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/2026-04-29-094735-1282740-1 -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 - Swish Layer
1740
1741typedef struct {
1742 ccv_cnnp_model_t super;
1743 ccv_nnc_tensor_symbol_t output;
1744 float beta;
1745} ccv_cnnp_model_swish_t;
1746
1747static 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)
1748{
1749 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);
1750 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"
, 1750, __extension__ __PRETTY_FUNCTION__); }));
1751 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", 1751, __extension__ __PRETTY_FUNCTION__
); }));
1752 ccv_cnnp_model_swish_t* const self = (ccv_cnnp_model_swish_t*)super;
1753 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1754 ccv_nnc_tensor_param_t output_params;
1755 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);
1756 ccv_nnc_hint_tensor_auto(swish, (ccv_nnc_tensor_param_t []){
1757 params,
1758 }, 1, ccv_nnc_no_hint, &output_params, 1);
1759 const ccv_nnc_tensor_symbol_t swish_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1760 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");
1761 outputs[0] = swish_output;
1762}
1763
1764static ccv_cnnp_model_t* _ccv_cnnp_swish_copy(const ccv_cnnp_model_t* const self, void* const context);
1765
1766static const ccv_cnnp_model_vtab_t ccv_cnnp_swish_isa = {
1767 .build = _ccv_cnnp_swish_build,
1768 .copy = _ccv_cnnp_swish_copy,
1769};
1770
1771ccv_cnnp_model_t* ccv_cnnp_swish(const float beta, const char* const name)
1772{
1773 ccv_cnnp_model_swish_t* const model_swish = (ccv_cnnp_model_swish_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_swish_t));
1774 model_swish->super.isa = &ccv_cnnp_swish_isa;
1775 model_swish->super.input_size = 1;
1776 model_swish->super.outputs = &model_swish->output;
1777 model_swish->super.output_size = 1;
1778 model_swish->beta = beta;
1779 ccv_cnnp_model_copy_name(&model_swish->super, name);
1780 return (ccv_cnnp_model_t*)model_swish;
1781}
1782
1783static ccv_cnnp_model_t* _ccv_cnnp_swish_copy(const ccv_cnnp_model_t* const self, void* const context)
1784{
1785 const ccv_cnnp_model_swish_t* const swish = (const ccv_cnnp_model_swish_t*)self;
1786 return ccv_cnnp_swish(swish->beta, self->name);
1787}
1788
1789// MARK - GELU Layer
1790
1791typedef struct {
1792 ccv_cnnp_model_t super;
1793 ccv_nnc_tensor_symbol_t output;
1794 int tanh;
1795} ccv_cnnp_model_gelu_t;
1796
1797static 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)
1798{
1799 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);
1800 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"
, 1800, __extension__ __PRETTY_FUNCTION__); }));
1801 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", 1801, __extension__ __PRETTY_FUNCTION__
); }));
1802 ccv_cnnp_model_gelu_t* const self = (ccv_cnnp_model_gelu_t*)super;
1803 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1804 ccv_nnc_tensor_param_t output_params;
1805 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);
1806 ccv_nnc_hint_tensor_auto(gelu, (ccv_nnc_tensor_param_t []){
1807 params,
1808 }, 1, ccv_nnc_no_hint, &output_params, 1);
1809 const ccv_nnc_tensor_symbol_t gelu_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1810 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");
1811 outputs[0] = gelu_output;
1812}
1813
1814static ccv_cnnp_model_t* _ccv_cnnp_gelu_copy(const ccv_cnnp_model_t* const self, void* const context);
1815
1816static const ccv_cnnp_model_vtab_t ccv_cnnp_gelu_isa = {
1817 .build = _ccv_cnnp_gelu_build,
1818 .copy = _ccv_cnnp_gelu_copy,
1819};
1820
1821ccv_cnnp_model_t* ccv_cnnp_gelu(const int tanh, const char* const name)
1822{
1823 ccv_cnnp_model_gelu_t* const model_gelu = (ccv_cnnp_model_gelu_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_gelu_t));
1824 model_gelu->super.isa = &ccv_cnnp_gelu_isa;
1825 model_gelu->super.input_size = 1;
1826 model_gelu->super.outputs = &model_gelu->output;
1827 model_gelu->super.output_size = 1;
1828 model_gelu->tanh = tanh;
1829 ccv_cnnp_model_copy_name(&model_gelu->super, name);
1830 return (ccv_cnnp_model_t*)model_gelu;
1831}
1832
1833static ccv_cnnp_model_t* _ccv_cnnp_gelu_copy(const ccv_cnnp_model_t* const super, void* const context)
1834{
1835 ccv_cnnp_model_gelu_t* const self = (ccv_cnnp_model_gelu_t*)super;
1836 return ccv_cnnp_gelu(self->tanh, self->super.name);
1837}
1838
1839// MARK - Leaky ReLU Layer
1840
1841typedef struct {
1842 ccv_cnnp_model_t super;
1843 ccv_nnc_tensor_symbol_t output;
1844 float negative_slope;
1845} ccv_cnnp_model_leaky_relu_t;
1846
1847static 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)
1848{
1849 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);
1850 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"
, 1850, __extension__ __PRETTY_FUNCTION__); }));
1851 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", 1851, __extension__ __PRETTY_FUNCTION__
); }));
1852 ccv_cnnp_model_leaky_relu_t* const self = (ccv_cnnp_model_leaky_relu_t*)super;
1853 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1854 ccv_nnc_tensor_param_t output_params;
1855 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);
1856 ccv_nnc_hint_tensor_auto(leaky_relu, (ccv_nnc_tensor_param_t []){
1857 params,
1858 }, 1, ccv_nnc_no_hint, &output_params, 1);
1859 const ccv_nnc_tensor_symbol_t leaky_relu_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1860 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");
1861 outputs[0] = leaky_relu_output;
1862}
1863
1864static ccv_cnnp_model_t* _ccv_cnnp_leaky_relu_copy(const ccv_cnnp_model_t* const self, void* const context);
1865
1866static const ccv_cnnp_model_vtab_t ccv_cnnp_leaky_relu_isa = {
1867 .build = _ccv_cnnp_leaky_relu_build,
1868 .copy = _ccv_cnnp_leaky_relu_copy,
1869};
1870
1871ccv_cnnp_model_t* ccv_cnnp_leaky_relu(const float negative_slope, const char* const name)
1872{
1873 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));
1874 model_leaky_relu->super.isa = &ccv_cnnp_leaky_relu_isa;
1875 model_leaky_relu->super.input_size = 1;
1876 model_leaky_relu->super.outputs = &model_leaky_relu->output;
1877 model_leaky_relu->super.output_size = 1;
1878 model_leaky_relu->negative_slope = negative_slope;
1879 ccv_cnnp_model_copy_name(&model_leaky_relu->super, name);
1880 return (ccv_cnnp_model_t*)model_leaky_relu;
1881}
1882
1883static ccv_cnnp_model_t* _ccv_cnnp_leaky_relu_copy(const ccv_cnnp_model_t* const super, void* const context)
1884{
1885 ccv_cnnp_model_leaky_relu_t* const self = (ccv_cnnp_model_leaky_relu_t*)super;
1886 return ccv_cnnp_leaky_relu(self->negative_slope, self->super.name);
1887}
1888
1889// MARK - Softmax Layer
1890
1891typedef struct {
1892 ccv_cnnp_model_t super;
1893 ccv_nnc_tensor_symbol_t output;
1894} ccv_cnnp_model_softmax_t;
1895
1896static 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)
1897{
1898 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);
1899 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"
, 1899, __extension__ __PRETTY_FUNCTION__); }));
1900 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", 1900, __extension__ __PRETTY_FUNCTION__
); }));
1901 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1902 ccv_nnc_tensor_param_t output_params;
1903 const ccv_nnc_cmd_t softmax = CMD_SOFTMAX_FORWARD()ccv_nnc_cmd(CCV_NNC_SOFTMAX_FORWARD, 0, ccv_nnc_cmd_auto, 0);
1904 ccv_nnc_hint_tensor_auto(softmax, (ccv_nnc_tensor_param_t []){
1905 params,
1906 }, 1, ccv_nnc_no_hint, &output_params, 1);
1907 const ccv_nnc_tensor_symbol_t softmax_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1908 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");
1909 outputs[0] = softmax_output;
1910}
1911
1912static ccv_cnnp_model_t* _ccv_cnnp_softmax_copy(const ccv_cnnp_model_t* const self, void* const context);
1913
1914static const ccv_cnnp_model_vtab_t ccv_cnnp_softmax_isa = {
1915 .build = _ccv_cnnp_softmax_build,
1916 .copy = _ccv_cnnp_softmax_copy,
1917};
1918
1919ccv_cnnp_model_t* ccv_cnnp_softmax(const char* const name)
1920{
1921 ccv_cnnp_model_softmax_t* const model_softmax = (ccv_cnnp_model_softmax_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_softmax_t));
1922 model_softmax->super.isa = &ccv_cnnp_softmax_isa;
1923 model_softmax->super.input_size = 1;
1924 model_softmax->super.outputs = &model_softmax->output;
1925 model_softmax->super.output_size = 1;
1926 ccv_cnnp_model_copy_name(&model_softmax->super, name);
1927 return (ccv_cnnp_model_t*)model_softmax;
1928}
1929
1930static ccv_cnnp_model_t* _ccv_cnnp_softmax_copy(const ccv_cnnp_model_t* const self, void* const context)
1931{
1932 return ccv_cnnp_softmax(self->name);
1933}
1934
1935// MARK - Add Layer
1936
1937typedef struct {
1938 ccv_cnnp_model_t super;
1939 float p;
1940 float q;
1941 ccv_nnc_tensor_symbol_t output;
1942} ccv_cnnp_model_add_t;
1943
1944static 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)
1945{
1946 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);
1947 const ccv_cnnp_model_add_t* const self = (const ccv_cnnp_model_add_t*)super;
1948 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"
, 1948, __extension__ __PRETTY_FUNCTION__); }));
1949 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", 1949, __extension__ __PRETTY_FUNCTION__
); }));
1950 ccv_nnc_tensor_param_t input_params[2];
1951 int i;
1952 for (i = 0; i < 2; i++)
1953 input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
1954 ccv_nnc_tensor_param_t output_params;
1955 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);
1956 ccv_nnc_hint_tensor_auto(add, input_params, 2, ccv_nnc_no_hint, &output_params, 1);
1957 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1958 ccv_nnc_graph_exec_symbol_new(graph, add, inputs, input_size, outputs, output_size, "add");
1959}
1960
1961static ccv_cnnp_model_t* _ccv_cnnp_add_copy(const ccv_cnnp_model_t* const self, void* const context);
1962
1963static const ccv_cnnp_model_vtab_t ccv_cnnp_add_isa = {
1964 .build = _ccv_cnnp_add_build,
1965 .copy = _ccv_cnnp_add_copy,
1966};
1967
1968ccv_cnnp_model_t* ccv_cnnp_add(const float p, const float q, const char* const name)
1969{
1970 ccv_cnnp_model_add_t* const model_add = (ccv_cnnp_model_add_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_add_t));
1971 model_add->super.isa = &ccv_cnnp_add_isa;
1972 model_add->super.input_size = 2;
1973 model_add->super.outputs = &model_add->output;
1974 model_add->super.output_size = 1;
1975 model_add->p = p;
1976 model_add->q = q;
1977 ccv_cnnp_model_copy_name(&model_add->super, name);
1978 return (ccv_cnnp_model_t*)model_add;
1979}
1980
1981static ccv_cnnp_model_t* _ccv_cnnp_add_copy(const ccv_cnnp_model_t* const super, void* const context)
1982{
1983 const ccv_cnnp_model_add_t* const self = (const ccv_cnnp_model_add_t*)super;
1984 return ccv_cnnp_add(self->p, self->q, self->super.name);
1985}
1986
1987// MARK - Mul Layer
1988
1989typedef struct {
1990 ccv_cnnp_model_t super;
1991 ccv_nnc_tensor_symbol_t output;
1992 float p;
1993} ccv_cnnp_model_mul_t;
1994
1995static 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)
1996{
1997 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);
1998 const ccv_cnnp_model_mul_t* const self = (const ccv_cnnp_model_mul_t*)super;
1999 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"
, 1999, __extension__ __PRETTY_FUNCTION__); }));
2000 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", 2000, __extension__ __PRETTY_FUNCTION__
); }));
2001 ccv_nnc_tensor_param_t input_params[2];
2002 int i;
2003 for (i = 0; i < 2; i++)
2004 input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
2005 ccv_nnc_tensor_param_t output_params;
2006 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);
2007 ccv_nnc_hint_tensor_auto(mul, input_params, 2, ccv_nnc_no_hint, &output_params, 1);
2008 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2009 ccv_nnc_graph_exec_symbol_new(graph, mul, inputs, input_size, outputs, output_size, "mul");
2010}
2011
2012static ccv_cnnp_model_t* _ccv_cnnp_mul_copy(const ccv_cnnp_model_t* const self, void* const context);
2013
2014static const ccv_cnnp_model_vtab_t ccv_cnnp_mul_isa = {
2015 .build = _ccv_cnnp_mul_build,
2016 .copy = _ccv_cnnp_mul_copy,
2017};
2018
2019ccv_cnnp_model_t* ccv_cnnp_mul(const float p, const char* const name)
2020{
2021 ccv_cnnp_model_mul_t* const model_mul = (ccv_cnnp_model_mul_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_mul_t));
2022 model_mul->super.isa = &ccv_cnnp_mul_isa;
2023 model_mul->super.input_size = 2;
2024 model_mul->super.outputs = &model_mul->output;
2025 model_mul->super.output_size = 1;
2026 model_mul->p = p;
2027 ccv_cnnp_model_copy_name(&model_mul->super, name);
2028 return (ccv_cnnp_model_t*)model_mul;
2029}
2030
2031static ccv_cnnp_model_t* _ccv_cnnp_mul_copy(const ccv_cnnp_model_t* const super, void* const context)
2032{
2033 const ccv_cnnp_model_mul_t* const self = (const ccv_cnnp_model_mul_t*)super;
2034 return ccv_cnnp_mul(self->p, self->super.name);
2035}
2036
2037// MARK - Scalar Mul Layer
2038
2039typedef struct {
2040 ccv_cnnp_model_t super;
2041 ccv_nnc_tensor_symbol_t output;
2042 float a;
2043} ccv_cnnp_model_scalar_mul_t;
2044
2045static 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)
2046{
2047 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);
2048 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"
, 2048, __extension__ __PRETTY_FUNCTION__); }));
2049 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", 2049, __extension__ __PRETTY_FUNCTION__
); }));
2050 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2051 ccv_nnc_tensor_param_t output_params;
2052 ccv_cnnp_model_scalar_mul_t* const self = (ccv_cnnp_model_scalar_mul_t*)super;
2053 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);
2054 ccv_nnc_hint_tensor_auto(scalar_mul, (ccv_nnc_tensor_param_t []){
2055 params,
2056 }, 1, ccv_nnc_no_hint, &output_params, 1);
2057 const ccv_nnc_tensor_symbol_t scalar_mul_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2058 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");
2059 outputs[0] = scalar_mul_output;
2060}
2061
2062static ccv_cnnp_model_t* _ccv_cnnp_scalar_mul_copy(const ccv_cnnp_model_t* const super, void* const context);
2063
2064static const ccv_cnnp_model_vtab_t ccv_cnnp_scalar_mul_isa = {
2065 .build = _ccv_cnnp_scalar_mul_build,
2066 .copy = _ccv_cnnp_scalar_mul_copy,
2067};
2068
2069ccv_cnnp_model_t* ccv_cnnp_scalar_mul(const float a, const char* const name)
2070{
2071 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));
2072 model_scalar_mul->super.isa = &ccv_cnnp_scalar_mul_isa;
2073 model_scalar_mul->super.input_size = 1;
2074 model_scalar_mul->super.outputs = &model_scalar_mul->output;
2075 model_scalar_mul->super.output_size = 1;
2076 model_scalar_mul->a = a;
2077 ccv_cnnp_model_copy_name(&model_scalar_mul->super, name);
2078 return (ccv_cnnp_model_t*)model_scalar_mul;
2079}
2080
2081static ccv_cnnp_model_t* _ccv_cnnp_scalar_mul_copy(const ccv_cnnp_model_t* const super, void* const context)
2082{
2083 const ccv_cnnp_model_scalar_mul_t* const self = (const ccv_cnnp_model_scalar_mul_t*)super;
2084 return ccv_cnnp_scalar_mul(self->a, self->super.name);
2085}
2086
2087// MARK - Div Layer
2088
2089typedef struct {
2090 ccv_cnnp_model_t super;
2091 ccv_nnc_tensor_symbol_t output;
2092 int reciprocal;
2093} ccv_cnnp_model_div_t;
2094
2095static 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)
2096{
2097 const ccv_cnnp_model_div_t* const self = (const ccv_cnnp_model_div_t*)super;
2098 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);
2099 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", 2099, __extension__ __PRETTY_FUNCTION__
); }));
2100 ccv_nnc_tensor_param_t input_params[2];
2101 int i;
2102 ccv_nnc_tensor_param_t output_params;
2103 const ccv_nnc_cmd_t div = CMD_EWDIV_FORWARD()ccv_nnc_cmd(CCV_NNC_EWDIV_FORWARD, 0, ccv_nnc_cmd_auto, 0);
2104 if (self->reciprocal)
2105 {
2106 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"
, 2106, __extension__ __PRETTY_FUNCTION__); }));
2107 input_params[0] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2108 input_params[1] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2109 ccv_nnc_hint_tensor_auto(div, input_params, 2, ccv_nnc_no_hint, &output_params, 1);
2110 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2111 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");
2112 } else {
2113 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"
, 2113, __extension__ __PRETTY_FUNCTION__); }));
2114 for (i = 0; i < 2; i++)
2115 input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
2116 ccv_nnc_hint_tensor_auto(div, input_params, input_size, ccv_nnc_no_hint, &output_params, 1);
2117 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2118 ccv_nnc_graph_exec_symbol_new(graph, div, inputs, input_size, outputs, output_size, "div");
2119 }
2120}
2121
2122static ccv_cnnp_model_t* _ccv_cnnp_div_copy(const ccv_cnnp_model_t* const self, void* const context);
2123
2124static const ccv_cnnp_model_vtab_t ccv_cnnp_div_isa = {
2125 .build = _ccv_cnnp_div_build,
2126 .copy = _ccv_cnnp_div_copy,
2127};
2128
2129ccv_cnnp_model_t* ccv_cnnp_div(const int reciprocal, const char* const name)
2130{
2131 ccv_cnnp_model_div_t* const model_div = (ccv_cnnp_model_div_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_div_t));
2132 model_div->super.isa = &ccv_cnnp_div_isa;
2133 model_div->super.input_size = reciprocal ? 1 : 2;
2134 model_div->super.outputs = &model_div->output;
2135 model_div->super.output_size = 1;
2136 model_div->reciprocal = reciprocal;
2137 ccv_cnnp_model_copy_name(&model_div->super, name);
2138 return (ccv_cnnp_model_t*)model_div;
2139}
2140
2141static ccv_cnnp_model_t* _ccv_cnnp_div_copy(const ccv_cnnp_model_t* const super, void* const context)
2142{
2143 const ccv_cnnp_model_div_t* const self = (const ccv_cnnp_model_div_t*)super;
2144 return ccv_cnnp_div(self->reciprocal, self->super.name);
2145}
2146
2147// MARK - Sqrt Layer
2148
2149typedef struct {
2150 ccv_cnnp_model_t super;
2151 ccv_nnc_tensor_symbol_t output;
2152} ccv_cnnp_model_sqrt_t;
2153
2154static 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)
2155{
2156 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);
2157 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", 2157, __extension__ __PRETTY_FUNCTION__
); }));
2158 ccv_nnc_tensor_param_t input_params[1];
2159 ccv_nnc_tensor_param_t output_params;
2160 const ccv_nnc_cmd_t sqrt = CMD_EWSQRT_FORWARD()ccv_nnc_cmd(CCV_NNC_EWSQRT_FORWARD, 0, ccv_nnc_cmd_auto, 0);
2161 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"
, 2161, __extension__ __PRETTY_FUNCTION__); }));
2162 input_params[0] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2163 ccv_nnc_hint_tensor_auto(sqrt, input_params, 1, ccv_nnc_no_hint, &output_params, 1);
2164 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2165 ccv_nnc_graph_exec_symbol_new(graph, sqrt, inputs, 1, outputs, output_size, "sqrt");
2166}
2167
2168static ccv_cnnp_model_t* _ccv_cnnp_sqrt_copy(const ccv_cnnp_model_t* const self, void* const context);
2169
2170static const ccv_cnnp_model_vtab_t ccv_cnnp_sqrt_isa = {
2171 .build = _ccv_cnnp_sqrt_build,
2172 .copy = _ccv_cnnp_sqrt_copy,
2173};
2174
2175ccv_cnnp_model_t* ccv_cnnp_sqrt(const char* const name)
2176{
2177 ccv_cnnp_model_sqrt_t* const model_sqrt = (ccv_cnnp_model_sqrt_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_sqrt_t));
2178 model_sqrt->super.isa = &ccv_cnnp_sqrt_isa;
2179 model_sqrt->super.input_size = 1;
2180 model_sqrt->super.outputs = &model_sqrt->output;
2181 model_sqrt->super.output_size = 1;
2182 ccv_cnnp_model_copy_name(&model_sqrt->super, name);
2183 return (ccv_cnnp_model_t*)model_sqrt;
2184}
2185
2186static ccv_cnnp_model_t* _ccv_cnnp_sqrt_copy(const ccv_cnnp_model_t* const super, void* const context)
2187{
2188 const ccv_cnnp_model_sqrt_t* const self = (const ccv_cnnp_model_sqrt_t*)super;
2189 return ccv_cnnp_sqrt(self->super.name);
2190}
2191
2192// MARK - Log Layer
2193
2194typedef struct {
2195 ccv_cnnp_model_t super;
2196 ccv_nnc_tensor_symbol_t output;
2197} ccv_cnnp_model_log_t;
2198
2199static 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)
2200{
2201 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);
2202 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", 2202, __extension__ __PRETTY_FUNCTION__
); }));
2203 ccv_nnc_tensor_param_t input_params[1];
2204 ccv_nnc_tensor_param_t output_params;
2205 const ccv_nnc_cmd_t log = CMD_EWLOG_FORWARD()ccv_nnc_cmd(CCV_NNC_EWLOG_FORWARD, 0, ccv_nnc_cmd_auto, 0);
2206 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"
, 2206, __extension__ __PRETTY_FUNCTION__); }));
2207 input_params[0] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2208 ccv_nnc_hint_tensor_auto(log, input_params, 1, 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, log, inputs, 1, outputs, output_size, "log");
2211}
2212
2213static ccv_cnnp_model_t* _ccv_cnnp_log_copy(const ccv_cnnp_model_t* const self, void* const context);
2214
2215static const ccv_cnnp_model_vtab_t ccv_cnnp_log_isa = {
2216 .build = _ccv_cnnp_log_build,
2217 .copy = _ccv_cnnp_log_copy,
2218};
2219
2220ccv_cnnp_model_t* ccv_cnnp_log(const char* const name)
2221{
2222 ccv_cnnp_model_log_t* const model_log = (ccv_cnnp_model_log_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_log_t));
2223 model_log->super.isa = &ccv_cnnp_log_isa;
2224 model_log->super.input_size = 1;
2225 model_log->super.outputs = &model_log->output;
2226 model_log->super.output_size = 1;
2227 ccv_cnnp_model_copy_name(&model_log->super, name);
2228 return (ccv_cnnp_model_t*)model_log;
2229}
2230
2231static ccv_cnnp_model_t* _ccv_cnnp_log_copy(const ccv_cnnp_model_t* const super, void* const context)
2232{
2233 return ccv_cnnp_log(super->name);
2234}
2235
2236// MARK - Pow Layer
2237
2238typedef struct {
2239 ccv_cnnp_model_t super;
2240 ccv_nnc_tensor_symbol_t output;
2241 ccv_nnc_cmd_param_t params;
2242} ccv_cnnp_model_pow_t;
2243
2244static 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)
2245{
2246 ccv_cnnp_model_pow_t* const self = (ccv_cnnp_model_pow_t*)super;
2247 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);
2248 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"
, 2248, __extension__ __PRETTY_FUNCTION__); }));
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 input_params[0] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2252 ccv_nnc_tensor_param_t output_params;
2253 const ccv_nnc_cmd_t pow = ccv_nnc_cmd(CCV_NNC_EWPOW_FORWARD, 0, self->params, 0);
2254 ccv_nnc_hint_tensor_auto(pow, input_params, 1, ccv_nnc_no_hint, &output_params, 1);
2255 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2256 ccv_nnc_graph_exec_symbol_new(graph, pow, inputs, input_size, outputs, output_size, "pow");
2257}
2258
2259static ccv_cnnp_model_t* _ccv_cnnp_pow_copy(const ccv_cnnp_model_t* const self, void* const context);
2260
2261static const ccv_cnnp_model_vtab_t ccv_cnnp_pow_isa = {
2262 .build = _ccv_cnnp_pow_build,
2263 .copy = _ccv_cnnp_pow_copy,
2264};
2265
2266ccv_cnnp_model_t* ccv_cnnp_pow(const float exponent, const char* const name)
2267{
2268 ccv_cnnp_model_pow_t* const model_pow = (ccv_cnnp_model_pow_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_pow_t));
2269 model_pow->super.isa = &ccv_cnnp_pow_isa;
2270 model_pow->super.input_size = 1;
2271 model_pow->super.outputs = &model_pow->output;
2272 model_pow->super.output_size = 1;
2273 model_pow->params = (ccv_nnc_cmd_param_t){
2274 .size = {
2275 .dim = { 1, 1, 1 }
2276 },
2277 .pow = {
2278 .exponent = exponent,
2279 },
2280 };
2281 ccv_cnnp_model_copy_name(&model_pow->super, name);
2282 return (ccv_cnnp_model_t*)model_pow;
2283}
2284
2285static ccv_cnnp_model_t* _ccv_cnnp_pow_copy(const ccv_cnnp_model_t* const super, void* const context)
2286{
2287 const ccv_cnnp_model_pow_t* const self = (const ccv_cnnp_model_pow_t*)super;
2288 return ccv_cnnp_pow(self->params.pow.exponent, super->name);
2289}
2290
2291// MARK - Sin Layer
2292
2293typedef struct {
2294 ccv_cnnp_model_t super;
2295 ccv_nnc_tensor_symbol_t output;
2296} ccv_cnnp_model_sin_t;
2297
2298static 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)
2299{
2300 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);
2301 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", 2301, __extension__ __PRETTY_FUNCTION__
); }));
2302 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"
, 2302, __extension__ __PRETTY_FUNCTION__); }));
2303 ccv_nnc_tensor_param_t input_params[1];
2304 ccv_nnc_tensor_param_t output_params;
2305 const ccv_nnc_cmd_t sin = CMD_EWSIN_FORWARD()ccv_nnc_cmd(CCV_NNC_EWSIN_FORWARD, 0, ccv_nnc_cmd_auto, 0);
2306 input_params[0] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2307 ccv_nnc_hint_tensor_auto(sin, input_params, 1, ccv_nnc_no_hint, &output_params, 1);
2308 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2309 ccv_nnc_graph_exec_symbol_new(graph, sin, inputs, 1, outputs, output_size, "sin");
2310}
2311
2312static ccv_cnnp_model_t* _ccv_cnnp_sin_copy(const ccv_cnnp_model_t* const self, void* const context);
2313
2314static const ccv_cnnp_model_vtab_t ccv_cnnp_sin_isa = {
2315 .build = _ccv_cnnp_sin_build,
2316 .copy = _ccv_cnnp_sin_copy,
2317};
2318
2319ccv_cnnp_model_t* ccv_cnnp_sin(const char* const name)
2320{
2321 ccv_cnnp_model_sin_t* const model_sin = (ccv_cnnp_model_sin_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_sin_t));
2322 model_sin->super.isa = &ccv_cnnp_sin_isa;
2323 model_sin->super.input_size = 1;
2324 model_sin->super.outputs = &model_sin->output;
2325 model_sin->super.output_size = 1;
2326 ccv_cnnp_model_copy_name(&model_sin->super, name);
2327 return (ccv_cnnp_model_t*)model_sin;
2328}
2329
2330static ccv_cnnp_model_t* _ccv_cnnp_sin_copy(const ccv_cnnp_model_t* const super, void* const context)
2331{
2332 return ccv_cnnp_sin(super->name);
2333}
2334
2335// MARK - Cos Layer
2336
2337typedef struct {
2338 ccv_cnnp_model_t super;
2339 ccv_nnc_tensor_symbol_t output;
2340} ccv_cnnp_model_cos_t;
2341
2342static 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)
2343{
2344 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);
2345 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", 2345, __extension__ __PRETTY_FUNCTION__
); }));
2346 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"
, 2346, __extension__ __PRETTY_FUNCTION__); }));
2347 ccv_nnc_tensor_param_t input_params[1];
2348 ccv_nnc_tensor_param_t output_params;
2349 const ccv_nnc_cmd_t cos = CMD_EWCOS_FORWARD()ccv_nnc_cmd(CCV_NNC_EWCOS_FORWARD, 0, ccv_nnc_cmd_auto, 0);
2350 input_params[0] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2351 ccv_nnc_hint_tensor_auto(cos, input_params, 1, ccv_nnc_no_hint, &output_params, 1);
2352 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2353 ccv_nnc_graph_exec_symbol_new(graph, cos, inputs, 1, outputs, output_size, "cos");
2354}
2355
2356static ccv_cnnp_model_t* _ccv_cnnp_cos_copy(const ccv_cnnp_model_t* const self, void* const context);
2357
2358static const ccv_cnnp_model_vtab_t ccv_cnnp_cos_isa = {
2359 .build = _ccv_cnnp_cos_build,
2360 .copy = _ccv_cnnp_cos_copy,
2361};
2362
2363ccv_cnnp_model_t* ccv_cnnp_cos(const char* const name)
2364{
2365 ccv_cnnp_model_cos_t* const model_cos = (ccv_cnnp_model_cos_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_cos_t));
2366 model_cos->super.isa = &ccv_cnnp_cos_isa;
2367 model_cos->super.input_size = 1;
2368 model_cos->super.outputs = &model_cos->output;
2369 model_cos->super.output_size = 1;
2370 ccv_cnnp_model_copy_name(&model_cos->super, name);
2371 return (ccv_cnnp_model_t*)model_cos;
2372}
2373
2374static ccv_cnnp_model_t* _ccv_cnnp_cos_copy(const ccv_cnnp_model_t* const super, void* const context)
2375{
2376 return ccv_cnnp_cos(super->name);
2377}
2378
2379// MARK - Rotate Half Layer
2380
2381typedef struct {
2382 ccv_cnnp_model_t super;
2383 ccv_nnc_tensor_symbol_t output;
2384} ccv_cnnp_model_rotate_half_t;
2385
2386static 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)
2387{
2388 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);
2389 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"
, 2389, __extension__ __PRETTY_FUNCTION__); }));
2390 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", 2390, __extension__ __PRETTY_FUNCTION__
); }));
2391 ccv_nnc_tensor_param_t input_params[1];
2392 ccv_nnc_tensor_param_t output_params;
2393 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);
2394 input_params[0] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2395 ccv_nnc_hint_tensor_auto(rotate_half, input_params, 1, ccv_nnc_no_hint, &output_params, 1);
2396 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2397 ccv_nnc_graph_exec_symbol_new(graph, rotate_half, inputs, 1, outputs, output_size, "rotate_half");
2398}
2399
2400static ccv_cnnp_model_t* _ccv_cnnp_rotate_half_copy(const ccv_cnnp_model_t* const self, void* const context);
2401
2402static const ccv_cnnp_model_vtab_t ccv_cnnp_rotate_half_isa = {
2403 .build = _ccv_cnnp_rotate_half_build,
2404 .copy = _ccv_cnnp_rotate_half_copy,
2405};
2406
2407ccv_cnnp_model_t* ccv_cnnp_rotate_half(const char* const name)
2408{
2409 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));
2410 model_rotate_half->super.isa = &ccv_cnnp_rotate_half_isa;
2411 model_rotate_half->super.input_size = 1;
2412 model_rotate_half->super.outputs = &model_rotate_half->output;
2413 model_rotate_half->super.output_size = 1;
2414 ccv_cnnp_model_copy_name(&model_rotate_half->super, name);
2415 return (ccv_cnnp_model_t*)model_rotate_half;
2416}
2417
2418static ccv_cnnp_model_t* _ccv_cnnp_rotate_half_copy(const ccv_cnnp_model_t* const super, void* const context)
2419{
2420 return ccv_cnnp_rotate_half(super->name);
2421}
2422
2423// MARK - Cmul Layer
2424
2425typedef struct {
2426 ccv_cnnp_model_t super;
2427 ccv_nnc_tensor_symbol_t output;
2428} ccv_cnnp_model_cmul_t;
2429
2430static 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)
2431{
2432 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);
2433 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"
, 2433, __extension__ __PRETTY_FUNCTION__); }));
2434 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", 2434, __extension__ __PRETTY_FUNCTION__
); }));
2435 ccv_nnc_tensor_param_t input_params[2];
2436 int i;
2437 for (i = 0; i < 2; i++)
2438 input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
2439 ccv_nnc_tensor_param_t output_params;
2440 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);
2441 ccv_nnc_hint_tensor_auto(mul, input_params, 2, ccv_nnc_no_hint, &output_params, 1);
2442 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2443 ccv_nnc_graph_exec_symbol_new(graph, mul, inputs, input_size, outputs, output_size, "cmul");
2444}
2445
2446static ccv_cnnp_model_t* _ccv_cnnp_cmul_copy(const ccv_cnnp_model_t* const self, void* const context);
2447
2448static const ccv_cnnp_model_vtab_t ccv_cnnp_cmul_isa = {
2449 .build = _ccv_cnnp_cmul_build,
2450 .copy = _ccv_cnnp_cmul_copy,
2451};
2452
2453ccv_cnnp_model_t* ccv_cnnp_cmul(const char* const name)
2454{
2455 ccv_cnnp_model_cmul_t* const model_cmul = (ccv_cnnp_model_cmul_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_cmul_t));
2456 model_cmul->super.isa = &ccv_cnnp_cmul_isa;
2457 model_cmul->super.input_size = 2;
2458 model_cmul->super.outputs = &model_cmul->output;
2459 model_cmul->super.output_size = 1;
2460 ccv_cnnp_model_copy_name(&model_cmul->super, name);
2461 return (ccv_cnnp_model_t*)model_cmul;
2462}
2463
2464static ccv_cnnp_model_t* _ccv_cnnp_cmul_copy(const ccv_cnnp_model_t* const super, void* const context)
2465{
2466 return ccv_cnnp_cmul(super->name);
2467}
2468
2469// MARK - Transpose Layer
2470
2471typedef struct {
2472 ccv_cnnp_model_t super;
2473 ccv_nnc_tensor_symbol_t output;
2474 int transpose[2];
2475} ccv_cnnp_model_transpose_t;
2476
2477static 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)
2478{
2479 ccv_cnnp_model_transpose_t* const self = (ccv_cnnp_model_transpose_t*)super;
2480 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);
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 if (self->transpose[0] == self->transpose[1])
2484 {
2485 outputs[0] = inputs[0];
2486 return;
2487 }
2488 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2489 ccv_nnc_tensor_param_t output_params;
2490 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);
2491 ccv_nnc_hint_tensor_auto(transpose, (ccv_nnc_tensor_param_t []){
2492 params,
2493 }, 1, ccv_nnc_no_hint, &output_params, 1);
2494 const ccv_nnc_tensor_symbol_t transpose_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2495 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");
2496 outputs[0] = transpose_output;
2497}
2498
2499static ccv_cnnp_model_t* _ccv_cnnp_transpose_copy(const ccv_cnnp_model_t* const super, void* const context);
2500
2501static const ccv_cnnp_model_vtab_t ccv_cnnp_transpose_isa = {
2502 .build = _ccv_cnnp_transpose_build,
2503 .copy = _ccv_cnnp_transpose_copy,
2504};
2505
2506ccv_cnnp_model_t* ccv_cnnp_transpose(const int axis_a, const int axis_b, const char* const name)
2507{
2508 ccv_cnnp_model_transpose_t* const model_transpose = (ccv_cnnp_model_transpose_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_transpose_t));
2509 model_transpose->super.isa = &ccv_cnnp_transpose_isa;
2510 model_transpose->super.input_size = 1;
2511 model_transpose->super.outputs = &model_transpose->output;
2512 model_transpose->super.output_size = 1;
2513 model_transpose->transpose[0] = axis_a;
2514 model_transpose->transpose[1] = axis_b;
2515 ccv_cnnp_model_copy_name(&model_transpose->super, name);
2516 return (ccv_cnnp_model_t*)model_transpose;
2517}
2518
2519static ccv_cnnp_model_t* _ccv_cnnp_transpose_copy(const ccv_cnnp_model_t* const super, void* const context)
2520{
2521 const ccv_cnnp_model_transpose_t* const self = (const ccv_cnnp_model_transpose_t*)super;
2522 return ccv_cnnp_transpose(self->transpose[0], self->transpose[1], self->super.name);
2523}
2524
2525// MARK - Layer Norm Layer
2526
2527typedef struct {
2528 ccv_cnnp_model_t super;
2529 ccv_nnc_tensor_symbol_t output;
2530 ccv_nnc_tensor_symbol_t bias;
2531 ccv_nnc_tensor_symbol_t scale;
2532 ccv_nnc_cmd_param_t params;
2533} ccv_cnnp_model_layer_norm_t;
2534
2535static 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)
2536{
2537 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);
2538 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"
, 2538, __extension__ __PRETTY_FUNCTION__); }));
2539 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", 2539, __extension__ __PRETTY_FUNCTION__
); }));
2540 ccv_cnnp_model_layer_norm_t* const self = (ccv_cnnp_model_layer_norm_t*)super;
2541 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2542 ccv_nnc_tensor_param_t bias_params = params;
2543 const int nd = ccv_nnc_tensor_nd(params.dim);
2544 int i;
2545 for (i = 0; i < nd; i++)
2546 bias_params.dim[i] = 1;
2547 for (i = 0; i < self->params.lnorm.count; i++)
2548 bias_params.dim[self->params.lnorm.axis[i]] = params.dim[self->params.lnorm.axis[i]];
2549 if (self->params.lnorm.elementwise_affine)
2550 {
2551 // Both scale and bias are shared between if this model is reused.
2552 if (!self->scale.graph)
2553 self->scale = ccv_nnc_tensor_symbol_new(graph, bias_params, "scale");
2554 if (!self->bias.graph)
2555 self->bias = ccv_nnc_tensor_symbol_new(graph, bias_params, "bias");
2556 }
2557 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
};
2558 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
};
2559 const ccv_nnc_cmd_t layer_norm = ccv_nnc_cmd(CCV_NNC_LAYER_NORM_FORWARD, 0, self->params, 0);
2560 ccv_nnc_tensor_param_t output_params[3];
2561 if (self->params.lnorm.elementwise_affine)
2562 ccv_nnc_hint_tensor_auto(layer_norm, (ccv_nnc_tensor_param_t []){
2563 params,
2564 bias_params,
2565 bias_params,
2566 }, 3, ccv_nnc_no_hint, output_params, 3);
2567 else
2568 ccv_nnc_hint_tensor_auto(layer_norm, (ccv_nnc_tensor_param_t []){
2569 params,
2570 }, 1, ccv_nnc_no_hint, output_params, 3);
2571 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0);
2572 const ccv_nnc_tensor_symbol_t saved_mean = ccv_nnc_tensor_symbol_new(graph, output_params[1], "saved_mean");
2573 const ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(graph, output_params[2], "saved_inv_std");
2574 if (self->params.lnorm.elementwise_affine)
2575 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");
2576 else
2577 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");
2578 outputs[0] = output;
2579}
2580
2581static 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)
2582{
2583 ccv_cnnp_model_layer_norm_t* const self = (ccv_cnnp_model_layer_norm_t*)super;
2584 if (self->scale.graph)
2585 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);
2586 if (self->bias.graph)
2587 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);
2588}
2589
2590static 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)
2591{
2592 ccv_cnnp_model_layer_norm_t* const self = (ccv_cnnp_model_layer_norm_t*)super;
2593 if (self->scale.graph)
2594 add_to_array(parameters, self->scale, is_trainable);
2595 if (self->bias.graph)
2596 add_to_array(parameters, self->bias, is_trainable);
2597}
2598
2599static ccv_cnnp_model_t* _ccv_cnnp_layer_norm_copy(const ccv_cnnp_model_t* const super, void* const context);
2600
2601static const ccv_cnnp_model_vtab_t ccv_cnnp_layer_norm_isa = {
2602 .build = _ccv_cnnp_layer_norm_build,
2603 .init_states = _ccv_cnnp_layer_norm_init_states,
2604 .add_to_parameter = _ccv_cnnp_layer_norm_add_to_parameter,
2605 .copy = _ccv_cnnp_layer_norm_copy,
2606};
2607
2608ccv_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)
2609{
2610 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));
2611 model_layer_norm->super.isa = &ccv_cnnp_layer_norm_isa;
2612 model_layer_norm->super.input_size = 1;
2613 model_layer_norm->super.outputs = &model_layer_norm->output;
2614 model_layer_norm->super.output_size = 1;
2615 model_layer_norm->super.is_trainable = is_trainable;
2616 ccv_cnnp_model_copy_name(&model_layer_norm->super, name);
2617 model_layer_norm->scale.d = CCV_NNC_NO_TENSOR_SYMBOL;
2618 model_layer_norm->scale.graph = 0;
2619 model_layer_norm->bias.d = CCV_NNC_NO_TENSOR_SYMBOL;
2620 model_layer_norm->bias.graph = 0;
2621 model_layer_norm->params.lnorm.epsilon = epsilon;
2622 model_layer_norm->params.lnorm.count = axis_count;
2623 model_layer_norm->params.lnorm.elementwise_affine = elementwise_affine;
2624 memcpy(model_layer_norm->params.lnorm.axis, axis, sizeof(int) * axis_count);
2625 return (ccv_cnnp_model_t*)model_layer_norm;
2626}
2627
2628static ccv_cnnp_model_t* _ccv_cnnp_layer_norm_copy(const ccv_cnnp_model_t* const super, void* const context)
2629{
2630 const ccv_cnnp_model_layer_norm_t* const self = (const ccv_cnnp_model_layer_norm_t*)super;
2631 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);
2632}
2633
2634// MARK - Group Norm Layer
2635
2636typedef struct {
2637 ccv_cnnp_model_t super;
2638 ccv_nnc_tensor_symbol_t output;
2639 ccv_nnc_tensor_symbol_t bias;
2640 ccv_nnc_tensor_symbol_t scale;
2641 ccv_nnc_cmd_param_t params;
2642} ccv_cnnp_model_group_norm_t;
2643
2644static 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)
2645{
2646 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);
2647 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"
, 2647, __extension__ __PRETTY_FUNCTION__); }));
2648 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", 2648, __extension__ __PRETTY_FUNCTION__
); }));
2649 ccv_cnnp_model_group_norm_t* const self = (ccv_cnnp_model_group_norm_t*)super;
2650 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2651 ccv_nnc_tensor_param_t bias_params = params;
2652 const int nd = ccv_nnc_tensor_nd(params.dim);
2653 int i;
2654 for (i = 0; i < nd; i++)
2655 bias_params.dim[i] = 1;
2656 bias_params.dim[self->params.gnorm.group_axis] = params.dim[self->params.gnorm.group_axis];
2657 if (self->params.gnorm.elementwise_affine)
2658 {
2659 // Both scale and bias are shared between if this model is reused.
2660 if (!self->scale.graph)
2661 self->scale = ccv_nnc_tensor_symbol_new(graph, bias_params, "scale");
2662 if (!self->bias.graph)
2663 self->bias = ccv_nnc_tensor_symbol_new(graph, bias_params, "bias");
2664 }
2665 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
};
2666 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
};
2667 const ccv_nnc_cmd_t group_norm = ccv_nnc_cmd(CCV_NNC_GROUP_NORM_FORWARD, 0, self->params, 0);
2668 ccv_nnc_tensor_param_t output_params[3];
2669 if (self->params.gnorm.elementwise_affine)
2670 ccv_nnc_hint_tensor_auto(group_norm, (ccv_nnc_tensor_param_t []){
2671 params,
2672 bias_params,
2673 bias_params,
2674 }, 3, ccv_nnc_no_hint, output_params, 3);
2675 else
2676 ccv_nnc_hint_tensor_auto(group_norm, (ccv_nnc_tensor_param_t []){
2677 params,
2678 }, 1, ccv_nnc_no_hint, output_params, 3);
2679 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0);
2680 const ccv_nnc_tensor_symbol_t saved_mean = ccv_nnc_tensor_symbol_new(graph, output_params[1], "saved_mean");
2681 const ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(graph, output_params[2], "saved_inv_std");
2682 if (self->params.gnorm.elementwise_affine)
2683 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");
2684 else
2685 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");
2686 outputs[0] = output;
2687}
2688
2689static 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)
2690{
2691 ccv_cnnp_model_group_norm_t* const self = (ccv_cnnp_model_group_norm_t*)super;
2692 if (self->scale.graph)
2693 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);
2694 if (self->bias.graph)
2695 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);
2696}
2697
2698static 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)
2699{
2700 ccv_cnnp_model_group_norm_t* const self = (ccv_cnnp_model_group_norm_t*)super;
2701 if (self->scale.graph)
2702 add_to_array(parameters, self->scale, is_trainable);
2703 if (self->bias.graph)
2704 add_to_array(parameters, self->bias, is_trainable);
2705}
2706
2707static ccv_cnnp_model_t* _ccv_cnnp_group_norm_copy(const ccv_cnnp_model_t* const super, void* const context);
2708
2709static const ccv_cnnp_model_vtab_t ccv_cnnp_group_norm_isa = {
2710 .build = _ccv_cnnp_group_norm_build,
2711 .init_states = _ccv_cnnp_group_norm_init_states,
2712 .add_to_parameter = _ccv_cnnp_group_norm_add_to_parameter,
2713 .copy = _ccv_cnnp_group_norm_copy,
2714};
2715
2716ccv_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)
2717{
2718 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));
2719 model_group_norm->super.isa = &ccv_cnnp_group_norm_isa;
2720 model_group_norm->super.input_size = 1;
2721 model_group_norm->super.outputs = &model_group_norm->output;
2722 model_group_norm->super.output_size = 1;
2723 model_group_norm->super.is_trainable = is_trainable;
2724 ccv_cnnp_model_copy_name(&model_group_norm->super, name);
2725 model_group_norm->scale.d = CCV_NNC_NO_TENSOR_SYMBOL;
2726 model_group_norm->scale.graph = 0;
2727 model_group_norm->bias.d = CCV_NNC_NO_TENSOR_SYMBOL;
2728 model_group_norm->bias.graph = 0;
2729 model_group_norm->params.gnorm.group_axis = group_axis;
2730 model_group_norm->params.gnorm.groups = groups;
2731 model_group_norm->params.gnorm.epsilon = epsilon;
2732 model_group_norm->params.gnorm.reduce_count = axis_count;
2733 model_group_norm->params.gnorm.elementwise_affine = elementwise_affine;
2734 memcpy(model_group_norm->params.gnorm.reduce_axis, reduce_axis, sizeof(int) * axis_count);
2735 return (ccv_cnnp_model_t*)model_group_norm;
2736}
2737
2738static ccv_cnnp_model_t* _ccv_cnnp_group_norm_copy(const ccv_cnnp_model_t* const super, void* const context)
2739{
2740 const ccv_cnnp_model_group_norm_t* const self = (const ccv_cnnp_model_group_norm_t*)super;
2741 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);
2742}
2743
2744// MARK - RMSNorm Layer
2745
2746typedef struct {
2747 ccv_cnnp_model_t super;
2748 ccv_nnc_tensor_symbol_t output;
2749 ccv_nnc_tensor_symbol_t scale;
2750 ccv_nnc_cmd_param_t params;
2751} ccv_cnnp_model_rmsnorm_t;
2752
2753static 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)
2754{
2755 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);
2756 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"
, 2756, __extension__ __PRETTY_FUNCTION__); }));
2757 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", 2757, __extension__ __PRETTY_FUNCTION__
); }));
2758 ccv_cnnp_model_rmsnorm_t* const self = (ccv_cnnp_model_rmsnorm_t*)super;
2759 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2760 ccv_nnc_tensor_param_t scale_params = params;
2761 const int nd = ccv_nnc_tensor_nd(params.dim);
2762 int i;
2763 for (i = 0; i < nd; i++)
2764 scale_params.dim[i] = 1;
2765 for (i = 0; i < self->params.rmsnorm.count; i++)
2766 scale_params.dim[self->params.rmsnorm.axis[i]] = params.dim[self->params.rmsnorm.axis[i]];
2767 // Both scale and bias are shared between if this model is reused.
2768 if (self->params.rmsnorm.elementwise_affine)
2769 {
2770 if (!self->scale.graph)
2771 self->scale = ccv_nnc_tensor_symbol_new(graph, scale_params, "scale");
2772 }
2773 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
};
2774 const ccv_nnc_cmd_t rmsnorm = ccv_nnc_cmd(CCV_NNC_RMSNORM_FORWARD, 0, self->params, 0);
2775 ccv_nnc_tensor_param_t output_params[2];
2776 if (self->params.rmsnorm.elementwise_affine)
2777 ccv_nnc_hint_tensor_auto(rmsnorm, (ccv_nnc_tensor_param_t []){
2778 params,
2779 scale_params,
2780 }, 2, ccv_nnc_no_hint, output_params, 2);
2781 else
2782 ccv_nnc_hint_tensor_auto(rmsnorm, (ccv_nnc_tensor_param_t []){
2783 params,
2784 }, 1, ccv_nnc_no_hint, output_params, 2);
2785 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0);
2786 const ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(graph, output_params[1], "saved_inv_std");
2787 if (self->params.rmsnorm.elementwise_affine)
2788 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");
2789 else
2790 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");
2791 outputs[0] = output;
2792}
2793
2794static 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)
2795{
2796 ccv_cnnp_model_rmsnorm_t* const self = (ccv_cnnp_model_rmsnorm_t*)super;
2797 if (self->scale.graph)
2798 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);
2799}
2800
2801static 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)
2802{
2803 ccv_cnnp_model_rmsnorm_t* const self = (ccv_cnnp_model_rmsnorm_t*)super;
2804 if (self->scale.graph)
2805 add_to_array(parameters, self->scale, is_trainable);
2806}
2807
2808static ccv_cnnp_model_t* _ccv_cnnp_rmsnorm_copy(const ccv_cnnp_model_t* const super, void* const context);
2809
2810static const ccv_cnnp_model_vtab_t ccv_cnnp_rmsnorm_isa = {
2811 .build = _ccv_cnnp_rmsnorm_build,
2812 .init_states = _ccv_cnnp_rmsnorm_init_states,
2813 .add_to_parameter = _ccv_cnnp_rmsnorm_add_to_parameter,
2814 .copy = _ccv_cnnp_rmsnorm_copy,
2815};
2816
2817ccv_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)
2818{
2819 ccv_cnnp_model_rmsnorm_t* const model_rmsnorm = (ccv_cnnp_model_rmsnorm_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_rmsnorm_t));
2820 model_rmsnorm->super.isa = &ccv_cnnp_rmsnorm_isa;
2821 model_rmsnorm->super.input_size = 1;
2822 model_rmsnorm->super.outputs = &model_rmsnorm->output;
2823 model_rmsnorm->super.output_size = 1;
2824 model_rmsnorm->super.is_trainable = is_trainable;
2825 ccv_cnnp_model_copy_name(&model_rmsnorm->super, name);
2826 model_rmsnorm->scale.d = CCV_NNC_NO_TENSOR_SYMBOL;
2827 model_rmsnorm->scale.graph = 0;
2828 model_rmsnorm->params.rmsnorm.epsilon = epsilon;
2829 model_rmsnorm->params.rmsnorm.count = axis_count;
2830 model_rmsnorm->params.rmsnorm.elementwise_affine = elementwise_affine;
2831 memcpy(model_rmsnorm->params.lnorm.axis, axis, sizeof(int) * axis_count);
2832 return (ccv_cnnp_model_t*)model_rmsnorm;
2833}
2834
2835static ccv_cnnp_model_t* _ccv_cnnp_rmsnorm_copy(const ccv_cnnp_model_t* const super, void* const context)
2836{
2837 const ccv_cnnp_model_rmsnorm_t* const self = (const ccv_cnnp_model_rmsnorm_t*)super;
2838 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);
2839}
2840
2841// MARK - Batched Matrix Mul Layer
2842
2843typedef struct {
2844 ccv_cnnp_model_t super;
2845 ccv_nnc_tensor_symbol_t output;
2846 int transpose_a[2];
2847 int transpose_b[2];
2848 int flags;
2849} ccv_cnnp_model_matmul_t;
2850
2851static 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)
2852{
2853 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);
2854 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"
, 2854, __extension__ __PRETTY_FUNCTION__); }));
2855 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", 2855, __extension__ __PRETTY_FUNCTION__
); }));
2856 ccv_cnnp_model_matmul_t* const self = (ccv_cnnp_model_matmul_t*)super;
2857 ccv_nnc_tensor_param_t a_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2858 ccv_nnc_tensor_param_t b_params = ccv_nnc_tensor_symbol_params(graph, inputs[1]);
2859 ccv_nnc_tensor_param_t output_params;
2860 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);
2861 matmul.info.blas.flags = self->flags;
2862 ccv_nnc_hint_tensor_auto(matmul, (ccv_nnc_tensor_param_t []){
2863 a_params,
2864 b_params,
2865 }, 2, ccv_nnc_no_hint, &output_params, 1);
2866 const ccv_nnc_tensor_symbol_t matmul_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2867 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");
2868 outputs[0] = matmul_output;
2869}
2870
2871static ccv_cnnp_model_t* _ccv_cnnp_matmul_copy(const ccv_cnnp_model_t* const super, void* const context);
2872
2873static const ccv_cnnp_model_vtab_t ccv_cnnp_matmul_isa = {
2874 .build = _ccv_cnnp_matmul_build,
2875 .copy = _ccv_cnnp_matmul_copy,
2876};
2877
2878ccv_cnnp_model_t* ccv_cnnp_matmul(const int transpose_a[2], const int transpose_b[2], const int flags, const char* const name)
2879{
2880 ccv_cnnp_model_matmul_t* const model_matmul = (ccv_cnnp_model_matmul_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_matmul_t));
2881 model_matmul->super.isa = &ccv_cnnp_matmul_isa;
2882 model_matmul->super.input_size = 2;
2883 model_matmul->super.outputs = &model_matmul->output;
2884 model_matmul->super.output_size = 1;
2885 model_matmul->transpose_a[0] = transpose_a[0];
2886 model_matmul->transpose_a[1] = transpose_a[1];
2887 model_matmul->transpose_b[0] = transpose_b[0];
2888 model_matmul->transpose_b[1] = transpose_b[1];
2889 model_matmul->flags = flags;
2890 ccv_cnnp_model_copy_name(&model_matmul->super, name);
2891 return (ccv_cnnp_model_t*)model_matmul;
2892}
2893
2894static ccv_cnnp_model_t* _ccv_cnnp_matmul_copy(const ccv_cnnp_model_t* const super, void* const context)
2895{
2896 const ccv_cnnp_model_matmul_t* const self = (const ccv_cnnp_model_matmul_t*)super;
2897 return ccv_cnnp_matmul(self->transpose_a, self->transpose_b, self->flags, self->super.name);
2898}
2899
2900// MARK - Dropout Layer
2901
2902typedef struct {
2903 ccv_cnnp_model_t super;
2904 ccv_nnc_tensor_symbol_t output;
2905 ccv_nnc_graph_exec_symbol_t dropout;
2906 float p;
2907 int entirety;
2908} ccv_cnnp_model_dropout_t;
2909
2910static 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)
2911{
2912 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);
2913 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"
, 2913, __extension__ __PRETTY_FUNCTION__); }));
2914 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", 2914, __extension__ __PRETTY_FUNCTION__
); }));
2915 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2916 ccv_nnc_tensor_param_t output_params[2];
2917 ccv_cnnp_model_dropout_t* const self = (ccv_cnnp_model_dropout_t*)super;
2918 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);
2919 ccv_nnc_hint_tensor_auto(dropout, (ccv_nnc_tensor_param_t []){
2920 params,
2921 }, 1, ccv_nnc_no_hint, output_params, 2);
2922 const ccv_nnc_tensor_symbol_t dropout_output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0);
2923 const ccv_nnc_tensor_symbol_t mask = ccv_nnc_tensor_symbol_new(graph, output_params[1], "mask");
2924 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");
2925 outputs[0] = dropout_output;
2926}
2927
2928static 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)
2929{
2930 ccv_cnnp_model_dropout_t* const self = (ccv_cnnp_model_dropout_t*)super;
2931 if (self->dropout.graph)
2932 {
2933 if (is_test)
2934 // During test, the dropout is not applied. Data transfer is perfect because if these are the same tensor, it will skip.
2935 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);
2936 else
2937 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);
2938 }
2939}
2940
2941static ccv_cnnp_model_t* _ccv_cnnp_dropout_copy(const ccv_cnnp_model_t* const super, void* const context);
2942
2943static const ccv_cnnp_model_vtab_t ccv_cnnp_dropout_isa = {
2944 .build = _ccv_cnnp_dropout_build,
2945 .set_is_test = _ccv_cnnp_dropout_set_is_test,
2946 .copy = _ccv_cnnp_dropout_copy,
2947};
2948
2949ccv_cnnp_model_t* ccv_cnnp_dropout(const float p, const int entirety, const char* const name)
2950{
2951 ccv_cnnp_model_dropout_t* const model_dropout = (ccv_cnnp_model_dropout_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_dropout_t));
2952 model_dropout->super.isa = &ccv_cnnp_dropout_isa;
2953 model_dropout->super.input_size = 1;
2954 model_dropout->super.outputs = &model_dropout->output;
2955 model_dropout->super.output_size = 1;
2956 model_dropout->p = p;
2957 model_dropout->entirety = entirety;
2958 ccv_cnnp_model_copy_name(&model_dropout->super, name);
2959 return (ccv_cnnp_model_t*)model_dropout;
2960}
2961
2962static ccv_cnnp_model_t* _ccv_cnnp_dropout_copy(const ccv_cnnp_model_t* const super, void* const context)
2963{
2964 const ccv_cnnp_model_dropout_t* const self = (const ccv_cnnp_model_dropout_t*)super;
2965 return ccv_cnnp_dropout(self->p, self->entirety, self->super.name);
2966}
2967
2968// MARK - Masked Fill Layer
2969
2970typedef struct {
2971 ccv_cnnp_model_t super;
2972 ccv_nnc_tensor_symbol_t output;
2973 float eq;
2974 float fill;
2975} ccv_cnnp_model_masked_fill_t;
2976
2977static 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)
2978{
2979 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);
2980 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"
, 2980, __extension__ __PRETTY_FUNCTION__); }));
2981 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", 2981, __extension__ __PRETTY_FUNCTION__
); }));
2982 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2983 ccv_cnnp_model_masked_fill_t* const self = (ccv_cnnp_model_masked_fill_t*)super;
2984 const ccv_nnc_tensor_symbol_t masked_fill_output = ccv_nnc_tensor_symbol_new(graph, params, 0);
2985 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");
2986 outputs[0] = masked_fill_output;
2987}
2988
2989static ccv_cnnp_model_t* _ccv_cnnp_masked_fill_copy(const ccv_cnnp_model_t* const super, void* const context);
2990
2991static const ccv_cnnp_model_vtab_t ccv_cnnp_masked_fill_isa = {
2992 .build = _ccv_cnnp_masked_fill_build,
2993 .copy = _ccv_cnnp_masked_fill_copy,
2994};
2995
2996ccv_cnnp_model_t* ccv_cnnp_masked_fill(const float eq, const float fill, const char* const name)
2997{
2998 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));
2999 model_masked_fill->super.isa = &ccv_cnnp_masked_fill_isa;
3000 model_masked_fill->super.input_size = 2;
3001 model_masked_fill->super.outputs = &model_masked_fill->output;
3002 model_masked_fill->super.output_size = 1;
3003 model_masked_fill->eq = eq;
3004 model_masked_fill->fill = fill;
3005 ccv_cnnp_model_copy_name(&model_masked_fill->super, name);
3006 return (ccv_cnnp_model_t*)model_masked_fill;
3007}
3008
3009static ccv_cnnp_model_t* _ccv_cnnp_masked_fill_copy(const ccv_cnnp_model_t* const super, void* const context)
3010{
3011 const ccv_cnnp_model_masked_fill_t* const self = (const ccv_cnnp_model_masked_fill_t*)super;
3012 return ccv_cnnp_masked_fill(self->eq, self->fill, self->super.name);
3013}
3014
3015// MARK - Index Select Layer
3016
3017typedef struct {
3018 ccv_cnnp_model_t super;
3019 ccv_nnc_tensor_symbol_t output;
3020} ccv_cnnp_model_index_select_t;
3021
3022static 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)
3023{
3024 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);
3025 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"
, 3025, __extension__ __PRETTY_FUNCTION__); }));
3026 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", 3026, __extension__ __PRETTY_FUNCTION__
); }));
3027 const ccv_nnc_tensor_param_t vocab_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3028 const ccv_nnc_tensor_param_t index_params = ccv_nnc_tensor_symbol_params(graph, inputs[1]);
3029 ccv_nnc_tensor_param_t output_params;
3030 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);
3031 ccv_nnc_hint_tensor_auto(index_select, (ccv_nnc_tensor_param_t []){
3032 vocab_params,
3033 index_params,
3034 }, 2, ccv_nnc_no_hint, &output_params, 1);
3035 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3036 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");
3037 outputs[0] = output;
3038}
3039
3040static ccv_cnnp_model_t* _ccv_cnnp_index_select_copy(const ccv_cnnp_model_t* const super, void* const context);
3041
3042static const ccv_cnnp_model_vtab_t ccv_cnnp_index_select_isa = {
3043 .build = _ccv_cnnp_index_select_build,
3044 .copy = _ccv_cnnp_index_select_copy,
3045};
3046
3047ccv_cnnp_model_t* ccv_cnnp_index_select(const char* const name)
3048{
3049 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));
3050 model_index_select->super.isa = &ccv_cnnp_index_select_isa;
3051 model_index_select->super.input_size = 2;
3052 model_index_select->super.outputs = &model_index_select->output;
3053 model_index_select->super.output_size = 1;
3054 ccv_cnnp_model_copy_name(&model_index_select->super, name);
3055 return (ccv_cnnp_model_t*)model_index_select;
3056}
3057
3058static ccv_cnnp_model_t* _ccv_cnnp_index_select_copy(const ccv_cnnp_model_t* const super, void* const context)
3059{
3060 ccv_cnnp_model_index_select_t* const self = (ccv_cnnp_model_index_select_t*)super;
3061 return ccv_cnnp_index_select(self->super.name);
3062}
3063
3064// MARK - Embedding Layer
3065
3066typedef struct {
3067 ccv_cnnp_model_t super;
3068 ccv_nnc_tensor_symbol_t output;
3069 ccv_nnc_tensor_symbol_t vocab;
3070 int datatype;
3071 int vocab_size;
3072 int embed_size;
3073} ccv_cnnp_model_embedding_t;
3074
3075static 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)
3076{
3077 ccv_cnnp_model_embedding_t* const self = (ccv_cnnp_model_embedding_t*)super;
3078 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);
3079 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"
, 3079, __extension__ __PRETTY_FUNCTION__); }));
3080 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", 3080, __extension__ __PRETTY_FUNCTION__
); }));
3081 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3082 ccv_nnc_tensor_param_t vocab_params = params;
3083 memset(vocab_params.dim, 0, sizeof(vocab_params.dim));
3084 vocab_params.datatype = self->datatype;
3085 vocab_params.dim[0] = self->vocab_size;
3086 vocab_params.dim[1] = self->embed_size;
3087 if (!self->vocab.graph)
3088 self->vocab = ccv_nnc_tensor_symbol_new(graph, vocab_params, "vocab");
3089 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", 3089
, __extension__ __PRETTY_FUNCTION__); }));
3090 const ccv_nnc_tensor_symbol_t vocab = ccv_cnnp_model_get_symbol(super, self->vocab);
3091 ccv_nnc_tensor_param_t output_params;
3092 const ccv_nnc_cmd_t embedding = CMD_INDEX_SELECT_FORWARD()ccv_nnc_cmd(CCV_NNC_INDEX_SELECT_FORWARD, 0, ccv_nnc_cmd_auto
, 0);
3093 ccv_nnc_hint_tensor_auto(embedding, (ccv_nnc_tensor_param_t []){
3094 vocab_params,
3095 params,
3096 }, 2, ccv_nnc_no_hint, &output_params, 1);
3097 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3098 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");
3099 outputs[0] = output;
3100}
3101
3102static 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)
3103{
3104 ccv_cnnp_model_embedding_t* const self = (ccv_cnnp_model_embedding_t*)super;
3105 const float std = sqrtf(2) / sqrtf(self->vocab_size + self->embed_size);
3106 const float bound = sqrtf(3) * std;
3107 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);
3108}
3109
3110static 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)
3111{
3112 ccv_cnnp_model_embedding_t* const self = (ccv_cnnp_model_embedding_t*)super;
3113 add_to_array(parameters, self->vocab, is_trainable);
3114}
3115
3116static ccv_cnnp_model_t* _ccv_cnnp_embedding_copy(const ccv_cnnp_model_t* const super, void* const context);
3117
3118static const ccv_cnnp_model_vtab_t ccv_cnnp_embedding_isa = {
3119 .build = _ccv_cnnp_embedding_build,
3120 .init_states = _ccv_cnnp_embedding_init_states,
3121 .add_to_parameter = _ccv_cnnp_embedding_add_to_parameter,
3122 .copy = _ccv_cnnp_embedding_copy,
3123};
3124
3125ccv_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)
3126{
3127 ccv_cnnp_model_embedding_t* const model_embedding = (ccv_cnnp_model_embedding_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_embedding_t));
3128 model_embedding->super.isa = &ccv_cnnp_embedding_isa;
3129 model_embedding->super.input_size = 1;
3130 model_embedding->super.outputs = &model_embedding->output;
3131 model_embedding->super.output_size = 1;
3132 model_embedding->super.is_trainable = is_trainable;
3133 ccv_cnnp_model_copy_name(&model_embedding->super, name);
3134 model_embedding->vocab.d = CCV_NNC_NO_TENSOR_SYMBOL;
3135 model_embedding->vocab.graph = 0;
3136 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", 3136, __extension__ __PRETTY_FUNCTION__
); }));
3137 model_embedding->datatype = datatype;
3138 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", 3138, __extension__ __PRETTY_FUNCTION__
); }));
3139 model_embedding->vocab_size = vocab_size;
3140 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", 3140, __extension__ __PRETTY_FUNCTION__
); }));
3141 model_embedding->embed_size = embed_size;
3142 return (ccv_cnnp_model_t*)model_embedding;
3143}
3144
3145static ccv_cnnp_model_t* _ccv_cnnp_embedding_copy(const ccv_cnnp_model_t* const super, void* const context)
3146{
3147 ccv_cnnp_model_embedding_t* const self = (ccv_cnnp_model_embedding_t*)super;
3148 return ccv_cnnp_embedding(self->datatype, self->vocab_size, self->embed_size, self->super.is_trainable, self->super.name);
3149}
3150
3151// MARK - Pool Layers
3152
3153typedef struct {
3154 ccv_cnnp_model_t super;
3155 ccv_nnc_tensor_symbol_t output;
3156 int type;
3157 float width_scale;
3158 float height_scale;
3159 int align_corners;
3160} ccv_cnnp_model_upsample_t;
3161
3162static 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)
3163{
3164 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);
3165 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"
, 3165, __extension__ __PRETTY_FUNCTION__); }));
3166 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", 3166, __extension__ __PRETTY_FUNCTION__
); }));
3167 ccv_cnnp_model_upsample_t* const self = (ccv_cnnp_model_upsample_t*)super;
3168 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3169 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);
3170 ccv_nnc_tensor_param_t output_params;
3171 ccv_nnc_hint_tensor_auto(cmd, &params, 1, ccv_nnc_no_hint, &output_params, 1);
3172 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3173 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");
3174 outputs[0] = output;
3175}
3176
3177static ccv_cnnp_model_t* _ccv_cnnp_upsample_copy(const ccv_cnnp_model_t* const super, void* const context);
3178
3179static const ccv_cnnp_model_vtab_t ccv_cnnp_upsample_isa = {
3180 .build = _ccv_cnnp_upsample_build,
3181 .copy = _ccv_cnnp_upsample_copy,
3182};
3183
3184ccv_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)
3185{
3186 ccv_cnnp_model_upsample_t* const model_upsample = (ccv_cnnp_model_upsample_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_upsample_t));
3187 model_upsample->super.isa = &ccv_cnnp_upsample_isa;
3188 model_upsample->super.input_size = 1;
3189 model_upsample->super.outputs = &model_upsample->output;
3190 model_upsample->super.output_size = 1;
3191 ccv_cnnp_model_copy_name(&model_upsample->super, name);
3192 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", 3192, __extension__ __PRETTY_FUNCTION__
); }));
3193 model_upsample->type = type;
3194 model_upsample->width_scale = width_scale;
3195 model_upsample->height_scale = height_scale;
3196 model_upsample->align_corners = align_corners;
3197 return (ccv_cnnp_model_t*)model_upsample;
3198}
3199
3200static ccv_cnnp_model_t* _ccv_cnnp_upsample_copy(const ccv_cnnp_model_t* const super, void* const context)
3201{
3202 const ccv_cnnp_model_upsample_t* const self = (const ccv_cnnp_model_upsample_t*)super;
3203 return ccv_cnnp_upsample(self->type, self->width_scale, self->height_scale, self->align_corners, self->super.name);
3204}
3205
3206// MARK - Reduce Sum Layer
3207
3208typedef struct {
3209 ccv_cnnp_model_t super;
3210 int axis[CCV_NNC_MAX_DIM_ALLOC(12)];
3211 int count;
3212 ccv_nnc_tensor_symbol_t output;
3213} ccv_cnnp_model_reduce_sum_t;
3214
3215static 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)
3216{
3217 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);
3218 const ccv_cnnp_model_reduce_sum_t* const self = (const ccv_cnnp_model_reduce_sum_t*)super;
3219 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"
, 3219, __extension__ __PRETTY_FUNCTION__); }));
3220 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", 3220, __extension__ __PRETTY_FUNCTION__
); }));
3221 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3222 ccv_nnc_tensor_param_t output_params;
3223 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);
3224 int i;
3225 for (i = 0; i < self->count; i++)
3226 reduce_sum.info.reduce.axis[i] = self->axis[i];
3227 reduce_sum.info.reduce.count = self->count;
3228 ccv_nnc_hint_tensor_auto(reduce_sum, &input_params, 1, ccv_nnc_no_hint, &output_params, 1);
3229 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3230 ccv_nnc_graph_exec_symbol_new(graph, reduce_sum, inputs, input_size, outputs, output_size, "reduce_sum");
3231}
3232
3233static ccv_cnnp_model_t* _ccv_cnnp_reduce_sum_copy(const ccv_cnnp_model_t* const self, void* const context);
3234
3235static const ccv_cnnp_model_vtab_t ccv_cnnp_reduce_sum_isa = {
3236 .build = _ccv_cnnp_reduce_sum_build,
3237 .copy = _ccv_cnnp_reduce_sum_copy,
3238};
3239
3240ccv_cnnp_model_t* ccv_cnnp_reduce_sum(const int* const axis, const int axis_count, const char* const name)
3241{
3242 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));
3243 model_reduce_sum->super.isa = &ccv_cnnp_reduce_sum_isa;
3244 model_reduce_sum->super.input_size = 1;
3245 model_reduce_sum->super.outputs = &model_reduce_sum->output;
3246 model_reduce_sum->super.output_size = 1;
3247 ccv_cnnp_model_copy_name(&model_reduce_sum->super, name);
3248 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", 3248, __extension__ __PRETTY_FUNCTION__
); }));
3249 int i;
3250 for (i = 0; i < axis_count; i++)
3251 model_reduce_sum->axis[i] = axis[i];
3252 model_reduce_sum->count = axis_count;
3253 return (ccv_cnnp_model_t*)model_reduce_sum;
3254}
3255
3256static ccv_cnnp_model_t* _ccv_cnnp_reduce_sum_copy(const ccv_cnnp_model_t* const super, void* const context)
3257{
3258 const ccv_cnnp_model_reduce_sum_t* const self = (const ccv_cnnp_model_reduce_sum_t*)super;
3259 return ccv_cnnp_reduce_sum(self->axis, self->count, self->super.name);
3260}
3261
3262// MARK - Reduce Mean Layer
3263
3264typedef struct {
3265 ccv_cnnp_model_t super;
3266 int axis[CCV_NNC_MAX_DIM_ALLOC(12)];
3267 int count;
3268 ccv_nnc_tensor_symbol_t output;
3269} ccv_cnnp_model_reduce_mean_t;
3270
3271static 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)
3272{
3273 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);
3274 const ccv_cnnp_model_reduce_mean_t* const self = (const ccv_cnnp_model_reduce_mean_t*)super;
3275 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"
, 3275, __extension__ __PRETTY_FUNCTION__); }));
3276 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", 3276, __extension__ __PRETTY_FUNCTION__
); }));
3277 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3278 ccv_nnc_tensor_param_t output_params;
3279 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);
3280 int i;
3281 for (i = 0; i < self->count; i++)
3282 reduce_mean.info.reduce.axis[i] = self->axis[i];
3283 reduce_mean.info.reduce.count = self->count;
3284 ccv_nnc_hint_tensor_auto(reduce_mean, &input_params, 1, ccv_nnc_no_hint, &output_params, 1);
3285 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3286 ccv_nnc_graph_exec_symbol_new(graph, reduce_mean, inputs, input_size, outputs, output_size, "reduce_mean");
3287}
3288
3289static ccv_cnnp_model_t* _ccv_cnnp_reduce_mean_copy(const ccv_cnnp_model_t* const self, void* const context);
3290
3291static const ccv_cnnp_model_vtab_t ccv_cnnp_reduce_mean_isa = {
3292 .build = _ccv_cnnp_reduce_mean_build,
3293 .copy = _ccv_cnnp_reduce_mean_copy,
3294};
3295
3296ccv_cnnp_model_t* ccv_cnnp_reduce_mean(const int* const axis, const int axis_count, const char* const name)
3297{
3298 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));
3299 model_reduce_mean->super.isa = &ccv_cnnp_reduce_mean_isa;
3300 model_reduce_mean->super.input_size = 1;
3301 model_reduce_mean->super.outputs = &model_reduce_mean->output;
3302 model_reduce_mean->super.output_size = 1;
3303 ccv_cnnp_model_copy_name(&model_reduce_mean->super, name);
3304 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", 3304, __extension__ __PRETTY_FUNCTION__
); }));
3305 int i;
3306 for (i = 0; i < axis_count; i++)
3307 model_reduce_mean->axis[i] = axis[i];
3308 model_reduce_mean->count = axis_count;
3309 return (ccv_cnnp_model_t*)model_reduce_mean;
3310}
3311
3312static ccv_cnnp_model_t* _ccv_cnnp_reduce_mean_copy(const ccv_cnnp_model_t* const super, void* const context)
3313{
3314 const ccv_cnnp_model_reduce_mean_t* const self = (const ccv_cnnp_model_reduce_mean_t*)super;
3315 return ccv_cnnp_reduce_mean(self->axis, self->count, self->super.name);
3316}
3317
3318// MARK - Reduce Max Layer
3319
3320typedef struct {
3321 ccv_cnnp_model_t super;
3322 int axis[CCV_NNC_MAX_DIM_ALLOC(12)];
3323 int count;
3324 ccv_nnc_tensor_symbol_t output;
3325} ccv_cnnp_model_reduce_max_t;
3326
3327static 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)
3328{
3329 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);
3330 const ccv_cnnp_model_reduce_max_t* const self = (const ccv_cnnp_model_reduce_max_t*)super;
3331 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"
, 3331, __extension__ __PRETTY_FUNCTION__); }));
3332 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", 3332, __extension__ __PRETTY_FUNCTION__
); }));
3333 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3334 ccv_nnc_tensor_param_t output_params;
3335 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);
3336 int i;
3337 for (i = 0; i < self->count; i++)
3338 reduce_max.info.reduce.axis[i] = self->axis[i];
3339 reduce_max.info.reduce.count = self->count;
3340 ccv_nnc_hint_tensor_auto(reduce_max, &input_params, 1, ccv_nnc_no_hint, &output_params, 1);
3341 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3342 ccv_nnc_graph_exec_symbol_new(graph, reduce_max, inputs, input_size, outputs, output_size, "reduce_max");
3343}
3344
3345static ccv_cnnp_model_t* _ccv_cnnp_reduce_max_copy(const ccv_cnnp_model_t* const self, void* const context);
3346
3347static const ccv_cnnp_model_vtab_t ccv_cnnp_reduce_max_isa = {
3348 .build = _ccv_cnnp_reduce_max_build,
3349 .copy = _ccv_cnnp_reduce_max_copy,
3350};
3351
3352ccv_cnnp_model_t* ccv_cnnp_reduce_max(const int* const axis, const int axis_count, const char* const name)
3353{
3354 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));
3355 model_reduce_max->super.isa = &ccv_cnnp_reduce_max_isa;
3356 model_reduce_max->super.input_size = 1;
3357 model_reduce_max->super.outputs = &model_reduce_max->output;
3358 model_reduce_max->super.output_size = 1;
3359 ccv_cnnp_model_copy_name(&model_reduce_max->super, name);
3360 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", 3360, __extension__ __PRETTY_FUNCTION__
); }));
3361 int i;
3362 for (i = 0; i < axis_count; i++)
3363 model_reduce_max->axis[i] = axis[i];
3364 model_reduce_max->count = axis_count;
3365 return (ccv_cnnp_model_t*)model_reduce_max;
3366}
3367
3368static ccv_cnnp_model_t* _ccv_cnnp_reduce_max_copy(const ccv_cnnp_model_t* const super, void* const context)
3369{
3370 const ccv_cnnp_model_reduce_max_t* const self = (const ccv_cnnp_model_reduce_max_t*)super;
3371 return ccv_cnnp_reduce_max(self->axis, self->count, self->super.name);
3372}
3373
3374// MARK - Reduce Min Layer
3375
3376typedef struct {
3377 ccv_cnnp_model_t super;
3378 int axis[CCV_NNC_MAX_DIM_ALLOC(12)];
3379 int count;
3380 ccv_nnc_tensor_symbol_t output;
3381} ccv_cnnp_model_reduce_min_t;
3382
3383static 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)
3384{
3385 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);
3386 const ccv_cnnp_model_reduce_min_t* const self = (const ccv_cnnp_model_reduce_min_t*)super;
3387 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"
, 3387, __extension__ __PRETTY_FUNCTION__); }));
3388 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", 3388, __extension__ __PRETTY_FUNCTION__
); }));
3389 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3390 ccv_nnc_tensor_param_t output_params;
3391 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);
3392 int i;
3393 for (i = 0; i < self->count; i++)
3394 reduce_min.info.reduce.axis[i] = self->axis[i];
3395 reduce_min.info.reduce.count = self->count;
3396 ccv_nnc_hint_tensor_auto(reduce_min, &input_params, 1, ccv_nnc_no_hint, &output_params, 1);
3397 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3398 ccv_nnc_graph_exec_symbol_new(graph, reduce_min, inputs, input_size, outputs, output_size, "reduce_min");
3399}
3400
3401static ccv_cnnp_model_t* _ccv_cnnp_reduce_min_copy(const ccv_cnnp_model_t* const self, void* const context);
3402
3403static const ccv_cnnp_model_vtab_t ccv_cnnp_reduce_min_isa = {
3404 .build = _ccv_cnnp_reduce_min_build,
3405 .copy = _ccv_cnnp_reduce_min_copy,
3406};
3407
3408ccv_cnnp_model_t* ccv_cnnp_reduce_min(const int* const axis, const int axis_count, const char* const name)
3409{
3410 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));
3411 model_reduce_min->super.isa = &ccv_cnnp_reduce_min_isa;
3412 model_reduce_min->super.input_size = 1;
3413 model_reduce_min->super.outputs = &model_reduce_min->output;
3414 model_reduce_min->super.output_size = 1;
3415 ccv_cnnp_model_copy_name(&model_reduce_min->super, name);
3416 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", 3416, __extension__ __PRETTY_FUNCTION__
); }));
3417 int i;
3418 for (i = 0; i < axis_count; i++)
3419 model_reduce_min->axis[i] = axis[i];
3420 model_reduce_min->count = axis_count;
3421 return (ccv_cnnp_model_t*)model_reduce_min;
3422}
3423
3424static ccv_cnnp_model_t* _ccv_cnnp_reduce_min_copy(const ccv_cnnp_model_t* const super, void* const context)
3425{
3426 const ccv_cnnp_model_reduce_min_t* const self = (const ccv_cnnp_model_reduce_min_t*)super;
3427 return ccv_cnnp_reduce_min(self->axis, self->count, self->super.name);
3428}
3429
3430// MARK - Reduce Norm2 Layer
3431
3432typedef struct {
3433 ccv_cnnp_model_t super;
3434 int axis[CCV_NNC_MAX_DIM_ALLOC(12)];
3435 int count;
3436 ccv_nnc_tensor_symbol_t output;
3437} ccv_cnnp_model_reduce_norm2_t;
3438
3439static 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)
3440{
3441 const ccv_cnnp_model_reduce_norm2_t* const self = (const ccv_cnnp_model_reduce_norm2_t*)super;
3442 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);
3443 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"
, 3443, __extension__ __PRETTY_FUNCTION__); }));
3444 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", 3444, __extension__ __PRETTY_FUNCTION__
); }));
3445 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3446 ccv_nnc_tensor_param_t output_params;
3447 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);
3448 int i;
3449 for (i = 0; i < self->count; i++)
3450 reduce_norm2.info.reduce.axis[i] = self->axis[i];
3451 reduce_norm2.info.reduce.count = self->count;
3452 ccv_nnc_hint_tensor_auto(reduce_norm2, &input_params, 1, ccv_nnc_no_hint, &output_params, 1);
3453 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3454 ccv_nnc_graph_exec_symbol_new(graph, reduce_norm2, inputs, input_size, outputs, output_size, "reduce_norm2");
3455}
3456
3457static ccv_cnnp_model_t* _ccv_cnnp_reduce_norm2_copy(const ccv_cnnp_model_t* const self, void* const context);
3458
3459static const ccv_cnnp_model_vtab_t ccv_cnnp_reduce_norm2_isa = {
3460 .build = _ccv_cnnp_reduce_norm2_build,
3461 .copy = _ccv_cnnp_reduce_norm2_copy,
3462};
3463
3464ccv_cnnp_model_t* ccv_cnnp_reduce_norm2(const int* const axis, const int axis_count, const char* const name)
3465{
3466 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));
3467 model_reduce_norm2->super.isa = &ccv_cnnp_reduce_norm2_isa;
3468 model_reduce_norm2->super.input_size = 1;
3469 model_reduce_norm2->super.outputs = &model_reduce_norm2->output;
3470 model_reduce_norm2->super.output_size = 1;
3471 ccv_cnnp_model_copy_name(&model_reduce_norm2->super, name);
3472 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", 3472, __extension__ __PRETTY_FUNCTION__
); }));
3473 int i;
3474 for (i = 0; i < axis_count; i++)
3475 model_reduce_norm2->axis[i] = axis[i];
3476 model_reduce_norm2->count = axis_count;
3477 return (ccv_cnnp_model_t*)model_reduce_norm2;
3478}
3479
3480static ccv_cnnp_model_t* _ccv_cnnp_reduce_norm2_copy(const ccv_cnnp_model_t* const super, void* const context)
3481{
3482 const ccv_cnnp_model_reduce_norm2_t* const self = (const ccv_cnnp_model_reduce_norm2_t*)super;
3483 return ccv_cnnp_reduce_norm2(self->axis, self->count, self->super.name);
3484}
3485
3486// MARK - Argmax Layer
3487
3488typedef struct {
3489 ccv_cnnp_model_t super;
3490 int axis;
3491 ccv_nnc_tensor_symbol_t output;
3492} ccv_cnnp_model_argmax_t;
3493
3494static 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)
3495{
3496 const ccv_cnnp_model_argmax_t* const self = (const ccv_cnnp_model_argmax_t*)super;
3497 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);
3498 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"
, 3498, __extension__ __PRETTY_FUNCTION__); }));
3499 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", 3499, __extension__ __PRETTY_FUNCTION__
); }));
3500 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3501 ccv_nnc_tensor_param_t output_params;
3502 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);
3503 argmax.info.reduce.axis[0] = self->axis;
3504 argmax.info.reduce.count = 1;
3505 ccv_nnc_hint_tensor_auto(argmax, &input_params, 1, ccv_nnc_no_hint, &output_params, 1);
3506 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3507 ccv_nnc_graph_exec_symbol_new(graph, argmax, inputs, input_size, outputs, output_size, "argmax");
3508}
3509
3510static ccv_cnnp_model_t* _ccv_cnnp_argmax_copy(const ccv_cnnp_model_t* const self, void* const context);
3511
3512static const ccv_cnnp_model_vtab_t ccv_cnnp_argmax_isa = {
3513 .build = _ccv_cnnp_argmax_build,
3514 .copy = _ccv_cnnp_argmax_copy,
3515};
3516
3517ccv_cnnp_model_t* ccv_cnnp_argmax(const int axis, const char* const name)
3518{
3519 ccv_cnnp_model_argmax_t* const model_argmax = (ccv_cnnp_model_argmax_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_argmax_t));
3520 model_argmax->super.isa = &ccv_cnnp_argmax_isa;
3521 model_argmax->super.input_size = 1;
3522 model_argmax->super.outputs = &model_argmax->output;
3523 model_argmax->super.output_size = 1;
3524 ccv_cnnp_model_copy_name(&model_argmax->super, name);
3525 model_argmax->axis = axis;
3526 return (ccv_cnnp_model_t*)model_argmax;
3527}
3528
3529static ccv_cnnp_model_t* _ccv_cnnp_argmax_copy(const ccv_cnnp_model_t* const super, void* const context)
3530{
3531 const ccv_cnnp_model_argmax_t* const self = (const ccv_cnnp_model_argmax_t*)super;
3532 return ccv_cnnp_argmax(self->axis, self->super.name);
3533}
3534
3535// MARK - Argmin Layer
3536
3537typedef struct {
3538 ccv_cnnp_model_t super;
3539 int axis;
3540 ccv_nnc_tensor_symbol_t output;
3541} ccv_cnnp_model_argmin_t;
3542
3543static 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)
3544{
3545 const ccv_cnnp_model_argmin_t* const self = (const ccv_cnnp_model_argmin_t*)super;
3546 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);
3547 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"
, 3547, __extension__ __PRETTY_FUNCTION__); }));
3548 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", 3548, __extension__ __PRETTY_FUNCTION__
); }));
3549 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3550 ccv_nnc_tensor_param_t output_params;
3551 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);
3552 argmin.info.reduce.axis[0] = self->axis;
3553 argmin.info.reduce.count = 1;
3554 ccv_nnc_hint_tensor_auto(argmin, &input_params, 1, ccv_nnc_no_hint, &output_params, 1);
3555 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3556 ccv_nnc_graph_exec_symbol_new(graph, argmin, inputs, input_size, outputs, output_size, "argmin");
3557}
3558
3559static ccv_cnnp_model_t* _ccv_cnnp_argmin_copy(const ccv_cnnp_model_t* const self, void* const context);
3560
3561static const ccv_cnnp_model_vtab_t ccv_cnnp_argmin_isa = {
3562 .build = _ccv_cnnp_argmin_build,
3563 .copy = _ccv_cnnp_argmin_copy,
3564};
3565
3566ccv_cnnp_model_t* ccv_cnnp_argmin(const int axis, const char* const name)
3567{
3568 ccv_cnnp_model_argmin_t* const model_argmin = (ccv_cnnp_model_argmin_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_argmin_t));
3569 model_argmin->super.isa = &ccv_cnnp_argmin_isa;
3570 model_argmin->super.input_size = 1;
3571 model_argmin->super.outputs = &model_argmin->output;
3572 model_argmin->super.output_size = 1;
3573 ccv_cnnp_model_copy_name(&model_argmin->super, name);
3574 model_argmin->axis = axis;
3575 return (ccv_cnnp_model_t*)model_argmin;
3576}
3577
3578static ccv_cnnp_model_t* _ccv_cnnp_argmin_copy(const ccv_cnnp_model_t* const super, void* const context)
3579{
3580 const ccv_cnnp_model_argmin_t* const self = (const ccv_cnnp_model_argmin_t*)super;
3581 return ccv_cnnp_argmin(self->axis, self->super.name);
3582}
3583
3584// MARK - Min Layer
3585
3586typedef struct {
3587 ccv_cnnp_model_t super;
3588 ccv_nnc_tensor_symbol_t output;
3589} ccv_cnnp_model_min_t;
3590
3591static 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)
3592{
3593 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);
3594 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"
, 3594, __extension__ __PRETTY_FUNCTION__); }));
3595 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", 3595, __extension__ __PRETTY_FUNCTION__
); }));
3596 ccv_nnc_tensor_param_t input_params[2];
3597 int i;
3598 for (i = 0; i < 2; i++)
3599 input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
3600 ccv_nnc_tensor_param_t output_params;
3601 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);
3602 ccv_nnc_hint_tensor_auto(min, input_params, 2, ccv_nnc_no_hint, &output_params, 1);
3603 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3604 ccv_nnc_graph_exec_symbol_new(graph, min, inputs, input_size, outputs, output_size, "min");
3605}
3606
3607static ccv_cnnp_model_t* _ccv_cnnp_min_copy(const ccv_cnnp_model_t* const self, void* const context);
3608
3609static const ccv_cnnp_model_vtab_t ccv_cnnp_min_isa = {
3610 .build = _ccv_cnnp_min_build,
3611 .copy = _ccv_cnnp_min_copy,
3612};
3613
3614ccv_cnnp_model_t* ccv_cnnp_min(const char* const name)
3615{
3616 ccv_cnnp_model_min_t* const model_min = (ccv_cnnp_model_min_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_min_t));
3617 model_min->super.isa = &ccv_cnnp_min_isa;
3618 model_min->super.input_size = 2;
3619 model_min->super.outputs = &model_min->output;
3620 model_min->super.output_size = 1;
3621 ccv_cnnp_model_copy_name(&model_min->super, name);
3622 return (ccv_cnnp_model_t*)model_min;
3623}
3624
3625static ccv_cnnp_model_t* _ccv_cnnp_min_copy(const ccv_cnnp_model_t* const super, void* const context)
3626{
3627 const ccv_cnnp_model_min_t* const self = (const ccv_cnnp_model_min_t*)super;
3628 return ccv_cnnp_min(self->super.name);
3629}
3630
3631// MARK - Max Layer
3632
3633typedef struct {
3634 ccv_cnnp_model_t super;
3635 ccv_nnc_tensor_symbol_t output;
3636} ccv_cnnp_model_max_t;
3637
3638static 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)
3639{
3640 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);
3641 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"
, 3641, __extension__ __PRETTY_FUNCTION__); }));
3642 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", 3642, __extension__ __PRETTY_FUNCTION__
); }));
3643 ccv_nnc_tensor_param_t input_params[2];
3644 int i;
3645 for (i = 0; i < 2; i++)
3646 input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
3647 ccv_nnc_tensor_param_t output_params;
3648 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);
3649 ccv_nnc_hint_tensor_auto(max, input_params, 2, ccv_nnc_no_hint, &output_params, 1);
3650 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3651 ccv_nnc_graph_exec_symbol_new(graph, max, inputs, input_size, outputs, output_size, "max");
3652}
3653
3654static ccv_cnnp_model_t* _ccv_cnnp_max_copy(const ccv_cnnp_model_t* const self, void* const context);
3655
3656static const ccv_cnnp_model_vtab_t ccv_cnnp_max_isa = {
3657 .build = _ccv_cnnp_max_build,
3658 .copy = _ccv_cnnp_max_copy,
3659};
3660
3661ccv_cnnp_model_t* ccv_cnnp_max(const char* const name)
3662{
3663 ccv_cnnp_model_max_t* const model_max = (ccv_cnnp_model_max_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_max_t));
3664 model_max->super.isa = &ccv_cnnp_max_isa;
3665 model_max->super.input_size = 2;
3666 model_max->super.outputs = &model_max->output;
3667 model_max->super.output_size = 1;
3668 ccv_cnnp_model_copy_name(&model_max->super, name);
3669 return (ccv_cnnp_model_t*)model_max;
3670}
3671
3672static ccv_cnnp_model_t* _ccv_cnnp_max_copy(const ccv_cnnp_model_t* const super, void* const context)
3673{
3674 const ccv_cnnp_model_max_t* const self = (const ccv_cnnp_model_max_t*)super;
3675 return ccv_cnnp_max(self->super.name);
3676}
3677
3678// MARK - LSTM Layer
3679
3680typedef struct {
3681 ccv_cnnp_model_t super;
3682 int masked;
3683 ccv_nnc_tensor_symbol_t output;
3684 ccv_nnc_tensor_symbol_t weights;
3685 ccv_nnc_tensor_symbol_t reserves;
3686 ccv_nnc_cmd_param_t params;
3687 ccv_nnc_graph_exec_symbol_t lstm;
3688} ccv_cnnp_model_lstm_t;
3689
3690static int _ccv_cnnp_lstm_weight_dim(int bidirectional, int num_layers, int input_size, int hidden_size, int proj_size, int bias)
3691{
3692 const int D = !!bidirectional + 1;
3693 if (hidden_size == proj_size)
3694 return (num_layers * (bias ? 8 : 0) + (num_layers - 1) * (hidden_size * 4 * D + hidden_size * 4) + input_size * 4 + hidden_size * 4) * D;
3695 else
3696 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;
3697}
3698
3699static 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)
3700{
3701 ccv_cnnp_model_lstm_t* const self = (ccv_cnnp_model_lstm_t*)super;
3702 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);
3703 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", 3703, __extension__ __PRETTY_FUNCTION__
); }));
3704 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", 3704, __extension__ __PRETTY_FUNCTION__
); }));
3705 const int proj_size = self->params.rnn.proj_size == 0 ? self->params.rnn.hidden_size : self->params.rnn.proj_size;
3706 ccv_nnc_tensor_param_t input_params[5];
3707 input_params[0]= ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3708 if (input_size == 2)
3709 input_params[1] = ccv_nnc_tensor_symbol_params(graph, inputs[1]);
3710 input_params[4] = input_params[0];
3711 memset(input_params[4].dim, 0, sizeof(input_params[4].dim));
3712 const int x_nd = ccv_nnc_tensor_nd(input_params[0].dim);
3713 const int feature_count = input_params[0].dim[x_nd - 1];
3714 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);
3715 input_params[4].dim[1] = self->params.rnn.hidden_size;
3716 const ccv_nnc_cmd_t lstm = ccv_nnc_cmd(CCV_NNC_LSTM_FORWARD, 0, self->params, 0);
3717 ccv_nnc_tensor_param_t output_params[4];
3718 ccv_nnc_hint_tensor_auto(lstm, input_params, 5, ccv_nnc_no_hint, output_params, 4);
3719 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0);
3720 if (!self->weights.graph)
3721 self->weights = ccv_nnc_tensor_symbol_new(graph, input_params[4], "weights");
3722 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"
, 3722, __extension__ __PRETTY_FUNCTION__); }));
3723 const ccv_nnc_tensor_symbol_t weights = ccv_cnnp_model_get_symbol(super, self->weights);
3724 if (!self->reserves.graph)
3725 self->reserves = ccv_nnc_tensor_symbol_new(graph, output_params[3], "reserves");
3726 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"
, 3726, __extension__ __PRETTY_FUNCTION__); }));
3727 const ccv_nnc_tensor_symbol_t reserves = ccv_cnnp_model_get_symbol(super, self->reserves);
3728 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
};
3729 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");
3730}
3731
3732static 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)
3733{
3734 ccv_cnnp_model_lstm_t* const self = (ccv_cnnp_model_lstm_t*)super;
3735 if (self->weights.graph)
3736 {
3737 const float stdv = 1.0 / sqrt(self->params.rnn.hidden_size);
3738 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);
3739 }
3740}
3741
3742static 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)
3743{
3744 ccv_cnnp_model_lstm_t* const self = (ccv_cnnp_model_lstm_t*)super;
3745 if (self->weights.graph)
3746 add_to_array(parameters, self->weights, is_trainable);
3747}
3748
3749static 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)
3750{
3751 ccv_cnnp_model_lstm_t* const self = (ccv_cnnp_model_lstm_t*)super;
3752 if (self->lstm.graph)
3753 {
3754 self->params.rnn.is_test = is_test;
3755 updater(context, self->lstm, ccv_nnc_cmd(CCV_NNC_LSTM_FORWARD, 0, self->params, 0), ccv_nnc_no_hint);
3756 }
3757}
3758
3759static ccv_cnnp_model_t* _ccv_cnnp_lstm_copy(const ccv_cnnp_model_t* const self, void* const context);
3760
3761static const ccv_cnnp_model_vtab_t ccv_cnnp_lstm_isa = {
3762 .build = _ccv_cnnp_lstm_build,
3763 .init_states = _ccv_cnnp_lstm_init_states,
3764 .add_to_parameter = _ccv_cnnp_lstm_add_to_parameter,
3765 .copy = _ccv_cnnp_lstm_copy,
3766 .set_is_test = _ccv_cnnp_lstm_set_is_test,
3767};
3768
3769ccv_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)
3770{
3771 ccv_cnnp_model_lstm_t* const model_lstm = (ccv_cnnp_model_lstm_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_lstm_t));
3772 model_lstm->super.isa = &ccv_cnnp_lstm_isa;
3773 model_lstm->super.input_size = masked ? 2 : 1;
3774 model_lstm->super.outputs = &model_lstm->output;
3775 model_lstm->super.output_size = 1;
3776 model_lstm->super.is_trainable = is_trainable;
3777 ccv_cnnp_model_copy_name(&model_lstm->super, name);
3778 model_lstm->masked = masked;
3779 model_lstm->weights.d = CCV_NNC_NO_TENSOR_SYMBOL;
3780 model_lstm->weights.graph = 0;
3781 model_lstm->params.rnn.hidden_size = hidden_size;
3782 model_lstm->params.rnn.proj_size = proj_size;
3783 model_lstm->params.rnn.num_layers = num_layers;
3784 model_lstm->params.rnn.bias = bias;
3785 model_lstm->params.rnn.batch_first = batch_first;
3786 model_lstm->params.rnn.bidirectional = bidirectional;
3787 model_lstm->params.rnn.dropout = dropout;
3788 return (ccv_cnnp_model_t*)model_lstm;
3789}
3790
3791static ccv_cnnp_model_t* _ccv_cnnp_lstm_copy(const ccv_cnnp_model_t* const super, void* const context)
3792{
3793 const ccv_cnnp_model_lstm_t* const self = (const ccv_cnnp_model_lstm_t*)super;
3794 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);
3795}
3796
3797/// MARK - Datatype conversion layer.
3798
3799typedef struct {
3800 ccv_cnnp_model_t super;
3801 ccv_nnc_tensor_symbol_t output;
3802 int datatype;
3803 int ref_to_last;
3804} ccv_cnnp_model_datatype_conversion_t;
3805
3806static 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)
3807{
3808 ccv_cnnp_model_datatype_conversion_t* const self = (ccv_cnnp_model_datatype_conversion_t*)super;
3809 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);
3810 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3811 if (self->ref_to_last)
3812 {
3813 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", 3813, __extension__ __PRETTY_FUNCTION__
); }));
3814 const ccv_nnc_tensor_param_t last_params = ccv_nnc_tensor_symbol_params(graph, inputs[input_size - 1]);
3815 params.datatype = last_params.datatype;
3816 } else
3817 params.datatype = self->datatype;
3818 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 3818, __extension__ __PRETTY_FUNCTION__
); }));
3819 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
3820 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);
3821}
3822
3823static ccv_cnnp_model_t* _ccv_cnnp_datatype_conversion_copy(const ccv_cnnp_model_t* const self, void* const context);
3824
3825static const ccv_cnnp_model_vtab_t ccv_cnnp_datatype_conversion_isa = {
3826 .build = _ccv_cnnp_datatype_conversion_build,
3827 .copy = _ccv_cnnp_datatype_conversion_copy,
3828};
3829
3830ccv_cnnp_model_t* ccv_cnnp_datatype_conversion(const int datatype, const int ref_to_last, const char* const name)
3831{
3832 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));
3833 model_datatype_conversion->super.isa = &ccv_cnnp_datatype_conversion_isa;
3834 model_datatype_conversion->super.input_size = 0;
3835 model_datatype_conversion->super.outputs = &model_datatype_conversion->output;
3836 model_datatype_conversion->super.output_size = 1;
3837 model_datatype_conversion->datatype = datatype;
3838 model_datatype_conversion->ref_to_last = ref_to_last;
3839 ccv_cnnp_model_copy_name(&model_datatype_conversion->super, name);
3840 return (ccv_cnnp_model_t*)model_datatype_conversion;
3841}
3842
3843static ccv_cnnp_model_t* _ccv_cnnp_datatype_conversion_copy(const ccv_cnnp_model_t* const super, void* const context)
3844{
3845 ccv_cnnp_model_datatype_conversion_t* const self = (ccv_cnnp_model_datatype_conversion_t*)super;
3846 return ccv_cnnp_datatype_conversion(self->datatype, self->ref_to_last, self->super.name);
3847}
3848
3849/// MARK - Clamp layer.
3850
3851typedef struct {
3852 ccv_cnnp_model_t super;
3853 ccv_nnc_tensor_symbol_t output;
3854 float min;
3855 float max;
3856} ccv_cnnp_model_clamp_t;
3857
3858static 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)
3859{
3860 ccv_cnnp_model_clamp_t* const self = (ccv_cnnp_model_clamp_t*)super;
3861 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);
3862 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3863 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", 3863, __extension__ __PRETTY_FUNCTION__
); }));
3864 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
3865 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);
3866}
3867
3868static ccv_cnnp_model_t* _ccv_cnnp_clamp_copy(const ccv_cnnp_model_t* const self, void* const context);
3869
3870static const ccv_cnnp_model_vtab_t ccv_cnnp_clamp_isa = {
3871 .build = _ccv_cnnp_clamp_build,
3872 .copy = _ccv_cnnp_clamp_copy,
3873};
3874
3875ccv_cnnp_model_t* ccv_cnnp_clamp(const float min, const float max, const char* const name)
3876{
3877 ccv_cnnp_model_clamp_t* const model_clamp = (ccv_cnnp_model_clamp_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_clamp_t));
3878 model_clamp->super.isa = &ccv_cnnp_clamp_isa;
3879 model_clamp->super.input_size = 0;
3880 model_clamp->super.outputs = &model_clamp->output;
3881 model_clamp->super.output_size = 1;
3882 model_clamp->min = min;
3883 model_clamp->max = max;
3884 ccv_cnnp_model_copy_name(&model_clamp->super, name);
3885 return (ccv_cnnp_model_t*)model_clamp;
3886}
3887
3888static ccv_cnnp_model_t* _ccv_cnnp_clamp_copy(const ccv_cnnp_model_t* const super, void* const context)
3889{
3890 ccv_cnnp_model_clamp_t* const self = (ccv_cnnp_model_clamp_t*)super;
3891 return ccv_cnnp_clamp(self->min, self->max, self->super.name);
3892}
3893
3894// MARK - Parameter Layer
3895
3896typedef struct {
3897 ccv_cnnp_model_t super;
3898 float init_bound;
3899 ccv_nnc_tensor_symbol_t weights;
3900 ccv_nnc_tensor_param_t weights_params;
3901 ccv_nnc_tensor_symbol_t output;
3902} ccv_cnnp_model_parameter_t;
3903
3904static 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)
3905{
3906 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);
3907 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", 3907, __extension__ __PRETTY_FUNCTION__
); }));
3908 ccv_cnnp_model_parameter_t* const self = (ccv_cnnp_model_parameter_t*)super;
3909 if (!self->weights.graph)
3910 self->weights = ccv_nnc_tensor_symbol_new(graph, self->weights_params, "weights");
3911 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"
, 3911, __extension__ __PRETTY_FUNCTION__); }));
3912 outputs[0] = ccv_cnnp_model_get_symbol(super, self->weights);
3913}
3914
3915static 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)
3916{
3917 ccv_cnnp_model_parameter_t* const self = (ccv_cnnp_model_parameter_t*)super;
3918 if (self->init_bound > 0)
3919 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);
3920 else
3921 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);
3922}
3923
3924static 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)
3925{
3926 ccv_cnnp_model_parameter_t* const self = (ccv_cnnp_model_parameter_t*)super;
3927 add_to_array(parameters, self->weights, is_trainable);
3928}
3929
3930static ccv_cnnp_model_t* _ccv_cnnp_parameter_copy(const ccv_cnnp_model_t* const super, void* const context);
3931
3932static const ccv_cnnp_model_vtab_t ccv_cnnp_parameter_isa = {
3933 .build = _ccv_cnnp_parameter_build,
3934 .init_states = _ccv_cnnp_parameter_init_states,
3935 .add_to_parameter = _ccv_cnnp_parameter_add_to_parameter,
3936 .copy = _ccv_cnnp_parameter_copy,
3937};
3938
3939ccv_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)
3940{
3941 ccv_cnnp_model_parameter_t* const model_parameter = (ccv_cnnp_model_parameter_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_parameter_t));
3942 model_parameter->super.isa = &ccv_cnnp_parameter_isa;
3943 model_parameter->super.input_size = 0;
3944 model_parameter->super.outputs = &model_parameter->output;
3945 model_parameter->super.output_size = 1;
3946 model_parameter->super.is_trainable = is_trainable;
3947 ccv_cnnp_model_copy_name(&model_parameter->super, name);
3948 model_parameter->weights.d = CCV_NNC_NO_TENSOR_SYMBOL;
3949 model_parameter->weights.graph = 0;
3950 model_parameter->weights_params = params;
3951 return (ccv_cnnp_model_t*)model_parameter;
3952}
3953
3954static ccv_cnnp_model_t* _ccv_cnnp_parameter_copy(const ccv_cnnp_model_t* const super, void* const context)
3955{
3956 const ccv_cnnp_model_parameter_t* const self = (const ccv_cnnp_model_parameter_t*)super;
3957 return ccv_cnnp_parameter(self->weights_params, self->init_bound, self->super.is_trainable, self->super.name);
3958}
3959
3960// MARK - Scalar Layer
3961
3962typedef struct {
3963 ccv_cnnp_model_t super;
3964 int type;
3965 int format;
3966 int datatype;
3967 float value;
3968 ccv_nnc_tensor_symbol_t output;
3969} ccv_cnnp_model_scalar_t;
3970
3971static 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)
3972{
3973 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);
3974 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", 3974, __extension__ __PRETTY_FUNCTION__
); }));
3975 ccv_cnnp_model_scalar_t* const self = (ccv_cnnp_model_scalar_t*)super;
3976 ccv_nnc_tensor_param_t params = {
3977 .type = self->type,
3978 .format = self->format,
3979 .datatype = self->datatype,
3980 .dim = {
3981 1
3982 }
3983 };
3984 if (input_size > 0)
3985 {
3986 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3987 params.type = input_params.type;
3988 params.format = input_params.format;
3989 params.datatype = input_params.datatype;
3990 }
3991 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
3992 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);
3993}
3994
3995static ccv_cnnp_model_t* _ccv_cnnp_scalar_copy(const ccv_cnnp_model_t* const super, void* const context);
3996
3997static const ccv_cnnp_model_vtab_t ccv_cnnp_scalar_isa = {
3998 .build = _ccv_cnnp_scalar_build,
3999 .copy = _ccv_cnnp_scalar_copy,
4000};
4001
4002ccv_cnnp_model_t* ccv_cnnp_scalar(const int type, const int format, const int datatype, const float value, const char* const name)
4003{
4004 ccv_cnnp_model_scalar_t* const model_scalar = (ccv_cnnp_model_scalar_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_scalar_t));
4005 model_scalar->super.isa = &ccv_cnnp_scalar_isa;
4006 model_scalar->super.input_size = 0;
4007 model_scalar->super.outputs = &model_scalar->output;
4008 model_scalar->super.output_size = 1;
4009 ccv_cnnp_model_copy_name(&model_scalar->super, name);
4010 model_scalar->type = type;
4011 model_scalar->format = format;
4012 model_scalar->datatype = datatype;
4013 model_scalar->value = value;
4014 return (ccv_cnnp_model_t*)model_scalar;
4015}
4016
4017static ccv_cnnp_model_t* _ccv_cnnp_scalar_copy(const ccv_cnnp_model_t* const super, void* const context)
4018{
4019 const ccv_cnnp_model_scalar_t* const self = (const ccv_cnnp_model_scalar_t*)super;
4020 return ccv_cnnp_scalar(self->type, self->format, self->datatype, self->value, self->super.name);
4021}
4022
4023// MARK - Variable Layer
4024
4025typedef struct {
4026 ccv_cnnp_model_t super;
4027 ccv_nnc_tensor_param_t params;
4028 ccv_nnc_tensor_symbol_t output;
4029} ccv_cnnp_model_variable_t;
4030
4031static 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)
4032{
4033 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);
4034 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"
, 4034, __extension__ __PRETTY_FUNCTION__); }));
4035 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", 4035, __extension__ __PRETTY_FUNCTION__
); }));
4036 ccv_cnnp_model_variable_t* const self = (ccv_cnnp_model_variable_t*)super;
4037 outputs[0] = ccv_nnc_tensor_symbol_new(graph, self->params, 0);
4038}
4039
4040static ccv_cnnp_model_t* _ccv_cnnp_variable_copy(const ccv_cnnp_model_t* const super, void* const context);
4041
4042static const ccv_cnnp_model_vtab_t ccv_cnnp_variable_isa = {
4043 .build = _ccv_cnnp_variable_build,
4044 .copy = _ccv_cnnp_variable_copy,
4045};
4046
4047ccv_cnnp_model_t* ccv_cnnp_variable(const ccv_nnc_tensor_param_t params, const char* const name)
4048{
4049 ccv_cnnp_model_variable_t* const model_variable = (ccv_cnnp_model_variable_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_variable_t));
4050 model_variable->super.isa = &ccv_cnnp_variable_isa;
4051 model_variable->super.input_size = 0;
4052 model_variable->super.outputs = &model_variable->output;
4053 model_variable->super.output_size = 1;
4054 ccv_cnnp_model_copy_name(&model_variable->super, name);
4055 model_variable->params = params;
4056 return (ccv_cnnp_model_t*)model_variable;
4057}
4058
4059static ccv_cnnp_model_t* _ccv_cnnp_variable_copy(const ccv_cnnp_model_t* const super, void* const context)
4060{
4061 const ccv_cnnp_model_variable_t* const self = (const ccv_cnnp_model_variable_t*)super;
4062 return ccv_cnnp_variable(self->params, self->super.name);
4063}
4064
4065// MARK - Send Layer
4066
4067typedef struct {
4068 ccv_cnnp_model_t super;
4069 int device_id;
4070 ccv_nnc_tensor_symbol_t output;
4071} ccv_cnnp_model_send_t;
4072
4073static 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)
4074{
4075 ccv_cnnp_model_send_t* const self = (ccv_cnnp_model_send_t*)super;
4076 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);
4077 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"
, 4077, __extension__ __PRETTY_FUNCTION__); }));
4078 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", 4078, __extension__ __PRETTY_FUNCTION__
); }));
4079 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4080 params.type = (params.type & ~0x3) | CCV_TENSOR_GPU_MEMORY;
4081 CCV_TENSOR_SET_DEVICE_ID(params.type, self->device_id)(params.type) = (((params.type) & ~0xfff00) | (((self->
device_id) & 0xfff) << 8));
4082 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4083 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");
4084}
4085
4086static ccv_cnnp_model_t* _ccv_cnnp_send_copy(const ccv_cnnp_model_t* const super, void* const context);
4087
4088static const ccv_cnnp_model_vtab_t ccv_cnnp_send_isa = {
4089 .build = _ccv_cnnp_send_build,
4090 .copy = _ccv_cnnp_send_copy,
4091};
4092
4093ccv_cnnp_model_t* ccv_cnnp_send(const int device_id, const char* const name)
4094{
4095 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", 4095, __extension__ __PRETTY_FUNCTION__
); }));
4096 ccv_cnnp_model_send_t* const model_send = (ccv_cnnp_model_send_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_send_t));
4097 model_send->super.isa = &ccv_cnnp_send_isa;
4098 model_send->super.input_size = 1;
4099 model_send->super.outputs = &model_send->output;
4100 model_send->super.output_size = 1;
4101 ccv_cnnp_model_copy_name(&model_send->super, name);
4102 model_send->device_id = device_id;
4103 return (ccv_cnnp_model_t*)model_send;
4104}
4105
4106static ccv_cnnp_model_t* _ccv_cnnp_send_copy(const ccv_cnnp_model_t* const super, void* const context)
4107{
4108 const ccv_cnnp_model_send_t* const self = (const ccv_cnnp_model_send_t*)super;
4109 return ccv_cnnp_send(self->device_id, self->super.name);
4110}
4111
4112// MARK - Move Layer
4113
4114typedef struct {
4115 ccv_cnnp_model_t super;
4116 ccv_nnc_tensor_symbol_t output;
4117} ccv_cnnp_model_move_t;
4118
4119static 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)
4120{
4121 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);
4122 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"
, 4122, __extension__ __PRETTY_FUNCTION__); }));
4123 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", 4123, __extension__ __PRETTY_FUNCTION__
); }));
4124 outputs[0] = inputs[1];
4125 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");
4126}
4127
4128static ccv_cnnp_model_t* _ccv_cnnp_move_copy(const ccv_cnnp_model_t* const super, void* const context);
4129
4130static const ccv_cnnp_model_vtab_t ccv_cnnp_move_isa = {
4131 .build = _ccv_cnnp_move_build,
4132 .copy = _ccv_cnnp_move_copy,
4133};
4134
4135ccv_cnnp_model_t* ccv_cnnp_move(const char* const name)
4136{
4137 ccv_cnnp_model_move_t* const model_move = (ccv_cnnp_model_move_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_move_t));
4138 model_move->super.isa = &ccv_cnnp_move_isa;
4139 model_move->super.input_size = 2;
4140 model_move->super.outputs = &model_move->output;
4141 model_move->super.output_size = 1;
4142 ccv_cnnp_model_copy_name(&model_move->super, name);
4143 return (ccv_cnnp_model_t*)model_move;
4144}
4145
4146static ccv_cnnp_model_t* _ccv_cnnp_move_copy(const ccv_cnnp_model_t* const super, void* const context)
4147{
4148 const ccv_cnnp_model_move_t* const self = (const ccv_cnnp_model_move_t*)super;
4149 return ccv_cnnp_move(self->super.name);
4150}
4151
4152// MARK - "Making" Contiguous Layer
4153
4154typedef struct {
4155 ccv_cnnp_model_t super;
4156 ccv_nnc_tensor_symbol_t output;
4157} ccv_cnnp_model_contiguous_t;
4158
4159static 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)
4160{
4161 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);
4162 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"
, 4162, __extension__ __PRETTY_FUNCTION__); }));
4163 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", 4163, __extension__ __PRETTY_FUNCTION__
); }));
4164 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4165 ccv_nnc_tensor_symbol_t to = ccv_nnc_tensor_symbol_alias_to(graph, inputs[0]);
4166 if (to.d == CCV_NNC_NO_TENSOR_SYMBOL) // If we are not reshape an alias, it is straightforward.
4167 {
4168 outputs[0] = inputs[0];
4169 return;
4170 }
4171 // Otherwise, we need to check its stride to know if it is contiguous.
4172 int old_stride[CCV_NNC_MAX_DIM_ALLOC(12)];
4173 ccv_nnc_tensor_symbol_alias_params(graph, inputs[0], 0, old_stride);
4174 // We identify permute by checking if the stride is not in descending order.
4175 // This also covered "permute" through reshape, rather than using ccv_cnnp_permute directly.
4176 if (ccv_nnc_is_tensor_stride_packed(old_stride, params.dim))
4177 {
4178 outputs[0] = inputs[0];
4179 return;
4180 }
4181 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4182 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");
4183 ccv_nnc_graph_exec_symbol_set_flags(graph, make_contiguous, CCV_NNC_GRAPH_EXEC_DISABLE_OPT);
4184}
4185
4186static ccv_cnnp_model_t* _ccv_cnnp_contiguous_copy(const ccv_cnnp_model_t* const super, void* const context);
4187
4188static const ccv_cnnp_model_vtab_t ccv_cnnp_contiguous_isa = {
4189 .build = _ccv_cnnp_contiguous_build,
4190 .copy = _ccv_cnnp_contiguous_copy,
4191};
4192
4193ccv_cnnp_model_t* ccv_cnnp_contiguous(const char* const name)
4194{
4195 ccv_cnnp_model_contiguous_t* const model_contiguous = (ccv_cnnp_model_contiguous_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_contiguous_t));
4196 model_contiguous->super.isa = &ccv_cnnp_contiguous_isa;
4197 model_contiguous->super.input_size = 1;
4198 model_contiguous->super.outputs = &model_contiguous->output;
4199 model_contiguous->super.output_size = 1;
4200 ccv_cnnp_model_copy_name(&model_contiguous->super, name);
4201 return (ccv_cnnp_model_t*)model_contiguous;
4202}
4203
4204static ccv_cnnp_model_t* _ccv_cnnp_contiguous_copy(const ccv_cnnp_model_t* const super, void* const context)
4205{
4206 const ccv_cnnp_model_contiguous_t* const self = (const ccv_cnnp_model_contiguous_t*)super;
4207 return ccv_cnnp_contiguous(self->super.name);
4208}
4209
4210// MARK - "Making" Copy Layer
4211
4212typedef struct {
4213 ccv_cnnp_model_t super;
4214 ccv_nnc_tensor_symbol_t output;
4215} ccv_cnnp_model_copy_t;
4216
4217static 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)
4218{
4219 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);
4220 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"
, 4220, __extension__ __PRETTY_FUNCTION__); }));
4221 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", 4221, __extension__ __PRETTY_FUNCTION__
); }));
4222 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4223 ccv_nnc_tensor_symbol_t to = ccv_nnc_tensor_symbol_alias_to(graph, inputs[0]);
4224 if (to.d == CCV_NNC_NO_TENSOR_SYMBOL) // If we are not reshape an alias, it is straightforward.
4225 {
4226 outputs[0] = inputs[0];
4227 return;
4228 }
4229 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4230 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");
4231 ccv_nnc_graph_exec_symbol_set_flags(graph, make_contiguous, CCV_NNC_GRAPH_EXEC_DISABLE_OPT);
4232}
4233
4234static ccv_cnnp_model_t* _ccv_cnnp_copy_copy(const ccv_cnnp_model_t* const super, void* const context);
4235
4236static const ccv_cnnp_model_vtab_t ccv_cnnp_copy_isa = {
4237 .build = _ccv_cnnp_copy_build,
4238 .copy = _ccv_cnnp_copy_copy,
4239};
4240
4241ccv_cnnp_model_t* ccv_cnnp_copy(const char* const name)
4242{
4243 ccv_cnnp_model_copy_t* const model_copy = (ccv_cnnp_model_copy_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_copy_t));
4244 model_copy->super.isa = &ccv_cnnp_copy_isa;
4245 model_copy->super.input_size = 1;
4246 model_copy->super.outputs = &model_copy->output;
4247 model_copy->super.output_size = 1;
4248 ccv_cnnp_model_copy_name(&model_copy->super, name);
4249 return (ccv_cnnp_model_t*)model_copy;
4250}
4251
4252static ccv_cnnp_model_t* _ccv_cnnp_copy_copy(const ccv_cnnp_model_t* const super, void* const context)
4253{
4254 const ccv_cnnp_model_copy_t* const self = (const ccv_cnnp_model_copy_t*)super;
4255 return ccv_cnnp_copy(self->super.name);
4256}
4257
4258// MARK - All-To-All Layer
4259
4260typedef struct {
4261 ccv_cnnp_model_t super;
4262 int axis;
4263 ccv_nnc_tensor_symbol_t outputs[1];
4264} ccv_cnnp_model_all_to_all_t;
4265
4266static 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)
4267{
4268 ccv_cnnp_model_all_to_all_t* const self = (ccv_cnnp_model_all_to_all_t*)super;
4269 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);
4270 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", 4270, __extension__ __PRETTY_FUNCTION__
); }));
4271 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", 4271, __extension__ __PRETTY_FUNCTION__
); }));
4272 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", 4272, __extension__ __PRETTY_FUNCTION__
); }));
4273 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", 4273, __extension__ __PRETTY_FUNCTION__
); }));
4274 const ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4275 const int nd = ccv_nnc_tensor_nd(input_params.dim);
4276 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", 4276, __extension__ __PRETTY_FUNCTION__
); }));
4277 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", 4277, __extension__ __PRETTY_FUNCTION__
); }));
4278 int i;
4279 for (i = 0; i < input_size; i++)
4280 {
4281 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
4282 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", 4282, __extension__ __PRETTY_FUNCTION__
); }));
4283 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", 4283, __extension__ __PRETTY_FUNCTION__
); }));
4284 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", 4284, __extension__ __PRETTY_FUNCTION__
); }));
4285 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", 4285, __extension__ __PRETTY_FUNCTION__
); }));
4286 outputs[i] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4287 }
4288 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");
4289}
4290
4291static ccv_cnnp_model_t* _ccv_cnnp_all_to_all_copy(const ccv_cnnp_model_t* const super, void* const context);
4292
4293static const ccv_cnnp_model_vtab_t ccv_cnnp_all_to_all_isa = {
4294 .build = _ccv_cnnp_all_to_all_build,
4295 .copy = _ccv_cnnp_all_to_all_copy,
4296};
4297
4298ccv_cnnp_model_t* ccv_cnnp_all_to_all(const int count, const int axis, const char* const name)
4299{
4300 assert(count > 0)((void) sizeof ((count > 0) ? 1 : 0), __extension__ ({ if (
count > 0) ; else __assert_fail ("count > 0", "ccv_cnnp_model_addons.c"
, 4300, __extension__ __PRETTY_FUNCTION__); }));
4301 assert(axis >= 0)((void) sizeof ((axis >= 0) ? 1 : 0), __extension__ ({ if (
axis >= 0) ; else __assert_fail ("axis >= 0", "ccv_cnnp_model_addons.c"
, 4301, __extension__ __PRETTY_FUNCTION__); }));
4302 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));
4303 model_all_to_all->super.isa = &ccv_cnnp_all_to_all_isa;
4304 model_all_to_all->super.input_size = count;
4305 model_all_to_all->super.outputs = model_all_to_all->outputs;
4306 model_all_to_all->super.output_size = count;
4307 model_all_to_all->axis = axis;
4308 ccv_cnnp_model_copy_name(&model_all_to_all->super, name);
4309 return (ccv_cnnp_model_t*)model_all_to_all;
4310}
4311
4312static ccv_cnnp_model_t* _ccv_cnnp_all_to_all_copy(const ccv_cnnp_model_t* const super, void* const context)
4313{
4314 const ccv_cnnp_model_all_to_all_t* const self = (const ccv_cnnp_model_all_to_all_t*)super;
4315 return ccv_cnnp_all_to_all(self->super.output_size, self->axis, self->super.name);
4316}
4317
4318// MARK - Scaled-Dot Product Attention Layer
4319
4320typedef struct {
4321 ccv_cnnp_model_t super;
4322 ccv_nnc_tensor_symbol_t output;
4323 ccv_nnc_tensor_symbol_t weights;
4324 ccv_nnc_tensor_symbol_t bias;
4325 float scale;
4326 int is_causal;
4327 int has_attn_mask;
4328 int flags;
4329 int fused_unify_head_weights;
4330 int no_bias;
4331} ccv_cnnp_model_scaled_dot_product_attention_t;
4332
4333static 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)
4334{
4335 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);
4336 assert(input_size == 3 || input_size == 4)((void) sizeof ((input_size == 3 || input_size == 4) ? 1 : 0)
, __extension__ ({ if (input_size == 3 || input_size == 4) ; else
__assert_fail ("input_size == 3 || input_size == 4", "ccv_cnnp_model_addons.c"
, 4336, __extension__ __PRETTY_FUNCTION__); }));
4337 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", 4337, __extension__ __PRETTY_FUNCTION__
); }));
4338 ccv_cnnp_model_scaled_dot_product_attention_t* const self = (ccv_cnnp_model_scaled_dot_product_attention_t*)super;
4339 const ccv_nnc_tensor_param_t q_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4340 const ccv_nnc_tensor_param_t k_params = ccv_nnc_tensor_symbol_params(graph, inputs[1]);
4341 const ccv_nnc_tensor_param_t v_params = ccv_nnc_tensor_symbol_params(graph, inputs[2]);
4342 const int v_nd = ccv_nnc_tensor_nd(v_params.dim);
4343 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", 4343, __extension__ __PRETTY_FUNCTION__
); }));
4344 const int hEv = (v_nd == 3 ? 1 : v_params.dim[2]) * v_params.dim[v_nd - 1];
4345 ccv_nnc_tensor_param_t weights_params = q_params;
4346 memset(weights_params.dim, 0, sizeof(weights_params.dim));
4347 weights_params.dim[0] = hEv;
4348 weights_params.dim[1] = hEv;
4349 ccv_nnc_tensor_param_t bias_params = q_params;
4350 memset(bias_params.dim, 0, sizeof(bias_params.dim));
4351 bias_params.dim[0] = hEv;
4352 ccv_nnc_cmd_t cmd = {0};
4353 cmd.cmd = CCV_NNC_SCALED_DOT_PRODUCT_ATTENTION_FORWARD;
4354 cmd.info.scaled_dot_product_attention.scale = self->scale;
4355 cmd.info.scaled_dot_product_attention.is_causal = self->is_causal;
4356 cmd.info.scaled_dot_product_attention.flags = self->flags;
4357 ccv_nnc_tensor_param_t output_params[3];
4358 ccv_nnc_tensor_symbol_t output;
4359 ccv_nnc_tensor_symbol_t saved_softmax_lse;
4360 ccv_nnc_tensor_symbol_t saved_v_proj = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
4361 ccv_nnc_tensor_symbol_t attn_mask = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
4362 ccv_nnc_tensor_symbol_t weights = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
4363 ccv_nnc_tensor_symbol_t bias = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
4364 if (self->has_attn_mask)
4365 attn_mask = inputs[3];
4366 if (self->fused_unify_head_weights)
4367 {
4368 if (!self->weights.graph)
4369 self->weights = ccv_nnc_tensor_symbol_new(graph, weights_params, "weights");
4370 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"
, 4370, __extension__ __PRETTY_FUNCTION__); }));
4371 weights = ccv_cnnp_model_get_symbol(super, self->weights);
4372 if (!self->no_bias)
4373 {
4374 if (!self->bias.graph)
4375 self->bias = ccv_nnc_tensor_symbol_new(graph, bias_params, "bias");
4376 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", 4376, __extension__ __PRETTY_FUNCTION__
); }));
4377 bias = ccv_cnnp_model_get_symbol(super, self->bias);
4378 }
4379 ccv_nnc_hint_tensor_auto(cmd, (ccv_nnc_tensor_param_t []){
4380 q_params,
4381 k_params,
4382 v_params,
4383 (ccv_nnc_tensor_param_t){},
4384 weights_params,
4385 bias_params,
4386 }, 6, ccv_nnc_no_hint, output_params, 3);
4387 output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0);
4388 saved_softmax_lse = ccv_nnc_tensor_symbol_new(graph, output_params[1], 0);
4389 saved_v_proj = ccv_nnc_tensor_symbol_new(graph, output_params[2], 0);
4390 } else {
4391 ccv_nnc_hint_tensor_auto(cmd, (ccv_nnc_tensor_param_t []){
4392 q_params,
4393 k_params,
4394 v_params,
4395 }, 3, ccv_nnc_no_hint, output_params, 2);
4396 output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0);
4397 saved_softmax_lse = ccv_nnc_tensor_symbol_new(graph, output_params[1], 0);
4398 }
4399 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");
4400 outputs[0] = output;
4401}
4402
4403static 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)
4404{
4405 ccv_cnnp_model_scaled_dot_product_attention_t* const self = (ccv_cnnp_model_scaled_dot_product_attention_t*)super;
4406 if (self->weights.graph)
4407 {
4408 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"
, 4408, __extension__ __PRETTY_FUNCTION__); }));
4409 const ccv_nnc_tensor_param_t weight_params = ccv_nnc_tensor_symbol_params(graph, self->weights);
4410 const int c = weight_params.dim[1];
4411 const float std = sqrtf(2) / sqrtf(c);
4412 const float bound = sqrtf(3) * std;
4413 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);
4414 if (self->bias.graph)
4415 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);
4416 }
4417}
4418
4419static 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)
4420{
4421 ccv_cnnp_model_scaled_dot_product_attention_t* const self = (ccv_cnnp_model_scaled_dot_product_attention_t*)super;
4422 if (self->weights.graph)
4423 {
4424 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"
, 4424, __extension__ __PRETTY_FUNCTION__); }));
4425 add_to_array(parameters, self->weights, is_trainable);
4426 if (self->bias.graph)
4427 add_to_array(parameters, self->bias, is_trainable);
4428 }
4429}
4430
4431static ccv_cnnp_model_t* _ccv_cnnp_scaled_dot_product_attention_copy(const ccv_cnnp_model_t* const super, void* const context);
4432
4433static const ccv_cnnp_model_vtab_t ccv_cnnp_scaled_dot_product_attention_isa = {
4434 .build = _ccv_cnnp_scaled_dot_product_attention_build,
4435 .copy = _ccv_cnnp_scaled_dot_product_attention_copy,
4436};
4437
4438static const ccv_cnnp_model_vtab_t ccv_cnnp_scaled_dot_product_attention_fused_isa = {
4439 .build = _ccv_cnnp_scaled_dot_product_attention_build,
4440 .init_states = _ccv_cnnp_scaled_dot_product_attention_init_states,
4441 .add_to_parameter = _ccv_cnnp_scaled_dot_product_attention_add_to_parameter,
4442 .copy = _ccv_cnnp_scaled_dot_product_attention_copy,
4443};
4444
4445ccv_cnnp_model_t* ccv_cnnp_scaled_dot_product_attention(const float scale, const int is_causal, const int has_attn_mask, const int flags, const int fused_unify_head_weights, const int no_bias, const int is_trainable, const char* const name)
4446{
4447 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));
4448 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;
4449 model_scaled_dot_product_attention->super.input_size = has_attn_mask ? 4 : 3;
4450 model_scaled_dot_product_attention->super.outputs = &model_scaled_dot_product_attention->output;
4451 model_scaled_dot_product_attention->super.output_size = 1;
4452 model_scaled_dot_product_attention->super.is_trainable = is_trainable;
4453 ccv_cnnp_model_copy_name(&model_scaled_dot_product_attention->super, name);
4454 model_scaled_dot_product_attention->weights.d = CCV_NNC_NO_TENSOR_SYMBOL;
4455 model_scaled_dot_product_attention->weights.graph = 0;
4456 model_scaled_dot_product_attention->bias.d = CCV_NNC_NO_TENSOR_SYMBOL;
4457 model_scaled_dot_product_attention->bias.graph = 0;
4458 model_scaled_dot_product_attention->scale = scale;
4459 model_scaled_dot_product_attention->is_causal = is_causal;
4460 model_scaled_dot_product_attention->has_attn_mask = has_attn_mask;
4461 model_scaled_dot_product_attention->flags = flags;
4462 model_scaled_dot_product_attention->fused_unify_head_weights = fused_unify_head_weights;
4463 model_scaled_dot_product_attention->no_bias = no_bias;
4464 return (ccv_cnnp_model_t*)model_scaled_dot_product_attention;
4465}
4466
4467static ccv_cnnp_model_t* _ccv_cnnp_scaled_dot_product_attention_copy(const ccv_cnnp_model_t* const super, void* const context)
4468{
4469 const ccv_cnnp_model_scaled_dot_product_attention_t* const self = (const ccv_cnnp_model_scaled_dot_product_attention_t*)super;
4470 return ccv_cnnp_scaled_dot_product_attention(self->scale, self->is_causal, self->has_attn_mask, self->flags, self->fused_unify_head_weights, self->no_bias, self->super.is_trainable, self->super.name);
4471}
4472
4473// MARK - Debug Layer
4474
4475typedef struct {
4476 ccv_cnnp_model_t super;
4477 ccv_nnc_tensor_symbol_t output;
4478 ccv_cnnp_model_debug_f debugger;
4479 ccv_cnnp_model_debug_context_deinit_f debug_deinit;
4480 ccv_cnnp_model_debug_context_copy_f debug_copy;
4481 void* debug_context;
4482} ccv_cnnp_model_debug_t;
4483
4484static 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)
4485{
4486 if (cmd.cmd == CCV_NNC_CUSTOM_BACKWARD)
4487 {
4488 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", 4488, __extension__ __PRETTY_FUNCTION__
); }));
4489 }
4490 ccv_cnnp_model_debug_t* const self = (ccv_cnnp_model_debug_t*)cmd.data;
4491 self->debugger(inputs, input_size, stream_context, self->debug_context);
4492 return CCV_NNC_EXEC_SUCCESS;
4493}
4494
4495static ccv_nnc_cmd_vtab_t ccv_cnnp_debug_exec_isa = {
4496 .exec = _ccv_cnnp_debug_exec
4497};
4498
4499static 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)
4500{
4501 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);
4502 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", 4502, __extension__ __PRETTY_FUNCTION__
); }));
4503 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", 4503, __extension__ __PRETTY_FUNCTION__
); }));
4504 ccv_nnc_tensor_symbol_t to = ccv_nnc_tensor_symbol_alias_to(graph, inputs[0]);
4505 ccv_nnc_tensor_param_t output_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4506 if (to.d == CCV_NNC_NO_TENSOR_SYMBOL) // If we are not reshape an alias, it is straightforward.
4507 {
4508 int ofs[CCV_NNC_MAX_DIM_ALLOC(12)] = {0};
4509 int stride[CCV_NNC_MAX_DIM_ALLOC(12)];
4510 ccv_nnc_tensor_get_stride(output_params.dim, stride);
4511 outputs[0] = ccv_nnc_tensor_symbol_alias_new(graph, inputs[0], ofs, stride, output_params, 0);
4512 } else {
4513 int old_ofs[CCV_NNC_MAX_DIM_ALLOC(12)];
4514 int old_stride[CCV_NNC_MAX_DIM_ALLOC(12)];
4515 ccv_nnc_tensor_symbol_alias_params(graph, inputs[0], old_ofs, old_stride);
4516 outputs[0] = ccv_nnc_tensor_symbol_alias_new(graph, to, old_ofs, old_stride, output_params, 0);
4517 }
4518 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);
4519 cmd.data = self;
4520 ccv_nnc_graph_exec_symbol_t make_debug = ccv_nnc_graph_exec_symbol_new(graph, cmd, inputs, input_size, outputs, 1, "debug");
4521 // Disable any optimizations.
4522 ccv_nnc_graph_exec_symbol_set_flags(graph, make_debug, CCV_NNC_GRAPH_EXEC_DISABLE_OPT);
4523}
4524
4525static void _ccv_cnnp_debug_deinit(ccv_cnnp_model_t* const super)
4526{
4527 const ccv_cnnp_model_debug_t* const self = (const ccv_cnnp_model_debug_t*)super;
4528 if (self->debug_deinit && self->debug_context)
4529 self->debug_deinit(self->debug_context);
4530}
4531
4532static ccv_cnnp_model_t* _ccv_cnnp_debug_copy(const ccv_cnnp_model_t* const super, void* const context);
4533
4534static const ccv_cnnp_model_vtab_t ccv_cnnp_debug_isa = {
4535 .build = _ccv_cnnp_debug_build,
4536 .deinit = _ccv_cnnp_debug_deinit,
4537 .copy = _ccv_cnnp_debug_copy,
4538};
4539
4540ccv_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)
4541{
4542 ccv_cnnp_model_debug_t* const model_debug = (ccv_cnnp_model_debug_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_debug_t));
4543 model_debug->super.isa = &ccv_cnnp_debug_isa;
4544 model_debug->super.input_size = 0;
4545 model_debug->super.outputs = &model_debug->output;
4546 model_debug->super.output_size = 1;
4547 model_debug->debugger = func;
4548 model_debug->debug_context = context;
4549 model_debug->debug_deinit = deinit;
4550 model_debug->debug_copy = copy;
4551 ccv_cnnp_model_copy_name(&model_debug->super, name);
4552 return (ccv_cnnp_model_t*)model_debug;
4553}
4554
4555static ccv_cnnp_model_t* _ccv_cnnp_debug_copy(const ccv_cnnp_model_t* const super, void* const context)
4556{
4557 const ccv_cnnp_model_debug_t* const self = (const ccv_cnnp_model_debug_t*)super;
4558 void* debug_context = self->debug_context;
4559 if (self->debug_copy && self->debug_context)
4560 debug_context = self->debug_copy(self->debug_context);
4561 return ccv_cnnp_debug(self->debugger, debug_context, self->debug_deinit, self->debug_copy, self->super.name);
4562}
4563
4564/// MARK - Sort layer.
4565
4566typedef struct {
4567 ccv_cnnp_model_t super;
4568 ccv_nnc_tensor_symbol_t outputs[2];
4569 int along_axis;
4570 int descending;
4571} ccv_cnnp_model_sort_t;
4572
4573static 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)
4574{
4575 ccv_cnnp_model_sort_t* const self = (ccv_cnnp_model_sort_t*)super;
4576 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);
4577 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4578 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", 4578, __extension__ __PRETTY_FUNCTION__
); }));
4579 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4580 params.datatype = CCV_32S;
4581 outputs[1] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4582 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");
4583}
4584
4585static ccv_cnnp_model_t* _ccv_cnnp_sort_copy(const ccv_cnnp_model_t* const self, void* const context);
4586
4587static const ccv_cnnp_model_vtab_t ccv_cnnp_sort_isa = {
4588 .build = _ccv_cnnp_sort_build,
4589 .copy = _ccv_cnnp_sort_copy,
4590};
4591
4592ccv_cnnp_model_t* ccv_cnnp_sort(const int along_axis, const int descending, const char* const name)
4593{
4594 ccv_cnnp_model_sort_t* const model_sort = (ccv_cnnp_model_sort_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_sort_t));
4595 model_sort->super.isa = &ccv_cnnp_sort_isa;
4596 model_sort->super.input_size = 0;
4597 model_sort->super.outputs = model_sort->outputs;
4598 model_sort->super.output_size = 2;
4599 model_sort->along_axis = along_axis;
4600 model_sort->descending = descending;
4601 ccv_cnnp_model_copy_name(&model_sort->super, name);
4602 return (ccv_cnnp_model_t*)model_sort;
4603}
4604
4605static ccv_cnnp_model_t* _ccv_cnnp_sort_copy(const ccv_cnnp_model_t* const super, void* const context)
4606{
4607 ccv_cnnp_model_sort_t* const self = (ccv_cnnp_model_sort_t*)super;
4608 return ccv_cnnp_sort(self->along_axis, self->descending, self->super.name);
4609}
4610
4611/// MARK - Partition layer.
4612
4613typedef struct {
4614 ccv_cnnp_model_t super;
4615 ccv_nnc_tensor_symbol_t outputs[2];
4616 int kth;
4617 int along_axis;
4618 int descending;
4619} ccv_cnnp_model_partition_t;
4620
4621static 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)
4622{
4623 ccv_cnnp_model_partition_t* const self = (ccv_cnnp_model_partition_t*)super;
4624 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);
4625 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4626 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", 4626, __extension__ __PRETTY_FUNCTION__
); }));
4627 if (self->kth > 0)
4628 params.dim[self->along_axis] = self->kth;
4629 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4630 params.datatype = CCV_32S;
4631 outputs[1] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4632 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");
4633}
4634
4635static ccv_cnnp_model_t* _ccv_cnnp_partition_copy(const ccv_cnnp_model_t* const self, void* const context);
4636
4637static const ccv_cnnp_model_vtab_t ccv_cnnp_partition_isa = {
4638 .build = _ccv_cnnp_partition_build,
4639 .copy = _ccv_cnnp_partition_copy,
4640};
4641
4642ccv_cnnp_model_t* ccv_cnnp_partition(const int kth, const int along_axis, const int descending, const char* const name)
4643{
4644 ccv_cnnp_model_partition_t* const model_partition = (ccv_cnnp_model_partition_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_partition_t));
4645 model_partition->super.isa = &ccv_cnnp_partition_isa;
4646 model_partition->super.input_size = 0;
4647 model_partition->super.outputs = model_partition->outputs;
4648 model_partition->super.output_size = 2;
4649 model_partition->kth = kth;
4650 model_partition->along_axis = along_axis;
4651 model_partition->descending = descending;
4652 ccv_cnnp_model_copy_name(&model_partition->super, name);
4653 return (ccv_cnnp_model_t*)model_partition;
4654}
4655
4656static ccv_cnnp_model_t* _ccv_cnnp_partition_copy(const ccv_cnnp_model_t* const super, void* const context)
4657{
4658 ccv_cnnp_model_partition_t* const self = (ccv_cnnp_model_partition_t*)super;
4659 return ccv_cnnp_partition(self->kth, self->along_axis, self->descending, self->super.name);
4660}
4661
4662/// MARK - Unique consecutive layer.
4663
4664typedef struct {
4665 ccv_cnnp_model_t super;
4666 ccv_nnc_tensor_symbol_t outputs[2];
4667 int bincount;
4668} ccv_cnnp_model_unique_consecutive_t;
4669
4670static 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)
4671{
4672 ccv_cnnp_model_unique_consecutive_t* const self = (ccv_cnnp_model_unique_consecutive_t*)super;
4673 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);
4674 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4675 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", 4675, __extension__ __PRETTY_FUNCTION__
); }));
4676 if (self->bincount > 0)
4677 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; }
);
4678 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4679 params.datatype = CCV_32S;
4680 outputs[1] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4681 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");
4682}
4683
4684static ccv_cnnp_model_t* _ccv_cnnp_unique_consecutive_copy(const ccv_cnnp_model_t* const self, void* const context);
4685
4686static const ccv_cnnp_model_vtab_t ccv_cnnp_unique_consecutive_isa = {
4687 .build = _ccv_cnnp_unique_consecutive_build,
4688 .copy = _ccv_cnnp_unique_consecutive_copy,
4689};
4690
4691ccv_cnnp_model_t* ccv_cnnp_unique_consecutive(const int bincount, const char* const name)
4692{
4693 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));
4694 model_unique_consecutive->super.isa = &ccv_cnnp_unique_consecutive_isa;
4695 model_unique_consecutive->super.input_size = 0;
4696 model_unique_consecutive->super.outputs = model_unique_consecutive->outputs;
4697 model_unique_consecutive->super.output_size = 2;
4698 model_unique_consecutive->bincount = bincount;
4699 ccv_cnnp_model_copy_name(&model_unique_consecutive->super, name);
4700 return (ccv_cnnp_model_t*)model_unique_consecutive;
4701}
4702
4703static ccv_cnnp_model_t* _ccv_cnnp_unique_consecutive_copy(const ccv_cnnp_model_t* const super, void* const context)
4704{
4705 ccv_cnnp_model_unique_consecutive_t* const self = (ccv_cnnp_model_unique_consecutive_t*)super;
4706 return ccv_cnnp_unique_consecutive(self->bincount, self->super.name);
4707}
4708
4709/// MARK - Scatter add layer.
4710
4711typedef struct {
4712 ccv_cnnp_model_t super;
4713 ccv_nnc_tensor_symbol_t output;
4714 int bincount;
4715} ccv_cnnp_model_scatter_add_t;
4716
4717static 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)
4718{
4719 ccv_cnnp_model_scatter_add_t* const self = (ccv_cnnp_model_scatter_add_t*)super;
4720 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);
4721 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4722 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", 4722, __extension__ __PRETTY_FUNCTION__
); }));
4723 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", 4723, __extension__ __PRETTY_FUNCTION__
); }));
4724 params.dim[0] = self->bincount;
4725 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4726 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");
4727}
4728
4729static ccv_cnnp_model_t* _ccv_cnnp_scatter_add_copy(const ccv_cnnp_model_t* const self, void* const context);
4730
4731static const ccv_cnnp_model_vtab_t ccv_cnnp_scatter_add_isa = {
4732 .build = _ccv_cnnp_scatter_add_build,
4733 .copy = _ccv_cnnp_scatter_add_copy,
4734};
4735
4736ccv_cnnp_model_t* ccv_cnnp_scatter_add(const int bincount, const char* const name)
4737{
4738 assert(bincount > 0)((void) sizeof ((bincount > 0) ? 1 : 0), __extension__ ({ if
(bincount > 0) ; else __assert_fail ("bincount > 0", "ccv_cnnp_model_addons.c"
, 4738, __extension__ __PRETTY_FUNCTION__); }));
4739 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));
4740 model_scatter_add->super.isa = &ccv_cnnp_scatter_add_isa;
4741 model_scatter_add->super.input_size = 0;
4742 model_scatter_add->super.outputs = &model_scatter_add->output;
4743 model_scatter_add->super.output_size = 1;
4744 model_scatter_add->bincount = bincount;
4745 ccv_cnnp_model_copy_name(&model_scatter_add->super, name);
4746 return (ccv_cnnp_model_t*)model_scatter_add;
4747}
4748
4749static ccv_cnnp_model_t* _ccv_cnnp_scatter_add_copy(const ccv_cnnp_model_t* const super, void* const context)
4750{
4751 ccv_cnnp_model_scatter_add_t* const self = (ccv_cnnp_model_scatter_add_t*)super;
4752 return ccv_cnnp_scatter_add(self->bincount, self->super.name);
4753}
4754
4755// MARK - Segmented Dense Layer
4756
4757typedef struct {
4758 ccv_cnnp_model_t super;
4759 ccv_nnc_tensor_symbol_t output;
4760 ccv_nnc_tensor_symbol_t weights;
4761 ccv_nnc_tensor_symbol_t bias;
4762 int segments;
4763 int count;
4764 int no_bias;
4765 int flags;
4766} ccv_cnnp_model_segmented_dense_t;
4767
4768static 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)
4769{
4770 ccv_cnnp_model_segmented_dense_t* const self = (ccv_cnnp_model_segmented_dense_t*)super;
4771 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);
4772 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"
, 4772, __extension__ __PRETTY_FUNCTION__); }));
4773 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", 4773, __extension__ __PRETTY_FUNCTION__
); }));
4774 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4775 const ccv_nnc_tensor_param_t indices_params = ccv_nnc_tensor_symbol_params(graph, inputs[1]);
4776 const ccv_nnc_tensor_param_t counts_params = ccv_nnc_tensor_symbol_params(graph, inputs[2]);
4777 ccv_nnc_tensor_param_t weights_params = params;
4778 memset(weights_params.dim, 0, sizeof(weights_params.dim));
4779 weights_params.dim[0] = self->segments;
4780 weights_params.dim[1] = self->count;
4781 weights_params.dim[2] = params.dim[ccv_nnc_tensor_nd(params.dim) - 1];
4782 if (!self->weights.graph)
4783 self->weights = ccv_nnc_tensor_symbol_new(graph, weights_params, "weights");
4784 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"
, 4784, __extension__ __PRETTY_FUNCTION__); }));
4785 const ccv_nnc_tensor_symbol_t weights = ccv_cnnp_model_get_symbol(super, self->weights);
4786 ccv_nnc_tensor_param_t bias_params = params;
4787 memset(bias_params.dim, 0, sizeof(bias_params.dim));
4788 bias_params.dim[0] = self->segments;
4789 bias_params.dim[1] = self->count;
4790 ccv_nnc_cmd_t cmd = {0};
4791 cmd.cmd = CCV_NNC_SEGMENTED_GEMM_FORWARD;
4792 cmd.info.blas.a[0] = 1;
4793 cmd.info.blas.a[1] = 1;
4794 cmd.info.blas.transpose_b[0] = 1;
4795 cmd.info.blas.transpose_b[1] = 2;
4796 cmd.info.blas.flags = self->flags;
4797 ccv_nnc_tensor_param_t output_params;
4798 ccv_nnc_hint_tensor_auto(cmd, (ccv_nnc_tensor_param_t []){
4799 params, indices_params, counts_params,
4800 weights_params,
4801 bias_params,
4802 }, 5, ccv_nnc_no_hint, &output_params, 1);
4803 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
4804 if (self->no_bias)
4805 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");
4806 else {
4807 if (!self->bias.graph)
4808 self->bias = ccv_nnc_tensor_symbol_new(graph, bias_params, "bias");
4809 const ccv_nnc_tensor_symbol_t bias = ccv_cnnp_model_get_symbol(super, self->bias);
4810 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");
4811 }
4812 outputs[0] = output;
4813}
4814
4815static 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)
4816{
4817 ccv_cnnp_model_segmented_dense_t* const self = (ccv_cnnp_model_segmented_dense_t*)super;
4818 const ccv_nnc_tensor_param_t weight_params = ccv_nnc_tensor_symbol_params(graph, self->weights);
4819 const int c = weight_params.dim[1];
4820 const float std = sqrtf(2) / sqrtf(c);
4821 const float bound = sqrtf(3) * std;
4822 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);
4823 if (self->bias.graph)
4824 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);
4825}
4826
4827static 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)
4828{
4829 ccv_cnnp_model_segmented_dense_t* const self = (ccv_cnnp_model_segmented_dense_t*)super;
4830 add_to_array(parameters, self->weights, is_trainable);
4831 if (self->bias.graph)
4832 add_to_array(parameters, self->bias, is_trainable);
4833}
4834
4835static ccv_cnnp_model_t* _ccv_cnnp_segmented_dense_copy(const ccv_cnnp_model_t* const super, void* const context);
4836
4837static const ccv_cnnp_model_vtab_t ccv_cnnp_segmented_dense_isa = {
4838 .build = _ccv_cnnp_segmented_dense_build,
4839 .init_states = _ccv_cnnp_segmented_dense_init_states,
4840 .add_to_parameter = _ccv_cnnp_segmented_dense_add_to_parameter,
4841 .copy = _ccv_cnnp_segmented_dense_copy,
4842};
4843
4844ccv_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)
4845{
4846 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));
4847 model_segmented_dense->super.isa = &ccv_cnnp_segmented_dense_isa;
4848 model_segmented_dense->super.input_size = 3;
4849 model_segmented_dense->super.outputs = &model_segmented_dense->output;
4850 model_segmented_dense->super.output_size = 1;
4851 model_segmented_dense->super.is_trainable = is_trainable;
4852 ccv_cnnp_model_copy_name(&model_segmented_dense->super, name);
4853 model_segmented_dense->weights.d = CCV_NNC_NO_TENSOR_SYMBOL;
4854 model_segmented_dense->weights.graph = 0;
4855 model_segmented_dense->bias.d = CCV_NNC_NO_TENSOR_SYMBOL;
4856 model_segmented_dense->bias.graph = 0;
4857 model_segmented_dense->segments = segments;
4858 model_segmented_dense->count = count;
4859 model_segmented_dense->no_bias = no_bias;
4860 model_segmented_dense->flags = flags;
4861 return (ccv_cnnp_model_t*)model_segmented_dense;
4862}
4863
4864static ccv_cnnp_model_t* _ccv_cnnp_segmented_dense_copy(const ccv_cnnp_model_t* const super, void* const context)
4865{
4866 const ccv_cnnp_model_segmented_dense_t* const self = (const ccv_cnnp_model_segmented_dense_t*)super;
4867 return ccv_cnnp_segmented_dense(self->segments, self->count, self->no_bias, self->flags, self->super.is_trainable, self->super.name);
4868}