK - Layer Norm Layer
2618
2619typedef struct {
2620 ccv_cnnp_model_t super;
2621 ccv_nnc_tensor_symbol_t output;
2622 ccv_nnc_tensor_symbol_t bias;
2623 ccv_nnc_tensor_symbol_t scale;
2624 ccv_nnc_cmd_param_t params;
2625} ccv_cnnp_model_layer_norm_t;
2626
2627static 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)
2628{
2629 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);
2630 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"
, 2630, __extension__ __PRETTY_FUNCTION__); }));
2631 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", 2631, __extension__ __PRETTY_FUNCTION__
); }));
2632 ccv_cnnp_model_layer_norm_t* const self = (ccv_cnnp_model_layer_norm_t*)super;
2633 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2634 ccv_nnc_tensor_param_t bias_params = params;
2635 const int nd = ccv_nnc_tensor_nd(params.dim);
2636 int i;
2637 for (i = 0; i < nd; i++)
2638 bias_params.dim[i] = 1;
2639 for (i = 0; i < self->params.lnorm.count; i++)
2640 bias_params.dim[self->params.lnorm.axis[i]] = params.dim[self->params.lnorm.axis[i]];
2641 if (self->params.lnorm.elementwise_affine)
2642 {
2643 // Both scale and bias are shared between if this model is reused.
2644 if (!self->scale.graph)
2645 self->scale = ccv_nnc_tensor_symbol_new(graph, bias_params, "scale");
2646 if (!self->bias.graph)
2647 self->bias = ccv_nnc_tensor_symbol_new(graph, bias_params, "bias");
2648 }
2649 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
};
2650 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
};
2651 const ccv_nnc_cmd_t layer_norm = ccv_nnc_cmd(CCV_NNC_LAYER_NORM_FORWARD, 0, self->params, 0);
2652 ccv_nnc_tensor_param_t output_params[3];
2653 if (self->params.lnorm.elementwise_affine)
2654 ccv_nnc_hint_tensor_auto(layer_norm, (ccv_nnc_tensor_param_t []){
2655 params,
2656 bias_params,
2657 bias_params,
2658 }, 3, ccv_nnc_no_hint, output_params, 3);
2659 else
2660 ccv_nnc_hint_tensor_auto(layer_norm, (ccv_nnc_tensor_param_t []){
2661 params,
2662 }, 1, ccv_nnc_no_hint, output_params, 3);
2663 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0);
2664 const ccv_nnc_tensor_symbol_t saved_mean = ccv_nnc_tensor_symbol_new(graph, output_params[1], "saved_mean");
2665 const ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(graph, output_params[2], "saved_inv_std");
2666 if (self->params.lnorm.elementwise_affine)
2667 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");
2668 else
2669 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");
2670 outputs[0] = output;
2671}
2672
2673static 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)
2674{
2675 ccv_cnnp_model_layer_norm_t* const self = (ccv_cnnp_model_layer_norm_t*)super;
2676 if (self->scale.graph)
2677 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);
2678 if (self->bias.graph)
2679 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);
2680}
2681
2682static 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)
2683{
2684 ccv_cnnp_model_layer_norm_t* const self = (ccv_cnnp_model_layer_norm_t*)super;
2685 if (self->scale.graph)
2686 add_to_array(parameters, self->scale, is_trainable);
2687 if (self->bias.graph)
2688 add_to_array(parameters, self->bias, is_trainable);
2689}
2690
2691static ccv_cnnp_model_t* _ccv_cnnp_layer_norm_copy(const ccv_cnnp_model_t* const super, void* const context);
2692
2693static const ccv_cnnp_model_vtab_t ccv_cnnp_layer_norm_isa = {
2694 .build = _ccv_cnnp_layer_norm_build,
2695 .init_states = _ccv_cnnp_layer_norm_init_states,
2696 .add_to_parameter = _ccv_cnnp_layer_norm_add_to_parameter,
2697 .copy = _ccv_cnnp_layer_norm_copy,
2698};
2699
2700ccv_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 int is_trainable, const char* const name)
2701{
2702 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));
2703 model_layer_norm->super.isa = &ccv_cnnp_layer_norm_isa;
2704 model_layer_norm->super.input_size = 1;
2705 model_layer_norm->super.outputs = &model_layer_norm->output;
2706 model_layer_norm->super.output_size = 1;
2707 model_layer_norm->super.is_trainable = is_trainable;
2708 ccv_cnnp_model_copy_name(&model_layer_norm->super, name);
2709 model_layer_norm->scale.d = CCV_NNC_NO_TENSOR_SYMBOL;
2710 model_layer_norm->scale.graph = 0;
2711 model_layer_norm->bias.d = CCV_NNC_NO_TENSOR_SYMBOL;
2712 model_layer_norm->bias.graph = 0;
2713 model_layer_norm->params.lnorm.epsilon = epsilon;
2714 model_layer_norm->params.lnorm.count = axis_count;
2715 model_layer_norm->params.lnorm.elementwise_affine = elementwise_affine;
2716 memcpy(model_layer_norm->params.lnorm.axis, axis, sizeof(int) * axis_count);
2717 return (ccv_cnnp_model_t*)model_layer_norm;
2718}
2719
2720static ccv_cnnp_model_t* _ccv_cnnp_layer_norm_copy(const ccv_cnnp_model_t* const super, void* const context)
2721{
2722 const ccv_cnnp_model_layer_norm_t* const self = (const ccv_cnnp_model_layer_norm_t*)super;
2723 return ccv_cnnp_layer_norm(self->params.lnorm.epsilon, self->params.lnorm.axis, self->params.lnorm.count, self->params.lnorm.elementwise_affine, self->super.is_trainable, self->super.name);
2724}
2725
2726// MARK - Group Norm Layer
2727
2728typedef struct {
2729 ccv_cnnp_model_t super;
2730 ccv_nnc_tensor_symbol_t output;
2731 ccv_nnc_tensor_symbol_t bias;
2732 ccv_nnc_tensor_symbol_t scale;
2733 ccv_nnc_cmd_param_t params;
2734} ccv_cnnp_model_group_norm_t;
2735
2736static 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)
2737{
2738 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);
2739 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"
, 2739, __extension__ __PRETTY_FUNCTION__); }));
2740 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", 2740, __extension__ __PRETTY_FUNCTION__
); }));
2741 ccv_cnnp_model_group_norm_t* const self = (ccv_cnnp_model_group_norm_t*)super;
2742 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2743 ccv_nnc_tensor_param_t bias_params = params;
2744 const int nd = ccv_nnc_tensor_nd(params.dim);
2745 int i;
2746 for (i = 0; i < nd; i++)
2747 bias_params.dim[i] = 1;
2748 bias_params.dim[self->params.gnorm.group_axis] = params.dim[self->params.gnorm.group_axis];
2749 if (self->params.gnorm.elementwise_affine)
2750 {
2751 // Both scale and bias are shared between if this model is reused.
2752 if (!self->scale.graph)
2753 self->scale = ccv_nnc_tensor_symbol_new(graph, bias_params, "scale");
2754 if (!self->bias.graph)
2755 self->bias = ccv_nnc_tensor_symbol_new(graph, bias_params, "bias");
2756 }
2757 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
};
2758 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
};
2759 const ccv_nnc_cmd_t group_norm = ccv_nnc_cmd(CCV_NNC_GROUP_NORM_FORWARD, 0, self->params, 0);
2760 ccv_nnc_tensor_param_t output_params[3];
2761 if (self->params.gnorm.elementwise_affine)
2762 ccv_nnc_hint_tensor_auto(group_norm, (ccv_nnc_tensor_param_t []){
2763 params,
2764 bias_params,
2765 bias_params,
2766 }, 3, ccv_nnc_no_hint, output_params, 3);
2767 else
2768 ccv_nnc_hint_tensor_auto(group_norm, (ccv_nnc_tensor_param_t []){
2769 params,
2770 }, 1, ccv_nnc_no_hint, output_params, 3);
2771 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0);
2772 const ccv_nnc_tensor_symbol_t saved_mean = ccv_nnc_tensor_symbol_new(graph, output_params[1], "saved_mean");
2773 const ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(graph, output_params[2], "saved_inv_std");
2774 if (self->params.gnorm.elementwise_affine)
2775 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");
2776 else
2777 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");
2778 outputs[0] = output;
2779}
2780
2781static 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)
2782{
2783 ccv_cnnp_model_group_norm_t* const self = (ccv_cnnp_model_group_norm_t*)super;
2784 if (self->scale.graph)
2785 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);
2786 if (self->bias.graph)
2787 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);
2788}
2789
2790static 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)
2791{
2792 ccv_cnnp_model_group_norm_t* const self = (ccv_cnnp_model_group_norm_t*)super;
2793 if (self->scale.graph)
2794 add_to_array(parameters, self->scale, is_trainable);
2795 if (self->bias.graph)
2796 add_to_array(parameters, self->bias, is_trainable);
2797}
2798
2799static ccv_cnnp_model_t* _ccv_cnnp_group_norm_copy(const ccv_cnnp_model_t* const super, void* const context);
2800
2801static const ccv_cnnp_model_vtab_t ccv_cnnp_group_norm_isa = {
2802 .build = _ccv_cnnp_group_norm_build,
2803 .init_states = _ccv_cnnp_group_norm_init_states,
2804 .add_to_parameter = _ccv_cnnp_group_norm_add_to_parameter,
2805 .copy = _ccv_cnnp_group_norm_copy,
2806};
2807
2808ccv_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)
2809{
2810 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));
2811 model_group_norm->super.isa = &ccv_cnnp_group_norm_isa;
2812 model_group_norm->super.input_size = 1;
2813 model_group_norm->super.outputs = &model_group_norm->output;
2814 model_group_norm->super.output_size = 1;
2815 model_group_norm->super.is_trainable = is_trainable;
2816 ccv_cnnp_model_copy_name(&model_group_norm->super, name);
2817 model_group_norm->scale.d = CCV_NNC_NO_TENSOR_SYMBOL;
2818 model_group_norm->scale.graph = 0;
2819 model_group_norm->bias.d = CCV_NNC_NO_TENSOR_SYMBOL;
2820 model_group_norm->bias.graph = 0;
2821 model_group_norm->params.gnorm.group_axis = group_axis;
2822 model_group_norm->params.gnorm.groups = groups;
2823 model_group_norm->params.gnorm.epsilon = epsilon;
2824 model_group_norm->params.gnorm.reduce_count = axis_count;
2825 model_group_norm->params.gnorm.elementwise_affine = elementwise_affine;
2826 memcpy(model_group_norm->params.gnorm.reduce_axis, reduce_axis, sizeof(int) * axis_count);
2827 return (ccv_cnnp_model_t*)model_group_norm;
2828}
2829
2830static ccv_cnnp_model_t* _ccv_cnnp_group_norm_copy(const ccv_cnnp_model_t* const super, void* const context)
2831{
2832 const ccv_cnnp_model_group_norm_t* const self = (const ccv_cnnp_model_group_norm_t*)super;
2833 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);
2834}
2835
2836// MARK - RMSNorm Layer
2837
2838typedef struct {
2839 ccv_cnnp_model_t super;
2840 ccv_nnc_tensor_symbol_t output;
2841 ccv_nnc_tensor_symbol_t scale;
2842 ccv_nnc_cmd_param_t params;
2843} ccv_cnnp_model_rmsnorm_t;
2844
2845static 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)
2846{
2847 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);
2848 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"
, 2848, __extension__ __PRETTY_FUNCTION__); }));
2849 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", 2849, __extension__ __PRETTY_FUNCTION__
); }));
2850 ccv_cnnp_model_rmsnorm_t* const self = (ccv_cnnp_model_rmsnorm_t*)super;
2851 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2852 ccv_nnc_tensor_param_t scale_params = params;
2853 const int nd = ccv_nnc_tensor_nd(params.dim);
2854 int i;
2855 for (i = 0; i < nd; i++)
2856 scale_params.dim[i] = 1;
2857 for (i = 0; i < self->params.rmsnorm.count; i++)
2858 scale_params.dim[self->params.rmsnorm.axis[i]] = params.dim[self->params.rmsnorm.axis[i]];
2859 // Both scale and bias are shared between if this model is reused.
2860 if (self->params.rmsnorm.elementwise_affine)
2861 {
2862 if (!self->scale.graph)
2863 self->scale = ccv_nnc_tensor_symbol_new(graph, scale_params, "scale");
2864 }
2865 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
};
2866 const ccv_nnc_cmd_t rmsnorm = ccv_nnc_cmd(CCV_NNC_RMSNORM_FORWARD, 0, self->params, 0);
2867 ccv_nnc_tensor_param_t output_params[2];
2868 if (self->params.rmsnorm.elementwise_affine)
2869 ccv_nnc_hint_tensor_auto(rmsnorm, (ccv_nnc_tensor_param_t []){
2870 params,
2871 scale_params,
2872 }, 2, ccv_nnc_no_hint, output_params, 2);
2873 else
2874 ccv_nnc_hint_tensor_auto(rmsnorm, (ccv_nnc_tensor_param_t []){
2875 params,
2876 }, 1, ccv_nnc_no_hint, output_params, 2);
2877 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0);
2878 const ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(graph, output_params[1], "saved_inv_std");
2879 if (self->params.rmsnorm.elementwise_affine)
2880 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");
2881 else
2882 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");
2883 outputs[0] = output;
2884}
2885
2886static 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)
2887{
2888 ccv_cnnp_model_rmsnorm_t* const self = (ccv_cnnp_model_rmsnorm_t*)super;
2889 if (self->scale.graph)
2890 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);
2891}
2892
2893static 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)
2894{
2895 ccv_cnnp_model_rmsnorm_t* const self = (ccv_cnnp_model_rmsnorm_t*)super;
2896 if (self->scale.graph)
2897 add_to_array(parameters, self->scale, is_trainable);
2898}
2899
2900static ccv_cnnp_model_t* _ccv_cnnp_rmsnorm_copy(const ccv_cnnp_model_t* const super, void* const context);
2901
2902static const ccv_cnnp_model_vtab_t ccv_cnnp_rmsnorm_isa = {
2903 .build = _ccv_cnnp_rmsnorm_build,
2904 .init_states = _ccv_cnnp_rmsnorm_init_states,
2905 .add_to_parameter = _ccv_cnnp_rmsnorm_add_to_parameter,
2906 .copy = _ccv_cnnp_rmsnorm_copy,
2907};
2908
2909ccv_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 int is_trainable, const char* const name)
2910{
2911 ccv_cnnp_model_rmsnorm_t* const model_rmsnorm = (ccv_cnnp_model_rmsnorm_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_rmsnorm_t));
2912 model_rmsnorm->super.isa = &ccv_cnnp_rmsnorm_isa;
2913 model_rmsnorm->super.input_size = 1;
2914 model_rmsnorm->super.outputs = &model_rmsnorm->output;
2915 model_rmsnorm->super.output_size = 1;
2916 model_rmsnorm->super.is_trainable = is_trainable;
2917 ccv_cnnp_model_copy_name(&model_rmsnorm->super, name);
2918 model_rmsnorm->scale.d = CCV_NNC_NO_TENSOR_SYMBOL;
2919 model_rmsnorm->scale.graph = 0;
2920 model_rmsnorm->params.rmsnorm.epsilon = epsilon;
2921 model_rmsnorm->params.rmsnorm.count = axis_count;
2922 model_rmsnorm->params.rmsnorm.elementwise_affine = elementwise_affine;
2923 memcpy(model_rmsnorm->params.lnorm.axis, axis, sizeof(int) * axis_count);
2924 return (ccv_cnnp_model_t*)model_rmsnorm;
2925}
2926
2927static ccv_cnnp_model_t* _ccv_cnnp_rmsnorm_copy(const ccv_cnnp_model_t* const super, void* const context)
2928{
2929 const ccv_cnnp_model_rmsnorm_t* const self = (const ccv_cnnp_model_rmsnorm_t*)super;
2930 return ccv_cnnp_rmsnorm(self->params.rmsnorm.epsilon, self->params.rmsnorm.axis, self->params.rmsnorm.count, self->params.rmsnorm.elementwise_affine, self->super.is_trainable, self->super.name);
2931}
2932
2933// MARK - Batched Matrix Mul Layer
2934
2935typedef struct {
2936 ccv_cnnp_model_t super;
2937 ccv_nnc_tensor_symbol_t output;
2938 int transpose_a[2];
2939 int transpose_b[2];
2940 int flags;
2941} ccv_cnnp_model_matmul_t;
2942
2943static 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)
2944{
2945 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);
2946 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"
, 2946, __extension__ __PRETTY_FUNCTION__); }));
2947 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", 2947, __extension__ __PRETTY_FUNCTION__
); }));
2948 ccv_cnnp_model_matmul_t* const self = (ccv_cnnp_model_matmul_t*)super;
2949 ccv_nnc_tensor_param_t a_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2950 ccv_nnc_tensor_param_t b_params = ccv_nnc_tensor_symbol_params(graph, inputs[1]);
2951 ccv_nnc_tensor_param_t output_params;
2952 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);
2953 matmul.info.blas.flags = self->flags;
2954 ccv_nnc_hint_tensor_auto(matmul, (ccv_nnc_tensor_param_t []){
2955 a_params,
2956 b_params,
2957 }, 2, ccv_nnc_no_hint, &output_params, 1);
2958 const ccv_nnc_tensor_symbol_t matmul_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2959 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");
2960 outputs[0] = matmul_output;
2961}
2962
2963static ccv_cnnp_model_t* _ccv_cnnp_matmul_copy(const ccv_cnnp_model_t* const super, void* const context);
2964
2965static const ccv_cnnp_model_vtab_t ccv_cnnp_matmul_isa = {
2966 .build = _ccv_cnnp_matmul_build,
2967 .copy = _ccv_cnnp_matmul_copy,
2968};
2969
2970ccv_cnnp_model_t* ccv_cnnp_matmul(const int transpose_a[2], const int transpose_b[2], const int flags, const char* const name)
2971{
2972 ccv_cnnp_model_matmul_t* const model_matmul = (ccv_cnnp_model_matmul_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_matmul_t));
2973 model_matmul->super.isa = &ccv_cnnp_matmul_isa;
2974 model_matmul->super.input_size = 2;
2975 model_matmul->super.outputs = &model_matmul->output;
2976 model_matmul->super.output_size = 1;
2977 model_matmul->transpose_a[0] = transpose_a[0];
2978 model_matmul->transpose_a[1] = transpose_a[1];
2979 model_matmul->transpose_b[0] = transpose_b[0];
2980 model_matmul->transpose_b[1] = transpose_b[1];
2981 model_matmul->flags = flags;
2982 ccv_cnnp_model_copy_name(&model_matmul->super, name);
2983 return (ccv_cnnp_model_t*)model_matmul;
2984}
2985
2986static ccv_cnnp_model_t* _ccv_cnnp_matmul_copy(const ccv_cnnp_model_t* const super, void* const context)
2987{
2988 const ccv_cnnp_model_matmul_t* const self = (const ccv_cnnp_model_matmul_t*)super;
2989 return ccv_cnnp_matmul(self->transpose_a, self->transpose_b, self->flags, self->super.name);
2990}
2991
2992// MARK - Dropout Layer
2993
2994typedef struct {
2995 ccv_cnnp_model_t super;
2996 ccv_nnc_tensor_symbol_t output;
2997 ccv_nnc_graph_exec_symbol_t dropout;
2998 float p;
2999 int entirety;
3000} ccv_cnnp_model_dropout_t;
3001
3002static 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)
3003{
3004 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);
3005 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"
, 3005, __extension__ __PRETTY_FUNCTION__); }));
3006 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", 3006, __extension__ __PRETTY_FUNCTION__
); }));
3007 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3008 ccv_nnc_tensor_param_t output_params[2];
3009 ccv_cnnp_model_dropout_t* const self = (ccv_cnnp_model_dropout_t*)super;
3010 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);
3011 ccv_nnc_hint_tensor_auto(dropout, (ccv_nnc_tensor_param_t []){
3012 params,
3013 }, 1, ccv_nnc_no_hint, output_params, 2);
3014 const ccv_nnc_tensor_symbol_t dropout_output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0);
3015 const ccv_nnc_tensor_symbol_t mask = ccv_nnc_tensor_symbol_new(graph, output_params[1], "mask");
3016 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");
3017 outputs[0] = dropout_output;
3018}
3019
3020static 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)
3021{
3022 ccv_cnnp_model_dropout_t* const self = (ccv_cnnp_model_dropout_t*)super;
3023 if (self->dropout.graph)
3024 {
3025 if (is_test)
3026 // During test, the dropout is not applied. Data transfer is perfect because if these are the same tensor, it will skip.
3027 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);
3028 else
3029 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);
3030 }
3031}
3032
3033static ccv_cnnp_model_t* _ccv_cnnp_dropout_copy(const ccv_cnnp_model_t* const super, void* const context);
3034
3035static const ccv_cnnp_model_vtab_t ccv_cnnp_dropout_isa = {
3036 .build = _ccv_cnnp_dropout_build,
3037 .set_is_test = _ccv_cnnp_dropout_set_is_test,
3038 .copy = _ccv_cnnp_dropout_copy,
3039};
3040
3041ccv_cnnp_model_t* ccv_cnnp_dropout(const float p, const int entirety, const char* const name)
3042{
3043 ccv_cnnp_model_dropout_t* const model_dropout = (ccv_cnnp_model_dropout_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_dropout_t));
3044 model_dropout->super.isa = &ccv_cnnp_dropout_isa;
3045 model_dropout->super.input_size = 1;
3046 model_dropout->super.outputs = &model_dropout->output;
3047 model_dropout->super.output_size = 1;
3048 model_dropout->p = p;
3049 model_dropout->entirety = entirety;
3050 ccv_cnnp_model_copy_name(&model_dropout->super, name);
3051 return (ccv_cnnp_model_t*)model_dropout;
3052}
3053
3054static ccv_cnnp_model_t* _ccv_cnnp_dropout_copy(const ccv_cnnp_model_t* const super, void* const context)
3055{
3056 const ccv_cnnp_model_dropout_t* const self = (const ccv_cnnp_model_dropout_t*)super;
3057 return ccv_cnnp_dropout(self->p, self->entirety, self->super.name);
3058}
3059
3060// MARK - Masked Fill Layer
3061
3062typedef struct {
3063 ccv_cnnp_model_t super;
3064 ccv_nnc_tensor_symbol_t output;
3065 float eq;
3066 float fill;
3067} ccv_cnnp_model_masked_fill_t;
3068
3069static 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)
3070{
3071 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);
3072 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"
, 3072, __extension__ __PRETTY_FUNCTION__); }));
3073 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", 3073, __extension__ __PRETTY_FUNCTION__
); }));
3074 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3075 ccv_cnnp_model_masked_fill_t* const self = (ccv_cnnp_model_masked_fill_t*)super;
3076 const ccv_nnc_tensor_symbol_t masked_fill_output = ccv_nnc_tensor_symbol_new(graph, params, 0);
3077 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");
3078 outputs[0] = masked_fill_output;
3079}
3080
3081static ccv_cnnp_model_t* _ccv_cnnp_masked_fill_copy(const ccv_cnnp_model_t* const super, void* const context);
3082
3083static const ccv_cnnp_model_vtab_t ccv_cnnp_masked_fill_isa = {
3084 .build = _ccv_cnnp_masked_fill_build,
3085 .copy = _ccv_cnnp_masked_fill_copy,
3086};
3087
3088ccv_cnnp_model_t* ccv_cnnp_masked_fill(const float eq, const float fill, const char* const name)
3089{
3090 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));
3091 model_masked_fill->super.isa = &ccv_cnnp_masked_fill_isa;
3092 model_masked_fill->super.input_size = 2;
3093 model_masked_fill->super.outputs = &model_masked_fill->output;
3094 model_masked_fill->super.output_size = 1;
3095 model_masked_fill->eq = eq;
3096 model_masked_fill->fill = fill;
3097 ccv_cnnp_model_copy_name(&model_masked_fill->super, name);
3098 return (ccv_cnnp_model_t*)model_masked_fill;
3099}
3100
3101static ccv_cnnp_model_t* _ccv_cnnp_masked_fill_copy(const ccv_cnnp_model_t* const super, void* const context)
3102{
3103 const ccv_cnnp_model_masked_fill_t* const self = (const ccv_cnnp_model_masked_fill_t*)super;
3104 return ccv_cnnp_masked_fill(self->eq, self->fill, self->super.name);
3105}
3106
3107// MARK - Index Select Layer
3108
3109typedef struct {
3110 ccv_cnnp_model_t super;
3111 ccv_nnc_tensor_symbol_t output;
3112} ccv_cnnp_model_index_select_t;
3113
3114static 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)
3115{
3116 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);
3117 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"
, 3117, __extension__ __PRETTY_FUNCTION__); }));
3118 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", 3118, __extension__ __PRETTY_FUNCTION__
); }));
3119 const ccv_nnc_tensor_param_t vocab_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3120 const ccv_nnc_tensor_param_t index_params = ccv_nnc_tensor_symbol_params(graph, inputs[1]);
3121 ccv_nnc_tensor_param_t output_params;
3122 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);
3123 ccv_nnc_hint_tensor_auto(index_select, (ccv_nnc_tensor_param_t []){
3124 vocab_params,
3125 index_params,
3126 }, 2, ccv_nnc_no_hint, &output_params, 1);
3127 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3128 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");
3129 outputs[0] = output;
3130}
3131
3132static ccv_cnnp_model_t* _ccv_cnnp_index_select_copy(const ccv_cnnp_model_t* const super, void* const context);
3133
3134static const ccv_cnnp_model_vtab_t ccv_cnnp_index_select_isa = {
3135 .build = _ccv_cnnp_index_select_build,
3136 .copy = _ccv_cnnp_index_select_copy,
3137};
3138
3139ccv_cnnp_model_t* ccv_cnnp_index_select(const char* const name)
3140{
3141 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));
3142 model_index_select->super.isa = &ccv_cnnp_index_select_isa;
3143 model_index_select->super.input_size = 2;
3144 model_index_select->super.outputs = &model_index_select->output;
3145 model_index_select->super.output_size = 1;
3146 ccv_cnnp_model_copy_name(&model_index_select->super, name);
3147 return (ccv_cnnp_model_t*)model_index_select;
3148}
3149
3150static ccv_cnnp_model_t* _ccv_cnnp_index_select_copy(const ccv_cnnp_model_t* const super, void* const context)
3151{
3152 ccv_cnnp_model_index_select_t* const self = (ccv_cnnp_model_index_select_t*)super;
3153 return ccv_cnnp_index_select(self->super.name);
3154}
3155
3156// MARK - Embedding Layer
3157
3158typedef struct {
3159 ccv_cnnp_model_t super;
3160 ccv_nnc_tensor_symbol_t output;
3161 ccv_nnc_tensor_symbol_t vocab;
3162 int datatype;
3163 int vocab_size;
3164 int embed_size;
3165} ccv_cnnp_model_embedding_t;
3166
3167static 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)
3168{
3169 ccv_cnnp_model_embedding_t* const self = (ccv_cnnp_model_embedding_t*)super;
3170 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);
3171 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"
, 3171, __extension__ __PRETTY_FUNCTION__); }));
3172 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", 3172, __extension__ __PRETTY_FUNCTION__
); }));
3173 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3174 ccv_nnc_tensor_param_t vocab_params = params;
3175 memset(vocab_params.dim, 0, sizeof(vocab_params.dim));
3176 vocab_params.datatype = self->datatype;
3177 vocab_params.dim[0] = self->vocab_size;
3178 vocab_params.dim[1] = self->embed_size;
3179 if (!self->vocab.graph)
3180 self->vocab = ccv_nnc_tensor_symbol_new(graph, vocab_params, "vocab");
3181 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", 3181
, __extension__ __PRETTY_FUNCTION__); }));
3182 const ccv_nnc_tensor_symbol_t vocab = ccv_cnnp_model_get_symbol(super, self->vocab);
3183 ccv_nnc_tensor_param_t output_params;
3184 const ccv_nnc_cmd_t embedding = CMD_INDEX_SELECT_FORWARD()ccv_nnc_cmd(CCV_NNC_INDEX_SELECT_FORWARD, 0, ccv_nnc_cmd_auto
, 0);
3185 ccv_nnc_hint_tensor_auto(embedding, (ccv_nnc_tensor_param_t []){
3186 vocab_params,
3187 params,
3188 }, 2, ccv_nnc_no_hint, &output_params, 1);
3189 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3190 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");
3191 outputs[0] = output;
3192}
3193
3194static 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)
3195{
3196 ccv_cnnp_model_embedding_t* const self = (ccv_cnnp_model_embedding_t*)super;
3197 const float std = sqrtf(2) / sqrtf(self->vocab_size + self->embed_size);
3198 const float bound = sqrtf(3) * std;
3199 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);
3200}
3201
3202static 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)
3203{
3204 ccv_cnnp_model_embedding_t* const self = (ccv_cnnp_model_embedding_t*)super;
3205 add_to_array(parameters, self->vocab, is_trainable);
3206}
3207
3208static ccv_cnnp_model_t* _ccv_cnnp_embedding_copy(const ccv_cnnp_model_t* const super, void* const context);
3209
3210static const ccv_cnnp_model_vtab_t ccv_cnnp_embedding_isa = {
3211 .build = _ccv_cnnp_embedding_build,
3212 .init_states = _ccv_cnnp_embedding_init_states,
3213 .add_to_parameter = _ccv_cnnp_embedding_add_to_parameter,
3214 .copy = _ccv_cnnp_embedding_copy,
3215};
3216
3217ccv_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)
3218{
3219 ccv_cnnp_model_embedding_t* const model_embedding = (ccv_cnnp_model_embedding_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_embedding_t));
3220 model_embedding->super.isa = &ccv_cnnp_embedding_isa;
3221 model_embedding->super.input_size = 1;
3222 model_embedding->super.outputs = &model_embedding->output;
3223 model_embedding->super.output_size = 1;
3224 model_embedding->super.is_trainable = is_trainable;
3225 ccv_cnnp_model_copy_name(&model_embedding->super, name);
3226 model_embedding->vocab.d = CCV_NNC_NO_TENSOR_SYMBOL;
3227 model_embedding->vocab.graph = 0;
3228 assert(datatype == CCV_32F || datatype == CCV_16F || datatype == CCV_16BF)((void) sizeof ((datatype == CCV_32F || datatype == CCV_16F ||
datatype == CCV_16BF) ? 1 : 0), __extension__ ({ if (datatype
== CCV_32F || datatype == CCV_16F || datatype == CCV_16BF) ;
else __assert_fail ("datatype == CCV_32F || datatype == CCV_16F || datatype == CCV_16BF"
, "ccv_cnnp_model_addons.c", 3228, __extension__ __PRETTY_FUNCTION__
); }));
3229 model_embedding->datatype = datatype;
3230 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", 3230, __extension__ __PRETTY_FUNCTION__
); }));
3231 model_embedding->vocab_size = vocab_size;
3232 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", 3232, __extension__ __PRETTY_FUNCTION__
); }));
3233 model_embedding->embed_size = embed_size;
3234 return (ccv_cnnp_model_t*)model_embedding;
3235}
3236
3237static ccv_cnnp_model_t* _ccv_cnnp_embedding_copy(const ccv_cnnp_model_t* const super, void* const context)
3238{
3239 ccv_cnnp_model_embedding_t* const self = (ccv_cnnp_model_embedding_t*)super;
3240 return ccv_cnnp_embedding(self->datatype, self->vocab_size, self->embed_size, self->super.is_trainable, self->super.name);
3241}
3242
3243// MARK - Pool Layers
3244
3245typedef struct {
3246 ccv_cnnp_model_t super;
3247 ccv_nnc_tensor_symbol_t output;
3248 int type;
3249 float width_scale;
3250 float height_scale;
3251 int align_corners;
3252} ccv_cnnp_model_upsample_t;
3253
3254static 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)
3255{
3256 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);
3257 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"
, 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_upsample_t* const self = (ccv_cnnp_model_upsample_t*)super;
3260 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3261 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);
3262 ccv_nnc_tensor_param_t output_params;
3263 ccv_nnc_hint_tensor_auto(cmd, &params, 1, ccv_nnc_no_hint, &output_params, 1);
3264 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3265 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");
3266 outputs[0] = output;
3267}
3268
3269static ccv_cnnp_model_t* _ccv_cnnp_upsample_copy(const ccv_cnnp_model_t* const super, void* const context);
3270
3271static const ccv_cnnp_model_vtab_t ccv_cnnp_upsample_isa = {
3272 .build = _ccv_cnnp_upsample_build,
3273 .copy = _ccv_cnnp_upsample_copy,
3274};
3275
3276ccv_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)
3277{
3278 ccv_cnnp_model_upsample_t* const model_upsample = (ccv_cnnp_model_upsample_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_upsample_t));
3279 model_upsample->super.isa = &ccv_cnnp_upsample_isa;
3280 model_upsample->super.input_size = 1;
3281 model_upsample->super.outputs = &model_upsample->output;
3282 model_upsample->super.output_size = 1;
3283 ccv_cnnp_model_copy_name(&model_upsample->super, name);
3284 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", 3284, __extension__ __PRETTY_FUNCTION__
); }));
3285 model_upsample->type = type;
3286 model_upsample->width_scale = width_scale;
3287 model_upsample->height_scale = height_scale;
3288 model_upsample->align_corners = align_corners;
3289 return (ccv_cnnp_model_t*)model_upsample;
3290}
3291
3292static ccv_cnnp_model_t* _ccv_cnnp_upsample_copy(const ccv_cnnp_model_t* const super, void* const context)
3293{
3294 const ccv_cnnp_model_upsample_t* const self = (const ccv_cnnp_model_upsample_t*)super;
3295 return ccv_cnnp_upsample(self->type, self->width_scale, self->height_scale, self->align_corners, self->super.name);
3296}
3297
3298// MARK - Reduce Sum Layer
3299
3300typedef struct {
3301 ccv_cnnp_model_t super;
3302 int axis[CCV_NNC_MAX_DIM_ALLOC(12)];
3303 int count;
3304 ccv_nnc_tensor_symbol_t output;
3305} ccv_cnnp_model_reduce_sum_t;
3306
3307static 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)
3308{
3309 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);
3310 const ccv_cnnp_model_reduce_sum_t* const self = (const ccv_cnnp_model_reduce_sum_t*)super;
3311 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"
, 3311, __extension__ __PRETTY_FUNCTION__); }));
3312 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", 3312, __extension__ __PRETTY_FUNCTION__
); }));
3313 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3314 ccv_nnc_tensor_param_t output_params;
3315 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);
3316 int i;
3317 for (i = 0; i < self->count; i++)
3318 reduce_sum.info.reduce.axis[i] = self->axis[i];
3319 reduce_sum.info.reduce.count = self->count;
3320 ccv_nnc_hint_tensor_auto(reduce_sum, &input_params, 1, ccv_nnc_no_hint, &output_params, 1);
3321 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3322 ccv_nnc_graph_exec_symbol_new(graph, reduce_sum, inputs, input_size, outputs, output_size, "reduce_sum");
3323}
3324
3325static ccv_cnnp_model_t* _ccv_cnnp_reduce_sum_copy(const ccv_cnnp_model_t* const self, void* const context);
3326
3327static const ccv_cnnp_model_vtab_t ccv_cnnp_reduce_sum_isa = {
3328 .build = _ccv_cnnp_reduce_sum_build,
3329 .copy = _ccv_cnnp_reduce_sum_copy,
3330};
3331
3332ccv_cnnp_model_t* ccv_cnnp_reduce_sum(const int* const axis, const int axis_count, const char* const name)
3333{
3334 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));
3335 model_reduce_sum->super.isa = &ccv_cnnp_reduce_sum_isa;
3336 model_reduce_sum->super.input_size = 1;
3337 model_reduce_sum->super.outputs = &model_reduce_sum->output;
3338 model_reduce_sum->super.output_size = 1;
3339 ccv_cnnp_model_copy_name(&model_reduce_sum->super, name);
3340 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", 3340, __extension__ __PRETTY_FUNCTION__
); }));
3341 int i;
3342 for (i = 0; i < axis_count; i++)
3343 model_reduce_sum->axis[i] = axis[i];
3344 model_reduce_sum->count = axis_count;
3345 return (ccv_cnnp_model_t*)model_reduce_sum;
3346}
3347
3348static ccv_cnnp_model_t* _ccv_cnnp_reduce_sum_copy(const ccv_cnnp_model_t* const super, void* const context)
3349{
3350 const ccv_cnnp_model_reduce_sum_t* const self = (const ccv_cnnp_model_reduce_sum_t*)super;
3351 return ccv_cnnp_reduce_sum(self->axis, self->count, self->super.name);
3352}
3353
3354// MARK - Reduce Mean Layer
3355
3356typedef struct {
3357 ccv_cnnp_model_t super;
3358 int axis[CCV_NNC_MAX_DIM_ALLOC(12)];
3359 int count;
3360 ccv_nnc_tensor_symbol_t output;
3361} ccv_cnnp_model_reduce_mean_t;
3362
3363static 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)
3364{
3365 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);
3366 const ccv_cnnp_model_reduce_mean_t* const self = (const ccv_cnnp_model_reduce_mean_t*)super;
3367 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"
, 3367, __extension__ __PRETTY_FUNCTION__); }));
3368 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", 3368, __extension__ __PRETTY_FUNCTION__
); }));
3369 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3370 ccv_nnc_tensor_param_t output_params;
3371 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);
3372 int i;
3373 for (i = 0; i < self->count; i++)
3374 reduce_mean.info.reduce.axis[i] = self->axis[i];
3375 reduce_mean.info.reduce.count = self->count;
3376 ccv_nnc_hint_tensor_auto(reduce_mean, &input_params, 1, ccv_nnc_no_hint, &output_params, 1);
3377 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3378 ccv_nnc_graph_exec_symbol_new(graph, reduce_mean, inputs, input_size, outputs, output_size, "reduce_mean");
3379}
3380
3381static ccv_cnnp_model_t* _ccv_cnnp_reduce_mean_copy(const ccv_cnnp_model_t* const self, void* const context);
3382
3383static const ccv_cnnp_model_vtab_t ccv_cnnp_reduce_mean_isa = {
3384 .build = _ccv_cnnp_reduce_mean_build,
3385 .copy = _ccv_cnnp_reduce_mean_copy,
3386};
3387
3388ccv_cnnp_model_t* ccv_cnnp_reduce_mean(const int* const axis, const int axis_count, const char* const name)
3389{
3390 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));
3391 model_reduce_mean->super.isa = &ccv_cnnp_reduce_mean_isa;
3392 model_reduce_mean->super.input_size = 1;
3393 model_reduce_mean->super.outputs = &model_reduce_mean->output;
3394 model_reduce_mean->super.output_size = 1;
3395 ccv_cnnp_model_copy_name(&model_reduce_mean->super, name);
3396 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", 3396, __extension__ __PRETTY_FUNCTION__
); }));
3397 int i;
3398 for (i = 0; i < axis_count; i++)
3399 model_reduce_mean->axis[i] = axis[i];
3400 model_reduce_mean->count = axis_count;
3401 return (ccv_cnnp_model_t*)model_reduce_mean;
3402}
3403
3404static ccv_cnnp_model_t* _ccv_cnnp_reduce_mean_copy(const ccv_cnnp_model_t* const super, void* const context)
3405{
3406 const ccv_cnnp_model_reduce_mean_t* const self = (const ccv_cnnp_model_reduce_mean_t*)super;
3407 return ccv_cnnp_reduce_mean(self->axis, self->count, self->super.name);
3408}
3409
3410// MARK - Reduce Max Layer
3411
3412typedef struct {
3413 ccv_cnnp_model_t super;
3414 int axis[CCV_NNC_MAX_DIM_ALLOC(12)];
3415 int count;
3416 ccv_nnc_tensor_symbol_t output;
3417} ccv_cnnp_model_reduce_max_t;
3418
3419static 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)
3420{
3421 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);
3422 const ccv_cnnp_model_reduce_max_t* const self = (const ccv_cnnp_model_reduce_max_t*)super;
3423 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"
, 3423, __extension__ __PRETTY_FUNCTION__); }));
3424 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", 3424, __extension__ __PRETTY_FUNCTION__
); }));
3425 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3426 ccv_nnc_tensor_param_t output_params;
3427 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);
3428 int i;
3429 for (i = 0; i < self->count; i++)
3430 reduce_max.info.reduce.axis[i] = self->axis[i];
3431 reduce_max.info.reduce.count = self->count;
3432 ccv_nnc_hint_tensor_auto(reduce_max, &input_params, 1, ccv_nnc_no_hint, &output_params, 1);
3433 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3434 ccv_nnc_graph_exec_symbol_new(graph, reduce_max, inputs, input_size, outputs, output_size, "reduce_max");
3435}
3436
3437static ccv_cnnp_model_t* _ccv_cnnp_reduce_max_copy(const ccv_cnnp_model_t* const self, void* const context);
3438
3439static const ccv_cnnp_model_vtab_t ccv_cnnp_reduce_max_isa = {
3440 .build = _ccv_cnnp_reduce_max_build,
3441 .copy = _ccv_cnnp_reduce_max_copy,
3442};
3443
3444ccv_cnnp_model_t* ccv_cnnp_reduce_max(const int* const axis, const int axis_count, const char* const name)
3445{
3446 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));
3447 model_reduce_max->super.isa = &ccv_cnnp_reduce_max_isa;
3448 model_reduce_max->super.input_size = 1;
3449 model_reduce_max->super.outputs = &model_reduce_max->output;
3450 model_reduce_max->super.output_size = 1;
3451 ccv_cnnp_model_copy_name(&model_reduce_max->super, name);
3452 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", 3452, __extension__ __PRETTY_FUNCTION__
); }));
3453 int i;
3454 for (i = 0; i < axis_count; i++)
3455 model_reduce_max->axis[i] = axis[i];
3456 model_reduce_max->count = axis_count;
3457 return (ccv_cnnp_model_t*)model_reduce_max;
3458}
3459
3460static ccv_cnnp_model_t* _ccv_cnnp_reduce_max_copy(const ccv_cnnp_model_t* const super, void* const context)
3461{
3462 const ccv_cnnp_model_reduce_max_t* const self = (const ccv_cnnp_model_reduce_max_t*)super;
3463 return ccv_cnnp_reduce_max(self->axis, self->count, self->super.name);
3464}
3465
3466// MARK - Reduce Min Layer
3467
3468typedef struct {
3469 ccv_cnnp_model_t super;
3470 int axis[CCV_NNC_MAX_DIM_ALLOC(12)];
3471 int count;
3472 ccv_nnc_tensor_symbol_t output;
3473} ccv_cnnp_model_reduce_min_t;
3474
3475static 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)
3476{
3477 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);
3478 const ccv_cnnp_model_reduce_min_t* const self = (const ccv_cnnp_model_reduce_min_t*)super;
3479 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"
, 3479, __extension__ __PRETTY_FUNCTION__); }));
3480 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", 3480, __extension__ __PRETTY_FUNCTION__
); }));
3481 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3482 ccv_nnc_tensor_param_t output_params;
3483 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);
3484 int i;
3485 for (i = 0; i < self->count; i++)
3486 reduce_min.info.reduce.axis[i] = self->axis[i];
3487 reduce_min.info.reduce.count = self->count;
3488 ccv_nnc_hint_tensor_auto(reduce_min, &input_params, 1, ccv_nnc_no_hint, &output_params, 1);
3489 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3490 ccv_nnc_graph_exec_symbol_new(graph, reduce_min, inputs, input_size, outputs, output_size, "reduce_min");
3491}
3492
3493static ccv_cnnp_model_t* _ccv_cnnp_reduce_min_copy(const ccv_cnnp_model_t* const self, void* const context);
3494
3495static const ccv_cnnp_model_vtab_t ccv_cnnp_reduce_min_isa = {
3496 .build = _ccv_cnnp_reduce_min_build,
3497 .copy = _ccv_cnnp_reduce_min_copy,
3498};
3499
3500ccv_cnnp_model_t* ccv_cnnp_reduce_min(const int* const axis, const int axis_count, const char* const name)
3501{
3502 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));
3503 model_reduce_min->super.isa = &ccv_cnnp_reduce_min_isa;
3504 model_reduce_min->super.input_size = 1;
3505 model_reduce_min->super.outputs = &model_reduce_min->output;
3506 model_reduce_min->super.output_size = 1;
3507 ccv_cnnp_model_copy_name(&model_reduce_min->super, name);
3508 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", 3508, __extension__ __PRETTY_FUNCTION__
); }));
3509 int i;
3510 for (i = 0; i < axis_count; i++)
3511 model_reduce_min->axis[i] = axis[i];
3512 model_reduce_min->count = axis_count;
3513 return (ccv_cnnp_model_t*)model_reduce_min;
3514}
3515
3516static ccv_cnnp_model_t* _ccv_cnnp_reduce_min_copy(const ccv_cnnp_model_t* const super, void* const context)
3517{
3518 const ccv_cnnp_model_reduce_min_t* const self = (const ccv_cnnp_model_reduce_min_t*)super;
3519 return ccv_cnnp_reduce_min(self->axis, self->count, self->super.name);
3520}
3521
3522// MARK - Reduce Norm2 Layer
3523
3524typedef struct {
3525 ccv_cnnp_model_t super;
3526 int axis[CCV_NNC_MAX_DIM_ALLOC(12)];
3527 int count;
3528 ccv_nnc_tensor_symbol_t output;
3529} ccv_cnnp_model_reduce_norm2_t;
3530
3531static 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)
3532{
3533 const ccv_cnnp_model_reduce_norm2_t* const self = (const ccv_cnnp_model_reduce_norm2_t*)super;
3534 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);
3535 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"
, 3535, __extension__ __PRETTY_FUNCTION__); }));
3536 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", 3536, __extension__ __PRETTY_FUNCTION__
); }));
3537 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3538 ccv_nnc_tensor_param_t output_params;
3539 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);
3540 int i;
3541 for (i = 0; i < self->count; i++)
3542 reduce_norm2.info.reduce.axis[i] = self->axis[i];
3543 reduce_norm2.info.reduce.count = self->count;
3544 ccv_nnc_hint_tensor_auto(reduce_norm2, &input_params, 1, ccv_nnc_no_hint, &output_params, 1);
3545 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3546 ccv_nnc_graph_exec_symbol_new(graph, reduce_norm2, inputs, input_size, outputs, output_size, "reduce_norm2");
3547}
3548
3549static ccv_cnnp_model_t* _ccv_cnnp_reduce_norm2_copy(const ccv_cnnp_model_t* const self, void* const context);
3550
3551static const ccv_cnnp_model_vtab_t ccv_cnnp_reduce_norm2_isa = {
3552 .build = _ccv_cnnp_reduce_norm2_build,
3553 .copy = _ccv_cnnp_reduce_norm2_copy,
3554};
3555
3556ccv_cnnp_model_t* ccv_cnnp_reduce_norm2(const int* const axis, const int axis_count, const char* const name)
3557{
3558 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));
3559 model_reduce_norm2->super.isa = &ccv_cnnp_reduce_norm2_isa;
3560 model_reduce_norm2->super.input_size = 1;
3561 model_reduce_norm2->super.outputs = &model_reduce_norm2->output;
3562 model_reduce_norm2->super.output_size = 1;
3563 ccv_cnnp_model_copy_name(&model_reduce_norm2->super, name);
3564 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", 3564, __extension__ __PRETTY_FUNCTION__
); }));
3565 int i;
3566 for (i = 0; i < axis_count; i++)
3567 model_reduce_norm2->axis[i] = axis[i];
3568 model_reduce_norm2->count = axis_count;
3569 return (ccv_cnnp_model_t*)model_reduce_norm2;
3570}
3571
3572static ccv_cnnp_model_t* _ccv_cnnp_reduce_norm2_copy(const ccv_cnnp_model_t* const super, void* const context)
3573{
3574 const ccv_cnnp_model_reduce_norm2_t* const self = (const ccv_cnnp_model_reduce_norm2_t*)super;
3575 return ccv_cnnp_reduce_norm2(self->axis, self->count, self->super.name);
3576}
3577
3578// MARK - Argmax Layer
3579
3580typedef struct {
3581 ccv_cnnp_model_t super;
3582 int axis;
3583 ccv_nnc_tensor_symbol_t output;
3584} ccv_cnnp_model_argmax_t;
3585
3586static 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)
3587{
3588 const ccv_cnnp_model_argmax_t* const self = (const ccv_cnnp_model_argmax_t*)super;
3589 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);
3590 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"
, 3590, __extension__ __PRETTY_FUNCTION__); }));
3591 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", 3591, __extension__ __PRETTY_FUNCTION__
); }));
3592 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3593 ccv_nnc_tensor_param_t output_params;
3594 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);
3595 argmax.info.reduce.axis[0] = self->axis;
3596 argmax.info.reduce.count = 1;
3597 ccv_nnc_hint_tensor_auto(argmax, &input_params, 1, ccv_nnc_no_hint, &output_params, 1);
3598 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3599 ccv_nnc_graph_exec_symbol_new(graph, argmax, inputs, input_size, outputs, output_size, "argmax");
3600}
3601
3602static ccv_cnnp_model_t* _ccv_cnnp_argmax_copy(const ccv_cnnp_model_t* const self, void* const context);
3603
3604static const ccv_cnnp_model_vtab_t ccv_cnnp_argmax_isa = {
3605 .build = _ccv_cnnp_argmax_build,
3606 .copy = _ccv_cnnp_argmax_copy,
3607};
3608
3609ccv_cnnp_model_t* ccv_cnnp_argmax(const int axis, const char* const name)
3610{
3611 ccv_cnnp_model_argmax_t* const model_argmax = (ccv_cnnp_model_argmax_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_argmax_t));
3612 model_argmax->super.isa = &ccv_cnnp_argmax_isa;
3613 model_argmax->super.input_size = 1;
3614 model_argmax->super.outputs = &model_argmax->output;
3615 model_argmax->super.output_size = 1;
3616 ccv_cnnp_model_copy_name(&model_argmax->super, name);
3617 model_argmax->axis = axis;
3618 return (ccv_cnnp_model_t*)model_argmax;
3619}
3620
3621static ccv_cnnp_model_t* _ccv_cnnp_argmax_copy(const ccv_cnnp_model_t* const super, void* const context)
3622{
3623 const ccv_cnnp_model_argmax_t* const self = (const ccv_cnnp_model_argmax_t*)super;
3624 return ccv_cnnp_argmax(self->axis, self->super.name);
3625}
3626
3627// MARK - Argmin Layer
3628
3629typedef struct {
3630 ccv_cnnp_model_t super;
3631 int axis;
3632 ccv_nnc_tensor_symbol_t output;
3633} ccv_cnnp_model_argmin_t;
3634
3635static 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)
3636{
3637 const ccv_cnnp_model_argmin_t* const self = (const ccv_cnnp_model_argmin_t*)super;
3638 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);
3639 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"
, 3639, __extension__ __PRETTY_FUNCTION__); }));
3640 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", 3640, __extension__ __PRETTY_FUNCTION__
); }));
3641 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3642 ccv_nnc_tensor_param_t output_params;
3643 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);
3644 argmin.info.reduce.axis[0] = self->axis;
3645 argmin.info.reduce.count = 1;
3646 ccv_nnc_hint_tensor_auto(argmin, &input_params, 1, ccv_nnc_no_hint, &output_params, 1);
3647 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3648 ccv_nnc_graph_exec_symbol_new(graph, argmin, inputs, input_size, outputs, output_size, "argmin");
3649}
3650
3651static ccv_cnnp_model_t* _ccv_cnnp_argmin_copy(const ccv_cnnp_model_t* const self, void* const context);
3652
3653static const ccv_cnnp_model_vtab_t ccv_cnnp_argmin_isa = {
3654 .build = _ccv_cnnp_argmin_build,
3655 .copy = _ccv_cnnp_argmin_copy,
3656};
3657
3658ccv_cnnp_model_t* ccv_cnnp_argmin(const int axis, const char* const name)
3659{
3660 ccv_cnnp_model_argmin_t* const model_argmin = (ccv_cnnp_model_argmin_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_argmin_t));
3661 model_argmin->super.isa = &ccv_cnnp_argmin_isa;
3662 model_argmin->super.input_size = 1;
3663 model_argmin->super.outputs = &model_argmin->output;
3664 model_argmin->super.output_size = 1;
3665 ccv_cnnp_model_copy_name(&model_argmin->super, name);
3666 model_argmin->axis = axis;
3667 return (ccv_cnnp_model_t*)model_argmin;
3668}
3669
3670static ccv_cnnp_model_t* _ccv_cnnp_argmin_copy(const ccv_cnnp_model_t* const super, void* const context)
3671{
3672 const ccv_cnnp_model_argmin_t* const self = (const ccv_cnnp_model_argmin_t*)super;
3673 return ccv_cnnp_argmin(self->axis, self->super.name);
3674}
3675
3676// MARK - Min Layer
3677
3678typedef struct {
3679 ccv_cnnp_model_t super;
3680 ccv_nnc_tensor_symbol_t output;
3681} ccv_cnnp_model_min_t;
3682
3683static 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)
3684{
3685 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);
3686 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"
, 3686, __extension__ __PRETTY_FUNCTION__); }));
3687 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", 3687, __extension__ __PRETTY_FUNCTION__
); }));
3688 ccv_nnc_tensor_param_t input_params[2];
3689 int i;
3690 for (i = 0; i < 2; i++)
3691 input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
3692 ccv_nnc_tensor_param_t output_params;
3693 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);
3694 ccv_nnc_hint_tensor_auto(min, input_params, 2, ccv_nnc_no_hint, &output_params, 1);
3695 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3696 ccv_nnc_graph_exec_symbol_new(graph, min, inputs, input_size, outputs, output_size, "min");
3697}
3698
3699static ccv_cnnp_model_t* _ccv_cnnp_min_copy(const ccv_cnnp_model_t* const self, void* const context);
3700
3701static const ccv_cnnp_model_vtab_t ccv_cnnp_min_isa = {
3702 .build = _ccv_cnnp_min_build,
3703 .copy = _ccv_cnnp_min_copy,
3704};
3705
3706ccv_cnnp_model_t* ccv_cnnp_min(const char* const name)
3707{
3708 ccv_cnnp_model_min_t* const model_min = (ccv_cnnp_model_min_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_min_t));
3709 model_min->super.isa = &ccv_cnnp_min_isa;
3710 model_min->super.input_size = 2;
3711 model_min->super.outputs = &model_min->output;
3712 model_min->super.output_size = 1;
3713 ccv_cnnp_model_copy_name(&model_min->super, name);
3714 return (ccv_cnnp_model_t*)model_min;
3715}
3716
3717static ccv_cnnp_model_t* _ccv_cnnp_min_copy(const ccv_cnnp_model_t* const super, void* const context)
3718{
3719 const ccv_cnnp_model_min_t* const self = (const ccv_cnnp_model_min_t*)super;
3720 return ccv_cnnp_min(self->super.name);
3721}
3722
3723// MARK - Max Layer
3724
3725typedef struct {
3726 ccv_cnnp_model_t super;
3727 ccv_nnc_tensor_symbol_t output;
3728} ccv_cnnp_model_max_t;
3729
3730static 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)
3731{
3732 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);
3733 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"
, 3733, __extension__ __PRETTY_FUNCTION__); }));
3734 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", 3734, __extension__ __PRETTY_FUNCTION__
); }));
3735 ccv_nnc_tensor_param_t input_params[2];
3736 int i;
3737 for (i = 0; i < 2; i++)
3738 input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
3739 ccv_nnc_tensor_param_t output_params;
3740 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);
3741 ccv_nnc_hint_tensor_auto(max, input_params, 2, ccv_nnc_no_hint, &output_params, 1);
3742 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3743 ccv_nnc_graph_exec_symbol_new(graph, max, inputs, input_size, outputs, output_size, "max");
3744}
3745
3746static ccv_cnnp_model_t* _ccv_cnnp_max_copy(const ccv_cnnp_model_t* const self, void* const context);
3747
3748static const ccv_cnnp_model_vtab_t ccv_cnnp_max_isa = {
3749 .build = _ccv_cnnp_max_build,
3750 .copy = _ccv_cnnp_max_copy,
3751};
3752
3753ccv_cnnp_model_t* ccv_cnnp_max(const char* const name)
3754{
3755 ccv_cnnp_model_max_t* const model_max = (ccv_cnnp_model_max_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_max_t));
3756 model_max->super.isa = &ccv_cnnp_max_isa;
3757 model_max->super.input_size = 2;
3758 model_max->super.outputs = &model_max->output;
3759 model_max->super.output_size = 1;
3760 ccv_cnnp_model_copy_name(&model_max->super, name);
3761 return (ccv_cnnp_model_t*)model_max;
3762}
3763
3764static ccv_cnnp_model_t* _ccv_cnnp_max_copy(const ccv_cnnp_model_t* const super, void* const context)
3765{
3766 const ccv_cnnp_model_max_t* const self = (const ccv_cnnp_model_max_t*)super;
3767 return ccv_cnnp_max(self->super.name);
3768}
3769
3770// MARK - LSTM Layer
3771
3772typedef struct {
3773 ccv_cnnp_model_t super;
3774 int masked;
3775 ccv_nnc_tensor_symbol_t output;
3776 ccv_nnc_tensor_symbol_t weights;
3777 ccv_nnc_tensor_symbol_t reserves;
3778 ccv_nnc_cmd_param_t params;
3779 ccv_nnc_graph_exec_symbol_t lstm;
3780} ccv_cnnp_model_lstm_t;
3781
3782static int _ccv_cnnp_lstm_weight_dim(int bidirectional, int num_layers, int input_size, int hidden_size, int proj_size, int bias)
3783{
3784 const int D = !!bidirectional + 1;
3785 if (hidden_size == proj_size)
3786 return (num_layers * (bias ? 8 : 0) + (num_layers - 1) * (hidden_size * 4 * D + hidden_size * 4) + input_size * 4 + hidden_size * 4) * D;
3787 else
3788 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;
3789}
3790
3791static 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)
3792{
3793 ccv_cnnp_model_lstm_t* const self = (ccv_cnnp_model_lstm_t*)super;
3794 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);
3795 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", 3795, __extension__ __PRETTY_FUNCTION__
); }));
3796 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", 3796, __extension__ __PRETTY_FUNCTION__
); }));
3797 const int proj_size = self->params.rnn.proj_size == 0 ? self->params.rnn.hidden_size : self->params.rnn.proj_size;
3798 ccv_nnc_tensor_param_t input_params[5];
3799 input_params[0]= ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3800 if (input_size == 2)
3801 input_params[1] = ccv_nnc_tensor_symbol_params(graph, inputs[1]);
3802 input_params[4] = input_params[0];
3803 memset(input_params[4].dim, 0, sizeof(input_params[4].dim));
3804 const int x_nd = ccv_nnc_tensor_nd(input_params[0].dim);
3805 const int feature_count = input_params[0].dim[x_nd - 1];
3806 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);
3807 input_params[4].dim[1] = self->params.rnn.hidden_size;
3808 const ccv_nnc_cmd_t lstm = ccv_nnc_cmd(CCV_NNC_LSTM_FORWARD, 0, self->params, 0);
3809 ccv_nnc_tensor_param_t output_params[4];
3810 ccv_nnc_hint_tensor_auto(lstm, input_params, 5, ccv_nnc_no_hint, output_params, 4);
3811 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0);
3812 if (!self->weights.graph)
3813 self->weights = ccv_nnc_tensor_symbol_new(graph, input_params[4], "weights");
3814 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"
, 3814, __extension__ __PRETTY_FUNCTION__); }));
3815 const ccv_nnc_tensor_symbol_t weights = ccv_cnnp_model_get_symbol(super, self->weights);
3816 if (!self->reserves.graph)
3817 self->reserves = ccv_nnc_tensor_symbol_new(graph, output_params[3], "reserves");
3818 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"
, 3818, __extension__ __PRETTY_FUNCTION__); }));
3819 const ccv_nnc_tensor_symbol_t reserves = ccv_cnnp_model_get_symbol(super, self->reserves);
3820 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
};
3821 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");
3822}
3823
3824static 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)
3825{
3826 ccv_cnnp_model_lstm_t* const self = (ccv_cnnp_model_lstm_t*)super;
3827 if (self->weights.graph)
3828 {
3829 const float stdv = 1.0 / sqrt(self->params.rnn.hidden_size);
3830 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);
3831 }
3832}
3833
3834static 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)
3835{
3836 ccv_cnnp_model_lstm_t* const self = (ccv_cnnp_model_lstm_t*)super;
3837 if (self->weights.graph)
3838 add_to_array(parameters, self->weights, is_trainable);
3839}
3840
3841static 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)
3842{
3843 ccv_cnnp_model_lstm_t* const self = (ccv_cnnp_model_lstm_t*)super;
3844 if (self->lstm.graph)
3845 {
3846 self->params.rnn.is_test = is_test;
3847 updater(context, self->lstm, ccv_nnc_cmd(CCV_NNC_LSTM_FORWARD, 0, self->params, 0), ccv_nnc_no_hint);
3848 }
3849}
3850
3851static ccv_cnnp_model_t* _ccv_cnnp_lstm_copy(const ccv_cnnp_model_t* const self, void* const context);
3852
3853static const ccv_cnnp_model_vtab_t ccv_cnnp_lstm_isa = {
3854 .build = _ccv_cnnp_lstm_build,
3855 .init_states = _ccv_cnnp_lstm_init_states,
3856 .add_to_parameter = _ccv_cnnp_lstm_add_to_parameter,
3857 .copy = _ccv_cnnp_lstm_copy,
3858 .set_is_test = _ccv_cnnp_lstm_set_is_test,
3859};
3860
3861ccv_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)
3862{
3863 ccv_cnnp_model_lstm_t* const model_lstm = (ccv_cnnp_model_lstm_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_lstm_t));
3864 model_lstm->super.isa = &ccv_cnnp_lstm_isa;
3865 model_lstm->super.input_size = masked ? 2 : 1;
3866 model_lstm->super.outputs = &model_lstm->output;
3867 model_lstm->super.output_size = 1;
3868 model_lstm->super.is_trainable = is_trainable;
3869 ccv_cnnp_model_copy_name(&model_lstm->super, name);
3870 model_lstm->masked = masked;
3871 model_lstm->weights.d = CCV_NNC_NO_TENSOR_SYMBOL;
3872 model_lstm->weights.graph = 0;
3873 model_lstm->params.rnn.hidden_size = hidden_size;
3874 model_lstm->params.rnn.proj_size = proj_size;
3875 model_lstm->params.rnn.num_layers = num_layers;
3876 model_lstm->params.rnn.bias = bias;
3877 model_lstm->params.rnn.batch_first = batch_first;
3878 model_lstm->params.rnn.bidirectional = bidirectional;
3879 model_lstm->params.rnn.dropout = dropout;
3880 return (ccv_cnnp_model_t*)model_lstm;
3881}
3882
3883static ccv_cnnp_model_t* _ccv_cnnp_lstm_copy(const ccv_cnnp_model_t* const super, void* const context)
3884{
3885 const ccv_cnnp_model_lstm_t* const self = (const ccv_cnnp_model_lstm_t*)super;
3886 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);
3887}
3888
3889/// MARK - Datatype conversion layer.
3890
3891typedef struct {
3892 ccv_cnnp_model_t super;
3893 ccv_nnc_tensor_symbol_t output;
3894 int datatype;
3895 int ref_to_last;
3896} ccv_cnnp_model_datatype_conversion_t;
3897
3898static 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)
3899{
3900 ccv_cnnp_model_datatype_conversion_t* const self = (ccv_cnnp_model_datatype_conversion_t*)super;
3901 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);
3902 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3903 if (self->ref_to_last)
3904 {
3905 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", 3905, __extension__ __PRETTY_FUNCTION__
); }));
3906 const ccv_nnc_tensor_param_t last_params = ccv_nnc_tensor_symbol_params(graph, inputs[input_size - 1]);
3907 params.datatype = last_params.datatype;
3908 } else
3909 params.datatype = self->datatype;
3910 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", 3910, __extension__ __PRETTY_FUNCTION__
); }));
3911 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
3912 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);
3913}
3914
3915static ccv_cnnp_model_t* _ccv_cnnp_datatype_conversion_copy(const ccv_cnnp_model_t* const self, void* const context);
3916
3917static const ccv_cnnp_model_vtab_t ccv_cnnp_datatype_conversion_isa = {
3918 .build = _ccv_cnnp_datatype_conversion_build,
3919 .copy = _ccv_cnnp_datatype_conversion_copy,
3920};
3921
3922ccv_cnnp_model_t* ccv_cnnp_datatype_conversion(const int datatype, const int ref_to_last, const char* const name)
3923{
3924 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));
3925 model_datatype_conversion->super.isa = &ccv_cnnp_datatype_conversion_isa;
3926 model_datatype_conversion->super.input_size = 0;
3927 model_datatype_conversion->super.outputs = &model_datatype_conversion->output;
3928 model_datatype_conversion->super.output_size = 1;
3929 model_datatype_conversion->datatype = datatype;
3930 model_datatype_conversion->ref_to_last = ref_to_last;
3931 ccv_cnnp_model_copy_name(&model_datatype_conversion->super, name);
3932 return (ccv_cnnp_model_t*)model_datatype_conversion;
3933}
3934
3935static ccv_cnnp_model_t* _ccv_cnnp_datatype_conversion_copy(const ccv_cnnp_model_t* const super, void* const context)
3936{
3937 ccv_cnnp_model_datatype_conversion_t* const self = (ccv_cnnp_model_datatype_conversion_t*)super;
3938 return ccv_cnnp_datatype_conversion(self->datatype, self->ref_to_last, self->super.name);
3939}
3940
3941/// MARK - Clamp layer.
3942
3943typedef struct {
3944 ccv_cnnp_model_t super;
3945 ccv_nnc_tensor_symbol_t output;
3946 float min;
3947 float max;
3948} ccv_cnnp_model_clamp_t;
3949
3950static 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)
3951{
3952 ccv_cnnp_model_clamp_t* const self = (ccv_cnnp_model_clamp_t*)super;
3953 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);
3954 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3955 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", 3955, __extension__ __PRETTY_FUNCTION__
); }));
3956 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
3957 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);
3958}
3959
3960static ccv_cnnp_model_t* _ccv_cnnp_clamp_copy(const ccv_cnnp_model_t* const self, void* const context);
3961
3962static const ccv_cnnp_model_vtab_t ccv_cnnp_clamp_isa = {
3963 .build = _ccv_cnnp_clamp_build,
3964 .copy = _ccv_cnnp_clamp_copy,
3965};
3966
3967ccv_cnnp_model_t* ccv_cnnp_clamp(const float min, const float max, const char* const name)
3968{
3969 ccv_cnnp_model_clamp_t* const model_clamp = (ccv_cnnp_model_clamp_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_clamp_t));
3970 model_clamp->super.isa = &ccv_cnnp_clamp_isa;
3971 model_clamp->super.input_size = 0;
3972 model_clamp->super.outputs = &model_clamp->output;
3973 model_clamp->super.output_size = 1;
3974 model_clamp->min = min;
3975 model_clamp->max = max;
3976 ccv_cnnp_model_copy_name(&model_clamp->super, name);
3977 return (ccv_cnnp_model_t*)model_clamp;
3978}
3979
3980static ccv_cnnp_model_t* _ccv_cnnp_clamp_copy(const ccv_cnnp_model_t* const super, void* const context)
3981{
3982 ccv_cnnp_model_clamp_t* const self = (ccv_cnnp_model_clamp_t*)super;
3983 return ccv_cnnp_clamp(self->min, self->max, self->super.name);
3984}
3985
3986// MARK - Parameter Layer
3987
3988typedef struct {
3989 ccv_cnnp_model_t super;
3990 float init_bound;
3991 ccv_nnc_tensor_symbol_t weights;
3992 ccv_nnc_tensor_param_t weights_params;
3993 ccv_nnc_tensor_symbol_t output;
3994} ccv_cnnp_model_parameter_t;
3995
3996static 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)
3997{
3998 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);
3999 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", 3999, __extension__ __PRETTY_FUNCTION__
); }));
4000 ccv_cnnp_model_parameter_t* const self = (ccv_cnnp_model_parameter_t*)super;
4001 if (!self->weights.graph)
4002 self->weights = ccv_nnc_tensor_symbol_new(graph, self->weights_params, "weights");
4003 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"
, 4003, __extension__ __PRETTY_FUNCTION__); }));
4004 outputs[0] = ccv_cnnp_model_get_symbol(super, self->weights);
4005}
4006
4007static 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)
4008{
4009 ccv_cnnp_model_parameter_t* const self = (ccv_cnnp_model_parameter_t*)super;
4010 if (self->init_bound > 0)
4011 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);
4012 else
4013 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);
4014}
4015
4016static 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)
4017{
4018 ccv_cnnp_model_parameter_t* const self = (ccv_cnnp_model_parameter_t*)super;
4019 add_to_array(parameters, self->weights, is_trainable);
4020}
4021
4022static ccv_cnnp_model_t* _ccv_cnnp_parameter_copy(const ccv_cnnp_model_t* const super, void* const context);
4023
4024static const ccv_cnnp_model_vtab_t ccv_cnnp_parameter_isa = {
4025 .build = _ccv_cnnp_parameter_build,
4026 .init_states = _ccv_cnnp_parameter_init_states,
4027 .add_to_parameter = _ccv_cnnp_parameter_add_to_parameter,
4028 .copy = _ccv_cnnp_parameter_copy,
4029};
4030
4031ccv_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)
4032{
4033 ccv_cnnp_model_parameter_t* const model_parameter = (ccv_cnnp_model_parameter_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_parameter_t));
4034 model_parameter->super.isa = &ccv_cnnp_parameter_isa;
4035 model_parameter->super.input_size = 0;
4036 model_parameter->super.outputs = &model_parameter->output;
4037 model_parameter->super.output_size = 1;
4038 model_parameter->super.is_trainable = is_trainable;
4039 ccv_cnnp_model_copy_name(&model_parameter->super, name);
4040 model_parameter->weights.d = CCV_NNC_NO_TENSOR_SYMBOL;
4041 model_parameter->weights.graph = 0;
4042 model_parameter->weights_params = params;
4043 return (ccv_cnnp_model_t*)model_parameter;
4044}
4045
4046static ccv_cnnp_model_t* _ccv_cnnp_parameter_copy(const ccv_cnnp_model_t* const super, void* const context)
4047{
4048 const ccv_cnnp_model_parameter_t* const self = (const ccv_cnnp_model_parameter_t*)super;
4049 return ccv_cnnp_parameter(self->weights_params, self->init_bound, self->super.is_trainable, self->super.name);
4050}
4051
4052// MARK - Scalar Layer
4053
4054typedef struct {
4055 ccv_cnnp_model_t super;
4056 int type;
4057 int format;
4058 int datatype;
4059 float value;
4060 ccv_nnc_tensor_symbol_t output;
4061} ccv_cnnp_model_scalar_t;
4062
4063static 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)
4064{
4065 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);
4066 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", 4066, __extension__ __PRETTY_FUNCTION__
); }));
4067 ccv_cnnp_model_scalar_t* const self = (ccv_cnnp_model_scalar_t*)super;
4068 ccv_nnc_tensor_param_t params = {
4069 .type = self->type,
4070 .format = self->format,
4071 .datatype = self->datatype,
4072 .dim = {
4073 1
4074 }
4075 };
4076 if (input_size > 0)
4077 {
4078 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4079 params.type = input_params.type;
4080 params.format = input_params.format;
4081 params.datatype = input_params.datatype;
4082 }
4083 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4084 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);
4085}
4086
4087static ccv_cnnp_model_t* _ccv_cnnp_scalar_copy(const ccv_cnnp_model_t* const super, void* const context);
4088
4089static const ccv_cnnp_model_vtab_t ccv_cnnp_scalar_isa = {
4090 .build = _ccv_cnnp_scalar_build,
4091 .copy = _ccv_cnnp_scalar_copy,
4092};
4093
4094ccv_cnnp_model_t* ccv_cnnp_scalar(const int type, const int format, const int datatype, const float value, const char* const name)
4095{
4096 ccv_cnnp_model_scalar_t* const model_scalar = (ccv_cnnp_model_scalar_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_scalar_t));
4097 model_scalar->super.isa = &ccv_cnnp_scalar_isa;
4098 model_scalar->super.input_size = 0;
4099 model_scalar->super.outputs = &model_scalar->output;
4100 model_scalar->super.output_size = 1;
4101 ccv_cnnp_model_copy_name(&model_scalar->super, name);
4102 model_scalar->type = type;
4103 model_scalar->format = format;
4104 model_scalar->datatype = datatype;
4105 model_scalar->value = value;
4106 return (ccv_cnnp_model_t*)model_scalar;
4107}
4108
4109static ccv_cnnp_model_t* _ccv_cnnp_scalar_copy(const ccv_cnnp_model_t* const super, void* const context)
4110{
4111 const ccv_cnnp_model_scalar_t* const self = (const ccv_cnnp_model_scalar_t*)super;
4112 return ccv_cnnp_scalar(self->type, self->format, self->datatype, self->value, self->super.name);
4113}
4114
4115// MARK - Variable Layer
4116
4117typedef struct {
4118 ccv_cnnp_model_t super;
4119 ccv_nnc_tensor_param_t params;
4120 ccv_nnc_tensor_symbol_t output;
4121} ccv_cnnp_model_variable_t;
4122
4123static 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)
4124{
4125 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);
4126 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"
, 4126, __extension__ __PRETTY_FUNCTION__); }));
4127 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", 4127, __extension__ __PRETTY_FUNCTION__
); }));
4128 ccv_cnnp_model_variable_t* const self = (ccv_cnnp_model_variable_t*)super;
4129 outputs[0] = ccv_nnc_tensor_symbol_new(graph, self->params, 0);
4130}
4131
4132static ccv_cnnp_model_t* _ccv_cnnp_variable_copy(const ccv_cnnp_model_t* const super, void* const context);
4133
4134static const ccv_cnnp_model_vtab_t ccv_cnnp_variable_isa = {
4135 .build = _ccv_cnnp_variable_build,
4136 .copy = _ccv_cnnp_variable_copy,
4137};
4138
4139ccv_cnnp_model_t* ccv_cnnp_variable(const ccv_nnc_tensor_param_t params, const char* const name)
4140{
4141 ccv_cnnp_model_variable_t* const model_variable = (ccv_cnnp_model_variable_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_variable_t));
4142 model_variable->super.isa = &ccv_cnnp_variable_isa;
4143 model_variable->super.input_size = 0;
4144 model_variable->super.outputs = &model_variable->output;
4145 model_variable->super.output_size = 1;
4146 ccv_cnnp_model_copy_name(&model_variable->super, name);
4147 model_variable->params = params;
4148 return (ccv_cnnp_model_t*)model_variable;
4149}
4150
4151static ccv_cnnp_model_t* _ccv_cnnp_variable_copy(const ccv_cnnp_model_t* const super, void* const context)
4152{
4153 const ccv_cnnp_model_variable_t* const self = (const ccv_cnnp_model_variable_t*)super;
4154 return ccv_cnnp_variable(self->params, self->super.name);
4155}
4156
4157// MARK - Send Layer
4158
4159typedef struct {
4160 ccv_cnnp_model_t super;
4161 int device_id;
4162 ccv_nnc_tensor_symbol_t output;
4163} ccv_cnnp_model_send_t;
4164
4165static 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)
4166{
4167 ccv_cnnp_model_send_t* const self = (ccv_cnnp_model_send_t*)super;
4168 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);
4169 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"
, 4169, __extension__ __PRETTY_FUNCTION__); }));
4170 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", 4170, __extension__ __PRETTY_FUNCTION__
); }));
4171 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4172 params.type = (params.type & ~0x3) | CCV_TENSOR_GPU_MEMORY;
4173 CCV_TENSOR_SET_DEVICE_ID(params.type, self->device_id)(params.type) = (((params.type) & ~0xfff00) | (((self->
device_id) & 0xfff) << 8));
4174 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4175 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");
4176}
4177
4178static ccv_cnnp_model_t* _ccv_cnnp_send_copy(const ccv_cnnp_model_t* const super, void* const context);
4179
4180static const ccv_cnnp_model_vtab_t ccv_cnnp_send_isa = {
4181 .build = _ccv_cnnp_send_build,
4182 .copy = _ccv_cnnp_send_copy,
4183};
4184
4185ccv_cnnp_model_t* ccv_cnnp_send(const int device_id, const char* const name)
4186{
4187 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", 4187, __extension__ __PRETTY_FUNCTION__
); }));
4188 ccv_cnnp_model_send_t* const model_send = (ccv_cnnp_model_send_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_send_t));
4189 model_send->super.isa = &ccv_cnnp_send_isa;
4190 model_send->super.input_size = 1;
4191 model_send->super.outputs = &model_send->output;
4192 model_send->super.output_size = 1;
4193 ccv_cnnp_model_copy_name(&model_send->super, name);
4194 model_send->device_id = device_id;
4195 return (ccv_cnnp_model_t*)model_send;
4196}
4197
4198static ccv_cnnp_model_t* _ccv_cnnp_send_copy(const ccv_cnnp_model_t* const super, void* const context)
4199{
4200 const ccv_cnnp_model_send_t* const self = (const ccv_cnnp_model_send_t*)super;
4201 return ccv_cnnp_send(self->device_id, self->super.name);
4202}
4203
4204// MARK - Move Layer
4205
4206typedef struct {
4207 ccv_cnnp_model_t super;
4208 ccv_nnc_tensor_symbol_t output;
4209} ccv_cnnp_model_move_t;
4210
4211static 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)
4212{
4213 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);
4214 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"
, 4214, __extension__ __PRETTY_FUNCTION__); }));
4215 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", 4215, __extension__ __PRETTY_FUNCTION__
); }));
4216 outputs[0] = inputs[1];
4217 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");
4218}
4219
4220static ccv_cnnp_model_t* _ccv_cnnp_move_copy(const ccv_cnnp_model_t* const super, void* const context);
4221
4222static const ccv_cnnp_model_vtab_t ccv_cnnp_move_isa = {
4223 .build = _ccv_cnnp_move_build,
4224 .copy = _ccv_cnnp_move_copy,
4225};
4226
4227ccv_cnnp_model_t* ccv_cnnp_move(const char* const name)
4228{
4229 ccv_cnnp_model_move_t* const model_move = (ccv_cnnp_model_move_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_move_t));
4230 model_move->super.isa = &ccv_cnnp_move_isa;
4231 model_move->super.input_size = 2;
4232 model_move->super.outputs = &model_move->output;
4233 model_move->super.output_size = 1;
4234 ccv_cnnp_model_copy_name(&model_move->super, name);
4235 return (ccv_cnnp_model_t*)model_move;
4236}
4237
4238static ccv_cnnp_model_t* _ccv_cnnp_move_copy(const ccv_cnnp_model_t* const super, void* const context)
4239{
4240 const ccv_cnnp_model_move_t* const self = (const ccv_cnnp_model_move_t*)super;
4241 return ccv_cnnp_move(self->super.name);
4242}
4243
4244// MARK - "Making" Contiguous Layer
4245
4246typedef struct {
4247 ccv_cnnp_model_t super;
4248 ccv_nnc_tensor_symbol_t output;
4249} ccv_cnnp_model_contiguous_t;
4250
4251static 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)
4252{
4253 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);
4254 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"
, 4254, __extension__ __PRETTY_FUNCTION__); }));
4255 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", 4255, __extension__ __PRETTY_FUNCTION__
); }));
4256 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4257 ccv_nnc_tensor_symbol_t to = ccv_nnc_tensor_symbol_alias_to(graph, inputs[0]);
4258 if (to.d == CCV_NNC_NO_TENSOR_SYMBOL) // If we are not reshape an alias, it is straightforward.
4259 {
4260 outputs[0] = inputs[0];
4261 return;
4262 }
4263 // Otherwise, we need to check its stride to know if it is contiguous.
4264 int old_stride[CCV_NNC_MAX_DIM_ALLOC(12)];
4265 ccv_nnc_tensor_symbol_alias_params(graph, inputs[0], 0, old_stride);
4266 // We identify permute by checking if the stride is not in descending order.
4267 // This also covered "permute" through reshape, rather than using ccv_cnnp_permute directly.
4268 if (ccv_nnc_is_tensor_stride_packed(old_stride, params.dim))
4269 {
4270 outputs[0] = inputs[0];
4271 return;
4272 }
4273 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4274 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");
4275 ccv_nnc_graph_exec_symbol_set_flags(graph, make_contiguous, CCV_NNC_GRAPH_EXEC_DISABLE_OPT);
4276}
4277
4278static ccv_cnnp_model_t* _ccv_cnnp_contiguous_copy(const ccv_cnnp_model_t* const super, void* const context);
4279
4280static const ccv_cnnp_model_vtab_t ccv_cnnp_contiguous_isa = {
4281 .build = _ccv_cnnp_contiguous_build,
4282 .copy = _ccv_cnnp_contiguous_copy,
4283};
4284
4285ccv_cnnp_model_t* ccv_cnnp_contiguous(const char* const name)
4286{
4287 ccv_cnnp_model_contiguous_t* const model_contiguous = (ccv_cnnp_model_contiguous_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_contiguous_t));
4288 model_contiguous->super.isa = &ccv_cnnp_contiguous_isa;
4289 model_contiguous->super.input_size = 1;
4290 model_contiguous->super.outputs = &model_contiguous->output;
4291 model_contiguous->super.output_size = 1;
4292 ccv_cnnp_model_copy_name(&model_contiguous->super, name);
4293 return (ccv_cnnp_model_t*)model_contiguous;
4294}
4295
4296static ccv_cnnp_model_t* _ccv_cnnp_contiguous_copy(const ccv_cnnp_model_t* const super, void* const context)
4297{
4298 const ccv_cnnp_model_contiguous_t* const self = (const ccv_cnnp_model_contiguous_t*)super;
4299 return ccv_cnnp_contiguous(self->super.name);
4300}
4301
4302// MARK - "Making" Copy Layer
4303
4304typedef struct {
4305 ccv_cnnp_model_t super;
4306 ccv_nnc_tensor_symbol_t output;
4307} ccv_cnnp_model_copy_t;
4308
4309static 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)
4310{
4311 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);
4312 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"
, 4312, __extension__ __PRETTY_FUNCTION__); }));
4313 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", 4313, __extension__ __PRETTY_FUNCTION__
); }));
4314 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4315 ccv_nnc_tensor_symbol_t to = ccv_nnc_tensor_symbol_alias_to(graph, inputs[0]);
4316 if (to.d == CCV_NNC_NO_TENSOR_SYMBOL) // If we are not reshape an alias, it is straightforward.
4317 {
4318 outputs[0] = inputs[0];
4319 return;
4320 }
4321 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4322 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");
4323 ccv_nnc_graph_exec_symbol_set_flags(graph, make_contiguous, CCV_NNC_GRAPH_EXEC_DISABLE_OPT);
4324}
4325
4326static ccv_cnnp_model_t* _ccv_cnnp_copy_copy(const ccv_cnnp_model_t* const super, void* const context);
4327
4328static const ccv_cnnp_model_vtab_t ccv_cnnp_copy_isa = {
4329 .build = _ccv_cnnp_copy_build,
4330 .copy = _ccv_cnnp_copy_copy,
4331};
4332
4333ccv_cnnp_model_t* ccv_cnnp_copy(const char* const name)
4334{
4335 ccv_cnnp_model_copy_t* const model_copy = (ccv_cnnp_model_copy_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_copy_t));
4336 model_copy->super.isa = &ccv_cnnp_copy_isa;
4337 model_copy->super.input_size = 1;
4338 model_copy->super.outputs = &model_copy->output;
4339 model_copy->super.output_size = 1;
4340 ccv_cnnp_model_copy_name(&model_copy->super, name);
4341 return (ccv_cnnp_model_t*)model_copy;
4342}
4343
4344static ccv_cnnp_model_t* _ccv_cnnp_copy_copy(const ccv_cnnp_model_t* const super, void* const context)
4345{
4346 const ccv_cnnp_model_copy_t* const self = (const ccv_cnnp_model_copy_t*)super;
4347 return ccv_cnnp_copy(self->super.name);
4348}
4349
4350// MARK - All-To-All Layer
4351
4352typedef struct {
4353 ccv_cnnp_model_t super;
4354 int axis;
4355 ccv_nnc_tensor_symbol_t outputs[1];
4356} ccv_cnnp_model_all_to_all_t;
4357
4358static 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)
4359{
4360 ccv_cnnp_model_all_to_all_t* const self = (ccv_cnnp_model_all_to_all_t*)super;
4361 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);
4362 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", 4362, __extension__ __PRETTY_FUNCTION__
); }));
4363 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", 4363, __extension__ __PRETTY_FUNCTION__
); }));
4364 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", 4364, __extension__ __PRETTY_FUNCTION__
); }));
4365 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", 4365, __extension__ __PRETTY_FUNCTION__
); }));
4366 const ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4367 const int nd = ccv_nnc_tensor_nd(input_params.dim);
4368 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", 4368, __extension__ __PRETTY_FUNCTION__
); }));
4369 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", 4369, __extension__ __PRETTY_FUNCTION__
); }));
4370 int i;
4371 for (i = 0; i < input_size; i++)
4372 {
4373 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
4374 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", 4374, __extension__ __PRETTY_FUNCTION__
); }));
4375 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", 4375, __extension__ __PRETTY_FUNCTION__
); }));
4376 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", 4376, __extension__ __PRETTY_FUNCTION__
); }));
4377 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", 4377, __extension__ __PRETTY_FUNCTION__
); }));
4378 outputs[i] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4379 }
4380 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");
4381}
4382
4383static ccv_cnnp_model_t* _ccv_cnnp_all_to_all_copy(const ccv_cnnp_model_t* const super, void* const context);
4384
4385static const ccv_cnnp_model_vtab_t ccv_cnnp_all_to_all_isa = {
4386 .build = _ccv_cnnp_all_to_all_build,
4387 .copy = _ccv_cnnp_all_to_all_copy,
4388};
4389
4390ccv_cnnp_model_t* ccv_cnnp_all_to_all(const int count, const int axis, const char* const name)
4391{
4392 assert(count > 0)((void) sizeof ((count > 0) ? 1 : 0), __extension__ ({ if (
count > 0) ; else __assert_fail ("count > 0", "ccv_cnnp_model_addons.c"
, 4392, __extension__ __PRETTY_FUNCTION__); }));
4393 assert(axis >= 0)((void) sizeof ((axis >= 0) ? 1 : 0), __extension__ ({ if (
axis >= 0) ; else __assert_fail ("axis >= 0", "ccv_cnnp_model_addons.c"
, 4393, __extension__ __PRETTY_FUNCTION__); }));
4394 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));
4395 model_all_to_all->super.isa = &ccv_cnnp_all_to_all_isa;
4396 model_all_to_all->super.input_size = count;
4397 model_all_to_all->super.outputs = model_all_to_all->outputs;
4398 model_all_to_all->super.output_size = count;
4399 model_all_to_all->axis = axis;
4400 ccv_cnnp_model_copy_name(&model_all_to_all->super, name);
4401 return (ccv_cnnp_model_t*)model_all_to_all;
4402}
4403
4404static ccv_cnnp_model_t* _ccv_cnnp_all_to_all_copy(const ccv_cnnp_model_t* const super, void* const context)
4405{
4406 const ccv_cnnp_model_all_to_all_t* const self = (const ccv_cnnp_model_all_to_all_t*)super;
4407 return ccv_cnnp_all_to_all(self->super.output_size, self->axis, self->super.name);
4408}
4409
4410// MARK - Scaled-Dot Product Attention Layer
4411
4412typedef struct {
4413 ccv_cnnp_model_t super;
4414 ccv_nnc_tensor_symbol_t output;
4415 ccv_nnc_tensor_symbol_t weights;
4416 ccv_nnc_tensor_symbol_t bias;
4417 float scale;
4418 int is_causal;
4419 int has_attn_mask;
4420 int is_varlen;
4421 int max_seqlen_q;
4422 int max_seqlen_kv;
4423 int flags;
4424 int fused_unify_head_weights;
4425 int no_bias;
4426} ccv_cnnp_model_scaled_dot_product_attention_t;
4427
4428static 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)
4429{
4430 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);
4431 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", 4431, __extension__ __PRETTY_FUNCTION__
); }));
4432 ccv_cnnp_model_scaled_dot_product_attention_t* const self = (ccv_cnnp_model_scaled_dot_product_attention_t*)super;
4433 assert(input_size == (self->is_varlen ? 5 : (self->has_attn_mask ? 4 : 3)))((void) sizeof ((input_size == (self->is_varlen ? 5 : (self
->has_attn_mask ? 4 : 3))) ? 1 : 0), __extension__ ({ if (
input_size == (self->is_varlen ? 5 : (self->has_attn_mask
? 4 : 3))) ; else __assert_fail ("input_size == (self->is_varlen ? 5 : (self->has_attn_mask ? 4 : 3))"
, "ccv_cnnp_model_addons.c", 4433, __extension__ __PRETTY_FUNCTION__
); }));
4434 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", 4434, __extension__ __PRETTY_FUNCTION__
); }));
4435 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", 4435, __extension__ __PRETTY_FUNCTION__
); }));
4436 const ccv_nnc_tensor_param_t q_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4437 const ccv_nnc_tensor_param_t k_params = ccv_nnc_tensor_symbol_params(graph, inputs[1]);
4438 const ccv_nnc_tensor_param_t v_params = ccv_nnc_tensor_symbol_params(graph, inputs[2]);
4439 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){};
4440 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){};
4441 const int v_nd = ccv_nnc_tensor_nd(v_params.dim);
4442 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", 4442, __extension__ __PRETTY_FUNCTION__
); }));
4443 const int hEv = (v_nd == 3 ? 1 : v_params.dim[2]) * v_params.dim[v_nd - 1];
4444 ccv_nnc_tensor_param_t weights_params = q_params;
4445 memset(weights_params.dim, 0, sizeof(weights_params.dim));
4446 weights_params.dim[0] = hEv;
4447 weights_params.dim[1] = hEv;
4448 ccv_nnc_tensor_param_t bias_params = q_params;
4449 memset(bias_params.dim, 0, sizeof(bias_params.dim));
4450 bias_params.dim[0] = hEv;
4451 ccv_nnc_cmd_t cmd = {0};
4452 cmd.cmd = CCV_NNC_SCALED_DOT_PRODUCT_ATTENTION_FORWARD;
4453 cmd.info.scaled_dot_product_attention.scale = self->scale;
4454 cmd.info.scaled_dot_product_attention.is_causal = self->is_causal;
4455 cmd.info.scaled_dot_product_attention.is_varlen = self->is_varlen;
4456 cmd.info.scaled_dot_product_attention.max_seqlen_q = self->max_seqlen_q;
4457 cmd.info.scaled_dot_product_attention.max_seqlen_kv = self->max_seqlen_kv;
4458 cmd.info.scaled_dot_product_attention.flags = self->flags;
4459 ccv_nnc_tensor_param_t output_params[3];
4460 ccv_nnc_tensor_symbol_t output;
4461 ccv_nnc_tensor_symbol_t saved_softmax_lse;
4462 ccv_nnc_tensor_symbol_t saved_v_proj = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
4463 ccv_nnc_tensor_symbol_t attn_mask = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
4464 ccv_nnc_tensor_symbol_t weights = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
4465 ccv_nnc_tensor_symbol_t bias = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
4466 if (self->has_attn_mask)
4467 attn_mask = inputs[3];
4468 if (self->is_varlen)
4469 {
4470 ccv_nnc_hint_tensor_auto(cmd, (ccv_nnc_tensor_param_t []){
4471 q_params,
4472 k_params,
4473 v_params,
4474 (ccv_nnc_tensor_param_t){},
4475 (ccv_nnc_tensor_param_t){},
4476 (ccv_nnc_tensor_param_t){},
4477 q_seq_offsets_params,
4478 kv_seq_offsets_params,
4479 }, 8, ccv_nnc_no_hint, output_params, 2);
4480 output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0);
4481 saved_softmax_lse = ccv_nnc_tensor_symbol_new(graph, output_params[1], 0);
4482 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");
4483 } else if (self->fused_unify_head_weights)
4484 {
4485 if (!self->weights.graph)
4486 self->weights = ccv_nnc_tensor_symbol_new(graph, weights_params, "weights");
4487 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"
, 4487, __extension__ __PRETTY_FUNCTION__); }));
4488 weights = ccv_cnnp_model_get_symbol(super, self->weights);
4489 if (!self->no_bias)
4490 {
4491 if (!self->bias.graph)
4492 self->bias = ccv_nnc_tensor_symbol_new(graph, bias_params, "bias");
4493 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", 4493, __extension__ __PRETTY_FUNCTION__
); }));
4494 bias = ccv_cnnp_model_get_symbol(super, self->bias);
4495 }
4496 ccv_nnc_hint_tensor_auto(cmd, (ccv_nnc_tensor_param_t []){
4497 q_params,
4498 k_params,
4499 v_params,
4500 (ccv_nnc_tensor_param_t){},
4501 weights_params,
4502 bias_params,
4503 }, 6, ccv_nnc_no_hint, output_params, 3);
4504 output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0);
4505 saved_softmax_lse = ccv_nnc_tensor_symbol_new(graph, output_params[1], 0);
4506 saved_v_proj = ccv_nnc_tensor_symbol_new(graph, output_params[2], 0);
4507 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");
4508 } else {
4509 ccv_nnc_hint_tensor_auto(cmd, (ccv_nnc_tensor_param_t []){
4510 q_params,
4511 k_params,
4512 v_params,
4513 }, 3, ccv_nnc_no_hint, output_params, 2);
4514 output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0);
4515 saved_softmax_lse = ccv_nnc_tensor_symbol_new(graph, output_params[1], 0);
4516 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");
4517 }
4518 outputs[0] = output;
4519}
4520
4521static 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)
4522{
4523 ccv_cnnp_model_scaled_dot_product_attention_t* const self = (ccv_cnnp_model_scaled_dot_product_attention_t*)super;
4524 if (self->weights.graph)
4525 {
4526 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"
, 4526, __extension__ __PRETTY_FUNCTION__); }));
4527 const ccv_nnc_tensor_param_t weight_params = ccv_nnc_tensor_symbol_params(graph, self->weights);
4528 const int c = weight_params.dim[1];
4529 const float std = sqrtf(2) / sqrtf(c);
4530 const float bound = sqrtf(3) * std;
4531 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);
4532 if (self->bias.graph)
4533 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);
4534 }
4535}
4536
4537static 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)
4538{
4539 ccv_cnnp_model_scaled_dot_product_attention_t* const self = (ccv_cnnp_model_scaled_dot_product_attention_t*)super;
4540 if (self->weights.graph)
4541 {
4542 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"
, 4542, __extension__ __PRETTY_FUNCTION__); }));
4543 add_to_array(parameters, self->weights, is_trainable);
4544 if (self->bias.graph)
4545 add_to_array(parameters, self->bias, is_trainable);
4546 }
4547}
4548
4549static ccv_cnnp_model_t* _ccv_cnnp_scaled_dot_product_attention_copy(const ccv_cnnp_model_t* const super, void* const context);
4550
4551static const ccv_cnnp_model_vtab_t ccv_cnnp_scaled_dot_product_attention_isa = {
4552 .build = _ccv_cnnp_scaled_dot_product_attention_build,
4553 .copy = _ccv_cnnp_scaled_dot_product_attention_copy,
4554};
4555
4556static const ccv_cnnp_model_vtab_t ccv_cnnp_scaled_dot_product_attention_fused_isa = {
4557 .build = _ccv_cnnp_scaled_dot_product_attention_build,
4558 .init_states = _ccv_cnnp_scaled_dot_product_attention_init_states,
4559 .add_to_parameter = _ccv_cnnp_scaled_dot_product_attention_add_to_parameter,
4560 .copy = _ccv_cnnp_scaled_dot_product_attention_copy,
4561};
4562
4563ccv_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 fused_unify_head_weights, const int no_bias, const int is_trainable, const char* const name)
4564{
4565 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", 4565, __extension__ __PRETTY_FUNCTION__
); }));
4566 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", 4566, __extension__ __PRETTY_FUNCTION__
); }));
4567 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));
4568 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;
4569 model_scaled_dot_product_attention->super.input_size = is_varlen ? 5 : (has_attn_mask ? 4 : 3);
4570 model_scaled_dot_product_attention->super.outputs = &model_scaled_dot_product_attention->output;
4571 model_scaled_dot_product_attention->super.output_size = 1;
4572 model_scaled_dot_product_attention->super.is_trainable = is_trainable;
4573 ccv_cnnp_model_copy_name(&model_scaled_dot_product_attention->super, name);
4574 model_scaled_dot_product_attention->weights.d = CCV_NNC_NO_TENSOR_SYMBOL;
4575 model_scaled_dot_product_attention->weights.graph = 0;
4576 model_scaled_dot_product_attention->bias.d = CCV_NNC_NO_TENSOR_SYMBOL;
4577 model_scaled_dot_product_attention->bias.graph = 0;
4578 model_scaled_dot_product_attention->scale = scale;
4579 model_scaled_dot_product_attention->is_causal = is_causal;
4580 model_scaled_dot_product_attention->has_attn_mask = has_attn_mask;
4581 model_scaled_dot_product_attention->is_varlen = is_varlen;
4582 model_scaled_dot_product_attention->max_seqlen_q = max_seqlen_q;
4583 model_scaled_dot_product_attention->max_seqlen_kv = max_seqlen_kv;
4584 model_scaled_dot_product_attention->flags = flags;
4585 model_scaled_dot_product_attention->fused_unify_head_weights = fused_unify_head_weights;
4586 model_scaled_dot_product_attention->no_bias = no_bias;
4587 return (ccv_cnnp_model_t*)model_scaled_dot_product_attention;
4588}
4589
4590static ccv_cnnp_model_t* _ccv_cnnp_scaled_dot_product_attention_copy(const ccv_cnnp_model_t* const super, void* const context)
4591{
4592 const ccv_cnnp_model_scaled_dot_product_attention_t* const self = (const ccv_cnnp_model_scaled_dot_product_attention_t*)super;
4593 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->fused_unify_head_weights, self->no_bias, self->super.is_trainable, self->super.name);
4594}
4595
4596// MARK - Debug Layer
4597
4598typedef struct {
4599 ccv_cnnp_model_t super;
4600 ccv_nnc_tensor_symbol_t output;
4601 ccv_cnnp_model_debug_f debugger;
4602 ccv_cnnp_model_debug_context_deinit_f debug_deinit;
4603 ccv_cnnp_model_debug_context_copy_f debug_copy;
4604 void* debug_context;
4605} ccv_cnnp_model_debug_t;
4606
4607static 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)
4608{
4609 if (cmd.cmd == CCV_NNC_CUSTOM_BACKWARD)
4610 {
4611 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", 4611, __extension__ __PRETTY_FUNCTION__
); }));
4612 }
4613 ccv_cnnp_model_debug_t* const self = (ccv_cnnp_model_debug_t*)cmd.data;
4614 self->debugger(inputs, input_size, stream_context, self->debug_context);
4615 return CCV_NNC_EXEC_SUCCESS;
4616}
4617
4618static ccv_nnc_cmd_vtab_t ccv_cnnp_debug_exec_isa = {
4619 .exec = _ccv_cnnp_debug_exec
4620};
4621
4622static 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)
4623{
4624 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);
4625 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", 4625, __extension__ __PRETTY_FUNCTION__
); }));
4626 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", 4626, __extension__ __PRETTY_FUNCTION__
); }));
4627 ccv_nnc_tensor_symbol_t to = ccv_nnc_tensor_symbol_alias_to(graph, inputs[0]);
4628 ccv_nnc_tensor_param_t output_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4629 if (to.d == CCV_NNC_NO_TENSOR_SYMBOL) // If we are not reshape an alias, it is straightforward.
4630 {
4631 int ofs[CCV_NNC_MAX_DIM_ALLOC(12)] = {0};
4632 int stride[CCV_NNC_MAX_DIM_ALLOC(12)];
4633 ccv_nnc_tensor_get_stride(output_params.dim, stride);
4634 outputs[0] = ccv_nnc_tensor_symbol_alias_new(graph, inputs[0], ofs, stride, output_params, 0);
4635 } else {
4636 int old_ofs[CCV_NNC_MAX_DIM_ALLOC(12)];
4637 int old_stride[CCV_NNC_MAX_DIM_ALLOC(12)];
4638 ccv_nnc_tensor_symbol_alias_params(graph, inputs[0], old_ofs, old_stride);
4639 outputs[0] = ccv_nnc_tensor_symbol_alias_new(graph, to, old_ofs, old_stride, output_params, 0);
4640 }
4641 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);
4642 cmd.data = self;
4643 ccv_nnc_graph_exec_symbol_t make_debug = ccv_nnc_graph_exec_symbol_new(graph, cmd, inputs, input_size, outputs, 1, "debug");
4644 // Disable any optimizations.
4645 ccv_nnc_graph_exec_symbol_set_flags(graph, make_debug, CCV_NNC_GRAPH_EXEC_DISABLE_OPT);
4646}
4647
4648static void _ccv_cnnp_debug_deinit(ccv_cnnp_model_t* const super)
4649{
4650 const ccv_cnnp_model_debug_t* const self = (const ccv_cnnp_model_debug_t*)super;
4651 if (self->debug_deinit && self->debug_context)
4652 self->debug_deinit(self->debug_context);
4653}
4654
4655static ccv_cnnp_model_t* _ccv_cnnp_debug_copy(const ccv_cnnp_model_t* const super, void* const context);
4656
4657static const ccv_cnnp_model_vtab_t ccv_cnnp_debug_isa = {
4658 .build = _ccv_cnnp_debug_build,
4659 .deinit = _ccv_cnnp_debug_deinit,
4660 .copy = _ccv_cnnp_debug_copy,
4661};
4662
4663ccv_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)
4664{
4665 ccv_cnnp_model_debug_t* const model_debug = (ccv_cnnp_model_debug_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_debug_t));
4666 model_debug->super.isa = &ccv_cnnp_debug_isa;
4667 model_debug->super.input_size = 0;
4668 model_debug->super.outputs = &model_debug->output;
4669 model_debug->super.output_size = 1;
4670 model_debug->debugger = func;
4671 model_debug->debug_context = context;
4672 model_debug->debug_deinit = deinit;
4673 model_debug->debug_copy = copy;
4674 ccv_cnnp_model_copy_name(&model_debug->super, name);
4675 return (ccv_cnnp_model_t*)model_debug;
4676}
4677
4678static ccv_cnnp_model_t* _ccv_cnnp_debug_copy(const ccv_cnnp_model_t* const super, void* const context)
4679{
4680 const ccv_cnnp_model_debug_t* const self = (const ccv_cnnp_model_debug_t*)super;
4681 void* debug_context = self->debug_context;
4682 if (self->debug_copy && self->debug_context)
4683 debug_context = self->debug_copy(self->debug_context);
4684 return ccv_cnnp_debug(self->debugger, debug_context, self->debug_deinit, self->debug_copy, self->super.name);
4685}
4686
4687/// MARK - Sort layer.
4688
4689typedef struct {
4690 ccv_cnnp_model_t super;
4691 ccv_nnc_tensor_symbol_t outputs[2];
4692 int along_axis;
4693 int descending;
4694} ccv_cnnp_model_sort_t;
4695
4696static 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)
4697{
4698 ccv_cnnp_model_sort_t* const self = (ccv_cnnp_model_sort_t*)super;
4699 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);
4700 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4701 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", 4701, __extension__ __PRETTY_FUNCTION__
); }));
4702 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4703 params.datatype = CCV_32S;
4704 outputs[1] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4705 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");
4706}
4707
4708static ccv_cnnp_model_t* _ccv_cnnp_sort_copy(const ccv_cnnp_model_t* const self, void* const context);
4709
4710static const ccv_cnnp_model_vtab_t ccv_cnnp_sort_isa = {
4711 .build = _ccv_cnnp_sort_build,
4712 .copy = _ccv_cnnp_sort_copy,
4713};
4714
4715ccv_cnnp_model_t* ccv_cnnp_sort(const int along_axis, const int descending, const char* const name)
4716{
4717 ccv_cnnp_model_sort_t* const model_sort = (ccv_cnnp_model_sort_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_sort_t));
4718 model_sort->super.isa = &ccv_cnnp_sort_isa;
4719 model_sort->super.input_size = 0;
4720 model_sort->super.outputs = model_sort->outputs;
4721 model_sort->super.output_size = 2;
4722 model_sort->along_axis = along_axis;
4723 model_sort->descending = descending;
4724 ccv_cnnp_model_copy_name(&model_sort->super, name);
4725 return (ccv_cnnp_model_t*)model_sort;
4726}
4727
4728static ccv_cnnp_model_t* _ccv_cnnp_sort_copy(const ccv_cnnp_model_t* const super, void* const context)
4729{
4730 ccv_cnnp_model_sort_t* const self = (ccv_cnnp_model_sort_t*)super;
4731 return ccv_cnnp_sort(self->along_axis, self->descending, self->super.name);
4732}
4733
4734/// MARK - Partition layer.
4735
4736typedef struct {
4737 ccv_cnnp_model_t super;
4738 ccv_nnc_tensor_symbol_t outputs[2];
4739 int kth;
4740 int along_axis;
4741 int descending;
4742} ccv_cnnp_model_partition_t;
4743
4744static 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)
4745{
4746 ccv_cnnp_model_partition_t* const self = (ccv_cnnp_model_partition_t*)super;
4747 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);
4748 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4749 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", 4749, __extension__ __PRETTY_FUNCTION__
); }));
4750 if (self->kth > 0)
4751 params.dim[self->along_axis] = self->kth;
4752 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4753 params.datatype = CCV_32S;
4754 outputs[1] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4755 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");
4756}
4757
4758static ccv_cnnp_model_t* _ccv_cnnp_partition_copy(const ccv_cnnp_model_t* const self, void* const context);
4759
4760static const ccv_cnnp_model_vtab_t ccv_cnnp_partition_isa = {
4761 .build = _ccv_cnnp_partition_build,
4762 .copy = _ccv_cnnp_partition_copy,
4763};
4764
4765ccv_cnnp_model_t* ccv_cnnp_partition(const int kth, const int along_axis, const int descending, const char* const name)
4766{
4767 ccv_cnnp_model_partition_t* const model_partition = (ccv_cnnp_model_partition_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_partition_t));
4768 model_partition->super.isa = &ccv_cnnp_partition_isa;
4769 model_partition->super.input_size = 0;
4770 model_partition->super.outputs = model_partition->outputs;
4771 model_partition->super.output_size = 2;
4772 model_partition->kth = kth;
4773 model_partition->along_axis = along_axis;
4774 model_partition->descending = descending;
4775 ccv_cnnp_model_copy_name(&model_partition->super, name);
4776 return (ccv_cnnp_model_t*)model_partition;
4777}
4778
4779static ccv_cnnp_model_t* _ccv_cnnp_partition_copy(const ccv_cnnp_model_t* const super, void* const context)
4780{
4781 ccv_cnnp_model_partition_t* const self = (ccv_cnnp_model_partition_t*)super;
4782 return ccv_cnnp_partition(self->kth, self->along_axis, self->descending, self->super.name);
4783}
4784
4785/// MARK - Unique consecutive layer.
4786
4787typedef struct {
4788 ccv_cnnp_model_t super;
4789 ccv_nnc_tensor_symbol_t outputs[2];
4790 int bincount;
4791} ccv_cnnp_model_unique_consecutive_t;
4792
4793static 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)
4794{
4795 ccv_cnnp_model_unique_consecutive_t* const self = (ccv_cnnp_model_unique_consecutive_t*)super;
4796 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);
4797 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4798 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", 4798, __extension__ __PRETTY_FUNCTION__
); }));
4799 if (self->bincount > 0)
4800 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; }
);
4801 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4802 params.datatype = CCV_32S;
4803 outputs[1] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4804 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");
4805}
4806
4807static ccv_cnnp_model_t* _ccv_cnnp_unique_consecutive_copy(const ccv_cnnp_model_t* const self, void* const context);
4808
4809static const ccv_cnnp_model_vtab_t ccv_cnnp_unique_consecutive_isa = {
4810 .build = _ccv_cnnp_unique_consecutive_build,
4811 .copy = _ccv_cnnp_unique_consecutive_copy,
4812};
4813
4814ccv_cnnp_model_t* ccv_cnnp_unique_consecutive(const int bincount, const char* const name)
4815{
4816 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));
4817 model_unique_consecutive->super.isa = &ccv_cnnp_unique_consecutive_isa;
4818 model_unique_consecutive->super.input_size = 0;
4819 model_unique_consecutive->super.outputs = model_unique_consecutive->outputs;
4820 model_unique_consecutive->super.output_size = 2;
4821 model_unique_consecutive->bincount = bincount;
4822 ccv_cnnp_model_copy_name(&model_unique_consecutive->super, name);
4823 return (ccv_cnnp_model_t*)model_unique_consecutive;
4824}
4825
4826static ccv_cnnp_model_t* _ccv_cnnp_unique_consecutive_copy(const ccv_cnnp_model_t* const super, void* const context)
4827{
4828 ccv_cnnp_model_unique_consecutive_t* const self = (ccv_cnnp_model_unique_consecutive_t*)super;
4829 return ccv_cnnp_unique_consecutive(self->bincount, self->super.name);
4830}
4831
4832/// MARK - Scatter add layer.
4833
4834typedef struct {
4835 ccv_cnnp_model_t super;
4836 ccv_nnc_tensor_symbol_t output;
4837 int bincount;
4838} ccv_cnnp_model_scatter_add_t;
4839
4840static 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)
4841{
4842 ccv_cnnp_model_scatter_add_t* const self = (ccv_cnnp_model_scatter_add_t*)super;
4843 PRINT(CCV_CLI_VERBOSE, "[cnnp_scatter_add_build] - bincount: %d\n", self->bincount)do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_scatter_add_build] - bincount: %d\n", self->
bincount); fflush(stdout); } } while (0);
4844 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4845 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", 4845, __extension__ __PRETTY_FUNCTION__
); }));
4846 assert(self->bincount > 0)((void) sizeof ((self->bincount > 0) ? 1 : 0), __extension__
({ if (self->bincount > 0) ; else __assert_fail ("self->bincount > 0"
, "ccv_cnnp_model_addons.c", 4846, __extension__ __PRETTY_FUNCTION__
); }));
4847 params.dim[0] = self->bincount;
4848 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4849 ccv_nnc_graph_exec_symbol_new(graph, CMD_SCATTER_ADD_FORWARD(self->bincount)ccv_nnc_cmd(CCV_NNC_SCATTER_ADD_FORWARD, 0, ((ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.scatter_add={.bincount=self->bincount
}}), 0), inputs, input_size, outputs, output_size, "scatter_add");
4850}
4851
4852static ccv_cnnp_model_t* _ccv_cnnp_scatter_add_copy(const ccv_cnnp_model_t* const self, void* const context);
4853
4854static const ccv_cnnp_model_vtab_t ccv_cnnp_scatter_add_isa = {
4855 .build = _ccv_cnnp_scatter_add_build,
4856 .copy = _ccv_cnnp_scatter_add_copy,
4857};
4858
4859ccv_cnnp_model_t* ccv_cnnp_scatter_add(const int bincount, const char* const name)
4860{
4861 assert(bincount > 0)((void) sizeof ((bincount > 0) ? 1 : 0), __extension__ ({ if
(bincount > 0) ; else __assert_fail ("bincount > 0", "ccv_cnnp_model_addons.c"
, 4861, __extension__ __PRETTY_FUNCTION__); }));
4862 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));
4863 model_scatter_add->super.isa = &ccv_cnnp_scatter_add_isa;
4864 model_scatter_add->super.input_size = 0;
4865 model_scatter_add->super.outputs = &model_scatter_add->output;
4866 model_scatter_add->super.output_size = 1;
4867 model_scatter_add->bincount = bincount;
4868 ccv_cnnp_model_copy_name(&model_scatter_add->super, name);
4869 return (ccv_cnnp_model_t*)model_scatter_add;
4870}
4871
4872static ccv_cnnp_model_t* _ccv_cnnp_scatter_add_copy(const ccv_cnnp_model_t* const super, void* const context)
4873{
4874 ccv_cnnp_model_scatter_add_t* const self = (ccv_cnnp_model_scatter_add_t*)super;
4875 return ccv_cnnp_scatter_add(self->bincount, self->super.name);
4876}
4877
4878// MARK - Segmented Dense Layer
4879
4880typedef struct {
4881 ccv_cnnp_model_t super;
4882 ccv_nnc_tensor_symbol_t output;
4883 ccv_nnc_tensor_symbol_t weights;
4884 ccv_nnc_tensor_symbol_t bias;
4885 int segments;
4886 int count;
4887 int no_bias;
4888 int flags;
4889} ccv_cnnp_model_segmented_dense_t;
4890
4891static 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)
4892{
4893 ccv_cnnp_model_segmented_dense_t* const self = (ccv_cnnp_model_segmented_dense_t*)super;
4894 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);
4895 assert(input_size == 3)((void) sizeof ((input_size == 3) ? 1 : 0), __extension__ ({ if
(input_size == 3) ; else __assert_fail ("input_size == 3", "ccv_cnnp_model_addons.c"
, 4895, __extension__ __PRETTY_FUNCTION__); }));
4896 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", 4896, __extension__ __PRETTY_FUNCTION__
); }));
4897 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4898 const ccv_nnc_tensor_param_t indices_params = ccv_nnc_tensor_symbol_params(graph, inputs[1]);
4899 const ccv_nnc_tensor_param_t counts_params = ccv_nnc_tensor_symbol_params(graph, inputs[2]);
4900 ccv_nnc_tensor_param_t weights_params = params;
4901 memset(weights_params.dim, 0, sizeof(weights_params.dim));
4902 weights_params.dim[0] = self->segments;
4903 weights_params.dim[1] = self->count;
4904 weights_params.dim[2] = params.dim[ccv_nnc_tensor_nd(params.dim) - 1];
4905 if (!self->weights.graph)
4906 self->weights = ccv_nnc_tensor_symbol_new(graph, weights_params, "weights");
4907 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"
, 4907, __extension__ __PRETTY_FUNCTION__); }));
4908 const ccv_nnc_tensor_symbol_t weights = ccv_cnnp_model_get_symbol(super, self->weights);
4909 ccv_nnc_tensor_param_t bias_params = params;
4910 memset(bias_params.dim, 0, sizeof(bias_params.dim));
4911 bias_params.dim[0] = self->segments;
4912 bias_params.dim[1] = self->count;
4913 ccv_nnc_cmd_t cmd = {0};
4914 cmd.cmd = CCV_NNC_SEGMENTED_GEMM_FORWARD;
4915 cmd.info.blas.a[0] = 1;
4916 cmd.info.blas.a[1] = 1;
4917 cmd.info.blas.transpose_b[0] = 1;
4918 cmd.info.blas.transpose_b[1] = 2;
4919 cmd.info.blas.flags = self->flags;
4920 ccv_nnc_tensor_param_t output_params;
4921 ccv_nnc_hint_tensor_auto(cmd, (ccv_nnc_tensor_param_t []){
4922 params, indices_params, counts_params,
4923 weights_params,
4924 bias_params,
4925 }, 5, ccv_nnc_no_hint, &output_params, 1);
4926 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
4927 if (self->no_bias)
4928 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");
4929 else {
4930 if (!self->bias.graph)
4931 self->bias = ccv_nnc_tensor_symbol_new(graph, bias_params, "bias");
4932 const ccv_nnc_tensor_symbol_t bias = ccv_cnnp_model_get_symbol(super, self->bias);
4933 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");
4934 }
4935 outputs[0] = output;
4936}
4937
4938static 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)
4939{
4940 ccv_cnnp_model_segmented_dense_t* const self = (ccv_cnnp_model_segmented_dense_t*)super;
4941 const ccv_nnc_tensor_param_t weight_params = ccv_nnc_tensor_symbol_params(graph, self->weights);
4942 const int c = weight_params.dim[1];
4943 const float std = sqrtf(2) / sqrtf(c);
4944 const float bound = sqrtf(3) * std;
4945 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);
4946 if (self->bias.graph)
4947 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);
4948}
4949
4950static 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)
4951{
4952 ccv_cnnp_model_segmented_dense_t* const self = (ccv_cnnp_model_segmented_dense_t*)super;
4953 add_to_array(parameters, self->weights, is_trainable);
4954 if (self->bias.graph)
4955 add_to_array(parameters, self->bias, is_trainable);
4956}
4957
4958static ccv_cnnp_model_t* _ccv_cnnp_segmented_dense_copy(const ccv_cnnp_model_t* const super, void* const context);
4959
4960static const ccv_cnnp_model_vtab_t ccv_cnnp_segmented_dense_isa = {
4961 .build = _ccv_cnnp_segmented_dense_build,
4962 .init_states = _ccv_cnnp_segmented_dense_init_states,
4963 .add_to_parameter = _ccv_cnnp_segmented_dense_add_to_parameter,
4964 .copy = _ccv_cnnp_segmented_dense_copy,
4965};
4966
4967ccv_cnnp_model_t* ccv_cnnp_segmented_dense(const int segments, const int count, const int no_bias, const int flags, const int is_trainable, const char* const name)
4968{
4969 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));
4970 model_segmented_dense->super.isa = &ccv_cnnp_segmented_dense_isa;
4971 model_segmented_dense->super.input_size = 3;
4972 model_segmented_dense->super.outputs = &model_segmented_dense->output;
4973 model_segmented_dense->super.output_size = 1;
4974 model_segmented_dense->super.is_trainable = is_trainable;
4975 ccv_cnnp_model_copy_name(&model_segmented_dense->super, name);
4976 model_segmented_dense->weights.d = CCV_NNC_NO_TENSOR_SYMBOL;
4977 model_segmented_dense->weights.graph = 0;
4978 model_segmented_dense->bias.d = CCV_NNC_NO_TENSOR_SYMBOL;
4979 model_segmented_dense->bias.graph = 0;
4980 model_segmented_dense->segments = segments;
4981 model_segmented_dense->count = count;
4982 model_segmented_dense->no_bias = no_bias;
4983 model_segmented_dense->flags = flags;
4984 return (ccv_cnnp_model_t*)model_segmented_dense;
4985}
4986
4987static ccv_cnnp_model_t* _ccv_cnnp_segmented_dense_copy(const ccv_cnnp_model_t* const super, void* const context)
4988{
4989 const ccv_cnnp_model_segmented_dense_t* const self = (const ccv_cnnp_model_segmented_dense_t*)super;
4990 return ccv_cnnp_segmented_dense(self->segments, self->count, self->no_bias, self->flags, self->super.is_trainable, self->super.name);
4991}