Bug Summary

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

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -cc1 -triple x86_64-unknown-linux-gnu -analyze -disable-free -clear-ast-before-backend -disable-llvm-verifier -discard-value-names -main-file-name ccv_cnnp_model_addons.c -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -setup-static-analyzer -mrelocation-model pic -pic-level 2 -pic-is-pie -mframe-pointer=none -fmath-errno -ffp-contract=on -fno-rounding-math -mconstructor-aliases -funwind-tables=2 -target-cpu x86-64 -target-feature +sse2 -tune-cpu generic -debugger-tuning=gdb -fdebug-compilation-dir=/home/liu/actions-runner/_work/ccv/ccv/lib/nnc -fcoverage-compilation-dir=/home/liu/actions-runner/_work/ccv/ccv/lib/nnc -resource-dir /usr/local/lib/clang/19 -I ../ -I /usr/local/cuda/include -D HAVE_CBLAS -D HAVE_LIBPNG -D HAVE_LIBJPEG -D HAVE_FFTW3 -D HAVE_PTHREAD -D HAVE_LIBLINEAR -D HAVE_TESSERACT -D HAVE_AVCODEC -D HAVE_AVFORMAT -D HAVE_AVUTIL -D HAVE_SWSCALE -D HAVE_SSE2 -D HAVE_GSL -D HAVE_CUDA -D HAVE_CUDNN -D HAVE_NCCL -D USE_SYSTEM_CUB -I /usr/local/include -internal-isystem /usr/local/lib/clang/19/include -internal-isystem /usr/local/include -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/12/../../../../x86_64-linux-gnu/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -O3 -ferror-limit 19 -fgnuc-version=4.2.1 -fskip-odr-check-in-gmf -vectorize-loops -vectorize-slp -analyzer-output=html -faddrsig -D__GCC_HAVE_DWARF2_CFI_ASM=1 -o /home/liu/actions-runner/_work/ccv/ccv/_analyze/2026-05-02-115646-1519401-1 -x c ccv_cnnp_model_addons.c
1#include "ccv_nnc.h"
2#include "ccv_nnc_easy.h"
3#include "ccv_nnc_internal.h"
4#include "ccv_internal.h"
5#include "_ccv_cnnp_model.h"
6
7// MARK - Add-on Functions
8
9static int _ccv_cnnp_model_clip_grad_norm_reduce_norm2(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint, const int flags, ccv_nnc_tensor_t* const* const inputs, const int input_size, ccv_nnc_tensor_t* const* const outputs, const int output_size, ccv_nnc_stream_context_t* const stream_context)
10{
11 const int device_id = CCV_TENSOR_GET_DEVICE_ID(inputs[0]->info.type)(((inputs[0]->info.type) & 0xfff00) >> 8);
12 ccv_nnc_tensor_t* const old_norm2 = outputs[1 + device_id * 2];
13 ccv_nnc_tensor_t* const norm2 = outputs[1 + device_id * 2 + 1];
14 const int tensor_count = ccv_nnc_tensor_count(inputs[0]->info);
15 if (tensor_count == 1)
16 ccv_nnc_cmd_exec(CMD_MUL_FORWARD(1)ccv_nnc_cmd(CCV_NNC_MUL_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={1,}}}, 0), hint, flags, TENSOR_LIST(inputs[0], inputs[0])(ccv_nnc_tensor_t* []){inputs[0], inputs[0]}, (1 +1 +1 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(norm2)(ccv_nnc_tensor_t* []){norm2}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context);
17 else {
18 ccv_nnc_cmd_exec(CMD_REDUCE_NORM2_FORWARD()ccv_nnc_cmd(CCV_NNC_REDUCE_NORM2_FORWARD, 0, ((ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.reduce={.count=(1 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1),.axis={}}}
), 0), hint, flags, TENSOR_LIST(inputs[0])(ccv_nnc_tensor_t* []){inputs[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(norm2)(ccv_nnc_tensor_t* []){norm2}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context);
19 ccv_nnc_cmd_exec(CMD_MUL_FORWARD(1)ccv_nnc_cmd(CCV_NNC_MUL_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={1,}}}, 0), hint, flags, TENSOR_LIST(norm2, norm2)(ccv_nnc_tensor_t* []){norm2, norm2}, (1 +1 +1 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(norm2)(ccv_nnc_tensor_t* []){norm2}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context);
20 }
21 ccv_nnc_cmd_exec(CMD_ADD_FORWARD(1, 1)ccv_nnc_cmd(CCV_NNC_ADD_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={1, 1}}}, 0), hint, flags, TENSOR_LIST(old_norm2, norm2)(ccv_nnc_tensor_t* []){old_norm2, norm2}, (1 +1 +1 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(old_norm2)(ccv_nnc_tensor_t* []){old_norm2}, (1 +1 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context);
22 return CCV_NNC_EXEC_SUCCESS;
23}
24
25static ccv_nnc_cmd_vtab_t clip_grad_norm_reduce_norm2_vtab = {
26 .exec = _ccv_cnnp_model_clip_grad_norm_reduce_norm2
27};
28
29static int _ccv_cnnp_model_clip_grad_norm_scatter_norm2(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint, const int flags, ccv_nnc_tensor_t* const* const inputs, const int input_size, ccv_nnc_tensor_t* const* const outputs, const int output_size, ccv_nnc_stream_context_t* const stream_context)
30{
31 const int device_id = CCV_TENSOR_GET_DEVICE_ID(inputs[0]->info.type)(((inputs[0]->info.type) & 0xfff00) >> 8);
32 ccv_nnc_tensor_t* const norm2 = inputs[1 + device_id * 2];
33 ccv_nnc_cmd_exec(CMD_MUL_FORWARD(1)ccv_nnc_cmd(CCV_NNC_MUL_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={1,}}}, 0), hint, flags, TENSOR_LIST(inputs[0], norm2)(ccv_nnc_tensor_t* []){inputs[0], norm2}, (1 +1 +1 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(outputs[0])(ccv_nnc_tensor_t* []){outputs[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context);
34 return CCV_NNC_EXEC_SUCCESS;
35}
36
37static ccv_nnc_cmd_vtab_t clip_grad_norm_scatter_norm2_vtab = {
38 .exec = _ccv_cnnp_model_clip_grad_norm_scatter_norm2
39};
40
41void ccv_cnnp_model_parameters_clip_grad_norm(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t parameters, int norm_type, float max_norm, ccv_nnc_stream_context_t* const stream_context)
42{
43 assert(norm_type == 2)((void) sizeof ((norm_type == 2) ? 1 : 0), __extension__ ({ if
(norm_type == 2) ; else __assert_fail ("norm_type == 2", "ccv_cnnp_model_addons.c"
, 43, __extension__ __PRETTY_FUNCTION__); }));
44 ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
45 assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
(compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model_addons.c"
, 45, __extension__ __PRETTY_FUNCTION__); }));
46 const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count
); typeof (1) _b = (1); (_a > _b) ? _a : _b; });
47 ccv_nnc_tensor_t* norm2[parallel_count * 2];
48 ccv_nnc_tensor_t* max_normt[parallel_count];
49 const int stream_type = model->compiled_data->stream_type;
50 int i;
51 if (stream_type == CCV_STREAM_CONTEXT_GPU)
52 {
53 for (i = 0; i < parallel_count; i++)
54 {
55 ccv_nnc_tensor_param_t info = {
56 .type = CCV_TENSOR_GPU_MEMORY,
57 .format = CCV_TENSOR_FORMAT_NHWC,
58 .datatype = CCV_32F,
59 .dim = {1},
60 };
61 CCV_TENSOR_SET_DEVICE_ID(info.type, i)(info.type) = (((info.type) & ~0xfff00) | (((i) & 0xfff
) << 8));
62 norm2[i * 2] = ccv_nnc_tensor_new(ccv_nnc_xpu_alloc(&compiled_data->xpu_alloc, i, stream_context, ccv_nnc_tensor_data_size(info)), info, 0);
63 norm2[i * 2 + 1] = ccv_nnc_tensor_new(ccv_nnc_xpu_alloc(&compiled_data->xpu_alloc, i, stream_context, ccv_nnc_tensor_data_size(info)), info, 0);
64 max_normt[i] = ccv_nnc_tensor_new(ccv_nnc_xpu_alloc(&compiled_data->xpu_alloc, i, stream_context, ccv_nnc_tensor_data_size(info)), info, 0);
65 }
66 } else {
67 for (i = 0; i < parallel_count; i++)
68 {
69 ccv_nnc_tensor_param_t info = {
70 .type = CCV_TENSOR_CPU_MEMORY,
71 .format = CCV_TENSOR_FORMAT_NHWC,
72 .datatype = CCV_32F,
73 .dim = {1},
74 };
75 norm2[i * 2] = ccv_nnc_tensor_new(0, info, 0);
76 norm2[i * 2 + 1] = ccv_nnc_tensor_new(0, info, 0);
77 max_normt[i] = ccv_nnc_tensor_new(0, info, 0);
78 }
79 }
80 // zero out old norm2.
81 if (parallel_count > 1)
82 {
83 ccv_nnc_stream_context_t* streams[parallel_count];
84 ccv_nnc_stream_signal_t* signal;
85 if (stream_context)
86 signal = ccv_nnc_stream_context_emit_signal_new(stream_context);
87 for (i = 0; i < parallel_count; i++)
88 {
89 const int stream_type = CCV_TENSOR_GET_MEMORY(norm2[i * 2]->info.type)((norm2[i * 2]->info.type) & 0x3) == CCV_TENSOR_GPU_MEMORY ? CCV_STREAM_CONTEXT_GPU : CCV_STREAM_CONTEXT_CPU;
90 const int device_id = CCV_TENSOR_GET_DEVICE_ID(norm2[i * 2]->info.type)(((norm2[i * 2]->info.type) & 0xfff00) >> 8);
91 int type = stream_type;
92 CCV_STREAM_SET_DEVICE_ID(type, device_id)(type) = (((type) & ~0xfff00) | (((device_id) & 0xfff
) << 8));
93 ccv_nnc_stream_context_t* const stream_0 = ccv_cnnp_compiled_data_get_stream(compiled_data, type);
94 // Wait signal to finish.
95 if (stream_context)
96 ccv_nnc_stream_context_wait_signal(stream_0, signal);
97 ccv_nnc_cmd_exec(CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, 0, TENSOR_LIST(norm2[i * 2])(ccv_nnc_tensor_t* []){norm2[i * 2]}, (1 +1 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_0);
98 if (stream_context)
99 {
100 ccv_nnc_stream_signal_t* const signal = ccv_nnc_stream_context_emit_signal_new(stream_0);
101 ccv_nnc_stream_context_wait_signal(stream_context, signal);
102 }
103 streams[i] = stream_0;
104 }
105 // If this should be blocking, blocking it.
106 if (!stream_context)
107 for (i = 0; i < parallel_count; i++)
108 if (streams[i])
109 ccv_nnc_stream_context_wait(streams[i]);
110 } else {
111 ccv_nnc_cmd_exec(CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, 0, TENSOR_LIST(norm2[0])(ccv_nnc_tensor_t* []){norm2[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context);
112 }
113 // Gather norm2.
114 ccv_nnc_cmd_t reduce_cmd = {
115 .cmd = CCV_NNC_CUSTOM_FORWARD,
116 .isa = &clip_grad_norm_reduce_norm2_vtab,
117 };
118 ccv_cnnp_model_parameter_gradients_map(model, parameters, reduce_cmd, ccv_nnc_no_hint, 0, 0, 0, norm2, parallel_count * 2, stream_context);
119 // Now compute max(max_norm / norm2, 1.0).
120 if (parallel_count > 1)
121 {
122 ccv_nnc_stream_context_t* streams[parallel_count];
123 ccv_nnc_stream_signal_t* signal;
124 if (stream_context)
125 signal = ccv_nnc_stream_context_emit_signal_new(stream_context);
126 for (i = 0; i < parallel_count; i++)
127 {
128 const int stream_type = CCV_TENSOR_GET_MEMORY(norm2[i * 2]->info.type)((norm2[i * 2]->info.type) & 0x3) == CCV_TENSOR_GPU_MEMORY ? CCV_STREAM_CONTEXT_GPU : CCV_STREAM_CONTEXT_CPU;
129 const int device_id = CCV_TENSOR_GET_DEVICE_ID(norm2[i * 2]->info.type)(((norm2[i * 2]->info.type) & 0xfff00) >> 8);
130 int type = stream_type;
131 CCV_STREAM_SET_DEVICE_ID(type, device_id)(type) = (((type) & ~0xfff00) | (((device_id) & 0xfff
) << 8));
132 ccv_nnc_stream_context_t* const stream_0 = ccv_cnnp_compiled_data_get_stream(compiled_data, type);
133 // Wait signal to finish.
134 if (stream_context)
135 ccv_nnc_stream_context_wait_signal(stream_0, signal);
136 ccv_nnc_cmd_exec(CMD_EWSQRT_FORWARD()ccv_nnc_cmd(CCV_NNC_EWSQRT_FORWARD, 0, ccv_nnc_cmd_auto, 0), ccv_nnc_no_hint, 0, TENSOR_LIST(norm2[i * 2])(ccv_nnc_tensor_t* []){norm2[i * 2]}, (1 +1 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(norm2[i * 2])(ccv_nnc_tensor_t* []){norm2[i * 2]}, (1 +1 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_0);
137 ccv_nnc_cmd_exec(CMD_SET_FORWARD(max_norm)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={max_norm,}}}, 0), ccv_nnc_no_hint, 0, 0, 0, TENSOR_LIST(max_normt[i])(ccv_nnc_tensor_t* []){max_normt[i]}, (1 +1 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_0);
138 ccv_nnc_cmd_exec(CMD_EWDIV_FORWARD()ccv_nnc_cmd(CCV_NNC_EWDIV_FORWARD, 0, ccv_nnc_cmd_auto, 0), ccv_nnc_no_hint, 0, TENSOR_LIST(max_normt[i], norm2[i * 2])(ccv_nnc_tensor_t* []){max_normt[i], norm2[i * 2]}, (1 +1 +1 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1
), TENSOR_LIST(norm2[i * 2])(ccv_nnc_tensor_t* []){norm2[i * 2]}, (1 +1 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_0);
139 ccv_nnc_cmd_exec(CMD_CLAMP_FORWARD(NAN, 1)ccv_nnc_cmd(CCV_NNC_CLAMP_FORWARD, 0, (ccv_nnc_cmd_param_t){.
size={.dim={1,1,1}},.clamp={.min=(__builtin_nanf ("")),.max=1
}}, 0), ccv_nnc_no_hint, 0, TENSOR_LIST(norm2[i * 2])(ccv_nnc_tensor_t* []){norm2[i * 2]}, (1 +1 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(norm2[i * 2])(ccv_nnc_tensor_t* []){norm2[i * 2]}, (1 +1 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_0);
140 if (stream_context)
141 {
142 ccv_nnc_stream_signal_t* const signal = ccv_nnc_stream_context_emit_signal_new(stream_0);
143 ccv_nnc_stream_context_wait_signal(stream_context, signal);
144 }
145 streams[i] = stream_0;
146 }
147 // If this should be blocking, blocking it.
148 if (!stream_context)
149 for (i = 0; i < parallel_count; i++)
150 if (streams[i])
151 ccv_nnc_stream_context_wait(streams[i]);
152 } else {
153 ccv_nnc_cmd_exec(CMD_EWSQRT_FORWARD()ccv_nnc_cmd(CCV_NNC_EWSQRT_FORWARD, 0, ccv_nnc_cmd_auto, 0), ccv_nnc_no_hint, 0, TENSOR_LIST(norm2[0])(ccv_nnc_tensor_t* []){norm2[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(norm2[0])(ccv_nnc_tensor_t* []){norm2[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context);
154 ccv_nnc_cmd_exec(CMD_SET_FORWARD(max_norm)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={max_norm,}}}, 0), ccv_nnc_no_hint, 0, 0, 0, TENSOR_LIST(max_normt[0])(ccv_nnc_tensor_t* []){max_normt[0]}, (1 +1 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context);
155 ccv_nnc_cmd_exec(CMD_EWDIV_FORWARD()ccv_nnc_cmd(CCV_NNC_EWDIV_FORWARD, 0, ccv_nnc_cmd_auto, 0), ccv_nnc_no_hint, 0, TENSOR_LIST(max_normt[0], norm2[0])(ccv_nnc_tensor_t* []){max_normt[0], norm2[0]}, (1 +1 +1 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(norm2[0])(ccv_nnc_tensor_t* []){norm2[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context);
156 ccv_nnc_cmd_exec(CMD_CLAMP_FORWARD(NAN, 1)ccv_nnc_cmd(CCV_NNC_CLAMP_FORWARD, 0, (ccv_nnc_cmd_param_t){.
size={.dim={1,1,1}},.clamp={.min=(__builtin_nanf ("")),.max=1
}}, 0), ccv_nnc_no_hint, 0, TENSOR_LIST(norm2[0])(ccv_nnc_tensor_t* []){norm2[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(norm2[0])(ccv_nnc_tensor_t* []){norm2[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context);
157 }
158 ccv_nnc_cmd_t scatter_cmd = {
159 .cmd = CCV_NNC_CUSTOM_FORWARD,
160 .isa = &clip_grad_norm_scatter_norm2_vtab,
161 };
162 ccv_cnnp_model_parameter_gradients_map(model, parameters, scatter_cmd, ccv_nnc_no_hint, 0, norm2, parallel_count * 2, 0, 0, stream_context);
163 if (stream_type == CCV_STREAM_CONTEXT_GPU)
164 for (i = 0; i < parallel_count; i++)
165 {
166 ccv_nnc_xpu_free(&compiled_data->xpu_alloc, norm2[i * 2]->data.u8);
167 ccv_nnc_xpu_free(&compiled_data->xpu_alloc, norm2[i * 2 + 1]->data.u8);
168 ccv_nnc_xpu_free(&compiled_data->xpu_alloc, max_normt[i]->data.u8);
169 }
170 for (i = 0; i < parallel_count; i++)
171 {
172 ccv_nnc_tensor_free(norm2[i * 2]);
173 ccv_nnc_tensor_free(norm2[i * 2 + 1]);
174 ccv_nnc_tensor_free(max_normt[i]);
175 }
176}
177
178// MARK - Add-on Functions
179
180static int _ccv_cnnp_model_isnan(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint, const int flags, ccv_nnc_tensor_t* const* const inputs, const int input_size, ccv_nnc_tensor_t* const* const outputs, const int output_size, ccv_nnc_stream_context_t* const stream_context)
181{
182 const int device_id = CCV_TENSOR_GET_DEVICE_ID(inputs[0]->info.type)(((inputs[0]->info.type) & 0xfff00) >> 8);
183 ccv_nnc_tensor_t* const old_isnanr = outputs[1 + device_id * 2];
184 ccv_nnc_tensor_t* const isnanr = outputs[1 + device_id * 2 + 1];
185 ccv_nnc_cmd_t reduce_cmd = CMD_REDUCE_ISNAN_FORWARD()ccv_nnc_cmd(CCV_NNC_REDUCE_ISNAN_FORWARD, 0, ((ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.reduce={.count=(1 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1),.axis={}}}
), 0);
186 reduce_cmd.info.reduce.count = ccv_nnc_tensor_nd(inputs[0]->info.dim);
187 int i;
188 for (i = 0; i < cmd.info.reduce.count; i++)
189 reduce_cmd.info.reduce.axis[i] = i;
190 ccv_nnc_cmd_exec(reduce_cmd, hint, flags, TENSOR_LIST(inputs[0])(ccv_nnc_tensor_t* []){inputs[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(isnanr)(ccv_nnc_tensor_t* []){isnanr}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context);
191 ccv_nnc_cmd_exec(CMD_EWSUM_FORWARD()ccv_nnc_cmd(CCV_NNC_EWSUM_FORWARD, 0, ccv_nnc_cmd_auto, 0), hint, flags, TENSOR_LIST(old_isnanr, isnanr)(ccv_nnc_tensor_t* []){old_isnanr, isnanr}, (1 +1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(old_isnanr)(ccv_nnc_tensor_t* []){old_isnanr}, (1 +1 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context);
192 return CCV_NNC_EXEC_SUCCESS;
193}
194
195static ccv_nnc_cmd_vtab_t reduce_isnan_vtab = {
196 .exec = _ccv_cnnp_model_isnan
197};
198
199int ccv_cnnp_model_parameter_gradients_isnan(ccv_cnnp_model_t* const model, const ccv_cnnp_model_io_t parameters, ccv_nnc_stream_context_t* const stream_context)
200{
201 ccv_cnnp_compiled_data_t* const compiled_data = model->compiled_data;
202 assert(compiled_data)((void) sizeof ((compiled_data) ? 1 : 0), __extension__ ({ if
(compiled_data) ; else __assert_fail ("compiled_data", "ccv_cnnp_model_addons.c"
, 202, __extension__ __PRETTY_FUNCTION__); }));
203 const int parallel_count = ccv_max(model->parallel_count, 1)({ typeof (model->parallel_count) _a = (model->parallel_count
); typeof (1) _b = (1); (_a > _b) ? _a : _b; });
204 ccv_nnc_tensor_t* isnanr[parallel_count * 2];
205 const int stream_type = model->compiled_data->stream_type;
206 int i;
207 if (stream_type == CCV_STREAM_CONTEXT_GPU)
208 {
209 for (i = 0; i < parallel_count; i++)
210 {
211 ccv_nnc_tensor_param_t info = {
212 .type = CCV_TENSOR_GPU_MEMORY,
213 .format = CCV_TENSOR_FORMAT_NHWC,
214 .datatype = CCV_32S,
215 .dim = {1},
216 };
217 CCV_TENSOR_SET_DEVICE_ID(info.type, i)(info.type) = (((info.type) & ~0xfff00) | (((i) & 0xfff
) << 8));
218 isnanr[i * 2] = ccv_nnc_tensor_new(ccv_nnc_xpu_alloc(&compiled_data->xpu_alloc, i, stream_context, ccv_nnc_tensor_data_size(info)), info, 0);
219 isnanr[i * 2 + 1] = ccv_nnc_tensor_new(ccv_nnc_xpu_alloc(&compiled_data->xpu_alloc, i, stream_context, ccv_nnc_tensor_data_size(info)), info, 0);
220 }
221 } else {
222 for (i = 0; i < parallel_count; i++)
223 {
224 ccv_nnc_tensor_param_t info = {
225 .type = CCV_TENSOR_CPU_MEMORY,
226 .format = CCV_TENSOR_FORMAT_NHWC,
227 .datatype = CCV_32S,
228 .dim = {1},
229 };
230 isnanr[i * 2] = ccv_nnc_tensor_new(0, info, 0);
231 isnanr[i * 2 + 1] = ccv_nnc_tensor_new(0, info, 0);
232 }
233 }
234 // zero out old isnanr.
235 if (parallel_count > 1)
236 {
237 ccv_nnc_stream_context_t* streams[parallel_count];
238 ccv_nnc_stream_signal_t* signal;
239 if (stream_context)
240 signal = ccv_nnc_stream_context_emit_signal_new(stream_context);
241 for (i = 0; i < parallel_count; i++)
242 {
243 const int stream_type = CCV_TENSOR_GET_MEMORY(isnanr[i * 2]->info.type)((isnanr[i * 2]->info.type) & 0x3) == CCV_TENSOR_GPU_MEMORY ? CCV_STREAM_CONTEXT_GPU : CCV_STREAM_CONTEXT_CPU;
244 const int device_id = CCV_TENSOR_GET_DEVICE_ID(isnanr[i * 2]->info.type)(((isnanr[i * 2]->info.type) & 0xfff00) >> 8);
245 int type = stream_type;
246 CCV_STREAM_SET_DEVICE_ID(type, device_id)(type) = (((type) & ~0xfff00) | (((device_id) & 0xfff
) << 8));
247 ccv_nnc_stream_context_t* const stream_0 = ccv_cnnp_compiled_data_get_stream(compiled_data, type);
248 // Wait signal to finish.
249 if (stream_context)
250 ccv_nnc_stream_context_wait_signal(stream_0, signal);
251 ccv_nnc_cmd_exec(CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, 0, TENSOR_LIST(isnanr[i * 2])(ccv_nnc_tensor_t* []){isnanr[i * 2]}, (1 +1 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_0);
252 if (stream_context)
253 {
254 ccv_nnc_stream_signal_t* const signal = ccv_nnc_stream_context_emit_signal_new(stream_0);
255 ccv_nnc_stream_context_wait_signal(stream_context, signal);
256 }
257 streams[i] = stream_0;
258 }
259 // If this should be blocking, blocking it.
260 if (!stream_context)
261 for (i = 0; i < parallel_count; i++)
262 if (streams[i])
263 ccv_nnc_stream_context_wait(streams[i]);
264 } else
265 ccv_nnc_cmd_exec(CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, 0, TENSOR_LIST(isnanr[0])(ccv_nnc_tensor_t* []){isnanr[0]}, (1 +1 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), stream_context);
266 // Gather isnanr.
267 ccv_nnc_cmd_t reduce_cmd = {
268 .cmd = CCV_NNC_CUSTOM_FORWARD,
269 .isa = &reduce_isnan_vtab,
270 };
271 ccv_cnnp_model_parameter_gradients_map(model, parameters, reduce_cmd, ccv_nnc_no_hint, 0, 0, 0, isnanr, parallel_count * 2, stream_context);
272 for (i = 0; i < parallel_count; i++)
273 ccv_nnc_tensor_free(isnanr[i * 2 + 1]);
274 int retval = 0;
275 if (stream_type == CCV_TENSOR_GPU_MEMORY)
276 {
277 ccv_nnc_tensor_param_t info = {
278 .type = CCV_TENSOR_CPU_MEMORY,
279 .format = CCV_TENSOR_FORMAT_NHWC,
280 .datatype = CCV_32S,
281 .dim = {1},
282 };
283 ccv_nnc_tensor_t* checknan = ccv_nnc_tensor_new(0, info, 0);
284 for (i = 0; i < parallel_count; i++)
285 {
286 ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD()ccv_nnc_cmd(CCV_NNC_DATA_TRANSFER_FORWARD, 0, ccv_nnc_cmd_auto
, 0), ccv_nnc_no_hint, 0, TENSOR_LIST(isnanr[i * 2])(ccv_nnc_tensor_t* []){isnanr[i * 2]}, (1 +1 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_LIST(checknan)(ccv_nnc_tensor_t* []){checknan}, (1 +1 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), 0);
287 if (checknan->data.i32[0] > 0)
288 {
289 retval = 1;
290 break;
291 }
292 }
293 ccv_nnc_tensor_free(checknan);
294 } else {
295 for (i = 0; i < parallel_count; i++)
296 if (isnanr[i * 2]->data.i32[0] > 0)
297 {
298 retval = 1;
299 break;
300 }
301 }
302 for (i = 0; i < parallel_count; i++)
303 ccv_nnc_tensor_free(isnanr[i * 2]);
304 return retval;
305}
306
307// MARK - Core Layers
308
309static void _ccv_cnnp_sum_build(ccv_cnnp_model_t* const self, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
310{
311 PRINT(CCV_CLI_VERBOSE, "[cnnp_sum_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_sum_build] -\n"); fflush(stdout); } } while (
0);
312 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 312, __extension__ __PRETTY_FUNCTION__
); }));
313 outputs[0] = ccv_nnc_tensor_symbol_new(graph, ccv_nnc_tensor_symbol_params(graph, inputs[0]), 0);
314 ccv_nnc_graph_exec_symbol_new(graph, CMD_EWSUM_FORWARD()ccv_nnc_cmd(CCV_NNC_EWSUM_FORWARD, 0, ccv_nnc_cmd_auto, 0), inputs, input_size, outputs, output_size, 0);
315}
316
317static ccv_cnnp_model_t* _ccv_cnnp_sum_copy(const ccv_cnnp_model_t* const self, void* const context);
318
319static const ccv_cnnp_model_vtab_t ccv_cnnp_sum_isa = {
320 .build = _ccv_cnnp_sum_build,
321 .copy = _ccv_cnnp_sum_copy,
322};
323
324typedef struct {
325 ccv_cnnp_model_t super;
326 ccv_nnc_tensor_symbol_t output;
327} ccv_cnnp_model_sum_t;
328
329ccv_cnnp_model_t* ccv_cnnp_sum(const char* const name)
330{
331 ccv_cnnp_model_sum_t* const model_sum = (ccv_cnnp_model_sum_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_sum_t));
332 model_sum->super.isa = &ccv_cnnp_sum_isa;
333 model_sum->super.input_size = 0;
334 model_sum->super.outputs = &model_sum->output;
335 model_sum->super.output_size = 1;
336 ccv_cnnp_model_copy_name(&model_sum->super, name);
337 return (ccv_cnnp_model_t*)model_sum;
338}
339
340static ccv_cnnp_model_t* _ccv_cnnp_sum_copy(const ccv_cnnp_model_t* const self, void* const context)
341{
342 return ccv_cnnp_sum(self->name);
343}
344
345typedef struct {
346 ccv_cnnp_model_t super;
347 int axis;
348 ccv_nnc_tensor_symbol_t output;
349} ccv_cnnp_model_concat_t;
350
351static void _ccv_cnnp_concat_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
352{
353 const ccv_cnnp_model_concat_t* const self = (const ccv_cnnp_model_concat_t*)super;
354 PRINT(CCV_CLI_VERBOSE, "[cnnp_concat_build] 1. -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_concat_build] 1. -\n"); fflush(stdout); } } while
(0);
1
Assuming the condition is false
→
2
←
Taking false branch
→
3
←
Loop condition is false. Exiting loop
→
355 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 355, __extension__ __PRETTY_FUNCTION__
); }));
4
←
Assuming 'output_size' is equal to 1
→
5
←
Taking true branch
→
356 ccv_nnc_tensor_param_t output_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
357 int i, j;
358 if (output_params.dim[0] == 0)
6
←
Assuming the condition is false
→
7
←
Taking false branch
→
359 for (i = 1; i < input_size; i++)
360 {
361 output_params = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
362 if (output_params.dim[0] != 0)
363 break;
364 }
365 const int nd = ccv_nnc_tensor_nd(output_params.dim);
366 const int axis = self->axis;
367 assert(axis < nd)((void) sizeof ((axis < nd) ? 1 : 0), __extension__ ({ if (
axis < nd) ; else __assert_fail ("axis < nd", "ccv_cnnp_model_addons.c"
, 367, __extension__ __PRETTY_FUNCTION__); }));
8
←
Assuming 'axis' is < 'nd'
→
9
←
Taking true branch
→
368 output_params.dim[axis] = 0;
369 int input_is_contiguous = 1;
370 for (i = 0; i < input_size; i++)
10
←
Assuming 'i' is >= 'input_size'
→
11
←
Loop condition is false. Execution continues on line 394
→
371 {
372 const ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
373 const int input_nd = ccv_nnc_tensor_nd(input_params.dim);
374 if (input_nd == 0)
375 {
376 PRINT(CCV_CLI_VERBOSE, "[cnnp_concat_build] %d. input[%d]: -\n", i + 2, i)do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_concat_build] %d. input[%d]: -\n", i + 2, i)
; fflush(stdout); } } while (0);
377 input_is_contiguous = 0;
378 continue;
379 }
380 if (CCV_CLI_OUTPUT_LEVEL_IS(CCV_CLI_VERBOSE)(CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
381 {
382 PRINT(CCV_CLI_VERBOSE, "[cnnp_concat_build] %d. input[%d]: (%d", i + 2, i, input_params.dim[0])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_concat_build] %d. input[%d]: (%d", i + 2, i,
input_params.dim[0]); fflush(stdout); } } while (0);
383 int i;
384 for (i = 1; i < CCV_NNC_MAX_DIM_ALLOC(12) && input_params.dim[i] > 0; i++)
385 PRINT(CCV_CLI_VERBOSE, ", %d", input_params.dim[i])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf(", %d", input_params.dim[i]); fflush(stdout); } } while
(0);
386 PRINT(CCV_CLI_VERBOSE, ")\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf(")\n"); fflush(stdout); } } while (0);
387 }
388 assert(input_nd == nd)((void) sizeof ((input_nd == nd) ? 1 : 0), __extension__ ({ if
(input_nd == nd) ; else __assert_fail ("input_nd == nd", "ccv_cnnp_model_addons.c"
, 388, __extension__ __PRETTY_FUNCTION__); }));
389 for (j = 0; j < nd; j++)
390 if (j != axis)
391 { assert(input_params.dim[j] == output_params.dim[j])((void) sizeof ((input_params.dim[j] == output_params.dim[j])
? 1 : 0), __extension__ ({ if (input_params.dim[j] == output_params
.dim[j]) ; else __assert_fail ("input_params.dim[j] == output_params.dim[j]"
, "ccv_cnnp_model_addons.c", 391, __extension__ __PRETTY_FUNCTION__
); })); }
392 output_params.dim[axis] += input_params.dim[axis];
393 }
394 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
395 int ofs[CCV_NNC_MAX_DIM_ALLOC(12)] = {};
396 int stride[CCV_NNC_MAX_DIM_ALLOC(12)] = {};
397 ccv_nnc_tensor_get_stride(output_params.dim, stride);
398 if (input_is_contiguous
11.1
'input_is_contiguous' is 1
)
12
←
Taking true branch
→
399 {
400 ccv_nnc_tensor_symbol_t aliases[input_size];
13
←
Declared variable-length array (VLA) has negative size
401 for (i = 0; i < input_size; i++)
402 {
403 const ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
404 aliases[i] = ccv_nnc_tensor_symbol_alias_new(graph, outputs[0], ofs, stride, input_params, 0);
405 ofs[axis] += input_params.dim[axis];
406 }
407 // Format transform is more flexible.
408 ccv_nnc_graph_exec_symbol_new(graph, CMD_FORMAT_TRANSFORM_FORWARD()ccv_nnc_cmd(CCV_NNC_FORMAT_TRANSFORM_FORWARD, 0, ccv_nnc_cmd_auto
, 0), inputs, input_size, aliases, input_size, "concat");
409 } else {
410 ccv_nnc_tensor_symbol_t aliases[input_size];
411 for (i = 0; i < input_size; i++)
412 {
413 const ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
414 if (input_params.dim[0] == 0)
415 {
416 // Create a new alias anyway, but not going to use it, in this way, the alias count will match during absorb.
417 aliases[i] = ccv_nnc_tensor_symbol_alias_new(graph, outputs[0], ofs, stride, input_params, 0);
418 continue;
419 }
420 aliases[i] = ccv_nnc_tensor_symbol_alias_new(graph, outputs[0], ofs, stride, input_params, 0);
421 ofs[axis] += input_params.dim[axis];
422 }
423 // Format transform is more flexible.
424 ccv_nnc_graph_exec_symbol_new(graph, CMD_FORMAT_TRANSFORM_FORWARD()ccv_nnc_cmd(CCV_NNC_FORMAT_TRANSFORM_FORWARD, 0, ccv_nnc_cmd_auto
, 0), inputs, input_size, aliases, input_size, "concat");
425 }
426}
427
428static ccv_cnnp_model_t* _ccv_cnnp_concat_copy(const ccv_cnnp_model_t* const self, void* const context);
429
430static const ccv_cnnp_model_vtab_t ccv_cnnp_concat_isa = {
431 .build = _ccv_cnnp_concat_build,
432 .copy = _ccv_cnnp_concat_copy,
433};
434
435ccv_cnnp_model_t* ccv_cnnp_concat(const int axis, const char* const name)
436{
437 ccv_cnnp_model_concat_t* const model_concat = (ccv_cnnp_model_concat_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_concat_t));
438 model_concat->super.isa = &ccv_cnnp_concat_isa;
439 model_concat->super.input_size = 0;
440 model_concat->super.outputs = &model_concat->output;
441 model_concat->super.output_size = 1;
442 model_concat->axis = axis;
443 ccv_cnnp_model_copy_name(&model_concat->super, name);
444 return (ccv_cnnp_model_t*)model_concat;
445}
446
447static ccv_cnnp_model_t* _ccv_cnnp_concat_copy(const ccv_cnnp_model_t* const super, void* const context)
448{
449 const ccv_cnnp_model_concat_t* const self = (const ccv_cnnp_model_concat_t*)super;
450 return ccv_cnnp_concat(self->axis, self->super.name);
451}
452
453typedef struct {
454 ccv_cnnp_model_t super;
455 int axis;
456 ccv_nnc_tensor_symbol_t outputs[1];
457} ccv_cnnp_model_chunk_t;
458
459static void _ccv_cnnp_chunk_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
460{
461 const ccv_cnnp_model_concat_t* const self = (const ccv_cnnp_model_concat_t*)super;
462 PRINT(CCV_CLI_VERBOSE, "[cnnp_chunk_build] 1. axis: %d\n", self->axis)do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_chunk_build] 1. axis: %d\n", self->axis);
fflush(stdout); } } while (0);
463 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 463, __extension__ __PRETTY_FUNCTION__); }));
464 const ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
465 if (CCV_CLI_OUTPUT_LEVEL_IS(CCV_CLI_VERBOSE)(CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
466 {
467 PRINT(CCV_CLI_VERBOSE, "[cnnp_chunk_build] 2. input: (%d", input_params.dim[0])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_chunk_build] 2. input: (%d", input_params.dim
[0]); fflush(stdout); } } while (0);
468 int i;
469 for (i = 1; i < CCV_NNC_MAX_DIM_ALLOC(12) && input_params.dim[i] > 0; i++)
470 PRINT(CCV_CLI_VERBOSE, ", %d", input_params.dim[i])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf(", %d", input_params.dim[i]); fflush(stdout); } } while
(0);
471 PRINT(CCV_CLI_VERBOSE, ")\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf(")\n"); fflush(stdout); } } while (0);
472 }
473 ccv_nnc_tensor_param_t output_params = input_params;
474 int i;
475 const int nd = ccv_nnc_tensor_nd(output_params.dim);
476 const int axis = self->axis;
477 assert(axis < nd)((void) sizeof ((axis < nd) ? 1 : 0), __extension__ ({ if (
axis < nd) ; else __assert_fail ("axis < nd", "ccv_cnnp_model_addons.c"
, 477, __extension__ __PRETTY_FUNCTION__); }));
478 const int n = self->super.output_size;
479 assert(n == output_size)((void) sizeof ((n == output_size) ? 1 : 0), __extension__ ({
if (n == output_size) ; else __assert_fail ("n == output_size"
, "ccv_cnnp_model_addons.c", 479, __extension__ __PRETTY_FUNCTION__
); }));
480 assert(output_params.dim[axis] % n == 0)((void) sizeof ((output_params.dim[axis] % n == 0) ? 1 : 0), __extension__
({ if (output_params.dim[axis] % n == 0) ; else __assert_fail
("output_params.dim[axis] % n == 0", "ccv_cnnp_model_addons.c"
, 480, __extension__ __PRETTY_FUNCTION__); }));
481 output_params.dim[axis] = output_params.dim[axis] / n;
482 int ofs[CCV_NNC_MAX_DIM_ALLOC(12)] = {};
483 int stride[CCV_NNC_MAX_DIM_ALLOC(12)] = {};
484 ccv_nnc_tensor_get_stride(input_params.dim, stride);
485 ccv_nnc_tensor_symbol_t to = ccv_nnc_tensor_symbol_alias_to(graph, inputs[0]);
486 if (to.d == CCV_NNC_NO_TENSOR_SYMBOL) // If we are not reshape an alias, it is straightforward.
487 {
488 for (i = 0; i < output_size; i++)
489 {
490 outputs[i] = ccv_nnc_tensor_symbol_alias_new(graph, inputs[0], ofs, stride, output_params, 0);
491 ofs[axis] += output_params.dim[axis];
492 }
493 } else {
494 // Otherwise, we need to check if it is permute. For permute, we cannot do alias directly.
495 // We need to first materialize the permute and then run reshape on top of it, otherwise it will be wrong.
496 int old_stride[CCV_NNC_MAX_DIM_ALLOC(12)];
497 ccv_nnc_tensor_symbol_alias_params(graph, inputs[0], 0, old_stride);
498 // We identify permute by checking if the stride is not in descending order.
499 // This also covered "permute" through reshape, rather than using ccv_cnnp_permute directly.
500 int i, no_permute = 1;
501 for (i = 1; no_permute && i < nd; i++)
502 if (old_stride[i - 1] < old_stride[i])
503 no_permute = 0;
504 if (no_permute)
505 { // Just straightforward reshape if there is no no permute.
506 for (i = 0; i < output_size; i++)
507 {
508 outputs[i] = ccv_nnc_tensor_symbol_alias_new(graph, inputs[0], ofs, old_stride, output_params, 0);
509 ofs[axis] += output_params.dim[axis];
510 }
511 } else {
512 // Otherwise, we first do format transform to plain tensor and then do reshape.
513 ccv_nnc_tensor_symbol_t permuted = ccv_nnc_tensor_symbol_new(graph, input_params, 0);
514 ccv_nnc_graph_exec_symbol_new(graph, CMD_FORMAT_TRANSFORM_FORWARD()ccv_nnc_cmd(CCV_NNC_FORMAT_TRANSFORM_FORWARD, 0, ccv_nnc_cmd_auto
, 0), TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(permuted)(const ccv_nnc_tensor_symbol_t []){permuted}, (1 +1 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "reshape");
515 for (i = 0; i < output_size; i++)
516 {
517 outputs[i] = ccv_nnc_tensor_symbol_alias_new(graph, permuted, ofs, stride, output_params, 0);
518 ofs[axis] += output_params.dim[axis];
519 }
520 }
521 }
522}
523
524static ccv_cnnp_model_t* _ccv_cnnp_chunk_copy(const ccv_cnnp_model_t* const self, void* const context);
525
526static const ccv_cnnp_model_vtab_t ccv_cnnp_chunk_isa = {
527 .build = _ccv_cnnp_chunk_build,
528 .copy = _ccv_cnnp_chunk_copy,
529};
530
531ccv_cnnp_model_t* ccv_cnnp_chunk(const int n, const int axis, const char* const name)
532{
533 assert(n >= 1)((void) sizeof ((n >= 1) ? 1 : 0), __extension__ ({ if (n >=
1) ; else __assert_fail ("n >= 1", "ccv_cnnp_model_addons.c"
, 533, __extension__ __PRETTY_FUNCTION__); }));
534 ccv_cnnp_model_chunk_t* const model_chunk = (ccv_cnnp_model_chunk_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_chunk_t) + sizeof(ccv_nnc_tensor_symbol_t) * (n - 1));
535 model_chunk->super.isa = &ccv_cnnp_chunk_isa;
536 model_chunk->super.input_size = 1;
537 model_chunk->super.outputs = model_chunk->outputs;
538 model_chunk->super.output_size = n;
539 model_chunk->axis = axis;
540 ccv_cnnp_model_copy_name(&model_chunk->super, name);
541 return (ccv_cnnp_model_t*)model_chunk;
542}
543
544static ccv_cnnp_model_t* _ccv_cnnp_chunk_copy(const ccv_cnnp_model_t* const super, void* const context)
545{
546 const ccv_cnnp_model_chunk_t* const self = (const ccv_cnnp_model_chunk_t*)super;
547 return ccv_cnnp_chunk(self->super.output_size, self->axis, self->super.name);
548}
549
550typedef struct {
551 ccv_cnnp_model_t super;
552 ccv_nnc_tensor_symbol_t output;
553 int format;
554 int dim[CCV_NNC_MAX_DIM_ALLOC(12)];
555 int ofs[CCV_NNC_MAX_DIM_ALLOC(12)];
556 int stride[CCV_NNC_MAX_DIM_ALLOC(12)];
557} ccv_cnnp_model_reshape_t;
558
559static void _ccv_cnnp_reshape_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
560{
561 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 561, __extension__ __PRETTY_FUNCTION__); }));
562 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 562, __extension__ __PRETTY_FUNCTION__
); }));
563 ccv_cnnp_model_reshape_t* const self = (ccv_cnnp_model_reshape_t*)super;
564 if (CCV_CLI_OUTPUT_LEVEL_IS(CCV_CLI_VERBOSE)(CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
565 {
566 PRINT(CCV_CLI_VERBOSE, "[cnnp_reshape_build] 1. dim: (%d", self->dim[0])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_reshape_build] 1. dim: (%d", self->dim[0]
); fflush(stdout); } } while (0);
567 int i;
568 for (i = 1; i < CCV_NNC_MAX_DIM_ALLOC(12) && self->dim[i] > 0; i++)
569 PRINT(CCV_CLI_VERBOSE, ", %d", self->dim[i])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf(", %d", self->dim[i]); fflush(stdout); } } while
(0);
570 const int count = i;
571 PRINT(CCV_CLI_VERBOSE, "), ofs: (%d", self->ofs[0])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("), ofs: (%d", self->ofs[0]); fflush(stdout); } }
while (0);
572 for (i = 1; i < count; i++)
573 PRINT(CCV_CLI_VERBOSE, ", %d", self->ofs[i])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf(", %d", self->ofs[i]); fflush(stdout); } } while
(0);
574 PRINT(CCV_CLI_VERBOSE, "), stride: (%d", self->stride[0])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("), stride: (%d", self->stride[0]); fflush(stdout
); } } while (0);
575 for (i = 1; i < count; i++)
576 PRINT(CCV_CLI_VERBOSE, ", %d", self->stride[i])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf(", %d", self->stride[i]); fflush(stdout); } } while
(0);
577 PRINT(CCV_CLI_VERBOSE, ")\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf(")\n"); fflush(stdout); } } while (0);
578 }
579 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
580 int dim[CCV_NNC_MAX_DIM_ALLOC(12)];
581 memcpy(dim, self->dim, sizeof(dim));
582 int i, auto_idx = -1;
583 size_t known = 1;
584 const size_t tensor_count = ccv_nnc_tensor_count(params);
585 for (i = 0; i < CCV_NNC_MAX_DIM_ALLOC(12) && dim[i]; i++)
586 if (dim[i] == -1)
587 auto_idx = i;
588 else
589 known *= dim[i];
590 if (auto_idx >= 0)
591 {
592 assert(known > 0 && tensor_count % known == 0)((void) sizeof ((known > 0 && tensor_count % known
== 0) ? 1 : 0), __extension__ ({ if (known > 0 &&
tensor_count % known == 0) ; else __assert_fail ("known > 0 && tensor_count % known == 0"
, "ccv_cnnp_model_addons.c", 592, __extension__ __PRETTY_FUNCTION__
); }));
593 dim[auto_idx] = tensor_count / known;
594 }
595 if (CCV_CLI_OUTPUT_LEVEL_IS(CCV_CLI_VERBOSE)(CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
596 {
597 PRINT(CCV_CLI_VERBOSE, "[cnnp_reshape_build] 2. input: (%d", params.dim[0])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_reshape_build] 2. input: (%d", params.dim[0]
); fflush(stdout); } } while (0);
598 int i;
599 for (i = 1; i < CCV_NNC_MAX_DIM_ALLOC(12) && params.dim[i] > 0; i++)
600 PRINT(CCV_CLI_VERBOSE, ", %d", params.dim[i])do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf(", %d", params.dim[i]); fflush(stdout); } } while (
0);
601 PRINT(CCV_CLI_VERBOSE, ")\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf(")\n"); fflush(stdout); } } while (0);
602 }
603 assert(ccv_nnc_dimension_count(dim) <= ccv_nnc_tensor_count(params))((void) sizeof ((ccv_nnc_dimension_count(dim) <= ccv_nnc_tensor_count
(params)) ? 1 : 0), __extension__ ({ if (ccv_nnc_dimension_count
(dim) <= ccv_nnc_tensor_count(params)) ; else __assert_fail
("ccv_nnc_dimension_count(dim) <= ccv_nnc_tensor_count(params)"
, "ccv_cnnp_model_addons.c", 603, __extension__ __PRETTY_FUNCTION__
); }));
604 ccv_nnc_tensor_symbol_t to = ccv_nnc_tensor_symbol_alias_to(graph, inputs[0]);
605 int stride_from_dim[CCV_NNC_MAX_DIM_ALLOC(12)];
606 if (to.d == CCV_NNC_NO_TENSOR_SYMBOL) // If we are not reshape an alias, it is straightforward.
607 {
608 memcpy(params.dim, dim, sizeof(params.dim));
609 int* stride;
610 if (self->stride[0] == 0)
611 {
612 ccv_nnc_tensor_get_stride(dim, stride_from_dim);
613 stride = stride_from_dim;
614 } else
615 stride = self->stride;
616 if (self->format > 0)
617 params.format = self->format;
618 outputs[0] = ccv_nnc_tensor_symbol_alias_new(graph, inputs[0], self->ofs, stride, params, 0);
619 } else {
620 // Otherwise, we need to check if it is permute. For permute, we cannot do alias directly.
621 // We need to first materialize the permute and then run reshape on top of it, otherwise it will be wrong.
622 int old_stride[CCV_NNC_MAX_DIM_ALLOC(12)];
623 ccv_nnc_tensor_symbol_alias_params(graph, inputs[0], 0, old_stride);
624 // We identify permute by checking if the stride is not in descending order.
625 // This also covered "permute" through reshape, rather than using ccv_cnnp_permute directly.
626 const int nd = ccv_nnc_tensor_nd(params.dim);
627 const int new_nd = ccv_nnc_tensor_nd(dim);
628 int i, no_permute = 1;
629 // If the new dim has different nd, or we actually have a stride, we need to check if it is no permute or not.
630 if (new_nd != nd || (self->stride[0] != 0 && memcmp(self->stride, old_stride, sizeof(self->stride))))
631 for (i = 1; no_permute && i < nd; i++)
632 if (old_stride[i - 1] < old_stride[i])
633 no_permute = 0;
634 if (no_permute)
635 { // Just straightforward reshape if there is no no permute.
636 memcpy(params.dim, dim, sizeof(params.dim));
637 int* stride;
638 if (self->stride[0] == 0)
639 {
640 if (new_nd != nd) // Cannot use old stride.
641 {
642 ccv_nnc_tensor_get_stride(dim, stride_from_dim);
643 stride = stride_from_dim;
644 } else
645 stride = old_stride;
646 } else
647 stride = self->stride;
648 if (self->format > 0)
649 params.format = self->format;
650 outputs[0] = ccv_nnc_tensor_symbol_alias_new(graph, inputs[0], self->ofs, stride, params, 0);
651 } else {
652 // Otherwise, we first do format transform to plain tensor and then do reshape.
653 ccv_nnc_tensor_symbol_t permuted = ccv_nnc_tensor_symbol_new(graph, params, 0);
654 ccv_nnc_graph_exec_symbol_new(graph, CMD_FORMAT_TRANSFORM_FORWARD()ccv_nnc_cmd(CCV_NNC_FORMAT_TRANSFORM_FORWARD, 0, ccv_nnc_cmd_auto
, 0), TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(permuted)(const ccv_nnc_tensor_symbol_t []){permuted}, (1 +1 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "reshape");
655 memcpy(params.dim, dim, sizeof(params.dim));
656 int* stride;
657 if (self->stride[0] == 0)
658 {
659 ccv_nnc_tensor_get_stride(dim, stride_from_dim);
660 stride = stride_from_dim;
661 } else
662 stride = self->stride;
663 if (self->format > 0)
664 params.format = self->format;
665 // And then we create alias against the permuted one.
666 outputs[0] = ccv_nnc_tensor_symbol_alias_new(graph, permuted, self->ofs, stride, params, 0);
667 }
668 }
669}
670
671static ccv_cnnp_model_t* _ccv_cnnp_reshape_copy(const ccv_cnnp_model_t* const super, void* const context);
672
673static const ccv_cnnp_model_vtab_t ccv_cnnp_reshape_isa = {
674 .build = _ccv_cnnp_reshape_build,
675 .copy = _ccv_cnnp_reshape_copy,
676};
677
678ccv_cnnp_model_t* ccv_cnnp_reshape(const int format, const int dim[CCV_NNC_MAX_DIM_ALLOC(12)], const int ofs[CCV_NNC_MAX_DIM_ALLOC(12)], const int stride[CCV_NNC_MAX_DIM_ALLOC(12)], const char* const name)
679{
680 ccv_cnnp_model_reshape_t* const model_reshape = (ccv_cnnp_model_reshape_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_reshape_t));
681 model_reshape->super.isa = &ccv_cnnp_reshape_isa;
682 model_reshape->super.input_size = 1;
683 model_reshape->super.outputs = &model_reshape->output;
684 model_reshape->super.output_size = 1;
685 ccv_cnnp_model_copy_name(&model_reshape->super, name);
686 model_reshape->format = format;
687 memcpy(model_reshape->dim, dim, sizeof(model_reshape->dim));
688 memcpy(model_reshape->ofs, ofs, sizeof(model_reshape->ofs));
689 if (stride[0] != 0)
690 memcpy(model_reshape->stride, stride, sizeof(model_reshape->stride));
691 return (ccv_cnnp_model_t*)model_reshape;
692}
693
694static ccv_cnnp_model_t* _ccv_cnnp_reshape_copy(const ccv_cnnp_model_t* const super, void* const context)
695{
696 const ccv_cnnp_model_reshape_t* const self = (const ccv_cnnp_model_reshape_t*)super;
697 return ccv_cnnp_reshape(self->format, self->dim, self->ofs, self->stride, self->super.name);
698}
699
700typedef struct {
701 ccv_cnnp_model_t super;
702 ccv_nnc_tensor_symbol_t output;
703 int type;
704 int begin[CCV_NNC_MAX_DIM_ALLOC(12)];
705 int end[CCV_NNC_MAX_DIM_ALLOC(12)];
706} ccv_cnnp_model_pad_t;
707
708static void _ccv_cnnp_pad_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
709{
710 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 710, __extension__ __PRETTY_FUNCTION__); }));
711 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 711, __extension__ __PRETTY_FUNCTION__
); }));
712 ccv_cnnp_model_pad_t* const self = (ccv_cnnp_model_pad_t*)super;
713 PRINT(CCV_CLI_VERBOSE, "[cnnp_pad_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_pad_build] -\n"); fflush(stdout); } } while (
0);
714 const ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
715 const int nd = ccv_nnc_tensor_nd(input_params.dim);
716 ccv_nnc_tensor_param_t params = input_params;
717 int i;
718 for (i = 0 ; i < nd; i++)
719 params.dim[i] += self->begin[i] + self->end[i];
720 const ccv_nnc_tensor_symbol_t padded = ccv_nnc_tensor_symbol_new(graph, params, 0);
721 ccv_nnc_cmd_t pad = CMD_PAD_FORWARD(self->type, (), ())ccv_nnc_cmd(CCV_NNC_PAD_FORWARD, 0, ((ccv_nnc_cmd_param_t){.size
={.dim={}},.pad={.type=self->type,.end={}}}), 0);
722 memcpy(pad.info.size.dim, self->begin, sizeof(pad.info.size.dim));
723 memcpy(pad.info.pad.end, self->end, sizeof(pad.info.pad.end));
724 ccv_nnc_graph_exec_symbol_new(graph, pad, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(padded)(const ccv_nnc_tensor_symbol_t []){padded}, (1 +1 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "pad");
725 outputs[0] = padded;
726}
727
728static ccv_cnnp_model_t* _ccv_cnnp_pad_copy(const ccv_cnnp_model_t* const super, void* const context);
729
730static const ccv_cnnp_model_vtab_t ccv_cnnp_pad_isa = {
731 .build = _ccv_cnnp_pad_build,
732 .copy = _ccv_cnnp_pad_copy,
733};
734
735ccv_cnnp_model_t* ccv_cnnp_pad(const int type, const int begin[CCV_NNC_MAX_DIM_ALLOC(12)], const int end[CCV_NNC_MAX_DIM_ALLOC(12)], const char* const name)
736{
737 ccv_cnnp_model_pad_t* const model_pad = (ccv_cnnp_model_pad_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_pad_t));
738 model_pad->super.isa = &ccv_cnnp_pad_isa;
739 model_pad->super.input_size = 1;
740 model_pad->super.outputs = &model_pad->output;
741 model_pad->super.output_size = 1;
742 ccv_cnnp_model_copy_name(&model_pad->super, name);
743 model_pad->type = type;
744 memcpy(model_pad->begin, begin, sizeof(model_pad->begin));
745 memcpy(model_pad->end, end, sizeof(model_pad->end));
746 return (ccv_cnnp_model_t*)model_pad;
747}
748
749static ccv_cnnp_model_t* _ccv_cnnp_pad_copy(const ccv_cnnp_model_t* const super, void* const context)
750{
751 const ccv_cnnp_model_pad_t* const self = (const ccv_cnnp_model_pad_t*)super;
752 return ccv_cnnp_pad(self->type, self->begin, self->end, self->super.name);
753}
754
755typedef struct {
756 ccv_cnnp_model_t super;
757 ccv_nnc_tensor_symbol_t output;
758} ccv_cnnp_model_identity_t;
759
760static void _ccv_cnnp_identity_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
761{
762 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 762, __extension__ __PRETTY_FUNCTION__); }));
763 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 763, __extension__ __PRETTY_FUNCTION__
); }));
764 PRINT(CCV_CLI_VERBOSE, "[cnnp_identity_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_identity_build] -\n"); fflush(stdout); } } while
(0);
765 outputs[0] = inputs[0];
766}
767
768static ccv_cnnp_model_t* _ccv_cnnp_identity_copy(const ccv_cnnp_model_t* const super, void* const context);
769
770static const ccv_cnnp_model_vtab_t ccv_cnnp_identity_isa = {
771 .build = _ccv_cnnp_identity_build,
772 .copy = _ccv_cnnp_identity_copy,
773};
774
775ccv_cnnp_model_t* ccv_cnnp_identity(const char* const name)
776{
777 ccv_cnnp_model_identity_t* const model_identity = (ccv_cnnp_model_identity_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_identity_t));
778 model_identity->super.isa = &ccv_cnnp_identity_isa;
779 model_identity->super.input_size = 1;
780 model_identity->super.outputs = &model_identity->output;
781 model_identity->super.output_size = 1;
782 ccv_cnnp_model_copy_name(&model_identity->super, name);
783 return (ccv_cnnp_model_t*)model_identity;
784}
785
786static ccv_cnnp_model_t* _ccv_cnnp_identity_copy(const ccv_cnnp_model_t* const super, void* const context)
787{
788 const ccv_cnnp_model_identity_t* const self = (const ccv_cnnp_model_identity_t*)super;
789 return ccv_cnnp_identity(self->super.name);
790}
791
792typedef struct {
793 ccv_cnnp_model_t super;
794 ccv_nnc_tensor_symbol_t output;
795 int index[CCV_NNC_MAX_DIM_ALLOC(12)];
796} ccv_cnnp_model_permute_t;
797
798static void _ccv_cnnp_permute_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
799{
800 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 800, __extension__ __PRETTY_FUNCTION__); }));
801 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 801, __extension__ __PRETTY_FUNCTION__
); }));
802 ccv_cnnp_model_permute_t* const self = (ccv_cnnp_model_permute_t*)super;
803 PRINT(CCV_CLI_VERBOSE, "[cnnp_permute_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_permute_build] -\n"); fflush(stdout); } } while
(0);
804 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
805 ccv_nnc_tensor_symbol_t to = ccv_nnc_tensor_symbol_alias_to(graph, inputs[0]);
806 const int nd = ccv_nnc_tensor_nd(params.dim);
807 int input_dim[CCV_NNC_MAX_DIM_ALLOC(12)];
808 memcpy(input_dim, params.dim, sizeof(params.dim));
809 int input_stride[CCV_NNC_MAX_DIM_ALLOC(12)] = {};
810 int output_stride[CCV_NNC_MAX_DIM_ALLOC(12)] = {};
811 if (to.d == CCV_NNC_NO_TENSOR_SYMBOL) // If it is not an alias. Find stride and permute.
812 {
813 ccv_nnc_tensor_get_stride(input_dim, input_stride);
814 int i;
815 for (i = 0; i < nd; i++)
816 {
817 const int idx = self->index[i];
818 assert(idx >= 0 && idx < nd)((void) sizeof ((idx >= 0 && idx < nd) ? 1 : 0)
, __extension__ ({ if (idx >= 0 && idx < nd) ; else
__assert_fail ("idx >= 0 && idx < nd", "ccv_cnnp_model_addons.c"
, 818, __extension__ __PRETTY_FUNCTION__); }));
819 params.dim[i] = input_dim[idx];
820 output_stride[i] = input_stride[idx];
821 }
822 outputs[0] = ccv_nnc_tensor_symbol_alias_new(graph, inputs[0], ccv_nnc_no_ofs, output_stride, params, 0);
823 } else {
824 // if it is an alias, we can get the stride from it and use that.
825 int input_ofs[CCV_NNC_MAX_DIM_ALLOC(12)];
826 ccv_nnc_tensor_symbol_alias_params(graph, inputs[0], input_ofs, input_stride);
827 assert(input_stride[0] != 0)((void) sizeof ((input_stride[0] != 0) ? 1 : 0), __extension__
({ if (input_stride[0] != 0) ; else __assert_fail ("input_stride[0] != 0"
, "ccv_cnnp_model_addons.c", 827, __extension__ __PRETTY_FUNCTION__
); }));
828 int output_ofs[CCV_NNC_MAX_DIM_ALLOC(12)] = {};
829 int i;
830 for (i = 0; i < nd; i++)
831 {
832 const int idx = self->index[i];
833 assert(idx >= 0 && idx < nd)((void) sizeof ((idx >= 0 && idx < nd) ? 1 : 0)
, __extension__ ({ if (idx >= 0 && idx < nd) ; else
__assert_fail ("idx >= 0 && idx < nd", "ccv_cnnp_model_addons.c"
, 833, __extension__ __PRETTY_FUNCTION__); }));
834 params.dim[i] = input_dim[idx];
835 output_stride[i] = input_stride[idx];
836 output_ofs[i] = input_ofs[idx];
837 }
838 outputs[0] = ccv_nnc_tensor_symbol_alias_new(graph, inputs[0], output_ofs, output_stride, params, 0);
839 }
840}
841
842static ccv_cnnp_model_t* _ccv_cnnp_permute_copy(const ccv_cnnp_model_t* const super, void* const context);
843
844static const ccv_cnnp_model_vtab_t ccv_cnnp_permute_isa = {
845 .build = _ccv_cnnp_permute_build,
846 .copy = _ccv_cnnp_permute_copy,
847};
848
849ccv_cnnp_model_t* ccv_cnnp_permute(const int index[CCV_NNC_MAX_DIM_ALLOC(12)], const char* const name)
850{
851 ccv_cnnp_model_permute_t* const model_permute = (ccv_cnnp_model_permute_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_permute_t));
852 model_permute->super.isa = &ccv_cnnp_permute_isa;
853 model_permute->super.input_size = 1;
854 model_permute->super.outputs = &model_permute->output;
855 model_permute->super.output_size = 1;
856 ccv_cnnp_model_copy_name(&model_permute->super, name);
857 memcpy(model_permute->index, index, sizeof(model_permute->index));
858 return (ccv_cnnp_model_t*)model_permute;
859}
860
861static ccv_cnnp_model_t* _ccv_cnnp_permute_copy(const ccv_cnnp_model_t* const super, void* const context)
862{
863 const ccv_cnnp_model_permute_t* const self = (const ccv_cnnp_model_permute_t*)super;
864 return ccv_cnnp_permute(self->index, self->super.name);
865}
866
867typedef struct {
868 ccv_cnnp_model_t super;
869 int index;
870 ccv_nnc_tensor_symbol_t output;
871} ccv_cnnp_model_extract_t;
872
873static void _ccv_cnnp_extract_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
874{
875 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 875, __extension__ __PRETTY_FUNCTION__
); }));
876 ccv_cnnp_model_extract_t* const self = (ccv_cnnp_model_extract_t*)super;
877 PRINT(CCV_CLI_VERBOSE, "[cnnp_extract_build] index: %d\n", self->index)do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_extract_build] index: %d\n", self->index)
; fflush(stdout); } } while (0);
878 outputs[0] = inputs[self->index];
879}
880
881static ccv_cnnp_model_t* _ccv_cnnp_extract_copy(const ccv_cnnp_model_t* const self, void* const context);
882
883static const ccv_cnnp_model_vtab_t ccv_cnnp_extract_isa = {
884 .build = _ccv_cnnp_extract_build,
885 .copy = _ccv_cnnp_extract_copy,
886};
887
888ccv_cnnp_model_t* ccv_cnnp_extract(const int index, const char* const name)
889{
890 ccv_cnnp_model_extract_t* const model_extract = (ccv_cnnp_model_extract_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_extract_t));
891 model_extract->index = index;
892 model_extract->super.isa = &ccv_cnnp_extract_isa;
893 model_extract->super.input_size = 0;
894 model_extract->super.outputs = &model_extract->output;
895 model_extract->super.output_size = 1;
896 ccv_cnnp_model_copy_name(&model_extract->super, name);
897 return (ccv_cnnp_model_t*)model_extract;
898}
899
900static ccv_cnnp_model_t* _ccv_cnnp_extract_copy(const ccv_cnnp_model_t* const super, void* const context)
901{
902 ccv_cnnp_model_extract_t* const self = (ccv_cnnp_model_extract_t*)super;
903 return ccv_cnnp_extract(self->index, self->super.name);
904}
905
906typedef struct {
907 ccv_cnnp_model_t super;
908 ccv_nnc_tensor_symbol_t output;
909} ccv_cnnp_model_flatten_t;
910
911static void _ccv_cnnp_flatten_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
912{
913 PRINT(CCV_CLI_VERBOSE, "[cnnp_flatten_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_flatten_build] -\n"); fflush(stdout); } } while
(0);
914 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 914, __extension__ __PRETTY_FUNCTION__); }));
915 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 915, __extension__ __PRETTY_FUNCTION__
); }));
916 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
917 ccv_nnc_tensor_param_t output_params = params;
918 memset(output_params.dim, 0, sizeof(output_params.dim));
919 output_params.dim[0] = ccv_nnc_tensor_get_n(params);
920 assert(output_params.dim[0] > 0)((void) sizeof ((output_params.dim[0] > 0) ? 1 : 0), __extension__
({ if (output_params.dim[0] > 0) ; else __assert_fail ("output_params.dim[0] > 0"
, "ccv_cnnp_model_addons.c", 920, __extension__ __PRETTY_FUNCTION__
); }));
921 output_params.dim[1] = ccv_nnc_tensor_count(params) / output_params.dim[0];
922 int stride[CCV_NNC_MAX_DIM_ALLOC(12)] = {};
923 ccv_nnc_tensor_get_stride(output_params.dim, stride);
924 outputs[0] = ccv_nnc_tensor_symbol_alias_new(graph, inputs[0], DIM_ALLOC()(int [(12)]){}, stride, output_params, 0);
925}
926
927static ccv_cnnp_model_t* _ccv_cnnp_flatten_copy(const ccv_cnnp_model_t* const self, void* const context);
928
929static const ccv_cnnp_model_vtab_t ccv_cnnp_flatten_isa = {
930 .build = _ccv_cnnp_flatten_build,
931 .copy = _ccv_cnnp_flatten_copy,
932};
933
934ccv_cnnp_model_t* ccv_cnnp_flatten(const char* const name)
935{
936 ccv_cnnp_model_flatten_t* const model_flatten = (ccv_cnnp_model_flatten_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_flatten_t));
937 model_flatten->super.isa = &ccv_cnnp_flatten_isa;
938 model_flatten->super.input_size = 1;
939 model_flatten->super.outputs = &model_flatten->output;
940 model_flatten->super.output_size = 1;
941 ccv_cnnp_model_copy_name(&model_flatten->super, name);
942 return (ccv_cnnp_model_t*)model_flatten;
943}
944
945static ccv_cnnp_model_t* _ccv_cnnp_flatten_copy(const ccv_cnnp_model_t* const self, void* const context)
946{
947 return ccv_cnnp_flatten(self->name);
948}
949
950// MARK - Batch Norm Layer
951
952typedef struct {
953 ccv_cnnp_model_t super;
954 ccv_nnc_tensor_symbol_t output;
955 ccv_nnc_tensor_symbol_t bias;
956 ccv_nnc_tensor_symbol_t scale;
957 ccv_nnc_graph_exec_symbol_t batch_norm;
958 ccv_nnc_cmd_param_t params;
959 ccv_array_t* zero_inits;
960 ccv_array_t* retainables;
961} ccv_cnnp_model_batch_norm_t;
962
963static void _ccv_cnnp_batch_norm_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
964{
965 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 965, __extension__ __PRETTY_FUNCTION__); }));
966 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 966, __extension__ __PRETTY_FUNCTION__
); }));
967 ccv_cnnp_model_batch_norm_t* const self = (ccv_cnnp_model_batch_norm_t*)super;
968 PRINT(CCV_CLI_VERBOSE, "[cnnp_batch_norm_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_batch_norm_build] -\n"); fflush(stdout); } }
while (0);
969 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
970 const int nd = ccv_nnc_tensor_nd(params.dim);
971 ccv_nnc_tensor_param_t bias_params = params;
972 memset(bias_params.dim, 0, sizeof(bias_params.dim));
973 // If the accuracy is not enough, bump it to 32-bit floating point.
974 if (bias_params.datatype != CCV_32F && bias_params.datatype != CCV_64F)
975 bias_params.datatype = CCV_32F;
976 bias_params.dim[0] = nd > 1 ? ccv_nnc_tensor_get_c(params) : params.dim[0];
977 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, params, 0);
978 // Both scale and bias are shared between if this model is reused.
979 if (!self->scale.graph)
980 self->scale = ccv_nnc_tensor_symbol_new(graph, bias_params, "scale");
981 if (!self->bias.graph)
982 self->bias = ccv_nnc_tensor_symbol_new(graph, bias_params, "bias");
983 const ccv_nnc_tensor_symbol_t scale = ccv_cnnp_model_get_symbol(super, self->scale);
984 const ccv_nnc_tensor_symbol_t bias = ccv_cnnp_model_get_symbol(super, self->bias);
985 const ccv_nnc_tensor_symbol_t mean = ccv_nnc_tensor_symbol_new(graph, bias_params, "mean");
986 const ccv_nnc_tensor_symbol_t var = ccv_nnc_tensor_symbol_new(graph, bias_params, "var");
987 // Otherwise, notice mean, var, saved_mean, saved_inv_std are not reused.
988 if (!self->zero_inits)
989 self->zero_inits = ccv_array_new(sizeof(ccv_nnc_tensor_symbol_t), 0, 0);
990 ccv_array_push(self->zero_inits, &mean);
991 ccv_array_push(self->zero_inits, &var);
992 const ccv_nnc_tensor_symbol_t out_mean = ccv_nnc_tensor_symbol_new(graph, bias_params, "out_mean");
993 const ccv_nnc_tensor_symbol_t out_var = ccv_nnc_tensor_symbol_new(graph, bias_params, "out_var");
994 if (!self->retainables)
995 self->retainables = ccv_array_new(sizeof(ccv_nnc_tensor_symbol_t), 0, 0);
996 ccv_array_push(self->retainables, &out_mean);
997 ccv_array_push(self->retainables, &out_var);
998 const ccv_nnc_tensor_symbol_t saved_mean = ccv_nnc_tensor_symbol_new(graph, bias_params, "saved_mean");
999 const ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(graph, bias_params, "saved_inv_std");
1000 const int hw = ccv_nnc_tensor_hw(params, ccv_nnc_tensor_nd(params.dim), CCV_NNC_MAX_DIM(2));
1001 ccv_nnc_cmd_param_t batch_norm = self->params;
1002 batch_norm.bnorm.count = hw >= 0 ? CCV_NNC_MAX_DIM(2) + 1 : 1;
1003 int i;
1004 batch_norm.bnorm.axis[0] = (params.format == CCV_TENSOR_FORMAT_CHWN) ? 3 : 0;
1005 if (hw >= 0)
1006 for (i = 0; i < CCV_NNC_MAX_DIM(2); i++)
1007 batch_norm.bnorm.axis[i + 1] = i + hw;
1008 self->params = batch_norm;
1009 self->batch_norm = ccv_nnc_graph_exec_symbol_new(graph, ccv_nnc_cmd(CCV_NNC_BATCH_NORM_FORWARD, 0, batch_norm, 0), TENSOR_SYMBOL_LIST(inputs[0], scale, bias, mean, var)(const ccv_nnc_tensor_symbol_t []){inputs[0], scale, bias, mean
, var}, (1 +1 +1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(output, out_mean, out_var, saved_mean, saved_inv_std)(const ccv_nnc_tensor_symbol_t []){output, out_mean, out_var,
saved_mean, saved_inv_std}, (1 +1 +1 +1 +1 +1 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "batch_norm");
1010 outputs[0] = output;
1011}
1012
1013static void _ccv_cnnp_batch_norm_init_states(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_cnnp_state_initializer_f initializer, void* const context)
1014{
1015 ccv_cnnp_model_batch_norm_t* const self = (ccv_cnnp_model_batch_norm_t*)super;
1016 if (self->scale.graph)
1017 initializer(context, CMD_RANDOM_UNIFORM_FORWARD(0, 1)ccv_nnc_cmd(CCV_NNC_RANDOM_UNIFORM_FORWARD, 0, (ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.blas={.a={0, 1}}}, 0), ccv_nnc_no_hint, 0, 0, self->scale);
1018 if (self->bias.graph)
1019 initializer(context, CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, self->bias);
1020 int i;
1021 if (self->zero_inits)
1022 for (i = 0; i < self->zero_inits->rnum; i++)
1023 initializer(context, CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, *(ccv_nnc_tensor_symbol_t*)ccv_array_get(self->zero_inits, i)((void*)(((char*)((self->zero_inits)->data)) + (size_t)
(self->zero_inits)->rsize * (size_t)(i))));
1024}
1025
1026static void _ccv_cnnp_batch_norm_add_to_parameter(ccv_cnnp_model_t* const super, const ccv_cnnp_add_to_array_f add_to_array, void* const parameters, const int is_trainable)
1027{
1028 ccv_cnnp_model_batch_norm_t* const self = (ccv_cnnp_model_batch_norm_t*)super;
1029 if (self->scale.graph)
1030 add_to_array(parameters, self->scale, is_trainable);
1031 if (self->bias.graph)
1032 add_to_array(parameters, self->bias, is_trainable);
1033}
1034
1035static void _ccv_cnnp_batch_norm_add_to_output(ccv_cnnp_model_t* const super, const ccv_cnnp_add_to_array_f add_to_array, void* const outputs)
1036{
1037 ccv_cnnp_model_batch_norm_t* const self = (ccv_cnnp_model_batch_norm_t*)super;
1038 int i;
1039 if (self->retainables)
1040 for (i = 0; i < self->retainables->rnum; i++)
1041 {
1042 const ccv_nnc_tensor_symbol_t symbol = *(ccv_nnc_tensor_symbol_t*)ccv_array_get(self->retainables, i)((void*)(((char*)((self->retainables)->data)) + (size_t
)(self->retainables)->rsize * (size_t)(i)));
1043 add_to_array(outputs, symbol, 0);
1044 }
1045}
1046
1047static void _ccv_cnnp_batch_norm_set_is_test(ccv_cnnp_model_t* const super, const int is_test, const ccv_cnnp_cmd_updater_f updater, void* const context)
1048{
1049 ccv_cnnp_model_batch_norm_t* const self = (ccv_cnnp_model_batch_norm_t*)super;
1050 if (self->batch_norm.graph)
1051 {
1052 self->params.bnorm.is_test = is_test;
1053 updater(context, self->batch_norm, ccv_nnc_cmd(CCV_NNC_BATCH_NORM_FORWARD, 0, self->params, 0), ccv_nnc_no_hint);
1054 }
1055}
1056
1057static void _ccv_cnnp_batch_norm_deinit(ccv_cnnp_model_t* const super)
1058{
1059 ccv_cnnp_model_batch_norm_t* const self = (ccv_cnnp_model_batch_norm_t*)super;
1060 if (self->zero_inits)
1061 ccv_array_free(self->zero_inits);
1062 if (self->retainables)
1063 ccv_array_free(self->retainables);
1064}
1065
1066static ccv_cnnp_model_t* _ccv_cnnp_batch_norm_copy(const ccv_cnnp_model_t* const super, void* const context);
1067
1068static const ccv_cnnp_model_vtab_t ccv_cnnp_batch_norm_isa = {
1069 .build = _ccv_cnnp_batch_norm_build,
1070 .init_states = _ccv_cnnp_batch_norm_init_states,
1071 .add_to_parameter = _ccv_cnnp_batch_norm_add_to_parameter,
1072 .add_to_output = _ccv_cnnp_batch_norm_add_to_output,
1073 .copy = _ccv_cnnp_batch_norm_copy,
1074 .set_is_test = _ccv_cnnp_batch_norm_set_is_test,
1075 .deinit = _ccv_cnnp_batch_norm_deinit,
1076};
1077
1078ccv_cnnp_model_t* ccv_cnnp_batch_norm(const float momentum, const float epsilon, const int is_trainable, const char* const name)
1079{
1080 ccv_cnnp_model_batch_norm_t* const model_batch_norm = (ccv_cnnp_model_batch_norm_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_batch_norm_t));
1081 model_batch_norm->super.isa = &ccv_cnnp_batch_norm_isa;
1082 model_batch_norm->super.input_size = 1;
1083 model_batch_norm->super.outputs = &model_batch_norm->output;
1084 model_batch_norm->super.output_size = 1;
1085 model_batch_norm->super.is_trainable = is_trainable;
1086 ccv_cnnp_model_copy_name(&model_batch_norm->super, name);
1087 model_batch_norm->scale.d = CCV_NNC_NO_TENSOR_SYMBOL;
1088 model_batch_norm->scale.graph = 0;
1089 model_batch_norm->bias.d = CCV_NNC_NO_TENSOR_SYMBOL;
1090 model_batch_norm->bias.graph = 0;
1091 model_batch_norm->params.bnorm.momentum = momentum;
1092 model_batch_norm->params.bnorm.epsilon = epsilon;
1093 return (ccv_cnnp_model_t*)model_batch_norm;
1094}
1095
1096static ccv_cnnp_model_t* _ccv_cnnp_batch_norm_copy(const ccv_cnnp_model_t* const super, void* const context)
1097{
1098 const ccv_cnnp_model_batch_norm_t* const self = (const ccv_cnnp_model_batch_norm_t*)super;
1099 return ccv_cnnp_batch_norm(self->params.bnorm.momentum, self->params.bnorm.epsilon, self->super.is_trainable, self->super.name);
1100}
1101
1102// MARK - Convolution Layer
1103
1104typedef struct {
1105 ccv_cnnp_model_t super;
1106 ccv_nnc_tensor_symbol_t output;
1107 ccv_nnc_tensor_symbol_t weights;
1108 ccv_nnc_tensor_symbol_t bias;
1109 int groups;
1110 int filters;
1111 int kdim[CCV_NNC_MAX_DIM_ALLOC(12)];
1112 int dilation[CCV_NNC_MAX_DIM_ALLOC(12)];
1113 int no_bias;
1114 int format;
1115 ccv_nnc_hint_t hint;
1116} ccv_cnnp_model_convolution_t;
1117
1118static void _ccv_cnnp_convolution_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
1119{
1120 ccv_cnnp_model_convolution_t* const self = (ccv_cnnp_model_convolution_t*)super;
1121 PRINT(CCV_CLI_VERBOSE, "[cnnp_convolution_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_convolution_build] -\n"); fflush(stdout); } }
while (0);
1122 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 1122, __extension__ __PRETTY_FUNCTION__); }));
1123 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 1123, __extension__ __PRETTY_FUNCTION__
); }));
1124 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1125 int i;
1126 const int k_nd = ccv_nnc_tensor_nd(self->kdim);
1127 const int nd = k_nd + 2;
1128 ccv_nnc_tensor_param_t weights_params = params;
1129 if (self->format)
1130 weights_params.format = self->format;
1131 ccv_nnc_tensor_set_n(&weights_params, self->filters);
1132 const int a_nd = ccv_nnc_tensor_nd(params.dim);
1133 int c;
1134 switch (params.format)
1135 {
1136 case CCV_TENSOR_FORMAT_NHWC:
1137 c = params.dim[a_nd - 1];
1138 break;
1139 case CCV_TENSOR_FORMAT_NCHW:
1140 if (a_nd == k_nd + 1)
1141 c = params.dim[0];
1142 else
1143 c = params.dim[a_nd <= 1 ? 0 : 1];
1144 break;
1145 case CCV_TENSOR_FORMAT_CHWN:
1146 c = params.dim[0];
1147 break;
1148 }
1149 assert(c % self->groups == 0)((void) sizeof ((c % self->groups == 0) ? 1 : 0), __extension__
({ if (c % self->groups == 0) ; else __assert_fail ("c % self->groups == 0"
, "ccv_cnnp_model_addons.c", 1149, __extension__ __PRETTY_FUNCTION__
); }));
1150 ccv_nnc_tensor_set_c(&weights_params, nd, c / self->groups);
1151 int hw = -1;
1152 if (weights_params.format == CCV_TENSOR_FORMAT_NHWC || weights_params.format == CCV_TENSOR_FORMAT_CHWN)
1153 hw = 1;
1154 else if (weights_params.format == CCV_TENSOR_FORMAT_NCHW)
1155 hw = 2;
1156 assert(hw >= 0)((void) sizeof ((hw >= 0) ? 1 : 0), __extension__ ({ if (hw
>= 0) ; else __assert_fail ("hw >= 0", "ccv_cnnp_model_addons.c"
, 1156, __extension__ __PRETTY_FUNCTION__); }));
1157 for (i = 0; i < k_nd; i++)
1158 weights_params.dim[i + hw] = self->kdim[i];
1159 if (!self->weights.graph)
1160 self->weights = ccv_nnc_tensor_symbol_new(graph, weights_params, "weights");
1161 assert(self->weights.graph == graph)((void) sizeof ((self->weights.graph == graph) ? 1 : 0), __extension__
({ if (self->weights.graph == graph) ; else __assert_fail
("self->weights.graph == graph", "ccv_cnnp_model_addons.c"
, 1161, __extension__ __PRETTY_FUNCTION__); }));
1162 const ccv_nnc_tensor_symbol_t weights = ccv_cnnp_model_get_symbol(super, self->weights);
1163 ccv_nnc_tensor_param_t bias_params = params;
1164 if (self->format)
1165 bias_params.format = self->format;
1166 memset(bias_params.dim, 0, sizeof(bias_params.dim));
1167 bias_params.dim[0] = self->filters;
1168 ccv_nnc_cmd_t cmd = CMD_CONVOLUTION_FORWARD(self->groups, self->filters)ccv_nnc_cmd(CCV_NNC_CONVOLUTION_FORWARD, 0, ((ccv_nnc_cmd_param_t
){.size={.dim={}},.convolution={.count=self->filters,.groups
=self->groups}}), 0);
1169 for (i = 0; i < k_nd; i++)
1170 cmd.info.size.dim[i] = self->kdim[i];
1171 cmd.info.size.dim[k_nd] = c;
1172 memcpy(cmd.info.convolution.dilation, self->dilation, sizeof(self->dilation));
1173 ccv_nnc_tensor_param_t output_params;
1174 // Dilate weight size based on the dilation factor.
1175 for (i = 0; i < k_nd; i++)
1176 weights_params.dim[i + hw] = (self->kdim[i] - 1) * ccv_max(self->dilation[i], 1)({ typeof (self->dilation[i]) _a = (self->dilation[i]);
typeof (1) _b = (1); (_a > _b) ? _a : _b; }) + 1;
1177 ccv_nnc_hint_tensor_auto(cmd, (ccv_nnc_tensor_param_t []){
1178 params,
1179 weights_params,
1180 bias_params,
1181 }, 3, self->hint, &output_params, 1);
1182 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1183 ccv_nnc_graph_exec_symbol_t convolution;
1184 if (self->no_bias)
1185 convolution = ccv_nnc_graph_exec_symbol_new(graph, cmd, TENSOR_SYMBOL_LIST(inputs[0], weights)(const ccv_nnc_tensor_symbol_t []){inputs[0], weights}, (1 +1
+1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 -1), TENSOR_SYMBOL_LIST(output)(const ccv_nnc_tensor_symbol_t []){output}, (1 +1 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "convolution");
1186 else {
1187 if (!self->bias.graph)
1188 self->bias = ccv_nnc_tensor_symbol_new(graph, bias_params, "bias");
1189 const ccv_nnc_tensor_symbol_t bias = ccv_cnnp_model_get_symbol(super, self->bias);
1190 convolution = ccv_nnc_graph_exec_symbol_new(graph, cmd, TENSOR_SYMBOL_LIST(inputs[0], weights, bias)(const ccv_nnc_tensor_symbol_t []){inputs[0], weights, bias},
(1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 -1), TENSOR_SYMBOL_LIST(output)(const ccv_nnc_tensor_symbol_t []){output}, (1 +1 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "convolution");
1191 }
1192 ccv_nnc_graph_exec_symbol_set_hint(graph, convolution, self->hint);
1193 outputs[0] = output;
1194}
1195
1196static void _ccv_cnnp_convolution_init_states(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_cnnp_state_initializer_f initializer, void* const context)
1197{
1198 ccv_cnnp_model_convolution_t* const self = (ccv_cnnp_model_convolution_t*)super;
1199 const ccv_nnc_tensor_param_t weight_params = ccv_nnc_tensor_symbol_params(graph, self->weights);
1200 const int n = ccv_max(ccv_nnc_tensor_get_n(weight_params), 1)({ typeof (ccv_nnc_tensor_get_n(weight_params)) _a = (ccv_nnc_tensor_get_n
(weight_params)); typeof (1) _b = (1); (_a > _b) ? _a : _b
; });
1201 const int count = ccv_nnc_tensor_count(weight_params);
1202 const float std = sqrtf(2) / sqrtf(count / n);
1203 const float bound = sqrtf(3) * std;
1204 initializer(context, CMD_RANDOM_UNIFORM_FORWARD(-bound, bound)ccv_nnc_cmd(CCV_NNC_RANDOM_UNIFORM_FORWARD, 0, (ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.blas={.a={-bound, bound}}}, 0), ccv_nnc_no_hint, 0, 0, self->weights);
1205 if (self->bias.graph)
1206 initializer(context, CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, self->bias);
1207}
1208
1209static void _ccv_cnnp_convolution_add_to_parameter(ccv_cnnp_model_t* const super, const ccv_cnnp_add_to_array_f add_to_array, void* const parameters, const int is_trainable)
1210{
1211 ccv_cnnp_model_convolution_t* const self = (ccv_cnnp_model_convolution_t*)super;
1212 add_to_array(parameters, self->weights, is_trainable);
1213 if (self->bias.graph)
1214 add_to_array(parameters, self->bias, is_trainable);
1215}
1216
1217static ccv_cnnp_model_t* _ccv_cnnp_convolution_copy(const ccv_cnnp_model_t* const super, void* const context);
1218
1219static const ccv_cnnp_model_vtab_t ccv_cnnp_convolution_isa = {
1220 .build = _ccv_cnnp_convolution_build,
1221 .init_states = _ccv_cnnp_convolution_init_states,
1222 .add_to_parameter = _ccv_cnnp_convolution_add_to_parameter,
1223 .copy = _ccv_cnnp_convolution_copy,
1224};
1225
1226ccv_cnnp_model_t* ccv_cnnp_convolution(const int groups, const int filters, const int kdim[CCV_NNC_MAX_DIM_ALLOC(12)], const int dilation[CCV_NNC_MAX_DIM_ALLOC(12)], const int no_bias, ccv_nnc_hint_t hint, const int format, const int is_trainable, const char* const name)
1227{
1228 ccv_cnnp_model_convolution_t* const model_convolution = (ccv_cnnp_model_convolution_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_convolution_t));
1229 model_convolution->super.isa = &ccv_cnnp_convolution_isa;
1230 model_convolution->super.input_size = 1;
1231 model_convolution->super.outputs = &model_convolution->output;
1232 model_convolution->super.output_size = 1;
1233 model_convolution->super.is_trainable = is_trainable;
1234 ccv_cnnp_model_copy_name(&model_convolution->super, name);
1235 model_convolution->weights.d = CCV_NNC_NO_TENSOR_SYMBOL;
1236 model_convolution->weights.graph = 0;
1237 model_convolution->bias.d = CCV_NNC_NO_TENSOR_SYMBOL;
1238 model_convolution->bias.graph = 0;
1239 model_convolution->groups = groups;
1240 model_convolution->filters = filters;
1241 memcpy(model_convolution->kdim, kdim, sizeof(model_convolution->kdim));
1242 memcpy(model_convolution->dilation, dilation, sizeof(model_convolution->dilation));
1243 model_convolution->no_bias = no_bias;
1244 model_convolution->hint = hint;
1245 model_convolution->format = format;
1246 return (ccv_cnnp_model_t*)model_convolution;
1247}
1248
1249static ccv_cnnp_model_t* _ccv_cnnp_convolution_copy(const ccv_cnnp_model_t* const super, void* const context)
1250{
1251 ccv_cnnp_model_convolution_t* const self = (ccv_cnnp_model_convolution_t*)super;
1252 return ccv_cnnp_convolution(self->groups, self->filters, self->kdim, self->dilation, self->no_bias, self->hint, self->format, self->super.is_trainable, self->super.name);
1253}
1254
1255// MARK - Convolution Transpose Layer
1256
1257typedef struct {
1258 ccv_cnnp_model_t super;
1259 ccv_nnc_tensor_symbol_t output;
1260 ccv_nnc_tensor_symbol_t weights;
1261 ccv_nnc_tensor_symbol_t bias;
1262 int groups;
1263 int filters;
1264 int kdim[CCV_NNC_MAX_DIM_ALLOC(12)];
1265 int dilation[CCV_NNC_MAX_DIM_ALLOC(12)];
1266 int output_padding;
1267 int no_bias;
1268 int format;
1269 ccv_nnc_hint_t hint;
1270} ccv_cnnp_model_convolution_transpose_t;
1271
1272static void _ccv_cnnp_convolution_transpose_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
1273{
1274 ccv_cnnp_model_convolution_transpose_t* const self = (ccv_cnnp_model_convolution_transpose_t*)super;
1275 PRINT(CCV_CLI_VERBOSE, "[cnnp_convolution_transpose_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_convolution_transpose_build] -\n"); fflush(stdout
); } } while (0);
1276 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 1276, __extension__ __PRETTY_FUNCTION__); }));
1277 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 1277, __extension__ __PRETTY_FUNCTION__
); }));
1278 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1279 int i;
1280 const int nd = CCV_NNC_MAX_DIM(2) + 2;
1281 ccv_nnc_tensor_param_t weights_params = params;
1282 if (self->format)
1283 weights_params.format = self->format;
1284 const int c = ccv_nnc_tensor_get_c(params);
1285 ccv_nnc_tensor_set_n(&weights_params, c);
1286 assert(c % self->groups == 0)((void) sizeof ((c % self->groups == 0) ? 1 : 0), __extension__
({ if (c % self->groups == 0) ; else __assert_fail ("c % self->groups == 0"
, "ccv_cnnp_model_addons.c", 1286, __extension__ __PRETTY_FUNCTION__
); }));
1287 ccv_nnc_tensor_set_c(&weights_params, nd, self->filters / self->groups);
1288 const int hw = ccv_nnc_tensor_hw(weights_params, nd, CCV_NNC_MAX_DIM(2));
1289 assert(hw >= 0)((void) sizeof ((hw >= 0) ? 1 : 0), __extension__ ({ if (hw
>= 0) ; else __assert_fail ("hw >= 0", "ccv_cnnp_model_addons.c"
, 1289, __extension__ __PRETTY_FUNCTION__); }));
1290 for (i = 0; i < CCV_NNC_MAX_DIM(2); i++)
1291 weights_params.dim[i + hw] = self->kdim[i];
1292 if (!self->weights.graph)
1293 self->weights = ccv_nnc_tensor_symbol_new(graph, weights_params, "weights");
1294 assert(self->weights.graph == graph)((void) sizeof ((self->weights.graph == graph) ? 1 : 0), __extension__
({ if (self->weights.graph == graph) ; else __assert_fail
("self->weights.graph == graph", "ccv_cnnp_model_addons.c"
, 1294, __extension__ __PRETTY_FUNCTION__); }));
1295 const ccv_nnc_tensor_symbol_t weights = ccv_cnnp_model_get_symbol(super, self->weights);
1296 ccv_nnc_tensor_param_t bias_params = params;
1297 if (self->format)
1298 bias_params.format = self->format;
1299 memset(bias_params.dim, 0, sizeof(bias_params.dim));
1300 bias_params.dim[0] = self->filters;
1301 ccv_nnc_cmd_t cmd = CMD_CONVOLUTION_TRANSPOSE_FORWARD(self->groups, self->filters, self->output_padding)ccv_nnc_cmd(CCV_NNC_CONVOLUTION_TRANSPOSE_FORWARD, 0, ((ccv_nnc_cmd_param_t
){.size={.dim={}},.convolution_transpose={.count=self->filters
,.groups=self->groups,.output_padding=self->output_padding
}}), 0);
1302 for (i = 0; i < CCV_NNC_MAX_DIM(2); i++)
1303 cmd.info.size.dim[i] = self->kdim[i];
1304 cmd.info.size.dim[CCV_NNC_MAX_DIM(2)] = c;
1305 memcpy(cmd.info.convolution_transpose.dilation, self->dilation, sizeof(self->dilation));
1306 ccv_nnc_tensor_param_t output_params;
1307 // Dilate weight size based on the dilation factor.
1308 for (i = 0; i < CCV_NNC_MAX_DIM(2); i++)
1309 weights_params.dim[i + hw] = (self->kdim[i] - 1) * ccv_max(self->dilation[i], 1)({ typeof (self->dilation[i]) _a = (self->dilation[i]);
typeof (1) _b = (1); (_a > _b) ? _a : _b; }) + 1;
1310 ccv_nnc_hint_tensor_auto(cmd, (ccv_nnc_tensor_param_t []){
1311 params,
1312 weights_params,
1313 bias_params,
1314 }, 3, self->hint, &output_params, 1);
1315 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1316 ccv_nnc_graph_exec_symbol_t convolution_transpose;
1317 if (self->no_bias)
1318 convolution_transpose = ccv_nnc_graph_exec_symbol_new(graph, cmd, TENSOR_SYMBOL_LIST(inputs[0], weights)(const ccv_nnc_tensor_symbol_t []){inputs[0], weights}, (1 +1
+1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 -1), TENSOR_SYMBOL_LIST(output)(const ccv_nnc_tensor_symbol_t []){output}, (1 +1 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "convolution_transpose");
1319 else {
1320 if (!self->bias.graph)
1321 self->bias = ccv_nnc_tensor_symbol_new(graph, bias_params, "bias");
1322 const ccv_nnc_tensor_symbol_t bias = ccv_cnnp_model_get_symbol(super, self->bias);
1323 convolution_transpose = ccv_nnc_graph_exec_symbol_new(graph, cmd, TENSOR_SYMBOL_LIST(inputs[0], weights, bias)(const ccv_nnc_tensor_symbol_t []){inputs[0], weights, bias},
(1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 -1), TENSOR_SYMBOL_LIST(output)(const ccv_nnc_tensor_symbol_t []){output}, (1 +1 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "convolution_transpose");
1324 }
1325 ccv_nnc_graph_exec_symbol_set_hint(graph, convolution_transpose, self->hint);
1326 outputs[0] = output;
1327}
1328
1329static void _ccv_cnnp_convolution_transpose_init_states(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_cnnp_state_initializer_f initializer, void* const context)
1330{
1331 ccv_cnnp_model_convolution_transpose_t* const self = (ccv_cnnp_model_convolution_transpose_t*)super;
1332 const ccv_nnc_tensor_param_t weight_params = ccv_nnc_tensor_symbol_params(graph, self->weights);
1333 const int n = ccv_max(ccv_nnc_tensor_get_n(weight_params), 1)({ typeof (ccv_nnc_tensor_get_n(weight_params)) _a = (ccv_nnc_tensor_get_n
(weight_params)); typeof (1) _b = (1); (_a > _b) ? _a : _b
; });
1334 const int count = ccv_nnc_tensor_count(weight_params);
1335 const float std = sqrtf(2) / sqrtf(count / n);
1336 const float bound = sqrtf(3) * std;
1337 initializer(context, CMD_RANDOM_UNIFORM_FORWARD(-bound, bound)ccv_nnc_cmd(CCV_NNC_RANDOM_UNIFORM_FORWARD, 0, (ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.blas={.a={-bound, bound}}}, 0), ccv_nnc_no_hint, 0, 0, self->weights);
1338 if (self->bias.graph)
1339 initializer(context, CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, self->bias);
1340}
1341
1342static void _ccv_cnnp_convolution_transpose_add_to_parameter(ccv_cnnp_model_t* const super, const ccv_cnnp_add_to_array_f add_to_array, void* const parameters, const int is_trainable)
1343{
1344 ccv_cnnp_model_convolution_transpose_t* const self = (ccv_cnnp_model_convolution_transpose_t*)super;
1345 add_to_array(parameters, self->weights, is_trainable);
1346 if (self->bias.graph)
1347 add_to_array(parameters, self->bias, is_trainable);
1348}
1349
1350static ccv_cnnp_model_t* _ccv_cnnp_convolution_transpose_copy(const ccv_cnnp_model_t* const super, void* const context);
1351
1352static const ccv_cnnp_model_vtab_t ccv_cnnp_convolution_transpose_isa = {
1353 .build = _ccv_cnnp_convolution_transpose_build,
1354 .init_states = _ccv_cnnp_convolution_transpose_init_states,
1355 .add_to_parameter = _ccv_cnnp_convolution_transpose_add_to_parameter,
1356 .copy = _ccv_cnnp_convolution_transpose_copy,
1357};
1358
1359ccv_cnnp_model_t* ccv_cnnp_convolution_transpose(const int groups, const int filters, const int kdim[CCV_NNC_MAX_DIM_ALLOC(12)], const int dilation[CCV_NNC_MAX_DIM_ALLOC(12)], const int output_padding, const int no_bias, ccv_nnc_hint_t hint, const int format, const int is_trainable, const char* const name)
1360{
1361 ccv_cnnp_model_convolution_transpose_t* const model_convolution_transpose = (ccv_cnnp_model_convolution_transpose_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_convolution_transpose_t));
1362 model_convolution_transpose->super.isa = &ccv_cnnp_convolution_transpose_isa;
1363 model_convolution_transpose->super.input_size = 1;
1364 model_convolution_transpose->super.outputs = &model_convolution_transpose->output;
1365 model_convolution_transpose->super.output_size = 1;
1366 model_convolution_transpose->super.is_trainable = is_trainable;
1367 ccv_cnnp_model_copy_name(&model_convolution_transpose->super, name);
1368 model_convolution_transpose->weights.d = CCV_NNC_NO_TENSOR_SYMBOL;
1369 model_convolution_transpose->weights.graph = 0;
1370 model_convolution_transpose->bias.d = CCV_NNC_NO_TENSOR_SYMBOL;
1371 model_convolution_transpose->bias.graph = 0;
1372 model_convolution_transpose->groups = groups;
1373 model_convolution_transpose->filters = filters;
1374 memcpy(model_convolution_transpose->kdim, kdim, sizeof(model_convolution_transpose->kdim));
1375 memcpy(model_convolution_transpose->dilation, dilation, sizeof(model_convolution_transpose->dilation));
1376 model_convolution_transpose->output_padding = output_padding;
1377 model_convolution_transpose->no_bias = no_bias;
1378 model_convolution_transpose->hint = hint;
1379 model_convolution_transpose->format = format;
1380 return (ccv_cnnp_model_t*)model_convolution_transpose;
1381}
1382
1383static ccv_cnnp_model_t* _ccv_cnnp_convolution_transpose_copy(const ccv_cnnp_model_t* const super, void* const context)
1384{
1385 ccv_cnnp_model_convolution_transpose_t* const self = (ccv_cnnp_model_convolution_transpose_t*)super;
1386 return ccv_cnnp_convolution_transpose(self->groups, self->filters, self->kdim, self->dilation, self->output_padding, self->no_bias, self->hint, self->format, self->super.is_trainable, self->super.name);
1387}
1388
1389// MARK - Dense Layer
1390
1391typedef struct {
1392 ccv_cnnp_model_t super;
1393 ccv_nnc_tensor_symbol_t output;
1394 ccv_nnc_tensor_symbol_t weights;
1395 ccv_nnc_tensor_symbol_t bias;
1396 int count;
1397 int no_bias;
1398 int flags;
1399} ccv_cnnp_model_dense_t;
1400
1401static void _ccv_cnnp_dense_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
1402{
1403 ccv_cnnp_model_dense_t* const self = (ccv_cnnp_model_dense_t*)super;
1404 PRINT(CCV_CLI_VERBOSE, "[cnnp_dense_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_dense_build] -\n"); fflush(stdout); } } while
(0);
1405 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 1405, __extension__ __PRETTY_FUNCTION__); }));
1406 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 1406, __extension__ __PRETTY_FUNCTION__
); }));
1407 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1408 ccv_nnc_tensor_param_t weights_params = params;
1409 memset(weights_params.dim, 0, sizeof(weights_params.dim));
1410 weights_params.dim[0] = self->count;
1411 weights_params.dim[1] = params.dim[ccv_nnc_tensor_nd(params.dim) - 1];
1412 if (!self->weights.graph)
1413 self->weights = ccv_nnc_tensor_symbol_new(graph, weights_params, "weights");
1414 assert(self->weights.graph == graph)((void) sizeof ((self->weights.graph == graph) ? 1 : 0), __extension__
({ if (self->weights.graph == graph) ; else __assert_fail
("self->weights.graph == graph", "ccv_cnnp_model_addons.c"
, 1414, __extension__ __PRETTY_FUNCTION__); }));
1415 const ccv_nnc_tensor_symbol_t weights = ccv_cnnp_model_get_symbol(super, self->weights);
1416 ccv_nnc_tensor_param_t bias_params = params;
1417 memset(bias_params.dim, 0, sizeof(bias_params.dim));
1418 bias_params.dim[0] = self->count;
1419 ccv_nnc_cmd_t cmd = {0};
1420 cmd.cmd = CCV_NNC_GEMM_FORWARD;
1421 cmd.info.blas.a[0] = 1;
1422 cmd.info.blas.a[1] = 1;
1423 cmd.info.blas.transpose_b[0] = 0;
1424 cmd.info.blas.transpose_b[1] = 1;
1425 cmd.info.blas.flags = self->flags;
1426 ccv_nnc_tensor_param_t output_params;
1427 ccv_nnc_hint_tensor_auto(cmd, (ccv_nnc_tensor_param_t []){
1428 params,
1429 weights_params,
1430 bias_params,
1431 }, 3, ccv_nnc_no_hint, &output_params, 1);
1432 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1433 if (self->no_bias)
1434 ccv_nnc_graph_exec_symbol_new(graph, cmd, TENSOR_SYMBOL_LIST(inputs[0], weights)(const ccv_nnc_tensor_symbol_t []){inputs[0], weights}, (1 +1
+1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 -1), TENSOR_SYMBOL_LIST(output)(const ccv_nnc_tensor_symbol_t []){output}, (1 +1 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "dense");
1435 else {
1436 if (!self->bias.graph)
1437 self->bias = ccv_nnc_tensor_symbol_new(graph, bias_params, "bias");
1438 const ccv_nnc_tensor_symbol_t bias = ccv_cnnp_model_get_symbol(super, self->bias);
1439 ccv_nnc_graph_exec_symbol_new(graph, cmd, TENSOR_SYMBOL_LIST(inputs[0], weights, bias)(const ccv_nnc_tensor_symbol_t []){inputs[0], weights, bias},
(1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 -1), TENSOR_SYMBOL_LIST(output)(const ccv_nnc_tensor_symbol_t []){output}, (1 +1 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "dense");
1440 }
1441 outputs[0] = output;
1442}
1443
1444static void _ccv_cnnp_dense_init_states(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_cnnp_state_initializer_f initializer, void* const context)
1445{
1446 ccv_cnnp_model_dense_t* const self = (ccv_cnnp_model_dense_t*)super;
1447 const ccv_nnc_tensor_param_t weight_params = ccv_nnc_tensor_symbol_params(graph, self->weights);
1448 const int c = weight_params.dim[1];
1449 const float std = sqrtf(2) / sqrtf(c);
1450 const float bound = sqrtf(3) * std;
1451 initializer(context, CMD_RANDOM_UNIFORM_FORWARD(-bound, bound)ccv_nnc_cmd(CCV_NNC_RANDOM_UNIFORM_FORWARD, 0, (ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.blas={.a={-bound, bound}}}, 0), ccv_nnc_no_hint, 0, 0, self->weights);
1452 if (self->bias.graph)
1453 initializer(context, CMD_SET_FORWARD(0)ccv_nnc_cmd(CCV_NNC_SET_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={0,}}}, 0), ccv_nnc_no_hint, 0, 0, self->bias);
1454}
1455
1456static void _ccv_cnnp_dense_add_to_parameter(ccv_cnnp_model_t* const super, const ccv_cnnp_add_to_array_f add_to_array, void* const parameters, const int is_trainable)
1457{
1458 ccv_cnnp_model_dense_t* const self = (ccv_cnnp_model_dense_t*)super;
1459 add_to_array(parameters, self->weights, is_trainable);
1460 if (self->bias.graph)
1461 add_to_array(parameters, self->bias, is_trainable);
1462}
1463
1464static ccv_cnnp_model_t* _ccv_cnnp_dense_copy(const ccv_cnnp_model_t* const super, void* const context);
1465
1466static const ccv_cnnp_model_vtab_t ccv_cnnp_dense_isa = {
1467 .build = _ccv_cnnp_dense_build,
1468 .init_states = _ccv_cnnp_dense_init_states,
1469 .add_to_parameter = _ccv_cnnp_dense_add_to_parameter,
1470 .copy = _ccv_cnnp_dense_copy,
1471};
1472
1473ccv_cnnp_model_t* ccv_cnnp_dense(const int count, const int no_bias, const int flags, const int is_trainable, const char* const name)
1474{
1475 ccv_cnnp_model_dense_t* const model_dense = (ccv_cnnp_model_dense_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_dense_t));
1476 model_dense->super.isa = &ccv_cnnp_dense_isa;
1477 model_dense->super.input_size = 1;
1478 model_dense->super.outputs = &model_dense->output;
1479 model_dense->super.output_size = 1;
1480 model_dense->super.is_trainable = is_trainable;
1481 ccv_cnnp_model_copy_name(&model_dense->super, name);
1482 model_dense->weights.d = CCV_NNC_NO_TENSOR_SYMBOL;
1483 model_dense->weights.graph = 0;
1484 model_dense->bias.d = CCV_NNC_NO_TENSOR_SYMBOL;
1485 model_dense->bias.graph = 0;
1486 model_dense->count = count;
1487 model_dense->no_bias = no_bias;
1488 model_dense->flags = flags;
1489 return (ccv_cnnp_model_t*)model_dense;
1490}
1491
1492static ccv_cnnp_model_t* _ccv_cnnp_dense_copy(const ccv_cnnp_model_t* const super, void* const context)
1493{
1494 const ccv_cnnp_model_dense_t* const self = (const ccv_cnnp_model_dense_t*)super;
1495 return ccv_cnnp_dense(self->count, self->no_bias, self->flags, self->super.is_trainable, self->super.name);
1496}
1497
1498// MARK - Pool Layers
1499
1500typedef struct {
1501 ccv_cnnp_model_t super;
1502 ccv_nnc_tensor_symbol_t output;
1503 int kdim[CCV_NNC_MAX_DIM_ALLOC(12)];
1504 ccv_nnc_hint_t hint;
1505} ccv_cnnp_model_pool_t;
1506
1507static void _ccv_cnnp_max_pool_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
1508{
1509 ccv_cnnp_model_pool_t* const self = (ccv_cnnp_model_pool_t*)super;
1510 PRINT(CCV_CLI_VERBOSE, "[cnnp_max_pool_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_max_pool_build] -\n"); fflush(stdout); } } while
(0);
1511 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 1511, __extension__ __PRETTY_FUNCTION__); }));
1512 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 1512, __extension__ __PRETTY_FUNCTION__
); }));
1513 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1514 const int hw = ccv_nnc_tensor_hw(params, ccv_nnc_tensor_nd(params.dim), CCV_NNC_MAX_DIM(2));
1515 ccv_nnc_cmd_t cmd;
1516 if (hw >= 0 && self->kdim[0] == 0 && self->kdim[1] == 0)
1517 cmd = CMD_MAX_POOL_FORWARD(params.dim[hw], params.dim[hw + 1])ccv_nnc_cmd(CCV_NNC_MAX_POOL_FORWARD, 0, ((ccv_nnc_cmd_param_t
){.size={.dim={params.dim[hw], params.dim[hw + 1],1}}}), 0);
1518 else
1519 cmd = CMD_MAX_POOL_FORWARD(self->kdim[0], self->kdim[1])ccv_nnc_cmd(CCV_NNC_MAX_POOL_FORWARD, 0, ((ccv_nnc_cmd_param_t
){.size={.dim={self->kdim[0], self->kdim[1],1}}}), 0);
1520 ccv_nnc_tensor_param_t output_params;
1521 ccv_nnc_hint_tensor_auto(cmd, &params, 1, self->hint, &output_params, 1);
1522 const ccv_nnc_tensor_symbol_t pool_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1523 const ccv_nnc_graph_exec_symbol_t exec = ccv_nnc_graph_exec_symbol_new(graph, cmd, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(pool_output)(const ccv_nnc_tensor_symbol_t []){pool_output}, (1 +1 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "max_pool");
1524 ccv_nnc_graph_exec_symbol_set_hint(graph, exec, self->hint);
1525 outputs[0] = pool_output;
1526}
1527
1528static ccv_cnnp_model_t* _ccv_cnnp_max_pool_copy(const ccv_cnnp_model_t* const super, void* const context);
1529
1530static const ccv_cnnp_model_vtab_t ccv_cnnp_max_pool_isa = {
1531 .build = _ccv_cnnp_max_pool_build,
1532 .copy = _ccv_cnnp_max_pool_copy,
1533};
1534
1535ccv_cnnp_model_t* ccv_cnnp_max_pool(const int kdim[CCV_NNC_MAX_DIM_ALLOC(12)], const ccv_nnc_hint_t hint, const char* const name)
1536{
1537 ccv_cnnp_model_pool_t* const model_pool = (ccv_cnnp_model_pool_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_pool_t));
1538 model_pool->super.isa = &ccv_cnnp_max_pool_isa;
1539 model_pool->super.input_size = 1;
1540 model_pool->super.outputs = &model_pool->output;
1541 model_pool->super.output_size = 1;
1542 ccv_cnnp_model_copy_name(&model_pool->super, name);
1543 memcpy(model_pool->kdim, kdim, sizeof(model_pool->kdim));
1544 model_pool->hint = hint;
1545 return (ccv_cnnp_model_t*)model_pool;
1546}
1547
1548static ccv_cnnp_model_t* _ccv_cnnp_max_pool_copy(const ccv_cnnp_model_t* const super, void* const context)
1549{
1550 const ccv_cnnp_model_pool_t* const self = (const ccv_cnnp_model_pool_t*)super;
1551 return ccv_cnnp_max_pool(self->kdim, self->hint, self->super.name);
1552}
1553
1554static void _ccv_cnnp_average_pool_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
1555{
1556 ccv_cnnp_model_pool_t* const self = (ccv_cnnp_model_pool_t*)super;
1557 PRINT(CCV_CLI_VERBOSE, "[cnnp_average_pool_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_average_pool_build] -\n"); fflush(stdout); }
} while (0);
1558 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 1558, __extension__ __PRETTY_FUNCTION__); }));
1559 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 1559, __extension__ __PRETTY_FUNCTION__
); }));
1560 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1561 const int hw = ccv_nnc_tensor_hw(params, ccv_nnc_tensor_nd(params.dim), CCV_NNC_MAX_DIM(2));
1562 ccv_nnc_cmd_t cmd;
1563 if (hw >= 0 && self->kdim[0] == 0 && self->kdim[1] == 0)
1564 cmd = CMD_AVERAGE_POOL_FORWARD(params.dim[hw], params.dim[hw + 1])ccv_nnc_cmd(CCV_NNC_AVERAGE_POOL_FORWARD, 0, ((ccv_nnc_cmd_param_t
){.size={.dim={params.dim[hw], params.dim[hw + 1],1}}}), 0);
1565 else
1566 cmd = CMD_AVERAGE_POOL_FORWARD(self->kdim[0], self->kdim[1])ccv_nnc_cmd(CCV_NNC_AVERAGE_POOL_FORWARD, 0, ((ccv_nnc_cmd_param_t
){.size={.dim={self->kdim[0], self->kdim[1],1}}}), 0);
1567 ccv_nnc_tensor_param_t output_params;
1568 ccv_nnc_hint_tensor_auto(cmd, &params, 1, self->hint, &output_params, 1);
1569 const ccv_nnc_tensor_symbol_t pool_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1570 const ccv_nnc_graph_exec_symbol_t exec = ccv_nnc_graph_exec_symbol_new(graph, cmd, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(pool_output)(const ccv_nnc_tensor_symbol_t []){pool_output}, (1 +1 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "average_pool");
1571 ccv_nnc_graph_exec_symbol_set_hint(graph, exec, self->hint);
1572 outputs[0] = pool_output;
1573}
1574
1575static ccv_cnnp_model_t* _ccv_cnnp_average_pool_copy(const ccv_cnnp_model_t* const super, void* const context);
1576
1577static const ccv_cnnp_model_vtab_t ccv_cnnp_average_pool_isa = {
1578 .build = _ccv_cnnp_average_pool_build,
1579 .copy = _ccv_cnnp_average_pool_copy,
1580};
1581
1582ccv_cnnp_model_t* ccv_cnnp_average_pool(const int kdim[CCV_NNC_MAX_DIM_ALLOC(12)], const ccv_nnc_hint_t hint, const char* const name)
1583{
1584 ccv_cnnp_model_pool_t* const model_pool = (ccv_cnnp_model_pool_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_pool_t));
1585 model_pool->super.isa = &ccv_cnnp_average_pool_isa;
1586 model_pool->super.input_size = 1;
1587 model_pool->super.outputs = &model_pool->output;
1588 model_pool->super.output_size = 1;
1589 ccv_cnnp_model_copy_name(&model_pool->super, name);
1590 memcpy(model_pool->kdim, kdim, sizeof(model_pool->kdim));
1591 model_pool->hint = hint;
1592 return (ccv_cnnp_model_t*)model_pool;
1593}
1594
1595static ccv_cnnp_model_t* _ccv_cnnp_average_pool_copy(const ccv_cnnp_model_t* const super, void* const context)
1596{
1597 const ccv_cnnp_model_pool_t* const self = (const ccv_cnnp_model_pool_t*)super;
1598 return ccv_cnnp_average_pool(self->kdim, self->hint, self->super.name);
1599}
1600
1601// MARK - RELU Layer
1602
1603typedef struct {
1604 ccv_cnnp_model_t super;
1605 ccv_nnc_tensor_symbol_t output;
1606} ccv_cnnp_model_relu_t;
1607
1608static void _ccv_cnnp_relu_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
1609{
1610 PRINT(CCV_CLI_VERBOSE, "[cnnp_relu_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_relu_build] -\n"); fflush(stdout); } } while
(0);
1611 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 1611, __extension__ __PRETTY_FUNCTION__); }));
1612 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 1612, __extension__ __PRETTY_FUNCTION__
); }));
1613 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1614 ccv_nnc_tensor_param_t output_params;
1615 const ccv_nnc_cmd_t relu = CMD_RELU_FORWARD()ccv_nnc_cmd(CCV_NNC_RELU_FORWARD, 0, ccv_nnc_cmd_auto, 0);
1616 ccv_nnc_hint_tensor_auto(relu, (ccv_nnc_tensor_param_t []){
1617 params,
1618 }, 1, ccv_nnc_no_hint, &output_params, 1);
1619 const ccv_nnc_tensor_symbol_t relu_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1620 ccv_nnc_graph_exec_symbol_new(graph, relu, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(relu_output)(const ccv_nnc_tensor_symbol_t []){relu_output}, (1 +1 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "relu");
1621 outputs[0] = relu_output;
1622}
1623
1624static ccv_cnnp_model_t* _ccv_cnnp_relu_copy(const ccv_cnnp_model_t* const self, void* const context);
1625
1626static const ccv_cnnp_model_vtab_t ccv_cnnp_relu_isa = {
1627 .build = _ccv_cnnp_relu_build,
1628 .copy = _ccv_cnnp_relu_copy,
1629};
1630
1631ccv_cnnp_model_t* ccv_cnnp_relu(const char* const name)
1632{
1633 ccv_cnnp_model_relu_t* const model_relu = (ccv_cnnp_model_relu_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_relu_t));
1634 model_relu->super.isa = &ccv_cnnp_relu_isa;
1635 model_relu->super.input_size = 1;
1636 model_relu->super.outputs = &model_relu->output;
1637 model_relu->super.output_size = 1;
1638 ccv_cnnp_model_copy_name(&model_relu->super, name);
1639 return (ccv_cnnp_model_t*)model_relu;
1640}
1641
1642static ccv_cnnp_model_t* _ccv_cnnp_relu_copy(const ccv_cnnp_model_t* const self, void* const context)
1643{
1644 return ccv_cnnp_relu(self->name);
1645}
1646
1647// MARK - Sigmoid Layer
1648
1649typedef struct {
1650 ccv_cnnp_model_t super;
1651 ccv_nnc_tensor_symbol_t output;
1652} ccv_cnnp_model_sigmoid_t;
1653
1654static void _ccv_cnnp_sigmoid_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
1655{
1656 PRINT(CCV_CLI_VERBOSE, "[cnnp_sigmoid_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_sigmoid_build] -\n"); fflush(stdout); } } while
(0);
1657 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 1657, __extension__ __PRETTY_FUNCTION__); }));
1658 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 1658, __extension__ __PRETTY_FUNCTION__
); }));
1659 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1660 ccv_nnc_tensor_param_t output_params;
1661 const ccv_nnc_cmd_t sigmoid = CMD_SIGMOID_FORWARD()ccv_nnc_cmd(CCV_NNC_SIGMOID_FORWARD, 0, ccv_nnc_cmd_auto, 0);
1662 ccv_nnc_hint_tensor_auto(sigmoid, (ccv_nnc_tensor_param_t []){
1663 params,
1664 }, 1, ccv_nnc_no_hint, &output_params, 1);
1665 const ccv_nnc_tensor_symbol_t sigmoid_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1666 ccv_nnc_graph_exec_symbol_new(graph, sigmoid, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(sigmoid_output)(const ccv_nnc_tensor_symbol_t []){sigmoid_output}, (1 +1 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1
), "sigmoid");
1667 outputs[0] = sigmoid_output;
1668}
1669
1670static ccv_cnnp_model_t* _ccv_cnnp_sigmoid_copy(const ccv_cnnp_model_t* const self, void* const context);
1671
1672static const ccv_cnnp_model_vtab_t ccv_cnnp_sigmoid_isa = {
1673 .build = _ccv_cnnp_sigmoid_build,
1674 .copy = _ccv_cnnp_sigmoid_copy,
1675};
1676
1677ccv_cnnp_model_t* ccv_cnnp_sigmoid(const char* const name)
1678{
1679 ccv_cnnp_model_sigmoid_t* const model_sigmoid = (ccv_cnnp_model_sigmoid_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_sigmoid_t));
1680 model_sigmoid->super.isa = &ccv_cnnp_sigmoid_isa;
1681 model_sigmoid->super.input_size = 1;
1682 model_sigmoid->super.outputs = &model_sigmoid->output;
1683 model_sigmoid->super.output_size = 1;
1684 ccv_cnnp_model_copy_name(&model_sigmoid->super, name);
1685 return (ccv_cnnp_model_t*)model_sigmoid;
1686}
1687
1688static ccv_cnnp_model_t* _ccv_cnnp_sigmoid_copy(const ccv_cnnp_model_t* const self, void* const context)
1689{
1690 return ccv_cnnp_sigmoid(self->name);
1691}
1692
1693// MARK - Tanh Layer
1694
1695typedef struct {
1696 ccv_cnnp_model_t super;
1697 ccv_nnc_tensor_symbol_t output;
1698} ccv_cnnp_model_tanh_t;
1699
1700static void _ccv_cnnp_tanh_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
1701{
1702 PRINT(CCV_CLI_VERBOSE, "[cnnp_tanh_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_tanh_build] -\n"); fflush(stdout); } } while
(0);
1703 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 1703, __extension__ __PRETTY_FUNCTION__); }));
1704 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 1704, __extension__ __PRETTY_FUNCTION__
); }));
1705 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1706 ccv_nnc_tensor_param_t output_params;
1707 const ccv_nnc_cmd_t tanh = CMD_TANH_FORWARD()ccv_nnc_cmd(CCV_NNC_TANH_FORWARD, 0, ccv_nnc_cmd_auto, 0);
1708 ccv_nnc_hint_tensor_auto(tanh, (ccv_nnc_tensor_param_t []){
1709 params,
1710 }, 1, ccv_nnc_no_hint, &output_params, 1);
1711 const ccv_nnc_tensor_symbol_t tanh_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1712 ccv_nnc_graph_exec_symbol_new(graph, tanh, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(tanh_output)(const ccv_nnc_tensor_symbol_t []){tanh_output}, (1 +1 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "tanh");
1713 outputs[0] = tanh_output;
1714}
1715
1716static ccv_cnnp_model_t* _ccv_cnnp_tanh_copy(const ccv_cnnp_model_t* const self, void* const context);
1717
1718static const ccv_cnnp_model_vtab_t ccv_cnnp_tanh_isa = {
1719 .build = _ccv_cnnp_tanh_build,
1720 .copy = _ccv_cnnp_tanh_copy,
1721};
1722
1723ccv_cnnp_model_t* ccv_cnnp_tanh(const char* const name)
1724{
1725 ccv_cnnp_model_tanh_t* const model_tanh = (ccv_cnnp_model_tanh_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_tanh_t));
1726 model_tanh->super.isa = &ccv_cnnp_tanh_isa;
1727 model_tanh->super.input_size = 1;
1728 model_tanh->super.outputs = &model_tanh->output;
1729 model_tanh->super.output_size = 1;
1730 ccv_cnnp_model_copy_name(&model_tanh->super, name);
1731 return (ccv_cnnp_model_t*)model_tanh;
1732}
1733
1734static ccv_cnnp_model_t* _ccv_cnnp_tanh_copy(const ccv_cnnp_model_t* const self, void* const context)
1735{
1736 return ccv_cnnp_tanh(self->name);
1737}
1738
1739// MARK - Exp Layer
1740
1741typedef struct {
1742 ccv_cnnp_model_t super;
1743 ccv_nnc_tensor_symbol_t output;
1744} ccv_cnnp_model_exp_t;
1745
1746static void _ccv_cnnp_exp_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
1747{
1748 PRINT(CCV_CLI_VERBOSE, "[cnnp_exp_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_exp_build] -\n"); fflush(stdout); } } while (
0);
1749 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 1749, __extension__ __PRETTY_FUNCTION__); }));
1750 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 1750, __extension__ __PRETTY_FUNCTION__
); }));
1751 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1752 ccv_nnc_tensor_param_t output_params;
1753 const ccv_nnc_cmd_t exp = CMD_EWEXP_FORWARD()ccv_nnc_cmd(CCV_NNC_EWEXP_FORWARD, 0, ccv_nnc_cmd_auto, 0);
1754 ccv_nnc_hint_tensor_auto(exp, (ccv_nnc_tensor_param_t []){
1755 params,
1756 }, 1, ccv_nnc_no_hint, &output_params, 1);
1757 const ccv_nnc_tensor_symbol_t exp_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1758 ccv_nnc_graph_exec_symbol_new(graph, exp, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(exp_output)(const ccv_nnc_tensor_symbol_t []){exp_output}, (1 +1 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "exp");
1759 outputs[0] = exp_output;
1760}
1761
1762static ccv_cnnp_model_t* _ccv_cnnp_exp_copy(const ccv_cnnp_model_t* const self, void* const context);
1763
1764static const ccv_cnnp_model_vtab_t ccv_cnnp_exp_isa = {
1765 .build = _ccv_cnnp_exp_build,
1766 .copy = _ccv_cnnp_exp_copy,
1767};
1768
1769ccv_cnnp_model_t* ccv_cnnp_exp(const char* const name)
1770{
1771 ccv_cnnp_model_exp_t* const model_exp = (ccv_cnnp_model_exp_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_exp_t));
1772 model_exp->super.isa = &ccv_cnnp_exp_isa;
1773 model_exp->super.input_size = 1;
1774 model_exp->super.outputs = &model_exp->output;
1775 model_exp->super.output_size = 1;
1776 ccv_cnnp_model_copy_name(&model_exp->super, name);
1777 return (ccv_cnnp_model_t*)model_exp;
1778}
1779
1780static ccv_cnnp_model_t* _ccv_cnnp_exp_copy(const ccv_cnnp_model_t* const self, void* const context)
1781{
1782 return ccv_cnnp_exp(self->name);
1783}
1784
1785// MARK - Softplus Layer
1786
1787typedef struct {
1788 ccv_cnnp_model_t super;
1789 ccv_nnc_tensor_symbol_t output;
1790} ccv_cnnp_model_softplus_t;
1791
1792static void _ccv_cnnp_softplus_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
1793{
1794 PRINT(CCV_CLI_VERBOSE, "[cnnp_softplus_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_softplus_build] -\n"); fflush(stdout); } } while
(0);
1795 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 1795, __extension__ __PRETTY_FUNCTION__); }));
1796 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 1796, __extension__ __PRETTY_FUNCTION__
); }));
1797 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1798 ccv_nnc_tensor_param_t output_params;
1799 const ccv_nnc_cmd_t softplus = CMD_EWSOFTPLUS_FORWARD()ccv_nnc_cmd(CCV_NNC_EWSOFTPLUS_FORWARD, 0, ccv_nnc_cmd_auto, 0
);
1800 ccv_nnc_hint_tensor_auto(softplus, (ccv_nnc_tensor_param_t []){
1801 params,
1802 }, 1, ccv_nnc_no_hint, &output_params, 1);
1803 const ccv_nnc_tensor_symbol_t softplus_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1804 ccv_nnc_graph_exec_symbol_new(graph, softplus, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(softplus_output)(const ccv_nnc_tensor_symbol_t []){softplus_output}, (1 +1 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -
1), "softplus");
1805 outputs[0] = softplus_output;
1806}
1807
1808static ccv_cnnp_model_t* _ccv_cnnp_softplus_copy(const ccv_cnnp_model_t* const self, void* const context);
1809
1810static const ccv_cnnp_model_vtab_t ccv_cnnp_softplus_isa = {
1811 .build = _ccv_cnnp_softplus_build,
1812 .copy = _ccv_cnnp_softplus_copy,
1813};
1814
1815ccv_cnnp_model_t* ccv_cnnp_softplus(const char* const name)
1816{
1817 ccv_cnnp_model_softplus_t* const model_softplus = (ccv_cnnp_model_softplus_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_softplus_t));
1818 model_softplus->super.isa = &ccv_cnnp_softplus_isa;
1819 model_softplus->super.input_size = 1;
1820 model_softplus->super.outputs = &model_softplus->output;
1821 model_softplus->super.output_size = 1;
1822 ccv_cnnp_model_copy_name(&model_softplus->super, name);
1823 return (ccv_cnnp_model_t*)model_softplus;
1824}
1825
1826static ccv_cnnp_model_t* _ccv_cnnp_softplus_copy(const ccv_cnnp_model_t* const self, void* const context)
1827{
1828 return ccv_cnnp_softplus(self->name);
1829}
1830
1831// MARK - Swish Layer
1832
1833typedef struct {
1834 ccv_cnnp_model_t super;
1835 ccv_nnc_tensor_symbol_t output;
1836 float beta;
1837} ccv_cnnp_model_swish_t;
1838
1839static void _ccv_cnnp_swish_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
1840{
1841 PRINT(CCV_CLI_VERBOSE, "[cnnp_swish_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_swish_build] -\n"); fflush(stdout); } } while
(0);
1842 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 1842, __extension__ __PRETTY_FUNCTION__); }));
1843 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 1843, __extension__ __PRETTY_FUNCTION__
); }));
1844 ccv_cnnp_model_swish_t* const self = (ccv_cnnp_model_swish_t*)super;
1845 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1846 ccv_nnc_tensor_param_t output_params;
1847 const ccv_nnc_cmd_t swish = CMD_SWISH_FORWARD(self->beta)ccv_nnc_cmd(CCV_NNC_SWISH_FORWARD, 0, ((ccv_nnc_cmd_param_t){
.size={.dim={1,1,1}},.swish={.beta=self->beta}}), 0);
1848 ccv_nnc_hint_tensor_auto(swish, (ccv_nnc_tensor_param_t []){
1849 params,
1850 }, 1, ccv_nnc_no_hint, &output_params, 1);
1851 const ccv_nnc_tensor_symbol_t swish_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1852 ccv_nnc_graph_exec_symbol_new(graph, swish, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(swish_output)(const ccv_nnc_tensor_symbol_t []){swish_output}, (1 +1 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "swish");
1853 outputs[0] = swish_output;
1854}
1855
1856static ccv_cnnp_model_t* _ccv_cnnp_swish_copy(const ccv_cnnp_model_t* const self, void* const context);
1857
1858static const ccv_cnnp_model_vtab_t ccv_cnnp_swish_isa = {
1859 .build = _ccv_cnnp_swish_build,
1860 .copy = _ccv_cnnp_swish_copy,
1861};
1862
1863ccv_cnnp_model_t* ccv_cnnp_swish(const float beta, const char* const name)
1864{
1865 ccv_cnnp_model_swish_t* const model_swish = (ccv_cnnp_model_swish_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_swish_t));
1866 model_swish->super.isa = &ccv_cnnp_swish_isa;
1867 model_swish->super.input_size = 1;
1868 model_swish->super.outputs = &model_swish->output;
1869 model_swish->super.output_size = 1;
1870 model_swish->beta = beta;
1871 ccv_cnnp_model_copy_name(&model_swish->super, name);
1872 return (ccv_cnnp_model_t*)model_swish;
1873}
1874
1875static ccv_cnnp_model_t* _ccv_cnnp_swish_copy(const ccv_cnnp_model_t* const self, void* const context)
1876{
1877 const ccv_cnnp_model_swish_t* const swish = (const ccv_cnnp_model_swish_t*)self;
1878 return ccv_cnnp_swish(swish->beta, self->name);
1879}
1880
1881// MARK - GELU Layer
1882
1883typedef struct {
1884 ccv_cnnp_model_t super;
1885 ccv_nnc_tensor_symbol_t output;
1886 int tanh;
1887} ccv_cnnp_model_gelu_t;
1888
1889static void _ccv_cnnp_gelu_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
1890{
1891 PRINT(CCV_CLI_VERBOSE, "[cnnp_gelu_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_gelu_build] -\n"); fflush(stdout); } } while
(0);
1892 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 1892, __extension__ __PRETTY_FUNCTION__); }));
1893 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 1893, __extension__ __PRETTY_FUNCTION__
); }));
1894 ccv_cnnp_model_gelu_t* const self = (ccv_cnnp_model_gelu_t*)super;
1895 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1896 ccv_nnc_tensor_param_t output_params;
1897 const ccv_nnc_cmd_t gelu = CMD_GELU_FORWARD(self->tanh)ccv_nnc_cmd(CCV_NNC_GELU_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.gelu={.tanh=self->tanh}}, 0);
1898 ccv_nnc_hint_tensor_auto(gelu, (ccv_nnc_tensor_param_t []){
1899 params,
1900 }, 1, ccv_nnc_no_hint, &output_params, 1);
1901 const ccv_nnc_tensor_symbol_t gelu_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1902 ccv_nnc_graph_exec_symbol_new(graph, gelu, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(gelu_output)(const ccv_nnc_tensor_symbol_t []){gelu_output}, (1 +1 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), "gelu");
1903 outputs[0] = gelu_output;
1904}
1905
1906static ccv_cnnp_model_t* _ccv_cnnp_gelu_copy(const ccv_cnnp_model_t* const self, void* const context);
1907
1908static const ccv_cnnp_model_vtab_t ccv_cnnp_gelu_isa = {
1909 .build = _ccv_cnnp_gelu_build,
1910 .copy = _ccv_cnnp_gelu_copy,
1911};
1912
1913ccv_cnnp_model_t* ccv_cnnp_gelu(const int tanh, const char* const name)
1914{
1915 ccv_cnnp_model_gelu_t* const model_gelu = (ccv_cnnp_model_gelu_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_gelu_t));
1916 model_gelu->super.isa = &ccv_cnnp_gelu_isa;
1917 model_gelu->super.input_size = 1;
1918 model_gelu->super.outputs = &model_gelu->output;
1919 model_gelu->super.output_size = 1;
1920 model_gelu->tanh = tanh;
1921 ccv_cnnp_model_copy_name(&model_gelu->super, name);
1922 return (ccv_cnnp_model_t*)model_gelu;
1923}
1924
1925static ccv_cnnp_model_t* _ccv_cnnp_gelu_copy(const ccv_cnnp_model_t* const super, void* const context)
1926{
1927 ccv_cnnp_model_gelu_t* const self = (ccv_cnnp_model_gelu_t*)super;
1928 return ccv_cnnp_gelu(self->tanh, self->super.name);
1929}
1930
1931// MARK - Leaky ReLU Layer
1932
1933typedef struct {
1934 ccv_cnnp_model_t super;
1935 ccv_nnc_tensor_symbol_t output;
1936 float negative_slope;
1937} ccv_cnnp_model_leaky_relu_t;
1938
1939static void _ccv_cnnp_leaky_relu_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
1940{
1941 PRINT(CCV_CLI_VERBOSE, "[cnnp_leaky_relu_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_leaky_relu_build] -\n"); fflush(stdout); } }
while (0);
1942 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 1942, __extension__ __PRETTY_FUNCTION__); }));
1943 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 1943, __extension__ __PRETTY_FUNCTION__
); }));
1944 ccv_cnnp_model_leaky_relu_t* const self = (ccv_cnnp_model_leaky_relu_t*)super;
1945 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1946 ccv_nnc_tensor_param_t output_params;
1947 const ccv_nnc_cmd_t leaky_relu = CMD_LEAKY_RELU_FORWARD(self->negative_slope)ccv_nnc_cmd(CCV_NNC_LEAKY_RELU_FORWARD, 0, (ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.leaky_relu={.negative_slope=self->
negative_slope}}, 0);
1948 ccv_nnc_hint_tensor_auto(leaky_relu, (ccv_nnc_tensor_param_t []){
1949 params,
1950 }, 1, ccv_nnc_no_hint, &output_params, 1);
1951 const ccv_nnc_tensor_symbol_t leaky_relu_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
1952 ccv_nnc_graph_exec_symbol_new(graph, leaky_relu, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(leaky_relu_output)(const ccv_nnc_tensor_symbol_t []){leaky_relu_output}, (1 +1 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
-1), "leaky_relu");
1953 outputs[0] = leaky_relu_output;
1954}
1955
1956static ccv_cnnp_model_t* _ccv_cnnp_leaky_relu_copy(const ccv_cnnp_model_t* const self, void* const context);
1957
1958static const ccv_cnnp_model_vtab_t ccv_cnnp_leaky_relu_isa = {
1959 .build = _ccv_cnnp_leaky_relu_build,
1960 .copy = _ccv_cnnp_leaky_relu_copy,
1961};
1962
1963ccv_cnnp_model_t* ccv_cnnp_leaky_relu(const float negative_slope, const char* const name)
1964{
1965 ccv_cnnp_model_leaky_relu_t* const model_leaky_relu = (ccv_cnnp_model_leaky_relu_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_leaky_relu_t));
1966 model_leaky_relu->super.isa = &ccv_cnnp_leaky_relu_isa;
1967 model_leaky_relu->super.input_size = 1;
1968 model_leaky_relu->super.outputs = &model_leaky_relu->output;
1969 model_leaky_relu->super.output_size = 1;
1970 model_leaky_relu->negative_slope = negative_slope;
1971 ccv_cnnp_model_copy_name(&model_leaky_relu->super, name);
1972 return (ccv_cnnp_model_t*)model_leaky_relu;
1973}
1974
1975static ccv_cnnp_model_t* _ccv_cnnp_leaky_relu_copy(const ccv_cnnp_model_t* const super, void* const context)
1976{
1977 ccv_cnnp_model_leaky_relu_t* const self = (ccv_cnnp_model_leaky_relu_t*)super;
1978 return ccv_cnnp_leaky_relu(self->negative_slope, self->super.name);
1979}
1980
1981// MARK - Softmax Layer
1982
1983typedef struct {
1984 ccv_cnnp_model_t super;
1985 ccv_nnc_tensor_symbol_t output;
1986} ccv_cnnp_model_softmax_t;
1987
1988static void _ccv_cnnp_softmax_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
1989{
1990 PRINT(CCV_CLI_VERBOSE, "[cnnp_softmax_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_softmax_build] -\n"); fflush(stdout); } } while
(0);
1991 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 1991, __extension__ __PRETTY_FUNCTION__); }));
1992 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 1992, __extension__ __PRETTY_FUNCTION__
); }));
1993 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
1994 ccv_nnc_tensor_param_t output_params;
1995 const ccv_nnc_cmd_t softmax = CMD_SOFTMAX_FORWARD()ccv_nnc_cmd(CCV_NNC_SOFTMAX_FORWARD, 0, ccv_nnc_cmd_auto, 0);
1996 ccv_nnc_hint_tensor_auto(softmax, (ccv_nnc_tensor_param_t []){
1997 params,
1998 }, 1, ccv_nnc_no_hint, &output_params, 1);
1999 const ccv_nnc_tensor_symbol_t softmax_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2000 ccv_nnc_graph_exec_symbol_new(graph, softmax, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(softmax_output)(const ccv_nnc_tensor_symbol_t []){softmax_output}, (1 +1 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1
), "softmax");
2001 outputs[0] = softmax_output;
2002}
2003
2004static ccv_cnnp_model_t* _ccv_cnnp_softmax_copy(const ccv_cnnp_model_t* const self, void* const context);
2005
2006static const ccv_cnnp_model_vtab_t ccv_cnnp_softmax_isa = {
2007 .build = _ccv_cnnp_softmax_build,
2008 .copy = _ccv_cnnp_softmax_copy,
2009};
2010
2011ccv_cnnp_model_t* ccv_cnnp_softmax(const char* const name)
2012{
2013 ccv_cnnp_model_softmax_t* const model_softmax = (ccv_cnnp_model_softmax_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_softmax_t));
2014 model_softmax->super.isa = &ccv_cnnp_softmax_isa;
2015 model_softmax->super.input_size = 1;
2016 model_softmax->super.outputs = &model_softmax->output;
2017 model_softmax->super.output_size = 1;
2018 ccv_cnnp_model_copy_name(&model_softmax->super, name);
2019 return (ccv_cnnp_model_t*)model_softmax;
2020}
2021
2022static ccv_cnnp_model_t* _ccv_cnnp_softmax_copy(const ccv_cnnp_model_t* const self, void* const context)
2023{
2024 return ccv_cnnp_softmax(self->name);
2025}
2026
2027// MARK - Add Layer
2028
2029typedef struct {
2030 ccv_cnnp_model_t super;
2031 float p;
2032 float q;
2033 ccv_nnc_tensor_symbol_t output;
2034} ccv_cnnp_model_add_t;
2035
2036static void _ccv_cnnp_add_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
2037{
2038 PRINT(CCV_CLI_VERBOSE, "[cnnp_add_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_add_build] -\n"); fflush(stdout); } } while (
0);
2039 const ccv_cnnp_model_add_t* const self = (const ccv_cnnp_model_add_t*)super;
2040 assert(input_size == 2)((void) sizeof ((input_size == 2) ? 1 : 0), __extension__ ({ if
(input_size == 2) ; else __assert_fail ("input_size == 2", "ccv_cnnp_model_addons.c"
, 2040, __extension__ __PRETTY_FUNCTION__); }));
2041 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 2041, __extension__ __PRETTY_FUNCTION__
); }));
2042 ccv_nnc_tensor_param_t input_params[2];
2043 int i;
2044 for (i = 0; i < 2; i++)
2045 input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
2046 ccv_nnc_tensor_param_t output_params;
2047 const ccv_nnc_cmd_t add = CMD_ADD_FORWARD(self->p, self->q)ccv_nnc_cmd(CCV_NNC_ADD_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={self->p, self->q}}}, 0);
2048 ccv_nnc_hint_tensor_auto(add, input_params, 2, ccv_nnc_no_hint, &output_params, 1);
2049 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2050 ccv_nnc_graph_exec_symbol_new(graph, add, inputs, input_size, outputs, output_size, "add");
2051}
2052
2053static ccv_cnnp_model_t* _ccv_cnnp_add_copy(const ccv_cnnp_model_t* const self, void* const context);
2054
2055static const ccv_cnnp_model_vtab_t ccv_cnnp_add_isa = {
2056 .build = _ccv_cnnp_add_build,
2057 .copy = _ccv_cnnp_add_copy,
2058};
2059
2060ccv_cnnp_model_t* ccv_cnnp_add(const float p, const float q, const char* const name)
2061{
2062 ccv_cnnp_model_add_t* const model_add = (ccv_cnnp_model_add_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_add_t));
2063 model_add->super.isa = &ccv_cnnp_add_isa;
2064 model_add->super.input_size = 2;
2065 model_add->super.outputs = &model_add->output;
2066 model_add->super.output_size = 1;
2067 model_add->p = p;
2068 model_add->q = q;
2069 ccv_cnnp_model_copy_name(&model_add->super, name);
2070 return (ccv_cnnp_model_t*)model_add;
2071}
2072
2073static ccv_cnnp_model_t* _ccv_cnnp_add_copy(const ccv_cnnp_model_t* const super, void* const context)
2074{
2075 const ccv_cnnp_model_add_t* const self = (const ccv_cnnp_model_add_t*)super;
2076 return ccv_cnnp_add(self->p, self->q, self->super.name);
2077}
2078
2079// MARK - Mul Layer
2080
2081typedef struct {
2082 ccv_cnnp_model_t super;
2083 ccv_nnc_tensor_symbol_t output;
2084 float p;
2085} ccv_cnnp_model_mul_t;
2086
2087static void _ccv_cnnp_mul_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
2088{
2089 PRINT(CCV_CLI_VERBOSE, "[cnnp_mul_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_mul_build] -\n"); fflush(stdout); } } while (
0);
2090 const ccv_cnnp_model_mul_t* const self = (const ccv_cnnp_model_mul_t*)super;
2091 assert(input_size == 2)((void) sizeof ((input_size == 2) ? 1 : 0), __extension__ ({ if
(input_size == 2) ; else __assert_fail ("input_size == 2", "ccv_cnnp_model_addons.c"
, 2091, __extension__ __PRETTY_FUNCTION__); }));
2092 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 2092, __extension__ __PRETTY_FUNCTION__
); }));
2093 ccv_nnc_tensor_param_t input_params[2];
2094 int i;
2095 for (i = 0; i < 2; i++)
2096 input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
2097 ccv_nnc_tensor_param_t output_params;
2098 const ccv_nnc_cmd_t mul = CMD_MUL_FORWARD(self->p)ccv_nnc_cmd(CCV_NNC_MUL_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}},.blas={.a={self->p,}}}, 0);
2099 ccv_nnc_hint_tensor_auto(mul, input_params, 2, ccv_nnc_no_hint, &output_params, 1);
2100 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2101 ccv_nnc_graph_exec_symbol_new(graph, mul, inputs, input_size, outputs, output_size, "mul");
2102}
2103
2104static ccv_cnnp_model_t* _ccv_cnnp_mul_copy(const ccv_cnnp_model_t* const self, void* const context);
2105
2106static const ccv_cnnp_model_vtab_t ccv_cnnp_mul_isa = {
2107 .build = _ccv_cnnp_mul_build,
2108 .copy = _ccv_cnnp_mul_copy,
2109};
2110
2111ccv_cnnp_model_t* ccv_cnnp_mul(const float p, const char* const name)
2112{
2113 ccv_cnnp_model_mul_t* const model_mul = (ccv_cnnp_model_mul_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_mul_t));
2114 model_mul->super.isa = &ccv_cnnp_mul_isa;
2115 model_mul->super.input_size = 2;
2116 model_mul->super.outputs = &model_mul->output;
2117 model_mul->super.output_size = 1;
2118 model_mul->p = p;
2119 ccv_cnnp_model_copy_name(&model_mul->super, name);
2120 return (ccv_cnnp_model_t*)model_mul;
2121}
2122
2123static ccv_cnnp_model_t* _ccv_cnnp_mul_copy(const ccv_cnnp_model_t* const super, void* const context)
2124{
2125 const ccv_cnnp_model_mul_t* const self = (const ccv_cnnp_model_mul_t*)super;
2126 return ccv_cnnp_mul(self->p, self->super.name);
2127}
2128
2129// MARK - Scalar Mul Layer
2130
2131typedef struct {
2132 ccv_cnnp_model_t super;
2133 ccv_nnc_tensor_symbol_t output;
2134 float a;
2135} ccv_cnnp_model_scalar_mul_t;
2136
2137static void _ccv_cnnp_scalar_mul_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
2138{
2139 PRINT(CCV_CLI_VERBOSE, "[cnnp_scalar_mul_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_scalar_mul_build] -\n"); fflush(stdout); } }
while (0);
2140 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 2140, __extension__ __PRETTY_FUNCTION__); }));
2141 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 2141, __extension__ __PRETTY_FUNCTION__
); }));
2142 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2143 ccv_nnc_tensor_param_t output_params;
2144 ccv_cnnp_model_scalar_mul_t* const self = (ccv_cnnp_model_scalar_mul_t*)super;
2145 const ccv_nnc_cmd_t scalar_mul = CMD_SCALAR_MUL_FORWARD(self->a)ccv_nnc_cmd(CCV_NNC_SCALAR_MUL_FORWARD, 0, (ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}},.blas={.a={self->a,}}}, 0);
2146 ccv_nnc_hint_tensor_auto(scalar_mul, (ccv_nnc_tensor_param_t []){
2147 params,
2148 }, 1, ccv_nnc_no_hint, &output_params, 1);
2149 const ccv_nnc_tensor_symbol_t scalar_mul_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2150 ccv_nnc_graph_exec_symbol_new(graph, scalar_mul, TENSOR_SYMBOL_LIST(inputs[0])(const ccv_nnc_tensor_symbol_t []){inputs[0]}, (1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(scalar_mul_output)(const ccv_nnc_tensor_symbol_t []){scalar_mul_output}, (1 +1 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
-1), "scalar_mul");
2151 outputs[0] = scalar_mul_output;
2152}
2153
2154static ccv_cnnp_model_t* _ccv_cnnp_scalar_mul_copy(const ccv_cnnp_model_t* const super, void* const context);
2155
2156static const ccv_cnnp_model_vtab_t ccv_cnnp_scalar_mul_isa = {
2157 .build = _ccv_cnnp_scalar_mul_build,
2158 .copy = _ccv_cnnp_scalar_mul_copy,
2159};
2160
2161ccv_cnnp_model_t* ccv_cnnp_scalar_mul(const float a, const char* const name)
2162{
2163 ccv_cnnp_model_scalar_mul_t* const model_scalar_mul = (ccv_cnnp_model_scalar_mul_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_scalar_mul_t));
2164 model_scalar_mul->super.isa = &ccv_cnnp_scalar_mul_isa;
2165 model_scalar_mul->super.input_size = 1;
2166 model_scalar_mul->super.outputs = &model_scalar_mul->output;
2167 model_scalar_mul->super.output_size = 1;
2168 model_scalar_mul->a = a;
2169 ccv_cnnp_model_copy_name(&model_scalar_mul->super, name);
2170 return (ccv_cnnp_model_t*)model_scalar_mul;
2171}
2172
2173static ccv_cnnp_model_t* _ccv_cnnp_scalar_mul_copy(const ccv_cnnp_model_t* const super, void* const context)
2174{
2175 const ccv_cnnp_model_scalar_mul_t* const self = (const ccv_cnnp_model_scalar_mul_t*)super;
2176 return ccv_cnnp_scalar_mul(self->a, self->super.name);
2177}
2178
2179// MARK - Div Layer
2180
2181typedef struct {
2182 ccv_cnnp_model_t super;
2183 ccv_nnc_tensor_symbol_t output;
2184 int reciprocal;
2185} ccv_cnnp_model_div_t;
2186
2187static void _ccv_cnnp_div_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
2188{
2189 const ccv_cnnp_model_div_t* const self = (const ccv_cnnp_model_div_t*)super;
2190 PRINT(CCV_CLI_VERBOSE, "[cnnp_div_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_div_build] -\n"); fflush(stdout); } } while (
0);
2191 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 2191, __extension__ __PRETTY_FUNCTION__
); }));
2192 ccv_nnc_tensor_param_t input_params[2];
2193 int i;
2194 ccv_nnc_tensor_param_t output_params;
2195 const ccv_nnc_cmd_t div = CMD_EWDIV_FORWARD()ccv_nnc_cmd(CCV_NNC_EWDIV_FORWARD, 0, ccv_nnc_cmd_auto, 0);
2196 if (self->reciprocal)
2197 {
2198 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 2198, __extension__ __PRETTY_FUNCTION__); }));
2199 input_params[0] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2200 input_params[1] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2201 ccv_nnc_hint_tensor_auto(div, input_params, 2, ccv_nnc_no_hint, &output_params, 1);
2202 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2203 ccv_nnc_graph_exec_symbol_new(graph, div, TENSOR_SYMBOL_LIST(NO_TENSOR_SYMBOL, inputs[0])(const ccv_nnc_tensor_symbol_t []){(const ccv_nnc_tensor_symbol_t
){.d = CCV_NNC_NO_TENSOR_SYMBOL}, inputs[0]}, (1 +1 +1 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), outputs, output_size, "div");
2204 } else {
2205 assert(input_size == 2)((void) sizeof ((input_size == 2) ? 1 : 0), __extension__ ({ if
(input_size == 2) ; else __assert_fail ("input_size == 2", "ccv_cnnp_model_addons.c"
, 2205, __extension__ __PRETTY_FUNCTION__); }));
2206 for (i = 0; i < 2; i++)
2207 input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
2208 ccv_nnc_hint_tensor_auto(div, input_params, input_size, ccv_nnc_no_hint, &output_params, 1);
2209 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2210 ccv_nnc_graph_exec_symbol_new(graph, div, inputs, input_size, outputs, output_size, "div");
2211 }
2212}
2213
2214static ccv_cnnp_model_t* _ccv_cnnp_div_copy(const ccv_cnnp_model_t* const self, void* const context);
2215
2216static const ccv_cnnp_model_vtab_t ccv_cnnp_div_isa = {
2217 .build = _ccv_cnnp_div_build,
2218 .copy = _ccv_cnnp_div_copy,
2219};
2220
2221ccv_cnnp_model_t* ccv_cnnp_div(const int reciprocal, const char* const name)
2222{
2223 ccv_cnnp_model_div_t* const model_div = (ccv_cnnp_model_div_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_div_t));
2224 model_div->super.isa = &ccv_cnnp_div_isa;
2225 model_div->super.input_size = reciprocal ? 1 : 2;
2226 model_div->super.outputs = &model_div->output;
2227 model_div->super.output_size = 1;
2228 model_div->reciprocal = reciprocal;
2229 ccv_cnnp_model_copy_name(&model_div->super, name);
2230 return (ccv_cnnp_model_t*)model_div;
2231}
2232
2233static ccv_cnnp_model_t* _ccv_cnnp_div_copy(const ccv_cnnp_model_t* const super, void* const context)
2234{
2235 const ccv_cnnp_model_div_t* const self = (const ccv_cnnp_model_div_t*)super;
2236 return ccv_cnnp_div(self->reciprocal, self->super.name);
2237}
2238
2239// MARK - Sqrt Layer
2240
2241typedef struct {
2242 ccv_cnnp_model_t super;
2243 ccv_nnc_tensor_symbol_t output;
2244} ccv_cnnp_model_sqrt_t;
2245
2246static void _ccv_cnnp_sqrt_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
2247{
2248 PRINT(CCV_CLI_VERBOSE, "[cnnp_sqrt_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_sqrt_build] -\n"); fflush(stdout); } } while
(0);
2249 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 2249, __extension__ __PRETTY_FUNCTION__
); }));
2250 ccv_nnc_tensor_param_t input_params[1];
2251 ccv_nnc_tensor_param_t output_params;
2252 const ccv_nnc_cmd_t sqrt = CMD_EWSQRT_FORWARD()ccv_nnc_cmd(CCV_NNC_EWSQRT_FORWARD, 0, ccv_nnc_cmd_auto, 0);
2253 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 2253, __extension__ __PRETTY_FUNCTION__); }));
2254 input_params[0] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2255 ccv_nnc_hint_tensor_auto(sqrt, input_params, 1, ccv_nnc_no_hint, &output_params, 1);
2256 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2257 ccv_nnc_graph_exec_symbol_new(graph, sqrt, inputs, 1, outputs, output_size, "sqrt");
2258}
2259
2260static ccv_cnnp_model_t* _ccv_cnnp_sqrt_copy(const ccv_cnnp_model_t* const self, void* const context);
2261
2262static const ccv_cnnp_model_vtab_t ccv_cnnp_sqrt_isa = {
2263 .build = _ccv_cnnp_sqrt_build,
2264 .copy = _ccv_cnnp_sqrt_copy,
2265};
2266
2267ccv_cnnp_model_t* ccv_cnnp_sqrt(const char* const name)
2268{
2269 ccv_cnnp_model_sqrt_t* const model_sqrt = (ccv_cnnp_model_sqrt_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_sqrt_t));
2270 model_sqrt->super.isa = &ccv_cnnp_sqrt_isa;
2271 model_sqrt->super.input_size = 1;
2272 model_sqrt->super.outputs = &model_sqrt->output;
2273 model_sqrt->super.output_size = 1;
2274 ccv_cnnp_model_copy_name(&model_sqrt->super, name);
2275 return (ccv_cnnp_model_t*)model_sqrt;
2276}
2277
2278static ccv_cnnp_model_t* _ccv_cnnp_sqrt_copy(const ccv_cnnp_model_t* const super, void* const context)
2279{
2280 const ccv_cnnp_model_sqrt_t* const self = (const ccv_cnnp_model_sqrt_t*)super;
2281 return ccv_cnnp_sqrt(self->super.name);
2282}
2283
2284// MARK - Log Layer
2285
2286typedef struct {
2287 ccv_cnnp_model_t super;
2288 ccv_nnc_tensor_symbol_t output;
2289} ccv_cnnp_model_log_t;
2290
2291static void _ccv_cnnp_log_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
2292{
2293 PRINT(CCV_CLI_VERBOSE, "[cnnp_log_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_log_build] -\n"); fflush(stdout); } } while (
0);
2294 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 2294, __extension__ __PRETTY_FUNCTION__
); }));
2295 ccv_nnc_tensor_param_t input_params[1];
2296 ccv_nnc_tensor_param_t output_params;
2297 const ccv_nnc_cmd_t log = CMD_EWLOG_FORWARD()ccv_nnc_cmd(CCV_NNC_EWLOG_FORWARD, 0, ccv_nnc_cmd_auto, 0);
2298 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 2298, __extension__ __PRETTY_FUNCTION__); }));
2299 input_params[0] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2300 ccv_nnc_hint_tensor_auto(log, input_params, 1, ccv_nnc_no_hint, &output_params, 1);
2301 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2302 ccv_nnc_graph_exec_symbol_new(graph, log, inputs, 1, outputs, output_size, "log");
2303}
2304
2305static ccv_cnnp_model_t* _ccv_cnnp_log_copy(const ccv_cnnp_model_t* const self, void* const context);
2306
2307static const ccv_cnnp_model_vtab_t ccv_cnnp_log_isa = {
2308 .build = _ccv_cnnp_log_build,
2309 .copy = _ccv_cnnp_log_copy,
2310};
2311
2312ccv_cnnp_model_t* ccv_cnnp_log(const char* const name)
2313{
2314 ccv_cnnp_model_log_t* const model_log = (ccv_cnnp_model_log_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_log_t));
2315 model_log->super.isa = &ccv_cnnp_log_isa;
2316 model_log->super.input_size = 1;
2317 model_log->super.outputs = &model_log->output;
2318 model_log->super.output_size = 1;
2319 ccv_cnnp_model_copy_name(&model_log->super, name);
2320 return (ccv_cnnp_model_t*)model_log;
2321}
2322
2323static ccv_cnnp_model_t* _ccv_cnnp_log_copy(const ccv_cnnp_model_t* const super, void* const context)
2324{
2325 return ccv_cnnp_log(super->name);
2326}
2327
2328// MARK - Pow Layer
2329
2330typedef struct {
2331 ccv_cnnp_model_t super;
2332 ccv_nnc_tensor_symbol_t output;
2333 ccv_nnc_cmd_param_t params;
2334} ccv_cnnp_model_pow_t;
2335
2336static void _ccv_cnnp_pow_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
2337{
2338 ccv_cnnp_model_pow_t* const self = (ccv_cnnp_model_pow_t*)super;
2339 PRINT(CCV_CLI_VERBOSE, "[cnnp_pow_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_pow_build] -\n"); fflush(stdout); } } while (
0);
2340 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 2340, __extension__ __PRETTY_FUNCTION__); }));
2341 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 2341, __extension__ __PRETTY_FUNCTION__
); }));
2342 ccv_nnc_tensor_param_t input_params[1];
2343 input_params[0] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2344 ccv_nnc_tensor_param_t output_params;
2345 const ccv_nnc_cmd_t pow = ccv_nnc_cmd(CCV_NNC_EWPOW_FORWARD, 0, self->params, 0);
2346 ccv_nnc_hint_tensor_auto(pow, input_params, 1, ccv_nnc_no_hint, &output_params, 1);
2347 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2348 ccv_nnc_graph_exec_symbol_new(graph, pow, inputs, input_size, outputs, output_size, "pow");
2349}
2350
2351static ccv_cnnp_model_t* _ccv_cnnp_pow_copy(const ccv_cnnp_model_t* const self, void* const context);
2352
2353static const ccv_cnnp_model_vtab_t ccv_cnnp_pow_isa = {
2354 .build = _ccv_cnnp_pow_build,
2355 .copy = _ccv_cnnp_pow_copy,
2356};
2357
2358ccv_cnnp_model_t* ccv_cnnp_pow(const float exponent, const char* const name)
2359{
2360 ccv_cnnp_model_pow_t* const model_pow = (ccv_cnnp_model_pow_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_pow_t));
2361 model_pow->super.isa = &ccv_cnnp_pow_isa;
2362 model_pow->super.input_size = 1;
2363 model_pow->super.outputs = &model_pow->output;
2364 model_pow->super.output_size = 1;
2365 model_pow->params = (ccv_nnc_cmd_param_t){
2366 .size = {
2367 .dim = { 1, 1, 1 }
2368 },
2369 .pow = {
2370 .exponent = exponent,
2371 },
2372 };
2373 ccv_cnnp_model_copy_name(&model_pow->super, name);
2374 return (ccv_cnnp_model_t*)model_pow;
2375}
2376
2377static ccv_cnnp_model_t* _ccv_cnnp_pow_copy(const ccv_cnnp_model_t* const super, void* const context)
2378{
2379 const ccv_cnnp_model_pow_t* const self = (const ccv_cnnp_model_pow_t*)super;
2380 return ccv_cnnp_pow(self->params.pow.exponent, super->name);
2381}
2382
2383// MARK - Sin Layer
2384
2385typedef struct {
2386 ccv_cnnp_model_t super;
2387 ccv_nnc_tensor_symbol_t output;
2388} ccv_cnnp_model_sin_t;
2389
2390static void _ccv_cnnp_sin_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
2391{
2392 PRINT(CCV_CLI_VERBOSE, "[cnnp_sin_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_sin_build] -\n"); fflush(stdout); } } while (
0);
2393 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 2393, __extension__ __PRETTY_FUNCTION__
); }));
2394 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 2394, __extension__ __PRETTY_FUNCTION__); }));
2395 ccv_nnc_tensor_param_t input_params[1];
2396 ccv_nnc_tensor_param_t output_params;
2397 const ccv_nnc_cmd_t sin = CMD_EWSIN_FORWARD()ccv_nnc_cmd(CCV_NNC_EWSIN_FORWARD, 0, ccv_nnc_cmd_auto, 0);
2398 input_params[0] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2399 ccv_nnc_hint_tensor_auto(sin, input_params, 1, ccv_nnc_no_hint, &output_params, 1);
2400 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2401 ccv_nnc_graph_exec_symbol_new(graph, sin, inputs, 1, outputs, output_size, "sin");
2402}
2403
2404static ccv_cnnp_model_t* _ccv_cnnp_sin_copy(const ccv_cnnp_model_t* const self, void* const context);
2405
2406static const ccv_cnnp_model_vtab_t ccv_cnnp_sin_isa = {
2407 .build = _ccv_cnnp_sin_build,
2408 .copy = _ccv_cnnp_sin_copy,
2409};
2410
2411ccv_cnnp_model_t* ccv_cnnp_sin(const char* const name)
2412{
2413 ccv_cnnp_model_sin_t* const model_sin = (ccv_cnnp_model_sin_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_sin_t));
2414 model_sin->super.isa = &ccv_cnnp_sin_isa;
2415 model_sin->super.input_size = 1;
2416 model_sin->super.outputs = &model_sin->output;
2417 model_sin->super.output_size = 1;
2418 ccv_cnnp_model_copy_name(&model_sin->super, name);
2419 return (ccv_cnnp_model_t*)model_sin;
2420}
2421
2422static ccv_cnnp_model_t* _ccv_cnnp_sin_copy(const ccv_cnnp_model_t* const super, void* const context)
2423{
2424 return ccv_cnnp_sin(super->name);
2425}
2426
2427// MARK - Cos Layer
2428
2429typedef struct {
2430 ccv_cnnp_model_t super;
2431 ccv_nnc_tensor_symbol_t output;
2432} ccv_cnnp_model_cos_t;
2433
2434static void _ccv_cnnp_cos_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
2435{
2436 PRINT(CCV_CLI_VERBOSE, "[cnnp_cos_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_cos_build] -\n"); fflush(stdout); } } while (
0);
2437 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 2437, __extension__ __PRETTY_FUNCTION__
); }));
2438 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 2438, __extension__ __PRETTY_FUNCTION__); }));
2439 ccv_nnc_tensor_param_t input_params[1];
2440 ccv_nnc_tensor_param_t output_params;
2441 const ccv_nnc_cmd_t cos = CMD_EWCOS_FORWARD()ccv_nnc_cmd(CCV_NNC_EWCOS_FORWARD, 0, ccv_nnc_cmd_auto, 0);
2442 input_params[0] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2443 ccv_nnc_hint_tensor_auto(cos, input_params, 1, ccv_nnc_no_hint, &output_params, 1);
2444 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2445 ccv_nnc_graph_exec_symbol_new(graph, cos, inputs, 1, outputs, output_size, "cos");
2446}
2447
2448static ccv_cnnp_model_t* _ccv_cnnp_cos_copy(const ccv_cnnp_model_t* const self, void* const context);
2449
2450static const ccv_cnnp_model_vtab_t ccv_cnnp_cos_isa = {
2451 .build = _ccv_cnnp_cos_build,
2452 .copy = _ccv_cnnp_cos_copy,
2453};
2454
2455ccv_cnnp_model_t* ccv_cnnp_cos(const char* const name)
2456{
2457 ccv_cnnp_model_cos_t* const model_cos = (ccv_cnnp_model_cos_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_cos_t));
2458 model_cos->super.isa = &ccv_cnnp_cos_isa;
2459 model_cos->super.input_size = 1;
2460 model_cos->super.outputs = &model_cos->output;
2461 model_cos->super.output_size = 1;
2462 ccv_cnnp_model_copy_name(&model_cos->super, name);
2463 return (ccv_cnnp_model_t*)model_cos;
2464}
2465
2466static ccv_cnnp_model_t* _ccv_cnnp_cos_copy(const ccv_cnnp_model_t* const super, void* const context)
2467{
2468 return ccv_cnnp_cos(super->name);
2469}
2470
2471// MARK - Rotate Half Layer
2472
2473typedef struct {
2474 ccv_cnnp_model_t super;
2475 ccv_nnc_tensor_symbol_t output;
2476} ccv_cnnp_model_rotate_half_t;
2477
2478static void _ccv_cnnp_rotate_half_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
2479{
2480 PRINT(CCV_CLI_VERBOSE, "[cnnp_rotate_half_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_rotate_half_build] -\n"); fflush(stdout); } }
while (0);
2481 assert(input_size == 1)((void) sizeof ((input_size == 1) ? 1 : 0), __extension__ ({ if
(input_size == 1) ; else __assert_fail ("input_size == 1", "ccv_cnnp_model_addons.c"
, 2481, __extension__ __PRETTY_FUNCTION__); }));
2482 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 2482, __extension__ __PRETTY_FUNCTION__
); }));
2483 ccv_nnc_tensor_param_t input_params[1];
2484 ccv_nnc_tensor_param_t output_params;
2485 const ccv_nnc_cmd_t rotate_half = CMD_ROTATE_HALF_FORWARD()ccv_nnc_cmd(CCV_NNC_ROTATE_HALF_FORWARD, 0, (ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}}}, 0);
2486 input_params[0] = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2487 ccv_nnc_hint_tensor_auto(rotate_half, input_params, 1, ccv_nnc_no_hint, &output_params, 1);
2488 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2489 ccv_nnc_graph_exec_symbol_new(graph, rotate_half, inputs, 1, outputs, output_size, "rotate_half");
2490}
2491
2492static ccv_cnnp_model_t* _ccv_cnnp_rotate_half_copy(const ccv_cnnp_model_t* const self, void* const context);
2493
2494static const ccv_cnnp_model_vtab_t ccv_cnnp_rotate_half_isa = {
2495 .build = _ccv_cnnp_rotate_half_build,
2496 .copy = _ccv_cnnp_rotate_half_copy,
2497};
2498
2499ccv_cnnp_model_t* ccv_cnnp_rotate_half(const char* const name)
2500{
2501 ccv_cnnp_model_rotate_half_t* const model_rotate_half = (ccv_cnnp_model_rotate_half_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_rotate_half_t));
2502 model_rotate_half->super.isa = &ccv_cnnp_rotate_half_isa;
2503 model_rotate_half->super.input_size = 1;
2504 model_rotate_half->super.outputs = &model_rotate_half->output;
2505 model_rotate_half->super.output_size = 1;
2506 ccv_cnnp_model_copy_name(&model_rotate_half->super, name);
2507 return (ccv_cnnp_model_t*)model_rotate_half;
2508}
2509
2510static ccv_cnnp_model_t* _ccv_cnnp_rotate_half_copy(const ccv_cnnp_model_t* const super, void* const context)
2511{
2512 return ccv_cnnp_rotate_half(super->name);
2513}
2514
2515// MARK - Gated Delta Layer
2516
2517typedef struct {
2518 ccv_cnnp_model_t super;
2519 ccv_nnc_tensor_symbol_t outputs[2];
2520} ccv_cnnp_model_gated_delta_t;
2521
2522static void _ccv_cnnp_gated_delta_build(ccv_cnnp_model_t* const super, ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t* const inputs, const int input_size, ccv_nnc_tensor_symbol_t* const outputs, const int output_size)
2523{
2524 PRINT(CCV_CLI_VERBOSE, "[cnnp_gated_delta_build] -\n")do { if ((CCV_CLI_VERBOSE & ccv_cli_get_output_levels()))
{ printf("[cnnp_gated_delta_build] -\n"); fflush(stdout); } }
while (0);
2525 assert(input_size == 6)((void) sizeof ((input_size == 6) ? 1 : 0), __extension__ ({ if
(input_size == 6) ; else __assert_fail ("input_size == 6", "ccv_cnnp_model_addons.c"
, 2525, __extension__ __PRETTY_FUNCTION__); }));
2526 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", 2526, __extension__ __PRETTY_FUNCTION__
); }));
2527 ccv_nnc_tensor_param_t input_params[6];
2528 int i;
2529 for (i = 0; i < 6; i++)
2530 input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
2531 ccv_nnc_tensor_param_t output_params[2];
2532 const ccv_nnc_cmd_t gated_delta = CMD_GATED_DELTA_FORWARD()ccv_nnc_cmd(CCV_NNC_GATED_DELTA_FORWARD, 0, ((ccv_nnc_cmd_param_t
){.size={.dim={1,1,1}}}), 0);
2533 ccv_nnc_hint_tensor_auto(gated_delta, input_params, 6, ccv_nnc_no_hint, output_params, 2);
2534 for (i = 0; i < 2; i++)
2535 outputs[i] = ccv_nnc_tensor_symbol_new(graph, output_params[i], 0);
2536 ccv_nnc_graph_exec_symbol_new(graph, gated_delta, inputs, input_size, outputs, output_size, "gated_delta");
2537}
2538
2539static ccv_cnnp_model_t* _ccv_cnnp_gated_delta_copy(const ccv_cnnp_model_t* const self, void* const context);
2540
2541static const ccv_cnnp_model_vtab_t ccv_cnnp_gated_delta_isa = {
2542 .build = _ccv_cnnp_gated_delta_build,
2543 .copy = _ccv_cnnp_gated_delta_copy,
2544};
2545
2546ccv_cnnp_model_t* ccv_cnnp_gated_delta(const char* const name)
2547{
2548 ccv_cnnp_model_gated_delta_t* const model_gated_delta = (ccv_cnnp_model_gated_delta_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_gated_delta_t));
2549 model_gated_delta->super.isa = &ccv_cnnp_gated_delta_isa;
2550 model_gated_delta->super.input_size = 6;
2551 model_gated_delta->super.outputs = model_gated_delta->outputs;
2552 model_gated_delta->super.output_size = 2;
2553 ccv_cnnp_model_copy_name(&model_gated_delta->super, name);
2554 return (ccv_cnnp_model_t*)model_gated_delta;
2555}
2556
2557static ccv_cnnp_model_t* _ccv_cnnp_gated_delta_copy(const ccv_cnnp_model_t* const super, void* const context)
2558{
2559 return ccv_cnnp_gated_delta(super->name);
2560}
2561
2562// MARK - Cmul Layer
2563
2564typedef struct {
2565 ccv_cnnp_model_t super;
2566 ccv_nnc_tensor_symbol_t output;
2567} ccv_cnnp_model_cmul_t;
2568
2569static 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)
2570{
2571 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);
2572 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"
, 2572, __extension__ __PRETTY_FUNCTION__); }));
2573 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 2573, __extension__ __PRETTY_FUNCTION__
); }));
2574 ccv_nnc_tensor_param_t input_params[2];
2575 int i;
2576 for (i = 0; i < 2; i++)
2577 input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
2578 ccv_nnc_tensor_param_t output_params;
2579 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);
2580 ccv_nnc_hint_tensor_auto(mul, input_params, 2, ccv_nnc_no_hint, &output_params, 1);
2581 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2582 ccv_nnc_graph_exec_symbol_new(graph, mul, inputs, input_size, outputs, output_size, "cmul");
2583}
2584
2585static ccv_cnnp_model_t* _ccv_cnnp_cmul_copy(const ccv_cnnp_model_t* const self, void* const context);
2586
2587static const ccv_cnnp_model_vtab_t ccv_cnnp_cmul_isa = {
2588 .build = _ccv_cnnp_cmul_build,
2589 .copy = _ccv_cnnp_cmul_copy,
2590};
2591
2592ccv_cnnp_model_t* ccv_cnnp_cmul(const char* const name)
2593{
2594 ccv_cnnp_model_cmul_t* const model_cmul = (ccv_cnnp_model_cmul_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_cmul_t));
2595 model_cmul->super.isa = &ccv_cnnp_cmul_isa;
2596 model_cmul->super.input_size = 2;
2597 model_cmul->super.outputs = &model_cmul->output;
2598 model_cmul->super.output_size = 1;
2599 ccv_cnnp_model_copy_name(&model_cmul->super, name);
2600 return (ccv_cnnp_model_t*)model_cmul;
2601}
2602
2603static ccv_cnnp_model_t* _ccv_cnnp_cmul_copy(const ccv_cnnp_model_t* const super, void* const context)
2604{
2605 return ccv_cnnp_cmul(super->name);
2606}
2607
2608// MARK - Transpose Layer
2609
2610typedef struct {
2611 ccv_cnnp_model_t super;
2612 ccv_nnc_tensor_symbol_t output;
2613 int transpose[2];
2614} ccv_cnnp_model_transpose_t;
2615
2616static 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)
2617{
2618 ccv_cnnp_model_transpose_t* const self = (ccv_cnnp_model_transpose_t*)super;
2619 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);
2620 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"
, 2620, __extension__ __PRETTY_FUNCTION__); }));
2621 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 2621, __extension__ __PRETTY_FUNCTION__
); }));
2622 if (self->transpose[0] == self->transpose[1])
2623 {
2624 outputs[0] = inputs[0];
2625 return;
2626 }
2627 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2628 ccv_nnc_tensor_param_t output_params;
2629 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);
2630 ccv_nnc_hint_tensor_auto(transpose, (ccv_nnc_tensor_param_t []){
2631 params,
2632 }, 1, ccv_nnc_no_hint, &output_params, 1);
2633 const ccv_nnc_tensor_symbol_t transpose_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
2634 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");
2635 outputs[0] = transpose_output;
2636}
2637
2638static ccv_cnnp_model_t* _ccv_cnnp_transpose_copy(const ccv_cnnp_model_t* const super, void* const context);
2639
2640static const ccv_cnnp_model_vtab_t ccv_cnnp_transpose_isa = {
2641 .build = _ccv_cnnp_transpose_build,
2642 .copy = _ccv_cnnp_transpose_copy,
2643};
2644
2645ccv_cnnp_model_t* ccv_cnnp_transpose(const int axis_a, const int axis_b, const char* const name)
2646{
2647 ccv_cnnp_model_transpose_t* const model_transpose = (ccv_cnnp_model_transpose_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_transpose_t));
2648 model_transpose->super.isa = &ccv_cnnp_transpose_isa;
2649 model_transpose->super.input_size = 1;
2650 model_transpose->super.outputs = &model_transpose->output;
2651 model_transpose->super.output_size = 1;
2652 model_transpose->transpose[0] = axis_a;
2653 model_transpose->transpose[1] = axis_b;
2654 ccv_cnnp_model_copy_name(&model_transpose->super, name);
2655 return (ccv_cnnp_model_t*)model_transpose;
2656}
2657
2658static ccv_cnnp_model_t* _ccv_cnnp_transpose_copy(const ccv_cnnp_model_t* const super, void* const context)
2659{
2660 const ccv_cnnp_model_transpose_t* const self = (const ccv_cnnp_model_transpose_t*)super;
2661 return ccv_cnnp_transpose(self->transpose[0], self->transpose[1], self->super.name);
2662}
2663
2664// MARK - Layer Norm Layer
2665
2666typedef struct {
2667 ccv_cnnp_model_t super;
2668 ccv_nnc_tensor_symbol_t output;
2669 ccv_nnc_tensor_symbol_t bias;
2670 ccv_nnc_tensor_symbol_t scale;
2671 ccv_nnc_cmd_param_t params;
2672} ccv_cnnp_model_layer_norm_t;
2673
2674static 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)
2675{
2676 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);
2677 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"
, 2677, __extension__ __PRETTY_FUNCTION__); }));
2678 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 2678, __extension__ __PRETTY_FUNCTION__
); }));
2679 ccv_cnnp_model_layer_norm_t* const self = (ccv_cnnp_model_layer_norm_t*)super;
2680 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2681 ccv_nnc_tensor_param_t bias_params = params;
2682 const int nd = ccv_nnc_tensor_nd(params.dim);
2683 int i;
2684 for (i = 0; i < nd; i++)
2685 bias_params.dim[i] = 1;
2686 for (i = 0; i < self->params.lnorm.count; i++)
2687 bias_params.dim[self->params.lnorm.axis[i]] = params.dim[self->params.lnorm.axis[i]];
2688 if (self->params.lnorm.elementwise_affine)
2689 {
2690 // Both scale and bias are shared between if this model is reused.
2691 if (!self->scale.graph)
2692 self->scale = ccv_nnc_tensor_symbol_new(graph, bias_params, "scale");
2693 if (!self->bias.graph)
2694 self->bias = ccv_nnc_tensor_symbol_new(graph, bias_params, "bias");
2695 }
2696 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
};
2697 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
};
2698 const ccv_nnc_cmd_t layer_norm = ccv_nnc_cmd(CCV_NNC_LAYER_NORM_FORWARD, 0, self->params, 0);
2699 ccv_nnc_tensor_param_t output_params[3];
2700 if (self->params.lnorm.elementwise_affine)
2701 ccv_nnc_hint_tensor_auto(layer_norm, (ccv_nnc_tensor_param_t []){
2702 params,
2703 bias_params,
2704 bias_params,
2705 }, 3, ccv_nnc_no_hint, output_params, 3);
2706 else
2707 ccv_nnc_hint_tensor_auto(layer_norm, (ccv_nnc_tensor_param_t []){
2708 params,
2709 }, 1, ccv_nnc_no_hint, output_params, 3);
2710 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0);
2711 const ccv_nnc_tensor_symbol_t saved_mean = ccv_nnc_tensor_symbol_new(graph, output_params[1], "saved_mean");
2712 const ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(graph, output_params[2], "saved_inv_std");
2713 if (self->params.lnorm.elementwise_affine)
2714 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");
2715 else
2716 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");
2717 outputs[0] = output;
2718}
2719
2720static 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)
2721{
2722 ccv_cnnp_model_layer_norm_t* const self = (ccv_cnnp_model_layer_norm_t*)super;
2723 if (self->scale.graph)
2724 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);
2725 if (self->bias.graph)
2726 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);
2727}
2728
2729static 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)
2730{
2731 ccv_cnnp_model_layer_norm_t* const self = (ccv_cnnp_model_layer_norm_t*)super;
2732 if (self->scale.graph)
2733 add_to_array(parameters, self->scale, is_trainable);
2734 if (self->bias.graph)
2735 add_to_array(parameters, self->bias, is_trainable);
2736}
2737
2738static ccv_cnnp_model_t* _ccv_cnnp_layer_norm_copy(const ccv_cnnp_model_t* const super, void* const context);
2739
2740static const ccv_cnnp_model_vtab_t ccv_cnnp_layer_norm_isa = {
2741 .build = _ccv_cnnp_layer_norm_build,
2742 .init_states = _ccv_cnnp_layer_norm_init_states,
2743 .add_to_parameter = _ccv_cnnp_layer_norm_add_to_parameter,
2744 .copy = _ccv_cnnp_layer_norm_copy,
2745};
2746
2747ccv_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)
2748{
2749 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));
2750 model_layer_norm->super.isa = &ccv_cnnp_layer_norm_isa;
2751 model_layer_norm->super.input_size = 1;
2752 model_layer_norm->super.outputs = &model_layer_norm->output;
2753 model_layer_norm->super.output_size = 1;
2754 model_layer_norm->super.is_trainable = is_trainable;
2755 ccv_cnnp_model_copy_name(&model_layer_norm->super, name);
2756 model_layer_norm->scale.d = CCV_NNC_NO_TENSOR_SYMBOL;
2757 model_layer_norm->scale.graph = 0;
2758 model_layer_norm->bias.d = CCV_NNC_NO_TENSOR_SYMBOL;
2759 model_layer_norm->bias.graph = 0;
2760 model_layer_norm->params.lnorm.epsilon = epsilon;
2761 model_layer_norm->params.lnorm.count = axis_count;
2762 model_layer_norm->params.lnorm.elementwise_affine = elementwise_affine;
2763 memcpy(model_layer_norm->params.lnorm.axis, axis, sizeof(int) * axis_count);
2764 return (ccv_cnnp_model_t*)model_layer_norm;
2765}
2766
2767static ccv_cnnp_model_t* _ccv_cnnp_layer_norm_copy(const ccv_cnnp_model_t* const super, void* const context)
2768{
2769 const ccv_cnnp_model_layer_norm_t* const self = (const ccv_cnnp_model_layer_norm_t*)super;
2770 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);
2771}
2772
2773// MARK - Group Norm Layer
2774
2775typedef struct {
2776 ccv_cnnp_model_t super;
2777 ccv_nnc_tensor_symbol_t output;
2778 ccv_nnc_tensor_symbol_t bias;
2779 ccv_nnc_tensor_symbol_t scale;
2780 ccv_nnc_cmd_param_t params;
2781} ccv_cnnp_model_group_norm_t;
2782
2783static 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)
2784{
2785 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);
2786 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"
, 2786, __extension__ __PRETTY_FUNCTION__); }));
2787 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 2787, __extension__ __PRETTY_FUNCTION__
); }));
2788 ccv_cnnp_model_group_norm_t* const self = (ccv_cnnp_model_group_norm_t*)super;
2789 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2790 ccv_nnc_tensor_param_t bias_params = params;
2791 const int nd = ccv_nnc_tensor_nd(params.dim);
2792 int i;
2793 for (i = 0; i < nd; i++)
2794 bias_params.dim[i] = 1;
2795 bias_params.dim[self->params.gnorm.group_axis] = params.dim[self->params.gnorm.group_axis];
2796 if (self->params.gnorm.elementwise_affine)
2797 {
2798 // Both scale and bias are shared between if this model is reused.
2799 if (!self->scale.graph)
2800 self->scale = ccv_nnc_tensor_symbol_new(graph, bias_params, "scale");
2801 if (!self->bias.graph)
2802 self->bias = ccv_nnc_tensor_symbol_new(graph, bias_params, "bias");
2803 }
2804 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
};
2805 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
};
2806 const ccv_nnc_cmd_t group_norm = ccv_nnc_cmd(CCV_NNC_GROUP_NORM_FORWARD, 0, self->params, 0);
2807 ccv_nnc_tensor_param_t output_params[3];
2808 if (self->params.gnorm.elementwise_affine)
2809 ccv_nnc_hint_tensor_auto(group_norm, (ccv_nnc_tensor_param_t []){
2810 params,
2811 bias_params,
2812 bias_params,
2813 }, 3, ccv_nnc_no_hint, output_params, 3);
2814 else
2815 ccv_nnc_hint_tensor_auto(group_norm, (ccv_nnc_tensor_param_t []){
2816 params,
2817 }, 1, ccv_nnc_no_hint, output_params, 3);
2818 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0);
2819 const ccv_nnc_tensor_symbol_t saved_mean = ccv_nnc_tensor_symbol_new(graph, output_params[1], "saved_mean");
2820 const ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(graph, output_params[2], "saved_inv_std");
2821 if (self->params.gnorm.elementwise_affine)
2822 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");
2823 else
2824 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");
2825 outputs[0] = output;
2826}
2827
2828static 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)
2829{
2830 ccv_cnnp_model_group_norm_t* const self = (ccv_cnnp_model_group_norm_t*)super;
2831 if (self->scale.graph)
2832 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);
2833 if (self->bias.graph)
2834 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);
2835}
2836
2837static 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)
2838{
2839 ccv_cnnp_model_group_norm_t* const self = (ccv_cnnp_model_group_norm_t*)super;
2840 if (self->scale.graph)
2841 add_to_array(parameters, self->scale, is_trainable);
2842 if (self->bias.graph)
2843 add_to_array(parameters, self->bias, is_trainable);
2844}
2845
2846static ccv_cnnp_model_t* _ccv_cnnp_group_norm_copy(const ccv_cnnp_model_t* const super, void* const context);
2847
2848static const ccv_cnnp_model_vtab_t ccv_cnnp_group_norm_isa = {
2849 .build = _ccv_cnnp_group_norm_build,
2850 .init_states = _ccv_cnnp_group_norm_init_states,
2851 .add_to_parameter = _ccv_cnnp_group_norm_add_to_parameter,
2852 .copy = _ccv_cnnp_group_norm_copy,
2853};
2854
2855ccv_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)
2856{
2857 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));
2858 model_group_norm->super.isa = &ccv_cnnp_group_norm_isa;
2859 model_group_norm->super.input_size = 1;
2860 model_group_norm->super.outputs = &model_group_norm->output;
2861 model_group_norm->super.output_size = 1;
2862 model_group_norm->super.is_trainable = is_trainable;
2863 ccv_cnnp_model_copy_name(&model_group_norm->super, name);
2864 model_group_norm->scale.d = CCV_NNC_NO_TENSOR_SYMBOL;
2865 model_group_norm->scale.graph = 0;
2866 model_group_norm->bias.d = CCV_NNC_NO_TENSOR_SYMBOL;
2867 model_group_norm->bias.graph = 0;
2868 model_group_norm->params.gnorm.group_axis = group_axis;
2869 model_group_norm->params.gnorm.groups = groups;
2870 model_group_norm->params.gnorm.epsilon = epsilon;
2871 model_group_norm->params.gnorm.reduce_count = axis_count;
2872 model_group_norm->params.gnorm.elementwise_affine = elementwise_affine;
2873 memcpy(model_group_norm->params.gnorm.reduce_axis, reduce_axis, sizeof(int) * axis_count);
2874 return (ccv_cnnp_model_t*)model_group_norm;
2875}
2876
2877static ccv_cnnp_model_t* _ccv_cnnp_group_norm_copy(const ccv_cnnp_model_t* const super, void* const context)
2878{
2879 const ccv_cnnp_model_group_norm_t* const self = (const ccv_cnnp_model_group_norm_t*)super;
2880 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);
2881}
2882
2883// MARK - RMSNorm Layer
2884
2885typedef struct {
2886 ccv_cnnp_model_t super;
2887 ccv_nnc_tensor_symbol_t output;
2888 ccv_nnc_tensor_symbol_t scale;
2889 ccv_nnc_cmd_param_t params;
2890} ccv_cnnp_model_rmsnorm_t;
2891
2892static 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)
2893{
2894 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);
2895 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"
, 2895, __extension__ __PRETTY_FUNCTION__); }));
2896 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 2896, __extension__ __PRETTY_FUNCTION__
); }));
2897 ccv_cnnp_model_rmsnorm_t* const self = (ccv_cnnp_model_rmsnorm_t*)super;
2898 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2899 ccv_nnc_tensor_param_t scale_params = params;
2900 const int nd = ccv_nnc_tensor_nd(params.dim);
2901 int i;
2902 for (i = 0; i < nd; i++)
2903 scale_params.dim[i] = 1;
2904 for (i = 0; i < self->params.rmsnorm.count; i++)
2905 scale_params.dim[self->params.rmsnorm.axis[i]] = params.dim[self->params.rmsnorm.axis[i]];
2906 // Both scale and bias are shared between if this model is reused.
2907 if (self->params.rmsnorm.elementwise_affine)
2908 {
2909 if (!self->scale.graph)
2910 self->scale = ccv_nnc_tensor_symbol_new(graph, scale_params, "scale");
2911 }
2912 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
};
2913 const ccv_nnc_cmd_t rmsnorm = ccv_nnc_cmd(CCV_NNC_RMSNORM_FORWARD, 0, self->params, 0);
2914 ccv_nnc_tensor_param_t output_params[2];
2915 if (self->params.rmsnorm.elementwise_affine)
2916 ccv_nnc_hint_tensor_auto(rmsnorm, (ccv_nnc_tensor_param_t []){
2917 params,
2918 scale_params,
2919 }, 2, ccv_nnc_no_hint, output_params, 2);
2920 else
2921 ccv_nnc_hint_tensor_auto(rmsnorm, (ccv_nnc_tensor_param_t []){
2922 params,
2923 }, 1, ccv_nnc_no_hint, output_params, 2);
2924 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0);
2925 const ccv_nnc_tensor_symbol_t saved_inv_std = ccv_nnc_tensor_symbol_new(graph, output_params[1], "saved_inv_std");
2926 if (self->params.rmsnorm.elementwise_affine)
2927 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");
2928 else
2929 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");
2930 outputs[0] = output;
2931}
2932
2933static 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)
2934{
2935 ccv_cnnp_model_rmsnorm_t* const self = (ccv_cnnp_model_rmsnorm_t*)super;
2936 if (self->scale.graph)
2937 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);
2938}
2939
2940static 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)
2941{
2942 ccv_cnnp_model_rmsnorm_t* const self = (ccv_cnnp_model_rmsnorm_t*)super;
2943 if (self->scale.graph)
2944 add_to_array(parameters, self->scale, is_trainable);
2945}
2946
2947static ccv_cnnp_model_t* _ccv_cnnp_rmsnorm_copy(const ccv_cnnp_model_t* const super, void* const context);
2948
2949static const ccv_cnnp_model_vtab_t ccv_cnnp_rmsnorm_isa = {
2950 .build = _ccv_cnnp_rmsnorm_build,
2951 .init_states = _ccv_cnnp_rmsnorm_init_states,
2952 .add_to_parameter = _ccv_cnnp_rmsnorm_add_to_parameter,
2953 .copy = _ccv_cnnp_rmsnorm_copy,
2954};
2955
2956ccv_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)
2957{
2958 ccv_cnnp_model_rmsnorm_t* const model_rmsnorm = (ccv_cnnp_model_rmsnorm_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_rmsnorm_t));
2959 model_rmsnorm->super.isa = &ccv_cnnp_rmsnorm_isa;
2960 model_rmsnorm->super.input_size = 1;
2961 model_rmsnorm->super.outputs = &model_rmsnorm->output;
2962 model_rmsnorm->super.output_size = 1;
2963 model_rmsnorm->super.is_trainable = is_trainable;
2964 ccv_cnnp_model_copy_name(&model_rmsnorm->super, name);
2965 model_rmsnorm->scale.d = CCV_NNC_NO_TENSOR_SYMBOL;
2966 model_rmsnorm->scale.graph = 0;
2967 model_rmsnorm->params.rmsnorm.epsilon = epsilon;
2968 model_rmsnorm->params.rmsnorm.count = axis_count;
2969 model_rmsnorm->params.rmsnorm.elementwise_affine = elementwise_affine;
2970 memcpy(model_rmsnorm->params.lnorm.axis, axis, sizeof(int) * axis_count);
2971 return (ccv_cnnp_model_t*)model_rmsnorm;
2972}
2973
2974static ccv_cnnp_model_t* _ccv_cnnp_rmsnorm_copy(const ccv_cnnp_model_t* const super, void* const context)
2975{
2976 const ccv_cnnp_model_rmsnorm_t* const self = (const ccv_cnnp_model_rmsnorm_t*)super;
2977 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);
2978}
2979
2980// MARK - Batched Matrix Mul Layer
2981
2982typedef struct {
2983 ccv_cnnp_model_t super;
2984 ccv_nnc_tensor_symbol_t output;
2985 int transpose_a[2];
2986 int transpose_b[2];
2987 int flags;
2988} ccv_cnnp_model_matmul_t;
2989
2990static 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)
2991{
2992 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);
2993 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"
, 2993, __extension__ __PRETTY_FUNCTION__); }));
2994 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 2994, __extension__ __PRETTY_FUNCTION__
); }));
2995 ccv_cnnp_model_matmul_t* const self = (ccv_cnnp_model_matmul_t*)super;
2996 ccv_nnc_tensor_param_t a_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
2997 ccv_nnc_tensor_param_t b_params = ccv_nnc_tensor_symbol_params(graph, inputs[1]);
2998 ccv_nnc_tensor_param_t output_params;
2999 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);
3000 matmul.info.blas.flags = self->flags;
3001 ccv_nnc_hint_tensor_auto(matmul, (ccv_nnc_tensor_param_t []){
3002 a_params,
3003 b_params,
3004 }, 2, ccv_nnc_no_hint, &output_params, 1);
3005 const ccv_nnc_tensor_symbol_t matmul_output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3006 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");
3007 outputs[0] = matmul_output;
3008}
3009
3010static ccv_cnnp_model_t* _ccv_cnnp_matmul_copy(const ccv_cnnp_model_t* const super, void* const context);
3011
3012static const ccv_cnnp_model_vtab_t ccv_cnnp_matmul_isa = {
3013 .build = _ccv_cnnp_matmul_build,
3014 .copy = _ccv_cnnp_matmul_copy,
3015};
3016
3017ccv_cnnp_model_t* ccv_cnnp_matmul(const int transpose_a[2], const int transpose_b[2], const int flags, const char* const name)
3018{
3019 ccv_cnnp_model_matmul_t* const model_matmul = (ccv_cnnp_model_matmul_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_matmul_t));
3020 model_matmul->super.isa = &ccv_cnnp_matmul_isa;
3021 model_matmul->super.input_size = 2;
3022 model_matmul->super.outputs = &model_matmul->output;
3023 model_matmul->super.output_size = 1;
3024 model_matmul->transpose_a[0] = transpose_a[0];
3025 model_matmul->transpose_a[1] = transpose_a[1];
3026 model_matmul->transpose_b[0] = transpose_b[0];
3027 model_matmul->transpose_b[1] = transpose_b[1];
3028 model_matmul->flags = flags;
3029 ccv_cnnp_model_copy_name(&model_matmul->super, name);
3030 return (ccv_cnnp_model_t*)model_matmul;
3031}
3032
3033static ccv_cnnp_model_t* _ccv_cnnp_matmul_copy(const ccv_cnnp_model_t* const super, void* const context)
3034{
3035 const ccv_cnnp_model_matmul_t* const self = (const ccv_cnnp_model_matmul_t*)super;
3036 return ccv_cnnp_matmul(self->transpose_a, self->transpose_b, self->flags, self->super.name);
3037}
3038
3039// MARK - Dropout Layer
3040
3041typedef struct {
3042 ccv_cnnp_model_t super;
3043 ccv_nnc_tensor_symbol_t output;
3044 ccv_nnc_graph_exec_symbol_t dropout;
3045 float p;
3046 int entirety;
3047} ccv_cnnp_model_dropout_t;
3048
3049static 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)
3050{
3051 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);
3052 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"
, 3052, __extension__ __PRETTY_FUNCTION__); }));
3053 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 3053, __extension__ __PRETTY_FUNCTION__
); }));
3054 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3055 ccv_nnc_tensor_param_t output_params[2];
3056 ccv_cnnp_model_dropout_t* const self = (ccv_cnnp_model_dropout_t*)super;
3057 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);
3058 ccv_nnc_hint_tensor_auto(dropout, (ccv_nnc_tensor_param_t []){
3059 params,
3060 }, 1, ccv_nnc_no_hint, output_params, 2);
3061 const ccv_nnc_tensor_symbol_t dropout_output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0);
3062 const ccv_nnc_tensor_symbol_t mask = ccv_nnc_tensor_symbol_new(graph, output_params[1], "mask");
3063 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");
3064 outputs[0] = dropout_output;
3065}
3066
3067static 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)
3068{
3069 ccv_cnnp_model_dropout_t* const self = (ccv_cnnp_model_dropout_t*)super;
3070 if (self->dropout.graph)
3071 {
3072 if (is_test)
3073 // During test, the dropout is not applied. Data transfer is perfect because if these are the same tensor, it will skip.
3074 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);
3075 else
3076 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);
3077 }
3078}
3079
3080static ccv_cnnp_model_t* _ccv_cnnp_dropout_copy(const ccv_cnnp_model_t* const super, void* const context);
3081
3082static const ccv_cnnp_model_vtab_t ccv_cnnp_dropout_isa = {
3083 .build = _ccv_cnnp_dropout_build,
3084 .set_is_test = _ccv_cnnp_dropout_set_is_test,
3085 .copy = _ccv_cnnp_dropout_copy,
3086};
3087
3088ccv_cnnp_model_t* ccv_cnnp_dropout(const float p, const int entirety, const char* const name)
3089{
3090 ccv_cnnp_model_dropout_t* const model_dropout = (ccv_cnnp_model_dropout_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_dropout_t));
3091 model_dropout->super.isa = &ccv_cnnp_dropout_isa;
3092 model_dropout->super.input_size = 1;
3093 model_dropout->super.outputs = &model_dropout->output;
3094 model_dropout->super.output_size = 1;
3095 model_dropout->p = p;
3096 model_dropout->entirety = entirety;
3097 ccv_cnnp_model_copy_name(&model_dropout->super, name);
3098 return (ccv_cnnp_model_t*)model_dropout;
3099}
3100
3101static ccv_cnnp_model_t* _ccv_cnnp_dropout_copy(const ccv_cnnp_model_t* const super, void* const context)
3102{
3103 const ccv_cnnp_model_dropout_t* const self = (const ccv_cnnp_model_dropout_t*)super;
3104 return ccv_cnnp_dropout(self->p, self->entirety, self->super.name);
3105}
3106
3107// MARK - Masked Fill Layer
3108
3109typedef struct {
3110 ccv_cnnp_model_t super;
3111 ccv_nnc_tensor_symbol_t output;
3112 float eq;
3113 float fill;
3114} ccv_cnnp_model_masked_fill_t;
3115
3116static 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)
3117{
3118 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);
3119 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"
, 3119, __extension__ __PRETTY_FUNCTION__); }));
3120 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 3120, __extension__ __PRETTY_FUNCTION__
); }));
3121 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3122 ccv_cnnp_model_masked_fill_t* const self = (ccv_cnnp_model_masked_fill_t*)super;
3123 const ccv_nnc_tensor_symbol_t masked_fill_output = ccv_nnc_tensor_symbol_new(graph, params, 0);
3124 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");
3125 outputs[0] = masked_fill_output;
3126}
3127
3128static ccv_cnnp_model_t* _ccv_cnnp_masked_fill_copy(const ccv_cnnp_model_t* const super, void* const context);
3129
3130static const ccv_cnnp_model_vtab_t ccv_cnnp_masked_fill_isa = {
3131 .build = _ccv_cnnp_masked_fill_build,
3132 .copy = _ccv_cnnp_masked_fill_copy,
3133};
3134
3135ccv_cnnp_model_t* ccv_cnnp_masked_fill(const float eq, const float fill, const char* const name)
3136{
3137 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));
3138 model_masked_fill->super.isa = &ccv_cnnp_masked_fill_isa;
3139 model_masked_fill->super.input_size = 2;
3140 model_masked_fill->super.outputs = &model_masked_fill->output;
3141 model_masked_fill->super.output_size = 1;
3142 model_masked_fill->eq = eq;
3143 model_masked_fill->fill = fill;
3144 ccv_cnnp_model_copy_name(&model_masked_fill->super, name);
3145 return (ccv_cnnp_model_t*)model_masked_fill;
3146}
3147
3148static ccv_cnnp_model_t* _ccv_cnnp_masked_fill_copy(const ccv_cnnp_model_t* const super, void* const context)
3149{
3150 const ccv_cnnp_model_masked_fill_t* const self = (const ccv_cnnp_model_masked_fill_t*)super;
3151 return ccv_cnnp_masked_fill(self->eq, self->fill, self->super.name);
3152}
3153
3154// MARK - Index Select Layer
3155
3156typedef struct {
3157 ccv_cnnp_model_t super;
3158 ccv_nnc_tensor_symbol_t output;
3159} ccv_cnnp_model_index_select_t;
3160
3161static 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)
3162{
3163 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);
3164 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"
, 3164, __extension__ __PRETTY_FUNCTION__); }));
3165 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 3165, __extension__ __PRETTY_FUNCTION__
); }));
3166 const ccv_nnc_tensor_param_t vocab_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3167 const ccv_nnc_tensor_param_t index_params = ccv_nnc_tensor_symbol_params(graph, inputs[1]);
3168 ccv_nnc_tensor_param_t output_params;
3169 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);
3170 ccv_nnc_hint_tensor_auto(index_select, (ccv_nnc_tensor_param_t []){
3171 vocab_params,
3172 index_params,
3173 }, 2, ccv_nnc_no_hint, &output_params, 1);
3174 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3175 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");
3176 outputs[0] = output;
3177}
3178
3179static ccv_cnnp_model_t* _ccv_cnnp_index_select_copy(const ccv_cnnp_model_t* const super, void* const context);
3180
3181static const ccv_cnnp_model_vtab_t ccv_cnnp_index_select_isa = {
3182 .build = _ccv_cnnp_index_select_build,
3183 .copy = _ccv_cnnp_index_select_copy,
3184};
3185
3186ccv_cnnp_model_t* ccv_cnnp_index_select(const char* const name)
3187{
3188 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));
3189 model_index_select->super.isa = &ccv_cnnp_index_select_isa;
3190 model_index_select->super.input_size = 2;
3191 model_index_select->super.outputs = &model_index_select->output;
3192 model_index_select->super.output_size = 1;
3193 ccv_cnnp_model_copy_name(&model_index_select->super, name);
3194 return (ccv_cnnp_model_t*)model_index_select;
3195}
3196
3197static ccv_cnnp_model_t* _ccv_cnnp_index_select_copy(const ccv_cnnp_model_t* const super, void* const context)
3198{
3199 ccv_cnnp_model_index_select_t* const self = (ccv_cnnp_model_index_select_t*)super;
3200 return ccv_cnnp_index_select(self->super.name);
3201}
3202
3203// MARK - Embedding Layer
3204
3205typedef struct {
3206 ccv_cnnp_model_t super;
3207 ccv_nnc_tensor_symbol_t output;
3208 ccv_nnc_tensor_symbol_t vocab;
3209 int datatype;
3210 int vocab_size;
3211 int embed_size;
3212} ccv_cnnp_model_embedding_t;
3213
3214static 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)
3215{
3216 ccv_cnnp_model_embedding_t* const self = (ccv_cnnp_model_embedding_t*)super;
3217 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);
3218 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"
, 3218, __extension__ __PRETTY_FUNCTION__); }));
3219 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 3219, __extension__ __PRETTY_FUNCTION__
); }));
3220 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3221 ccv_nnc_tensor_param_t vocab_params = params;
3222 memset(vocab_params.dim, 0, sizeof(vocab_params.dim));
3223 vocab_params.datatype = self->datatype;
3224 vocab_params.dim[0] = self->vocab_size;
3225 vocab_params.dim[1] = self->embed_size;
3226 if (!self->vocab.graph)
3227 self->vocab = ccv_nnc_tensor_symbol_new(graph, vocab_params, "vocab");
3228 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", 3228
, __extension__ __PRETTY_FUNCTION__); }));
3229 const ccv_nnc_tensor_symbol_t vocab = ccv_cnnp_model_get_symbol(super, self->vocab);
3230 ccv_nnc_tensor_param_t output_params;
3231 const ccv_nnc_cmd_t embedding = CMD_INDEX_SELECT_FORWARD()ccv_nnc_cmd(CCV_NNC_INDEX_SELECT_FORWARD, 0, ccv_nnc_cmd_auto
, 0);
3232 ccv_nnc_hint_tensor_auto(embedding, (ccv_nnc_tensor_param_t []){
3233 vocab_params,
3234 params,
3235 }, 2, ccv_nnc_no_hint, &output_params, 1);
3236 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3237 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");
3238 outputs[0] = output;
3239}
3240
3241static 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)
3242{
3243 ccv_cnnp_model_embedding_t* const self = (ccv_cnnp_model_embedding_t*)super;
3244 const float std = sqrtf(2) / sqrtf(self->vocab_size + self->embed_size);
3245 const float bound = sqrtf(3) * std;
3246 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);
3247}
3248
3249static 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)
3250{
3251 ccv_cnnp_model_embedding_t* const self = (ccv_cnnp_model_embedding_t*)super;
3252 add_to_array(parameters, self->vocab, is_trainable);
3253}
3254
3255static ccv_cnnp_model_t* _ccv_cnnp_embedding_copy(const ccv_cnnp_model_t* const super, void* const context);
3256
3257static const ccv_cnnp_model_vtab_t ccv_cnnp_embedding_isa = {
3258 .build = _ccv_cnnp_embedding_build,
3259 .init_states = _ccv_cnnp_embedding_init_states,
3260 .add_to_parameter = _ccv_cnnp_embedding_add_to_parameter,
3261 .copy = _ccv_cnnp_embedding_copy,
3262};
3263
3264ccv_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)
3265{
3266 ccv_cnnp_model_embedding_t* const model_embedding = (ccv_cnnp_model_embedding_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_embedding_t));
3267 model_embedding->super.isa = &ccv_cnnp_embedding_isa;
3268 model_embedding->super.input_size = 1;
3269 model_embedding->super.outputs = &model_embedding->output;
3270 model_embedding->super.output_size = 1;
3271 model_embedding->super.is_trainable = is_trainable;
3272 ccv_cnnp_model_copy_name(&model_embedding->super, name);
3273 model_embedding->vocab.d = CCV_NNC_NO_TENSOR_SYMBOL;
3274 model_embedding->vocab.graph = 0;
3275 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", 3275, __extension__ __PRETTY_FUNCTION__
); }));
3276 model_embedding->datatype = datatype;
3277 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", 3277, __extension__ __PRETTY_FUNCTION__
); }));
3278 model_embedding->vocab_size = vocab_size;
3279 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", 3279, __extension__ __PRETTY_FUNCTION__
); }));
3280 model_embedding->embed_size = embed_size;
3281 return (ccv_cnnp_model_t*)model_embedding;
3282}
3283
3284static ccv_cnnp_model_t* _ccv_cnnp_embedding_copy(const ccv_cnnp_model_t* const super, void* const context)
3285{
3286 ccv_cnnp_model_embedding_t* const self = (ccv_cnnp_model_embedding_t*)super;
3287 return ccv_cnnp_embedding(self->datatype, self->vocab_size, self->embed_size, self->super.is_trainable, self->super.name);
3288}
3289
3290// MARK - Pool Layers
3291
3292typedef struct {
3293 ccv_cnnp_model_t super;
3294 ccv_nnc_tensor_symbol_t output;
3295 int type;
3296 float width_scale;
3297 float height_scale;
3298 int align_corners;
3299} ccv_cnnp_model_upsample_t;
3300
3301static 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)
3302{
3303 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);
3304 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"
, 3304, __extension__ __PRETTY_FUNCTION__); }));
3305 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 3305, __extension__ __PRETTY_FUNCTION__
); }));
3306 ccv_cnnp_model_upsample_t* const self = (ccv_cnnp_model_upsample_t*)super;
3307 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3308 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);
3309 ccv_nnc_tensor_param_t output_params;
3310 ccv_nnc_hint_tensor_auto(cmd, &params, 1, ccv_nnc_no_hint, &output_params, 1);
3311 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3312 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");
3313 outputs[0] = output;
3314}
3315
3316static ccv_cnnp_model_t* _ccv_cnnp_upsample_copy(const ccv_cnnp_model_t* const super, void* const context);
3317
3318static const ccv_cnnp_model_vtab_t ccv_cnnp_upsample_isa = {
3319 .build = _ccv_cnnp_upsample_build,
3320 .copy = _ccv_cnnp_upsample_copy,
3321};
3322
3323ccv_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)
3324{
3325 ccv_cnnp_model_upsample_t* const model_upsample = (ccv_cnnp_model_upsample_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_upsample_t));
3326 model_upsample->super.isa = &ccv_cnnp_upsample_isa;
3327 model_upsample->super.input_size = 1;
3328 model_upsample->super.outputs = &model_upsample->output;
3329 model_upsample->super.output_size = 1;
3330 ccv_cnnp_model_copy_name(&model_upsample->super, name);
3331 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", 3331, __extension__ __PRETTY_FUNCTION__
); }));
3332 model_upsample->type = type;
3333 model_upsample->width_scale = width_scale;
3334 model_upsample->height_scale = height_scale;
3335 model_upsample->align_corners = align_corners;
3336 return (ccv_cnnp_model_t*)model_upsample;
3337}
3338
3339static ccv_cnnp_model_t* _ccv_cnnp_upsample_copy(const ccv_cnnp_model_t* const super, void* const context)
3340{
3341 const ccv_cnnp_model_upsample_t* const self = (const ccv_cnnp_model_upsample_t*)super;
3342 return ccv_cnnp_upsample(self->type, self->width_scale, self->height_scale, self->align_corners, self->super.name);
3343}
3344
3345// MARK - Reduce Sum Layer
3346
3347typedef struct {
3348 ccv_cnnp_model_t super;
3349 int axis[CCV_NNC_MAX_DIM_ALLOC(12)];
3350 int count;
3351 ccv_nnc_tensor_symbol_t output;
3352} ccv_cnnp_model_reduce_sum_t;
3353
3354static 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)
3355{
3356 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);
3357 const ccv_cnnp_model_reduce_sum_t* const self = (const ccv_cnnp_model_reduce_sum_t*)super;
3358 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"
, 3358, __extension__ __PRETTY_FUNCTION__); }));
3359 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 3359, __extension__ __PRETTY_FUNCTION__
); }));
3360 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3361 ccv_nnc_tensor_param_t output_params;
3362 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);
3363 int i;
3364 for (i = 0; i < self->count; i++)
3365 reduce_sum.info.reduce.axis[i] = self->axis[i];
3366 reduce_sum.info.reduce.count = self->count;
3367 ccv_nnc_hint_tensor_auto(reduce_sum, &input_params, 1, ccv_nnc_no_hint, &output_params, 1);
3368 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3369 ccv_nnc_graph_exec_symbol_new(graph, reduce_sum, inputs, input_size, outputs, output_size, "reduce_sum");
3370}
3371
3372static ccv_cnnp_model_t* _ccv_cnnp_reduce_sum_copy(const ccv_cnnp_model_t* const self, void* const context);
3373
3374static const ccv_cnnp_model_vtab_t ccv_cnnp_reduce_sum_isa = {
3375 .build = _ccv_cnnp_reduce_sum_build,
3376 .copy = _ccv_cnnp_reduce_sum_copy,
3377};
3378
3379ccv_cnnp_model_t* ccv_cnnp_reduce_sum(const int* const axis, const int axis_count, const char* const name)
3380{
3381 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));
3382 model_reduce_sum->super.isa = &ccv_cnnp_reduce_sum_isa;
3383 model_reduce_sum->super.input_size = 1;
3384 model_reduce_sum->super.outputs = &model_reduce_sum->output;
3385 model_reduce_sum->super.output_size = 1;
3386 ccv_cnnp_model_copy_name(&model_reduce_sum->super, name);
3387 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", 3387, __extension__ __PRETTY_FUNCTION__
); }));
3388 int i;
3389 for (i = 0; i < axis_count; i++)
3390 model_reduce_sum->axis[i] = axis[i];
3391 model_reduce_sum->count = axis_count;
3392 return (ccv_cnnp_model_t*)model_reduce_sum;
3393}
3394
3395static ccv_cnnp_model_t* _ccv_cnnp_reduce_sum_copy(const ccv_cnnp_model_t* const super, void* const context)
3396{
3397 const ccv_cnnp_model_reduce_sum_t* const self = (const ccv_cnnp_model_reduce_sum_t*)super;
3398 return ccv_cnnp_reduce_sum(self->axis, self->count, self->super.name);
3399}
3400
3401// MARK - Reduce Mean Layer
3402
3403typedef struct {
3404 ccv_cnnp_model_t super;
3405 int axis[CCV_NNC_MAX_DIM_ALLOC(12)];
3406 int count;
3407 ccv_nnc_tensor_symbol_t output;
3408} ccv_cnnp_model_reduce_mean_t;
3409
3410static 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)
3411{
3412 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);
3413 const ccv_cnnp_model_reduce_mean_t* const self = (const ccv_cnnp_model_reduce_mean_t*)super;
3414 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"
, 3414, __extension__ __PRETTY_FUNCTION__); }));
3415 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 3415, __extension__ __PRETTY_FUNCTION__
); }));
3416 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3417 ccv_nnc_tensor_param_t output_params;
3418 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);
3419 int i;
3420 for (i = 0; i < self->count; i++)
3421 reduce_mean.info.reduce.axis[i] = self->axis[i];
3422 reduce_mean.info.reduce.count = self->count;
3423 ccv_nnc_hint_tensor_auto(reduce_mean, &input_params, 1, ccv_nnc_no_hint, &output_params, 1);
3424 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3425 ccv_nnc_graph_exec_symbol_new(graph, reduce_mean, inputs, input_size, outputs, output_size, "reduce_mean");
3426}
3427
3428static ccv_cnnp_model_t* _ccv_cnnp_reduce_mean_copy(const ccv_cnnp_model_t* const self, void* const context);
3429
3430static const ccv_cnnp_model_vtab_t ccv_cnnp_reduce_mean_isa = {
3431 .build = _ccv_cnnp_reduce_mean_build,
3432 .copy = _ccv_cnnp_reduce_mean_copy,
3433};
3434
3435ccv_cnnp_model_t* ccv_cnnp_reduce_mean(const int* const axis, const int axis_count, const char* const name)
3436{
3437 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));
3438 model_reduce_mean->super.isa = &ccv_cnnp_reduce_mean_isa;
3439 model_reduce_mean->super.input_size = 1;
3440 model_reduce_mean->super.outputs = &model_reduce_mean->output;
3441 model_reduce_mean->super.output_size = 1;
3442 ccv_cnnp_model_copy_name(&model_reduce_mean->super, name);
3443 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", 3443, __extension__ __PRETTY_FUNCTION__
); }));
3444 int i;
3445 for (i = 0; i < axis_count; i++)
3446 model_reduce_mean->axis[i] = axis[i];
3447 model_reduce_mean->count = axis_count;
3448 return (ccv_cnnp_model_t*)model_reduce_mean;
3449}
3450
3451static ccv_cnnp_model_t* _ccv_cnnp_reduce_mean_copy(const ccv_cnnp_model_t* const super, void* const context)
3452{
3453 const ccv_cnnp_model_reduce_mean_t* const self = (const ccv_cnnp_model_reduce_mean_t*)super;
3454 return ccv_cnnp_reduce_mean(self->axis, self->count, self->super.name);
3455}
3456
3457// MARK - Reduce Max Layer
3458
3459typedef struct {
3460 ccv_cnnp_model_t super;
3461 int axis[CCV_NNC_MAX_DIM_ALLOC(12)];
3462 int count;
3463 ccv_nnc_tensor_symbol_t output;
3464} ccv_cnnp_model_reduce_max_t;
3465
3466static 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)
3467{
3468 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);
3469 const ccv_cnnp_model_reduce_max_t* const self = (const ccv_cnnp_model_reduce_max_t*)super;
3470 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"
, 3470, __extension__ __PRETTY_FUNCTION__); }));
3471 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 3471, __extension__ __PRETTY_FUNCTION__
); }));
3472 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3473 ccv_nnc_tensor_param_t output_params;
3474 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);
3475 int i;
3476 for (i = 0; i < self->count; i++)
3477 reduce_max.info.reduce.axis[i] = self->axis[i];
3478 reduce_max.info.reduce.count = self->count;
3479 ccv_nnc_hint_tensor_auto(reduce_max, &input_params, 1, ccv_nnc_no_hint, &output_params, 1);
3480 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3481 ccv_nnc_graph_exec_symbol_new(graph, reduce_max, inputs, input_size, outputs, output_size, "reduce_max");
3482}
3483
3484static ccv_cnnp_model_t* _ccv_cnnp_reduce_max_copy(const ccv_cnnp_model_t* const self, void* const context);
3485
3486static const ccv_cnnp_model_vtab_t ccv_cnnp_reduce_max_isa = {
3487 .build = _ccv_cnnp_reduce_max_build,
3488 .copy = _ccv_cnnp_reduce_max_copy,
3489};
3490
3491ccv_cnnp_model_t* ccv_cnnp_reduce_max(const int* const axis, const int axis_count, const char* const name)
3492{
3493 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));
3494 model_reduce_max->super.isa = &ccv_cnnp_reduce_max_isa;
3495 model_reduce_max->super.input_size = 1;
3496 model_reduce_max->super.outputs = &model_reduce_max->output;
3497 model_reduce_max->super.output_size = 1;
3498 ccv_cnnp_model_copy_name(&model_reduce_max->super, name);
3499 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", 3499, __extension__ __PRETTY_FUNCTION__
); }));
3500 int i;
3501 for (i = 0; i < axis_count; i++)
3502 model_reduce_max->axis[i] = axis[i];
3503 model_reduce_max->count = axis_count;
3504 return (ccv_cnnp_model_t*)model_reduce_max;
3505}
3506
3507static ccv_cnnp_model_t* _ccv_cnnp_reduce_max_copy(const ccv_cnnp_model_t* const super, void* const context)
3508{
3509 const ccv_cnnp_model_reduce_max_t* const self = (const ccv_cnnp_model_reduce_max_t*)super;
3510 return ccv_cnnp_reduce_max(self->axis, self->count, self->super.name);
3511}
3512
3513// MARK - Reduce Min Layer
3514
3515typedef struct {
3516 ccv_cnnp_model_t super;
3517 int axis[CCV_NNC_MAX_DIM_ALLOC(12)];
3518 int count;
3519 ccv_nnc_tensor_symbol_t output;
3520} ccv_cnnp_model_reduce_min_t;
3521
3522static 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)
3523{
3524 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);
3525 const ccv_cnnp_model_reduce_min_t* const self = (const ccv_cnnp_model_reduce_min_t*)super;
3526 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"
, 3526, __extension__ __PRETTY_FUNCTION__); }));
3527 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 3527, __extension__ __PRETTY_FUNCTION__
); }));
3528 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3529 ccv_nnc_tensor_param_t output_params;
3530 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);
3531 int i;
3532 for (i = 0; i < self->count; i++)
3533 reduce_min.info.reduce.axis[i] = self->axis[i];
3534 reduce_min.info.reduce.count = self->count;
3535 ccv_nnc_hint_tensor_auto(reduce_min, &input_params, 1, ccv_nnc_no_hint, &output_params, 1);
3536 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3537 ccv_nnc_graph_exec_symbol_new(graph, reduce_min, inputs, input_size, outputs, output_size, "reduce_min");
3538}
3539
3540static ccv_cnnp_model_t* _ccv_cnnp_reduce_min_copy(const ccv_cnnp_model_t* const self, void* const context);
3541
3542static const ccv_cnnp_model_vtab_t ccv_cnnp_reduce_min_isa = {
3543 .build = _ccv_cnnp_reduce_min_build,
3544 .copy = _ccv_cnnp_reduce_min_copy,
3545};
3546
3547ccv_cnnp_model_t* ccv_cnnp_reduce_min(const int* const axis, const int axis_count, const char* const name)
3548{
3549 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));
3550 model_reduce_min->super.isa = &ccv_cnnp_reduce_min_isa;
3551 model_reduce_min->super.input_size = 1;
3552 model_reduce_min->super.outputs = &model_reduce_min->output;
3553 model_reduce_min->super.output_size = 1;
3554 ccv_cnnp_model_copy_name(&model_reduce_min->super, name);
3555 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", 3555, __extension__ __PRETTY_FUNCTION__
); }));
3556 int i;
3557 for (i = 0; i < axis_count; i++)
3558 model_reduce_min->axis[i] = axis[i];
3559 model_reduce_min->count = axis_count;
3560 return (ccv_cnnp_model_t*)model_reduce_min;
3561}
3562
3563static ccv_cnnp_model_t* _ccv_cnnp_reduce_min_copy(const ccv_cnnp_model_t* const super, void* const context)
3564{
3565 const ccv_cnnp_model_reduce_min_t* const self = (const ccv_cnnp_model_reduce_min_t*)super;
3566 return ccv_cnnp_reduce_min(self->axis, self->count, self->super.name);
3567}
3568
3569// MARK - Reduce Norm2 Layer
3570
3571typedef struct {
3572 ccv_cnnp_model_t super;
3573 int axis[CCV_NNC_MAX_DIM_ALLOC(12)];
3574 int count;
3575 ccv_nnc_tensor_symbol_t output;
3576} ccv_cnnp_model_reduce_norm2_t;
3577
3578static 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)
3579{
3580 const ccv_cnnp_model_reduce_norm2_t* const self = (const ccv_cnnp_model_reduce_norm2_t*)super;
3581 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);
3582 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"
, 3582, __extension__ __PRETTY_FUNCTION__); }));
3583 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 3583, __extension__ __PRETTY_FUNCTION__
); }));
3584 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3585 ccv_nnc_tensor_param_t output_params;
3586 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);
3587 int i;
3588 for (i = 0; i < self->count; i++)
3589 reduce_norm2.info.reduce.axis[i] = self->axis[i];
3590 reduce_norm2.info.reduce.count = self->count;
3591 ccv_nnc_hint_tensor_auto(reduce_norm2, &input_params, 1, ccv_nnc_no_hint, &output_params, 1);
3592 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3593 ccv_nnc_graph_exec_symbol_new(graph, reduce_norm2, inputs, input_size, outputs, output_size, "reduce_norm2");
3594}
3595
3596static ccv_cnnp_model_t* _ccv_cnnp_reduce_norm2_copy(const ccv_cnnp_model_t* const self, void* const context);
3597
3598static const ccv_cnnp_model_vtab_t ccv_cnnp_reduce_norm2_isa = {
3599 .build = _ccv_cnnp_reduce_norm2_build,
3600 .copy = _ccv_cnnp_reduce_norm2_copy,
3601};
3602
3603ccv_cnnp_model_t* ccv_cnnp_reduce_norm2(const int* const axis, const int axis_count, const char* const name)
3604{
3605 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));
3606 model_reduce_norm2->super.isa = &ccv_cnnp_reduce_norm2_isa;
3607 model_reduce_norm2->super.input_size = 1;
3608 model_reduce_norm2->super.outputs = &model_reduce_norm2->output;
3609 model_reduce_norm2->super.output_size = 1;
3610 ccv_cnnp_model_copy_name(&model_reduce_norm2->super, name);
3611 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", 3611, __extension__ __PRETTY_FUNCTION__
); }));
3612 int i;
3613 for (i = 0; i < axis_count; i++)
3614 model_reduce_norm2->axis[i] = axis[i];
3615 model_reduce_norm2->count = axis_count;
3616 return (ccv_cnnp_model_t*)model_reduce_norm2;
3617}
3618
3619static ccv_cnnp_model_t* _ccv_cnnp_reduce_norm2_copy(const ccv_cnnp_model_t* const super, void* const context)
3620{
3621 const ccv_cnnp_model_reduce_norm2_t* const self = (const ccv_cnnp_model_reduce_norm2_t*)super;
3622 return ccv_cnnp_reduce_norm2(self->axis, self->count, self->super.name);
3623}
3624
3625// MARK - Argmax Layer
3626
3627typedef struct {
3628 ccv_cnnp_model_t super;
3629 int axis;
3630 ccv_nnc_tensor_symbol_t output;
3631} ccv_cnnp_model_argmax_t;
3632
3633static 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)
3634{
3635 const ccv_cnnp_model_argmax_t* const self = (const ccv_cnnp_model_argmax_t*)super;
3636 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);
3637 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"
, 3637, __extension__ __PRETTY_FUNCTION__); }));
3638 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 3638, __extension__ __PRETTY_FUNCTION__
); }));
3639 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3640 ccv_nnc_tensor_param_t output_params;
3641 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);
3642 argmax.info.reduce.axis[0] = self->axis;
3643 argmax.info.reduce.count = 1;
3644 ccv_nnc_hint_tensor_auto(argmax, &input_params, 1, ccv_nnc_no_hint, &output_params, 1);
3645 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3646 ccv_nnc_graph_exec_symbol_new(graph, argmax, inputs, input_size, outputs, output_size, "argmax");
3647}
3648
3649static ccv_cnnp_model_t* _ccv_cnnp_argmax_copy(const ccv_cnnp_model_t* const self, void* const context);
3650
3651static const ccv_cnnp_model_vtab_t ccv_cnnp_argmax_isa = {
3652 .build = _ccv_cnnp_argmax_build,
3653 .copy = _ccv_cnnp_argmax_copy,
3654};
3655
3656ccv_cnnp_model_t* ccv_cnnp_argmax(const int axis, const char* const name)
3657{
3658 ccv_cnnp_model_argmax_t* const model_argmax = (ccv_cnnp_model_argmax_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_argmax_t));
3659 model_argmax->super.isa = &ccv_cnnp_argmax_isa;
3660 model_argmax->super.input_size = 1;
3661 model_argmax->super.outputs = &model_argmax->output;
3662 model_argmax->super.output_size = 1;
3663 ccv_cnnp_model_copy_name(&model_argmax->super, name);
3664 model_argmax->axis = axis;
3665 return (ccv_cnnp_model_t*)model_argmax;
3666}
3667
3668static ccv_cnnp_model_t* _ccv_cnnp_argmax_copy(const ccv_cnnp_model_t* const super, void* const context)
3669{
3670 const ccv_cnnp_model_argmax_t* const self = (const ccv_cnnp_model_argmax_t*)super;
3671 return ccv_cnnp_argmax(self->axis, self->super.name);
3672}
3673
3674// MARK - Argmin Layer
3675
3676typedef struct {
3677 ccv_cnnp_model_t super;
3678 int axis;
3679 ccv_nnc_tensor_symbol_t output;
3680} ccv_cnnp_model_argmin_t;
3681
3682static 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)
3683{
3684 const ccv_cnnp_model_argmin_t* const self = (const ccv_cnnp_model_argmin_t*)super;
3685 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);
3686 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"
, 3686, __extension__ __PRETTY_FUNCTION__); }));
3687 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 3687, __extension__ __PRETTY_FUNCTION__
); }));
3688 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3689 ccv_nnc_tensor_param_t output_params;
3690 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);
3691 argmin.info.reduce.axis[0] = self->axis;
3692 argmin.info.reduce.count = 1;
3693 ccv_nnc_hint_tensor_auto(argmin, &input_params, 1, ccv_nnc_no_hint, &output_params, 1);
3694 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3695 ccv_nnc_graph_exec_symbol_new(graph, argmin, inputs, input_size, outputs, output_size, "argmin");
3696}
3697
3698static ccv_cnnp_model_t* _ccv_cnnp_argmin_copy(const ccv_cnnp_model_t* const self, void* const context);
3699
3700static const ccv_cnnp_model_vtab_t ccv_cnnp_argmin_isa = {
3701 .build = _ccv_cnnp_argmin_build,
3702 .copy = _ccv_cnnp_argmin_copy,
3703};
3704
3705ccv_cnnp_model_t* ccv_cnnp_argmin(const int axis, const char* const name)
3706{
3707 ccv_cnnp_model_argmin_t* const model_argmin = (ccv_cnnp_model_argmin_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_argmin_t));
3708 model_argmin->super.isa = &ccv_cnnp_argmin_isa;
3709 model_argmin->super.input_size = 1;
3710 model_argmin->super.outputs = &model_argmin->output;
3711 model_argmin->super.output_size = 1;
3712 ccv_cnnp_model_copy_name(&model_argmin->super, name);
3713 model_argmin->axis = axis;
3714 return (ccv_cnnp_model_t*)model_argmin;
3715}
3716
3717static ccv_cnnp_model_t* _ccv_cnnp_argmin_copy(const ccv_cnnp_model_t* const super, void* const context)
3718{
3719 const ccv_cnnp_model_argmin_t* const self = (const ccv_cnnp_model_argmin_t*)super;
3720 return ccv_cnnp_argmin(self->axis, self->super.name);
3721}
3722
3723// MARK - Min Layer
3724
3725typedef struct {
3726 ccv_cnnp_model_t super;
3727 ccv_nnc_tensor_symbol_t output;
3728} ccv_cnnp_model_min_t;
3729
3730static 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)
3731{
3732 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);
3733 assert(input_size == 2)((void) sizeof ((input_size == 2) ? 1 : 0), __extension__ ({ if
(input_size == 2) ; else __assert_fail ("input_size == 2", "ccv_cnnp_model_addons.c"
, 3733, __extension__ __PRETTY_FUNCTION__); }));
3734 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 3734, __extension__ __PRETTY_FUNCTION__
); }));
3735 ccv_nnc_tensor_param_t input_params[2];
3736 int i;
3737 for (i = 0; i < 2; i++)
3738 input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
3739 ccv_nnc_tensor_param_t output_params;
3740 const ccv_nnc_cmd_t min = CMD_MIN_FORWARD()ccv_nnc_cmd(CCV_NNC_MIN_FORWARD, 0, (ccv_nnc_cmd_param_t){.size
={.dim={1,1,1}}}, 0);
3741 ccv_nnc_hint_tensor_auto(min, input_params, 2, ccv_nnc_no_hint, &output_params, 1);
3742 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3743 ccv_nnc_graph_exec_symbol_new(graph, min, inputs, input_size, outputs, output_size, "min");
3744}
3745
3746static ccv_cnnp_model_t* _ccv_cnnp_min_copy(const ccv_cnnp_model_t* const self, void* const context);
3747
3748static const ccv_cnnp_model_vtab_t ccv_cnnp_min_isa = {
3749 .build = _ccv_cnnp_min_build,
3750 .copy = _ccv_cnnp_min_copy,
3751};
3752
3753ccv_cnnp_model_t* ccv_cnnp_min(const char* const name)
3754{
3755 ccv_cnnp_model_min_t* const model_min = (ccv_cnnp_model_min_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_min_t));
3756 model_min->super.isa = &ccv_cnnp_min_isa;
3757 model_min->super.input_size = 2;
3758 model_min->super.outputs = &model_min->output;
3759 model_min->super.output_size = 1;
3760 ccv_cnnp_model_copy_name(&model_min->super, name);
3761 return (ccv_cnnp_model_t*)model_min;
3762}
3763
3764static ccv_cnnp_model_t* _ccv_cnnp_min_copy(const ccv_cnnp_model_t* const super, void* const context)
3765{
3766 const ccv_cnnp_model_min_t* const self = (const ccv_cnnp_model_min_t*)super;
3767 return ccv_cnnp_min(self->super.name);
3768}
3769
3770// MARK - Max Layer
3771
3772typedef struct {
3773 ccv_cnnp_model_t super;
3774 ccv_nnc_tensor_symbol_t output;
3775} ccv_cnnp_model_max_t;
3776
3777static 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)
3778{
3779 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);
3780 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"
, 3780, __extension__ __PRETTY_FUNCTION__); }));
3781 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 3781, __extension__ __PRETTY_FUNCTION__
); }));
3782 ccv_nnc_tensor_param_t input_params[2];
3783 int i;
3784 for (i = 0; i < 2; i++)
3785 input_params[i] = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
3786 ccv_nnc_tensor_param_t output_params;
3787 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);
3788 ccv_nnc_hint_tensor_auto(max, input_params, 2, ccv_nnc_no_hint, &output_params, 1);
3789 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
3790 ccv_nnc_graph_exec_symbol_new(graph, max, inputs, input_size, outputs, output_size, "max");
3791}
3792
3793static ccv_cnnp_model_t* _ccv_cnnp_max_copy(const ccv_cnnp_model_t* const self, void* const context);
3794
3795static const ccv_cnnp_model_vtab_t ccv_cnnp_max_isa = {
3796 .build = _ccv_cnnp_max_build,
3797 .copy = _ccv_cnnp_max_copy,
3798};
3799
3800ccv_cnnp_model_t* ccv_cnnp_max(const char* const name)
3801{
3802 ccv_cnnp_model_max_t* const model_max = (ccv_cnnp_model_max_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_max_t));
3803 model_max->super.isa = &ccv_cnnp_max_isa;
3804 model_max->super.input_size = 2;
3805 model_max->super.outputs = &model_max->output;
3806 model_max->super.output_size = 1;
3807 ccv_cnnp_model_copy_name(&model_max->super, name);
3808 return (ccv_cnnp_model_t*)model_max;
3809}
3810
3811static ccv_cnnp_model_t* _ccv_cnnp_max_copy(const ccv_cnnp_model_t* const super, void* const context)
3812{
3813 const ccv_cnnp_model_max_t* const self = (const ccv_cnnp_model_max_t*)super;
3814 return ccv_cnnp_max(self->super.name);
3815}
3816
3817// MARK - LSTM Layer
3818
3819typedef struct {
3820 ccv_cnnp_model_t super;
3821 int masked;
3822 ccv_nnc_tensor_symbol_t output;
3823 ccv_nnc_tensor_symbol_t weights;
3824 ccv_nnc_tensor_symbol_t reserves;
3825 ccv_nnc_cmd_param_t params;
3826 ccv_nnc_graph_exec_symbol_t lstm;
3827} ccv_cnnp_model_lstm_t;
3828
3829static int _ccv_cnnp_lstm_weight_dim(int bidirectional, int num_layers, int input_size, int hidden_size, int proj_size, int bias)
3830{
3831 const int D = !!bidirectional + 1;
3832 if (hidden_size == proj_size)
3833 return (num_layers * (bias ? 8 : 0) + (num_layers - 1) * (hidden_size * 4 * D + hidden_size * 4) + input_size * 4 + hidden_size * 4) * D;
3834 else
3835 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;
3836}
3837
3838static 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)
3839{
3840 ccv_cnnp_model_lstm_t* const self = (ccv_cnnp_model_lstm_t*)super;
3841 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);
3842 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", 3842, __extension__ __PRETTY_FUNCTION__
); }));
3843 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 3843, __extension__ __PRETTY_FUNCTION__
); }));
3844 const int proj_size = self->params.rnn.proj_size == 0 ? self->params.rnn.hidden_size : self->params.rnn.proj_size;
3845 ccv_nnc_tensor_param_t input_params[5];
3846 input_params[0]= ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3847 if (input_size == 2)
3848 input_params[1] = ccv_nnc_tensor_symbol_params(graph, inputs[1]);
3849 input_params[4] = input_params[0];
3850 memset(input_params[4].dim, 0, sizeof(input_params[4].dim));
3851 const int x_nd = ccv_nnc_tensor_nd(input_params[0].dim);
3852 const int feature_count = input_params[0].dim[x_nd - 1];
3853 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);
3854 input_params[4].dim[1] = self->params.rnn.hidden_size;
3855 const ccv_nnc_cmd_t lstm = ccv_nnc_cmd(CCV_NNC_LSTM_FORWARD, 0, self->params, 0);
3856 ccv_nnc_tensor_param_t output_params[4];
3857 ccv_nnc_hint_tensor_auto(lstm, input_params, 5, ccv_nnc_no_hint, output_params, 4);
3858 outputs[0] = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0);
3859 if (!self->weights.graph)
3860 self->weights = ccv_nnc_tensor_symbol_new(graph, input_params[4], "weights");
3861 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"
, 3861, __extension__ __PRETTY_FUNCTION__); }));
3862 const ccv_nnc_tensor_symbol_t weights = ccv_cnnp_model_get_symbol(super, self->weights);
3863 if (!self->reserves.graph)
3864 self->reserves = ccv_nnc_tensor_symbol_new(graph, output_params[3], "reserves");
3865 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"
, 3865, __extension__ __PRETTY_FUNCTION__); }));
3866 const ccv_nnc_tensor_symbol_t reserves = ccv_cnnp_model_get_symbol(super, self->reserves);
3867 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
};
3868 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");
3869}
3870
3871static 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)
3872{
3873 ccv_cnnp_model_lstm_t* const self = (ccv_cnnp_model_lstm_t*)super;
3874 if (self->weights.graph)
3875 {
3876 const float stdv = 1.0 / sqrt(self->params.rnn.hidden_size);
3877 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);
3878 }
3879}
3880
3881static 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)
3882{
3883 ccv_cnnp_model_lstm_t* const self = (ccv_cnnp_model_lstm_t*)super;
3884 if (self->weights.graph)
3885 add_to_array(parameters, self->weights, is_trainable);
3886}
3887
3888static 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)
3889{
3890 ccv_cnnp_model_lstm_t* const self = (ccv_cnnp_model_lstm_t*)super;
3891 if (self->lstm.graph)
3892 {
3893 self->params.rnn.is_test = is_test;
3894 updater(context, self->lstm, ccv_nnc_cmd(CCV_NNC_LSTM_FORWARD, 0, self->params, 0), ccv_nnc_no_hint);
3895 }
3896}
3897
3898static ccv_cnnp_model_t* _ccv_cnnp_lstm_copy(const ccv_cnnp_model_t* const self, void* const context);
3899
3900static const ccv_cnnp_model_vtab_t ccv_cnnp_lstm_isa = {
3901 .build = _ccv_cnnp_lstm_build,
3902 .init_states = _ccv_cnnp_lstm_init_states,
3903 .add_to_parameter = _ccv_cnnp_lstm_add_to_parameter,
3904 .copy = _ccv_cnnp_lstm_copy,
3905 .set_is_test = _ccv_cnnp_lstm_set_is_test,
3906};
3907
3908ccv_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)
3909{
3910 ccv_cnnp_model_lstm_t* const model_lstm = (ccv_cnnp_model_lstm_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_lstm_t));
3911 model_lstm->super.isa = &ccv_cnnp_lstm_isa;
3912 model_lstm->super.input_size = masked ? 2 : 1;
3913 model_lstm->super.outputs = &model_lstm->output;
3914 model_lstm->super.output_size = 1;
3915 model_lstm->super.is_trainable = is_trainable;
3916 ccv_cnnp_model_copy_name(&model_lstm->super, name);
3917 model_lstm->masked = masked;
3918 model_lstm->weights.d = CCV_NNC_NO_TENSOR_SYMBOL;
3919 model_lstm->weights.graph = 0;
3920 model_lstm->params.rnn.hidden_size = hidden_size;
3921 model_lstm->params.rnn.proj_size = proj_size;
3922 model_lstm->params.rnn.num_layers = num_layers;
3923 model_lstm->params.rnn.bias = bias;
3924 model_lstm->params.rnn.batch_first = batch_first;
3925 model_lstm->params.rnn.bidirectional = bidirectional;
3926 model_lstm->params.rnn.dropout = dropout;
3927 return (ccv_cnnp_model_t*)model_lstm;
3928}
3929
3930static ccv_cnnp_model_t* _ccv_cnnp_lstm_copy(const ccv_cnnp_model_t* const super, void* const context)
3931{
3932 const ccv_cnnp_model_lstm_t* const self = (const ccv_cnnp_model_lstm_t*)super;
3933 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);
3934}
3935
3936/// MARK - Datatype conversion layer.
3937
3938typedef struct {
3939 ccv_cnnp_model_t super;
3940 ccv_nnc_tensor_symbol_t output;
3941 int datatype;
3942 int ref_to_last;
3943} ccv_cnnp_model_datatype_conversion_t;
3944
3945static 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)
3946{
3947 ccv_cnnp_model_datatype_conversion_t* const self = (ccv_cnnp_model_datatype_conversion_t*)super;
3948 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);
3949 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
3950 if (self->ref_to_last)
3951 {
3952 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", 3952, __extension__ __PRETTY_FUNCTION__
); }));
3953 const ccv_nnc_tensor_param_t last_params = ccv_nnc_tensor_symbol_params(graph, inputs[input_size - 1]);
3954 params.datatype = last_params.datatype;
3955 } else
3956 params.datatype = self->datatype;
3957 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 3957, __extension__ __PRETTY_FUNCTION__
); }));
3958 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
3959 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);
3960}
3961
3962static ccv_cnnp_model_t* _ccv_cnnp_datatype_conversion_copy(const ccv_cnnp_model_t* const self, void* const context);
3963
3964static const ccv_cnnp_model_vtab_t ccv_cnnp_datatype_conversion_isa = {
3965 .build = _ccv_cnnp_datatype_conversion_build,
3966 .copy = _ccv_cnnp_datatype_conversion_copy,
3967};
3968
3969ccv_cnnp_model_t* ccv_cnnp_datatype_conversion(const int datatype, const int ref_to_last, const char* const name)
3970{
3971 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));
3972 model_datatype_conversion->super.isa = &ccv_cnnp_datatype_conversion_isa;
3973 model_datatype_conversion->super.input_size = 0;
3974 model_datatype_conversion->super.outputs = &model_datatype_conversion->output;
3975 model_datatype_conversion->super.output_size = 1;
3976 model_datatype_conversion->datatype = datatype;
3977 model_datatype_conversion->ref_to_last = ref_to_last;
3978 ccv_cnnp_model_copy_name(&model_datatype_conversion->super, name);
3979 return (ccv_cnnp_model_t*)model_datatype_conversion;
3980}
3981
3982static ccv_cnnp_model_t* _ccv_cnnp_datatype_conversion_copy(const ccv_cnnp_model_t* const super, void* const context)
3983{
3984 ccv_cnnp_model_datatype_conversion_t* const self = (ccv_cnnp_model_datatype_conversion_t*)super;
3985 return ccv_cnnp_datatype_conversion(self->datatype, self->ref_to_last, self->super.name);
3986}
3987
3988/// MARK - Clamp layer.
3989
3990typedef struct {
3991 ccv_cnnp_model_t super;
3992 ccv_nnc_tensor_symbol_t output;
3993 float min;
3994 float max;
3995} ccv_cnnp_model_clamp_t;
3996
3997static 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)
3998{
3999 ccv_cnnp_model_clamp_t* const self = (ccv_cnnp_model_clamp_t*)super;
4000 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);
4001 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4002 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 4002, __extension__ __PRETTY_FUNCTION__
); }));
4003 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4004 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);
4005}
4006
4007static ccv_cnnp_model_t* _ccv_cnnp_clamp_copy(const ccv_cnnp_model_t* const self, void* const context);
4008
4009static const ccv_cnnp_model_vtab_t ccv_cnnp_clamp_isa = {
4010 .build = _ccv_cnnp_clamp_build,
4011 .copy = _ccv_cnnp_clamp_copy,
4012};
4013
4014ccv_cnnp_model_t* ccv_cnnp_clamp(const float min, const float max, const char* const name)
4015{
4016 ccv_cnnp_model_clamp_t* const model_clamp = (ccv_cnnp_model_clamp_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_clamp_t));
4017 model_clamp->super.isa = &ccv_cnnp_clamp_isa;
4018 model_clamp->super.input_size = 0;
4019 model_clamp->super.outputs = &model_clamp->output;
4020 model_clamp->super.output_size = 1;
4021 model_clamp->min = min;
4022 model_clamp->max = max;
4023 ccv_cnnp_model_copy_name(&model_clamp->super, name);
4024 return (ccv_cnnp_model_t*)model_clamp;
4025}
4026
4027static ccv_cnnp_model_t* _ccv_cnnp_clamp_copy(const ccv_cnnp_model_t* const super, void* const context)
4028{
4029 ccv_cnnp_model_clamp_t* const self = (ccv_cnnp_model_clamp_t*)super;
4030 return ccv_cnnp_clamp(self->min, self->max, self->super.name);
4031}
4032
4033// MARK - Parameter Layer
4034
4035typedef struct {
4036 ccv_cnnp_model_t super;
4037 float init_bound;
4038 ccv_nnc_tensor_symbol_t weights;
4039 ccv_nnc_tensor_param_t weights_params;
4040 ccv_nnc_tensor_symbol_t output;
4041} ccv_cnnp_model_parameter_t;
4042
4043static 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)
4044{
4045 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);
4046 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 4046, __extension__ __PRETTY_FUNCTION__
); }));
4047 ccv_cnnp_model_parameter_t* const self = (ccv_cnnp_model_parameter_t*)super;
4048 if (!self->weights.graph)
4049 self->weights = ccv_nnc_tensor_symbol_new(graph, self->weights_params, "weights");
4050 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"
, 4050, __extension__ __PRETTY_FUNCTION__); }));
4051 outputs[0] = ccv_cnnp_model_get_symbol(super, self->weights);
4052}
4053
4054static 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)
4055{
4056 ccv_cnnp_model_parameter_t* const self = (ccv_cnnp_model_parameter_t*)super;
4057 if (self->init_bound > 0)
4058 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);
4059 else
4060 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);
4061}
4062
4063static 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)
4064{
4065 ccv_cnnp_model_parameter_t* const self = (ccv_cnnp_model_parameter_t*)super;
4066 add_to_array(parameters, self->weights, is_trainable);
4067}
4068
4069static ccv_cnnp_model_t* _ccv_cnnp_parameter_copy(const ccv_cnnp_model_t* const super, void* const context);
4070
4071static const ccv_cnnp_model_vtab_t ccv_cnnp_parameter_isa = {
4072 .build = _ccv_cnnp_parameter_build,
4073 .init_states = _ccv_cnnp_parameter_init_states,
4074 .add_to_parameter = _ccv_cnnp_parameter_add_to_parameter,
4075 .copy = _ccv_cnnp_parameter_copy,
4076};
4077
4078ccv_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)
4079{
4080 ccv_cnnp_model_parameter_t* const model_parameter = (ccv_cnnp_model_parameter_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_parameter_t));
4081 model_parameter->super.isa = &ccv_cnnp_parameter_isa;
4082 model_parameter->super.input_size = 0;
4083 model_parameter->super.outputs = &model_parameter->output;
4084 model_parameter->super.output_size = 1;
4085 model_parameter->super.is_trainable = is_trainable;
4086 ccv_cnnp_model_copy_name(&model_parameter->super, name);
4087 model_parameter->weights.d = CCV_NNC_NO_TENSOR_SYMBOL;
4088 model_parameter->weights.graph = 0;
4089 model_parameter->weights_params = params;
4090 return (ccv_cnnp_model_t*)model_parameter;
4091}
4092
4093static ccv_cnnp_model_t* _ccv_cnnp_parameter_copy(const ccv_cnnp_model_t* const super, void* const context)
4094{
4095 const ccv_cnnp_model_parameter_t* const self = (const ccv_cnnp_model_parameter_t*)super;
4096 return ccv_cnnp_parameter(self->weights_params, self->init_bound, self->super.is_trainable, self->super.name);
4097}
4098
4099// MARK - Scalar Layer
4100
4101typedef struct {
4102 ccv_cnnp_model_t super;
4103 int type;
4104 int format;
4105 int datatype;
4106 float value;
4107 ccv_nnc_tensor_symbol_t output;
4108} ccv_cnnp_model_scalar_t;
4109
4110static 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)
4111{
4112 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);
4113 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 4113, __extension__ __PRETTY_FUNCTION__
); }));
4114 ccv_cnnp_model_scalar_t* const self = (ccv_cnnp_model_scalar_t*)super;
4115 ccv_nnc_tensor_param_t params = {
4116 .type = self->type,
4117 .format = self->format,
4118 .datatype = self->datatype,
4119 .dim = {
4120 1
4121 }
4122 };
4123 if (input_size > 0)
4124 {
4125 ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4126 params.type = input_params.type;
4127 params.format = input_params.format;
4128 params.datatype = input_params.datatype;
4129 }
4130 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4131 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);
4132}
4133
4134static ccv_cnnp_model_t* _ccv_cnnp_scalar_copy(const ccv_cnnp_model_t* const super, void* const context);
4135
4136static const ccv_cnnp_model_vtab_t ccv_cnnp_scalar_isa = {
4137 .build = _ccv_cnnp_scalar_build,
4138 .copy = _ccv_cnnp_scalar_copy,
4139};
4140
4141ccv_cnnp_model_t* ccv_cnnp_scalar(const int type, const int format, const int datatype, const float value, const char* const name)
4142{
4143 ccv_cnnp_model_scalar_t* const model_scalar = (ccv_cnnp_model_scalar_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_scalar_t));
4144 model_scalar->super.isa = &ccv_cnnp_scalar_isa;
4145 model_scalar->super.input_size = 0;
4146 model_scalar->super.outputs = &model_scalar->output;
4147 model_scalar->super.output_size = 1;
4148 ccv_cnnp_model_copy_name(&model_scalar->super, name);
4149 model_scalar->type = type;
4150 model_scalar->format = format;
4151 model_scalar->datatype = datatype;
4152 model_scalar->value = value;
4153 return (ccv_cnnp_model_t*)model_scalar;
4154}
4155
4156static ccv_cnnp_model_t* _ccv_cnnp_scalar_copy(const ccv_cnnp_model_t* const super, void* const context)
4157{
4158 const ccv_cnnp_model_scalar_t* const self = (const ccv_cnnp_model_scalar_t*)super;
4159 return ccv_cnnp_scalar(self->type, self->format, self->datatype, self->value, self->super.name);
4160}
4161
4162// MARK - Variable Layer
4163
4164typedef struct {
4165 ccv_cnnp_model_t super;
4166 ccv_nnc_tensor_param_t params;
4167 ccv_nnc_tensor_symbol_t output;
4168} ccv_cnnp_model_variable_t;
4169
4170static 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)
4171{
4172 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);
4173 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"
, 4173, __extension__ __PRETTY_FUNCTION__); }));
4174 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 4174, __extension__ __PRETTY_FUNCTION__
); }));
4175 ccv_cnnp_model_variable_t* const self = (ccv_cnnp_model_variable_t*)super;
4176 outputs[0] = ccv_nnc_tensor_symbol_new(graph, self->params, 0);
4177}
4178
4179static ccv_cnnp_model_t* _ccv_cnnp_variable_copy(const ccv_cnnp_model_t* const super, void* const context);
4180
4181static const ccv_cnnp_model_vtab_t ccv_cnnp_variable_isa = {
4182 .build = _ccv_cnnp_variable_build,
4183 .copy = _ccv_cnnp_variable_copy,
4184};
4185
4186ccv_cnnp_model_t* ccv_cnnp_variable(const ccv_nnc_tensor_param_t params, const char* const name)
4187{
4188 ccv_cnnp_model_variable_t* const model_variable = (ccv_cnnp_model_variable_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_variable_t));
4189 model_variable->super.isa = &ccv_cnnp_variable_isa;
4190 model_variable->super.input_size = 0;
4191 model_variable->super.outputs = &model_variable->output;
4192 model_variable->super.output_size = 1;
4193 ccv_cnnp_model_copy_name(&model_variable->super, name);
4194 model_variable->params = params;
4195 return (ccv_cnnp_model_t*)model_variable;
4196}
4197
4198static ccv_cnnp_model_t* _ccv_cnnp_variable_copy(const ccv_cnnp_model_t* const super, void* const context)
4199{
4200 const ccv_cnnp_model_variable_t* const self = (const ccv_cnnp_model_variable_t*)super;
4201 return ccv_cnnp_variable(self->params, self->super.name);
4202}
4203
4204// MARK - Send Layer
4205
4206typedef struct {
4207 ccv_cnnp_model_t super;
4208 int device_id;
4209 ccv_nnc_tensor_symbol_t output;
4210} ccv_cnnp_model_send_t;
4211
4212static 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)
4213{
4214 ccv_cnnp_model_send_t* const self = (ccv_cnnp_model_send_t*)super;
4215 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);
4216 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"
, 4216, __extension__ __PRETTY_FUNCTION__); }));
4217 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 4217, __extension__ __PRETTY_FUNCTION__
); }));
4218 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4219 params.type = (params.type & ~0x3) | CCV_TENSOR_GPU_MEMORY;
4220 CCV_TENSOR_SET_DEVICE_ID(params.type, self->device_id)(params.type) = (((params.type) & ~0xfff00) | (((self->
device_id) & 0xfff) << 8));
4221 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4222 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");
4223}
4224
4225static ccv_cnnp_model_t* _ccv_cnnp_send_copy(const ccv_cnnp_model_t* const super, void* const context);
4226
4227static const ccv_cnnp_model_vtab_t ccv_cnnp_send_isa = {
4228 .build = _ccv_cnnp_send_build,
4229 .copy = _ccv_cnnp_send_copy,
4230};
4231
4232ccv_cnnp_model_t* ccv_cnnp_send(const int device_id, const char* const name)
4233{
4234 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", 4234, __extension__ __PRETTY_FUNCTION__
); }));
4235 ccv_cnnp_model_send_t* const model_send = (ccv_cnnp_model_send_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_send_t));
4236 model_send->super.isa = &ccv_cnnp_send_isa;
4237 model_send->super.input_size = 1;
4238 model_send->super.outputs = &model_send->output;
4239 model_send->super.output_size = 1;
4240 ccv_cnnp_model_copy_name(&model_send->super, name);
4241 model_send->device_id = device_id;
4242 return (ccv_cnnp_model_t*)model_send;
4243}
4244
4245static ccv_cnnp_model_t* _ccv_cnnp_send_copy(const ccv_cnnp_model_t* const super, void* const context)
4246{
4247 const ccv_cnnp_model_send_t* const self = (const ccv_cnnp_model_send_t*)super;
4248 return ccv_cnnp_send(self->device_id, self->super.name);
4249}
4250
4251// MARK - Move Layer
4252
4253typedef struct {
4254 ccv_cnnp_model_t super;
4255 ccv_nnc_tensor_symbol_t output;
4256} ccv_cnnp_model_move_t;
4257
4258static 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)
4259{
4260 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);
4261 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"
, 4261, __extension__ __PRETTY_FUNCTION__); }));
4262 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 4262, __extension__ __PRETTY_FUNCTION__
); }));
4263 outputs[0] = inputs[1];
4264 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");
4265}
4266
4267static ccv_cnnp_model_t* _ccv_cnnp_move_copy(const ccv_cnnp_model_t* const super, void* const context);
4268
4269static const ccv_cnnp_model_vtab_t ccv_cnnp_move_isa = {
4270 .build = _ccv_cnnp_move_build,
4271 .copy = _ccv_cnnp_move_copy,
4272};
4273
4274ccv_cnnp_model_t* ccv_cnnp_move(const char* const name)
4275{
4276 ccv_cnnp_model_move_t* const model_move = (ccv_cnnp_model_move_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_move_t));
4277 model_move->super.isa = &ccv_cnnp_move_isa;
4278 model_move->super.input_size = 2;
4279 model_move->super.outputs = &model_move->output;
4280 model_move->super.output_size = 1;
4281 ccv_cnnp_model_copy_name(&model_move->super, name);
4282 return (ccv_cnnp_model_t*)model_move;
4283}
4284
4285static ccv_cnnp_model_t* _ccv_cnnp_move_copy(const ccv_cnnp_model_t* const super, void* const context)
4286{
4287 const ccv_cnnp_model_move_t* const self = (const ccv_cnnp_model_move_t*)super;
4288 return ccv_cnnp_move(self->super.name);
4289}
4290
4291// MARK - "Making" Contiguous Layer
4292
4293typedef struct {
4294 ccv_cnnp_model_t super;
4295 ccv_nnc_tensor_symbol_t output;
4296} ccv_cnnp_model_contiguous_t;
4297
4298static 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)
4299{
4300 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);
4301 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"
, 4301, __extension__ __PRETTY_FUNCTION__); }));
4302 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 4302, __extension__ __PRETTY_FUNCTION__
); }));
4303 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4304 ccv_nnc_tensor_symbol_t to = ccv_nnc_tensor_symbol_alias_to(graph, inputs[0]);
4305 if (to.d == CCV_NNC_NO_TENSOR_SYMBOL) // If we are not reshape an alias, it is straightforward.
4306 {
4307 outputs[0] = inputs[0];
4308 return;
4309 }
4310 // Otherwise, we need to check its stride to know if it is contiguous.
4311 int old_stride[CCV_NNC_MAX_DIM_ALLOC(12)];
4312 ccv_nnc_tensor_symbol_alias_params(graph, inputs[0], 0, old_stride);
4313 // We identify permute by checking if the stride is not in descending order.
4314 // This also covered "permute" through reshape, rather than using ccv_cnnp_permute directly.
4315 if (ccv_nnc_is_tensor_stride_packed(old_stride, params.dim))
4316 {
4317 outputs[0] = inputs[0];
4318 return;
4319 }
4320 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4321 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");
4322 ccv_nnc_graph_exec_symbol_set_flags(graph, make_contiguous, CCV_NNC_GRAPH_EXEC_DISABLE_OPT);
4323}
4324
4325static ccv_cnnp_model_t* _ccv_cnnp_contiguous_copy(const ccv_cnnp_model_t* const super, void* const context);
4326
4327static const ccv_cnnp_model_vtab_t ccv_cnnp_contiguous_isa = {
4328 .build = _ccv_cnnp_contiguous_build,
4329 .copy = _ccv_cnnp_contiguous_copy,
4330};
4331
4332ccv_cnnp_model_t* ccv_cnnp_contiguous(const char* const name)
4333{
4334 ccv_cnnp_model_contiguous_t* const model_contiguous = (ccv_cnnp_model_contiguous_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_contiguous_t));
4335 model_contiguous->super.isa = &ccv_cnnp_contiguous_isa;
4336 model_contiguous->super.input_size = 1;
4337 model_contiguous->super.outputs = &model_contiguous->output;
4338 model_contiguous->super.output_size = 1;
4339 ccv_cnnp_model_copy_name(&model_contiguous->super, name);
4340 return (ccv_cnnp_model_t*)model_contiguous;
4341}
4342
4343static ccv_cnnp_model_t* _ccv_cnnp_contiguous_copy(const ccv_cnnp_model_t* const super, void* const context)
4344{
4345 const ccv_cnnp_model_contiguous_t* const self = (const ccv_cnnp_model_contiguous_t*)super;
4346 return ccv_cnnp_contiguous(self->super.name);
4347}
4348
4349// MARK - "Making" Copy Layer
4350
4351typedef struct {
4352 ccv_cnnp_model_t super;
4353 ccv_nnc_tensor_symbol_t output;
4354} ccv_cnnp_model_copy_t;
4355
4356static 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)
4357{
4358 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);
4359 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"
, 4359, __extension__ __PRETTY_FUNCTION__); }));
4360 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 4360, __extension__ __PRETTY_FUNCTION__
); }));
4361 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4362 ccv_nnc_tensor_symbol_t to = ccv_nnc_tensor_symbol_alias_to(graph, inputs[0]);
4363 if (to.d == CCV_NNC_NO_TENSOR_SYMBOL) // If we are not reshape an alias, it is straightforward.
4364 {
4365 outputs[0] = inputs[0];
4366 return;
4367 }
4368 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4369 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");
4370 ccv_nnc_graph_exec_symbol_set_flags(graph, make_contiguous, CCV_NNC_GRAPH_EXEC_DISABLE_OPT);
4371}
4372
4373static ccv_cnnp_model_t* _ccv_cnnp_copy_copy(const ccv_cnnp_model_t* const super, void* const context);
4374
4375static const ccv_cnnp_model_vtab_t ccv_cnnp_copy_isa = {
4376 .build = _ccv_cnnp_copy_build,
4377 .copy = _ccv_cnnp_copy_copy,
4378};
4379
4380ccv_cnnp_model_t* ccv_cnnp_copy(const char* const name)
4381{
4382 ccv_cnnp_model_copy_t* const model_copy = (ccv_cnnp_model_copy_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_copy_t));
4383 model_copy->super.isa = &ccv_cnnp_copy_isa;
4384 model_copy->super.input_size = 1;
4385 model_copy->super.outputs = &model_copy->output;
4386 model_copy->super.output_size = 1;
4387 ccv_cnnp_model_copy_name(&model_copy->super, name);
4388 return (ccv_cnnp_model_t*)model_copy;
4389}
4390
4391static ccv_cnnp_model_t* _ccv_cnnp_copy_copy(const ccv_cnnp_model_t* const super, void* const context)
4392{
4393 const ccv_cnnp_model_copy_t* const self = (const ccv_cnnp_model_copy_t*)super;
4394 return ccv_cnnp_copy(self->super.name);
4395}
4396
4397// MARK - All-To-All Layer
4398
4399typedef struct {
4400 ccv_cnnp_model_t super;
4401 int axis;
4402 ccv_nnc_tensor_symbol_t outputs[1];
4403} ccv_cnnp_model_all_to_all_t;
4404
4405static 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)
4406{
4407 ccv_cnnp_model_all_to_all_t* const self = (ccv_cnnp_model_all_to_all_t*)super;
4408 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);
4409 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", 4409, __extension__ __PRETTY_FUNCTION__
); }));
4410 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", 4410, __extension__ __PRETTY_FUNCTION__
); }));
4411 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", 4411, __extension__ __PRETTY_FUNCTION__
); }));
4412 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", 4412, __extension__ __PRETTY_FUNCTION__
); }));
4413 const ccv_nnc_tensor_param_t input_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4414 const int nd = ccv_nnc_tensor_nd(input_params.dim);
4415 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", 4415, __extension__ __PRETTY_FUNCTION__
); }));
4416 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", 4416, __extension__ __PRETTY_FUNCTION__
); }));
4417 int i;
4418 for (i = 0; i < input_size; i++)
4419 {
4420 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[i]);
4421 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", 4421, __extension__ __PRETTY_FUNCTION__
); }));
4422 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", 4422, __extension__ __PRETTY_FUNCTION__
); }));
4423 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", 4423, __extension__ __PRETTY_FUNCTION__
); }));
4424 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", 4424, __extension__ __PRETTY_FUNCTION__
); }));
4425 outputs[i] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4426 }
4427 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");
4428}
4429
4430static ccv_cnnp_model_t* _ccv_cnnp_all_to_all_copy(const ccv_cnnp_model_t* const super, void* const context);
4431
4432static const ccv_cnnp_model_vtab_t ccv_cnnp_all_to_all_isa = {
4433 .build = _ccv_cnnp_all_to_all_build,
4434 .copy = _ccv_cnnp_all_to_all_copy,
4435};
4436
4437ccv_cnnp_model_t* ccv_cnnp_all_to_all(const int count, const int axis, const char* const name)
4438{
4439 assert(count > 0)((void) sizeof ((count > 0) ? 1 : 0), __extension__ ({ if (
count > 0) ; else __assert_fail ("count > 0", "ccv_cnnp_model_addons.c"
, 4439, __extension__ __PRETTY_FUNCTION__); }));
4440 assert(axis >= 0)((void) sizeof ((axis >= 0) ? 1 : 0), __extension__ ({ if (
axis >= 0) ; else __assert_fail ("axis >= 0", "ccv_cnnp_model_addons.c"
, 4440, __extension__ __PRETTY_FUNCTION__); }));
4441 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));
4442 model_all_to_all->super.isa = &ccv_cnnp_all_to_all_isa;
4443 model_all_to_all->super.input_size = count;
4444 model_all_to_all->super.outputs = model_all_to_all->outputs;
4445 model_all_to_all->super.output_size = count;
4446 model_all_to_all->axis = axis;
4447 ccv_cnnp_model_copy_name(&model_all_to_all->super, name);
4448 return (ccv_cnnp_model_t*)model_all_to_all;
4449}
4450
4451static ccv_cnnp_model_t* _ccv_cnnp_all_to_all_copy(const ccv_cnnp_model_t* const super, void* const context)
4452{
4453 const ccv_cnnp_model_all_to_all_t* const self = (const ccv_cnnp_model_all_to_all_t*)super;
4454 return ccv_cnnp_all_to_all(self->super.output_size, self->axis, self->super.name);
4455}
4456
4457// MARK - Scaled-Dot Product Attention Layer
4458
4459typedef struct {
4460 ccv_cnnp_model_t super;
4461 ccv_nnc_tensor_symbol_t output;
4462 ccv_nnc_tensor_symbol_t weights;
4463 ccv_nnc_tensor_symbol_t bias;
4464 float scale;
4465 int is_causal;
4466 int has_attn_mask;
4467 int is_varlen;
4468 int max_seqlen_q;
4469 int max_seqlen_kv;
4470 int flags;
4471 int fused_unify_head_weights;
4472 int no_bias;
4473} ccv_cnnp_model_scaled_dot_product_attention_t;
4474
4475static 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)
4476{
4477 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);
4478 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 4478, __extension__ __PRETTY_FUNCTION__
); }));
4479 ccv_cnnp_model_scaled_dot_product_attention_t* const self = (ccv_cnnp_model_scaled_dot_product_attention_t*)super;
4480 assert(input_size == (self->is_varlen ? 5 : (self->has_attn_mask ? 4 : 3)))((void) sizeof ((input_size == (self->is_varlen ? 5 : (self
->has_attn_mask ? 4 : 3))) ? 1 : 0), __extension__ ({ if (
input_size == (self->is_varlen ? 5 : (self->has_attn_mask
? 4 : 3))) ; else __assert_fail ("input_size == (self->is_varlen ? 5 : (self->has_attn_mask ? 4 : 3))"
, "ccv_cnnp_model_addons.c", 4480, __extension__ __PRETTY_FUNCTION__
); }));
4481 assert(!self->is_varlen || !self->has_attn_mask)((void) sizeof ((!self->is_varlen || !self->has_attn_mask
) ? 1 : 0), __extension__ ({ if (!self->is_varlen || !self
->has_attn_mask) ; else __assert_fail ("!self->is_varlen || !self->has_attn_mask"
, "ccv_cnnp_model_addons.c", 4481, __extension__ __PRETTY_FUNCTION__
); }));
4482 assert(!self->is_varlen || !self->fused_unify_head_weights)((void) sizeof ((!self->is_varlen || !self->fused_unify_head_weights
) ? 1 : 0), __extension__ ({ if (!self->is_varlen || !self
->fused_unify_head_weights) ; else __assert_fail ("!self->is_varlen || !self->fused_unify_head_weights"
, "ccv_cnnp_model_addons.c", 4482, __extension__ __PRETTY_FUNCTION__
); }));
4483 const ccv_nnc_tensor_param_t q_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4484 const ccv_nnc_tensor_param_t k_params = ccv_nnc_tensor_symbol_params(graph, inputs[1]);
4485 const ccv_nnc_tensor_param_t v_params = ccv_nnc_tensor_symbol_params(graph, inputs[2]);
4486 const ccv_nnc_tensor_param_t q_seq_offsets_params = self->is_varlen ? ccv_nnc_tensor_symbol_params(graph, inputs[3]) : (ccv_nnc_tensor_param_t){};
4487 const ccv_nnc_tensor_param_t kv_seq_offsets_params = self->is_varlen ? ccv_nnc_tensor_symbol_params(graph, inputs[4]) : (ccv_nnc_tensor_param_t){};
4488 const int v_nd = ccv_nnc_tensor_nd(v_params.dim);
4489 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", 4489, __extension__ __PRETTY_FUNCTION__
); }));
4490 const int hEv = (v_nd == 3 ? 1 : v_params.dim[2]) * v_params.dim[v_nd - 1];
4491 ccv_nnc_tensor_param_t weights_params = q_params;
4492 memset(weights_params.dim, 0, sizeof(weights_params.dim));
4493 weights_params.dim[0] = hEv;
4494 weights_params.dim[1] = hEv;
4495 ccv_nnc_tensor_param_t bias_params = q_params;
4496 memset(bias_params.dim, 0, sizeof(bias_params.dim));
4497 bias_params.dim[0] = hEv;
4498 ccv_nnc_cmd_t cmd = {0};
4499 cmd.cmd = CCV_NNC_SCALED_DOT_PRODUCT_ATTENTION_FORWARD;
4500 cmd.info.scaled_dot_product_attention.scale = self->scale;
4501 cmd.info.scaled_dot_product_attention.is_causal = self->is_causal;
4502 cmd.info.scaled_dot_product_attention.is_varlen = self->is_varlen;
4503 cmd.info.scaled_dot_product_attention.max_seqlen_q = self->max_seqlen_q;
4504 cmd.info.scaled_dot_product_attention.max_seqlen_kv = self->max_seqlen_kv;
4505 cmd.info.scaled_dot_product_attention.flags = self->flags;
4506 ccv_nnc_tensor_param_t output_params[3];
4507 ccv_nnc_tensor_symbol_t output;
4508 ccv_nnc_tensor_symbol_t saved_softmax_lse;
4509 ccv_nnc_tensor_symbol_t saved_v_proj = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
4510 ccv_nnc_tensor_symbol_t attn_mask = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
4511 ccv_nnc_tensor_symbol_t weights = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
4512 ccv_nnc_tensor_symbol_t bias = NO_TENSOR_SYMBOL(const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
};
4513 if (self->has_attn_mask)
4514 attn_mask = inputs[3];
4515 if (self->is_varlen)
4516 {
4517 ccv_nnc_hint_tensor_auto(cmd, (ccv_nnc_tensor_param_t []){
4518 q_params,
4519 k_params,
4520 v_params,
4521 (ccv_nnc_tensor_param_t){},
4522 (ccv_nnc_tensor_param_t){},
4523 (ccv_nnc_tensor_param_t){},
4524 q_seq_offsets_params,
4525 kv_seq_offsets_params,
4526 }, 8, ccv_nnc_no_hint, output_params, 2);
4527 output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0);
4528 saved_softmax_lse = ccv_nnc_tensor_symbol_new(graph, output_params[1], 0);
4529 ccv_nnc_graph_exec_symbol_new(graph, cmd, TENSOR_SYMBOL_LIST(inputs[0], inputs[1], inputs[2], NO_TENSOR_SYMBOL, NO_TENSOR_SYMBOL, NO_TENSOR_SYMBOL, inputs[3], inputs[4])(const ccv_nnc_tensor_symbol_t []){inputs[0], inputs[1], inputs
[2], (const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
}, (const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
}, (const ccv_nnc_tensor_symbol_t){.d = CCV_NNC_NO_TENSOR_SYMBOL
}, inputs[3], inputs[4]}, (1 +1 +1 +1 +1 +1 +1 +1 +1 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 -1), TENSOR_SYMBOL_LIST(output, saved_softmax_lse, saved_v_proj)(const ccv_nnc_tensor_symbol_t []){output, saved_softmax_lse,
saved_v_proj}, (1 +1 +1 +1 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0
+0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +
0 +0 +0 +0 +0 +0 +0 +0 -1), "scaled_dot_product_attention");
4530 } else if (self->fused_unify_head_weights)
4531 {
4532 if (!self->weights.graph)
4533 self->weights = ccv_nnc_tensor_symbol_new(graph, weights_params, "weights");
4534 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"
, 4534, __extension__ __PRETTY_FUNCTION__); }));
4535 weights = ccv_cnnp_model_get_symbol(super, self->weights);
4536 if (!self->no_bias)
4537 {
4538 if (!self->bias.graph)
4539 self->bias = ccv_nnc_tensor_symbol_new(graph, bias_params, "bias");
4540 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", 4540, __extension__ __PRETTY_FUNCTION__
); }));
4541 bias = ccv_cnnp_model_get_symbol(super, self->bias);
4542 }
4543 ccv_nnc_hint_tensor_auto(cmd, (ccv_nnc_tensor_param_t []){
4544 q_params,
4545 k_params,
4546 v_params,
4547 (ccv_nnc_tensor_param_t){},
4548 weights_params,
4549 bias_params,
4550 }, 6, ccv_nnc_no_hint, output_params, 3);
4551 output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0);
4552 saved_softmax_lse = ccv_nnc_tensor_symbol_new(graph, output_params[1], 0);
4553 saved_v_proj = ccv_nnc_tensor_symbol_new(graph, output_params[2], 0);
4554 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");
4555 } else {
4556 ccv_nnc_hint_tensor_auto(cmd, (ccv_nnc_tensor_param_t []){
4557 q_params,
4558 k_params,
4559 v_params,
4560 }, 3, ccv_nnc_no_hint, output_params, 2);
4561 output = ccv_nnc_tensor_symbol_new(graph, output_params[0], 0);
4562 saved_softmax_lse = ccv_nnc_tensor_symbol_new(graph, output_params[1], 0);
4563 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");
4564 }
4565 outputs[0] = output;
4566}
4567
4568static 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)
4569{
4570 ccv_cnnp_model_scaled_dot_product_attention_t* const self = (ccv_cnnp_model_scaled_dot_product_attention_t*)super;
4571 if (self->weights.graph)
4572 {
4573 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"
, 4573, __extension__ __PRETTY_FUNCTION__); }));
4574 const ccv_nnc_tensor_param_t weight_params = ccv_nnc_tensor_symbol_params(graph, self->weights);
4575 const int c = weight_params.dim[1];
4576 const float std = sqrtf(2) / sqrtf(c);
4577 const float bound = sqrtf(3) * std;
4578 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);
4579 if (self->bias.graph)
4580 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);
4581 }
4582}
4583
4584static 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)
4585{
4586 ccv_cnnp_model_scaled_dot_product_attention_t* const self = (ccv_cnnp_model_scaled_dot_product_attention_t*)super;
4587 if (self->weights.graph)
4588 {
4589 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"
, 4589, __extension__ __PRETTY_FUNCTION__); }));
4590 add_to_array(parameters, self->weights, is_trainable);
4591 if (self->bias.graph)
4592 add_to_array(parameters, self->bias, is_trainable);
4593 }
4594}
4595
4596static ccv_cnnp_model_t* _ccv_cnnp_scaled_dot_product_attention_copy(const ccv_cnnp_model_t* const super, void* const context);
4597
4598static const ccv_cnnp_model_vtab_t ccv_cnnp_scaled_dot_product_attention_isa = {
4599 .build = _ccv_cnnp_scaled_dot_product_attention_build,
4600 .copy = _ccv_cnnp_scaled_dot_product_attention_copy,
4601};
4602
4603static const ccv_cnnp_model_vtab_t ccv_cnnp_scaled_dot_product_attention_fused_isa = {
4604 .build = _ccv_cnnp_scaled_dot_product_attention_build,
4605 .init_states = _ccv_cnnp_scaled_dot_product_attention_init_states,
4606 .add_to_parameter = _ccv_cnnp_scaled_dot_product_attention_add_to_parameter,
4607 .copy = _ccv_cnnp_scaled_dot_product_attention_copy,
4608};
4609
4610ccv_cnnp_model_t* ccv_cnnp_scaled_dot_product_attention(const float scale, const int is_causal, const int has_attn_mask, const int is_varlen, const int max_seqlen_q, const int max_seqlen_kv, const int flags, const int fused_unify_head_weights, const int no_bias, const int is_trainable, const char* const name)
4611{
4612 assert(!is_varlen || !has_attn_mask)((void) sizeof ((!is_varlen || !has_attn_mask) ? 1 : 0), __extension__
({ if (!is_varlen || !has_attn_mask) ; else __assert_fail ("!is_varlen || !has_attn_mask"
, "ccv_cnnp_model_addons.c", 4612, __extension__ __PRETTY_FUNCTION__
); }));
4613 assert(!is_varlen || !fused_unify_head_weights)((void) sizeof ((!is_varlen || !fused_unify_head_weights) ? 1
: 0), __extension__ ({ if (!is_varlen || !fused_unify_head_weights
) ; else __assert_fail ("!is_varlen || !fused_unify_head_weights"
, "ccv_cnnp_model_addons.c", 4613, __extension__ __PRETTY_FUNCTION__
); }));
4614 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));
4615 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;
4616 model_scaled_dot_product_attention->super.input_size = is_varlen ? 5 : (has_attn_mask ? 4 : 3);
4617 model_scaled_dot_product_attention->super.outputs = &model_scaled_dot_product_attention->output;
4618 model_scaled_dot_product_attention->super.output_size = 1;
4619 model_scaled_dot_product_attention->super.is_trainable = is_trainable;
4620 ccv_cnnp_model_copy_name(&model_scaled_dot_product_attention->super, name);
4621 model_scaled_dot_product_attention->weights.d = CCV_NNC_NO_TENSOR_SYMBOL;
4622 model_scaled_dot_product_attention->weights.graph = 0;
4623 model_scaled_dot_product_attention->bias.d = CCV_NNC_NO_TENSOR_SYMBOL;
4624 model_scaled_dot_product_attention->bias.graph = 0;
4625 model_scaled_dot_product_attention->scale = scale;
4626 model_scaled_dot_product_attention->is_causal = is_causal;
4627 model_scaled_dot_product_attention->has_attn_mask = has_attn_mask;
4628 model_scaled_dot_product_attention->is_varlen = is_varlen;
4629 model_scaled_dot_product_attention->max_seqlen_q = max_seqlen_q;
4630 model_scaled_dot_product_attention->max_seqlen_kv = max_seqlen_kv;
4631 model_scaled_dot_product_attention->flags = flags;
4632 model_scaled_dot_product_attention->fused_unify_head_weights = fused_unify_head_weights;
4633 model_scaled_dot_product_attention->no_bias = no_bias;
4634 return (ccv_cnnp_model_t*)model_scaled_dot_product_attention;
4635}
4636
4637static ccv_cnnp_model_t* _ccv_cnnp_scaled_dot_product_attention_copy(const ccv_cnnp_model_t* const super, void* const context)
4638{
4639 const ccv_cnnp_model_scaled_dot_product_attention_t* const self = (const ccv_cnnp_model_scaled_dot_product_attention_t*)super;
4640 return ccv_cnnp_scaled_dot_product_attention(self->scale, self->is_causal, self->has_attn_mask, self->is_varlen, self->max_seqlen_q, self->max_seqlen_kv, self->flags, self->fused_unify_head_weights, self->no_bias, self->super.is_trainable, self->super.name);
4641}
4642
4643// MARK - Debug Layer
4644
4645typedef struct {
4646 ccv_cnnp_model_t super;
4647 ccv_nnc_tensor_symbol_t output;
4648 ccv_cnnp_model_debug_f debugger;
4649 ccv_cnnp_model_debug_context_deinit_f debug_deinit;
4650 ccv_cnnp_model_debug_context_copy_f debug_copy;
4651 void* debug_context;
4652} ccv_cnnp_model_debug_t;
4653
4654static 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)
4655{
4656 if (cmd.cmd == CCV_NNC_CUSTOM_BACKWARD)
4657 {
4658 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", 4658, __extension__ __PRETTY_FUNCTION__
); }));
4659 }
4660 ccv_cnnp_model_debug_t* const self = (ccv_cnnp_model_debug_t*)cmd.data;
4661 self->debugger(inputs, input_size, stream_context, self->debug_context);
4662 return CCV_NNC_EXEC_SUCCESS;
4663}
4664
4665static ccv_nnc_cmd_vtab_t ccv_cnnp_debug_exec_isa = {
4666 .exec = _ccv_cnnp_debug_exec
4667};
4668
4669static 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)
4670{
4671 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);
4672 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", 4672, __extension__ __PRETTY_FUNCTION__
); }));
4673 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 4673, __extension__ __PRETTY_FUNCTION__
); }));
4674 ccv_nnc_tensor_symbol_t to = ccv_nnc_tensor_symbol_alias_to(graph, inputs[0]);
4675 ccv_nnc_tensor_param_t output_params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4676 if (to.d == CCV_NNC_NO_TENSOR_SYMBOL) // If we are not reshape an alias, it is straightforward.
4677 {
4678 int ofs[CCV_NNC_MAX_DIM_ALLOC(12)] = {0};
4679 int stride[CCV_NNC_MAX_DIM_ALLOC(12)];
4680 ccv_nnc_tensor_get_stride(output_params.dim, stride);
4681 outputs[0] = ccv_nnc_tensor_symbol_alias_new(graph, inputs[0], ofs, stride, output_params, 0);
4682 } else {
4683 int old_ofs[CCV_NNC_MAX_DIM_ALLOC(12)];
4684 int old_stride[CCV_NNC_MAX_DIM_ALLOC(12)];
4685 ccv_nnc_tensor_symbol_alias_params(graph, inputs[0], old_ofs, old_stride);
4686 outputs[0] = ccv_nnc_tensor_symbol_alias_new(graph, to, old_ofs, old_stride, output_params, 0);
4687 }
4688 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);
4689 cmd.data = self;
4690 ccv_nnc_graph_exec_symbol_t make_debug = ccv_nnc_graph_exec_symbol_new(graph, cmd, inputs, input_size, outputs, 1, "debug");
4691 // Disable any optimizations.
4692 ccv_nnc_graph_exec_symbol_set_flags(graph, make_debug, CCV_NNC_GRAPH_EXEC_DISABLE_OPT);
4693}
4694
4695static void _ccv_cnnp_debug_deinit(ccv_cnnp_model_t* const super)
4696{
4697 const ccv_cnnp_model_debug_t* const self = (const ccv_cnnp_model_debug_t*)super;
4698 if (self->debug_deinit && self->debug_context)
4699 self->debug_deinit(self->debug_context);
4700}
4701
4702static ccv_cnnp_model_t* _ccv_cnnp_debug_copy(const ccv_cnnp_model_t* const super, void* const context);
4703
4704static const ccv_cnnp_model_vtab_t ccv_cnnp_debug_isa = {
4705 .build = _ccv_cnnp_debug_build,
4706 .deinit = _ccv_cnnp_debug_deinit,
4707 .copy = _ccv_cnnp_debug_copy,
4708};
4709
4710ccv_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)
4711{
4712 ccv_cnnp_model_debug_t* const model_debug = (ccv_cnnp_model_debug_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_debug_t));
4713 model_debug->super.isa = &ccv_cnnp_debug_isa;
4714 model_debug->super.input_size = 0;
4715 model_debug->super.outputs = &model_debug->output;
4716 model_debug->super.output_size = 1;
4717 model_debug->debugger = func;
4718 model_debug->debug_context = context;
4719 model_debug->debug_deinit = deinit;
4720 model_debug->debug_copy = copy;
4721 ccv_cnnp_model_copy_name(&model_debug->super, name);
4722 return (ccv_cnnp_model_t*)model_debug;
4723}
4724
4725static ccv_cnnp_model_t* _ccv_cnnp_debug_copy(const ccv_cnnp_model_t* const super, void* const context)
4726{
4727 const ccv_cnnp_model_debug_t* const self = (const ccv_cnnp_model_debug_t*)super;
4728 void* debug_context = self->debug_context;
4729 if (self->debug_copy && self->debug_context)
4730 debug_context = self->debug_copy(self->debug_context);
4731 return ccv_cnnp_debug(self->debugger, debug_context, self->debug_deinit, self->debug_copy, self->super.name);
4732}
4733
4734/// MARK - Sort layer.
4735
4736typedef struct {
4737 ccv_cnnp_model_t super;
4738 ccv_nnc_tensor_symbol_t outputs[2];
4739 int along_axis;
4740 int descending;
4741} ccv_cnnp_model_sort_t;
4742
4743static 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)
4744{
4745 ccv_cnnp_model_sort_t* const self = (ccv_cnnp_model_sort_t*)super;
4746 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);
4747 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4748 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", 4748, __extension__ __PRETTY_FUNCTION__
); }));
4749 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4750 params.datatype = CCV_32S;
4751 outputs[1] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4752 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");
4753}
4754
4755static ccv_cnnp_model_t* _ccv_cnnp_sort_copy(const ccv_cnnp_model_t* const self, void* const context);
4756
4757static const ccv_cnnp_model_vtab_t ccv_cnnp_sort_isa = {
4758 .build = _ccv_cnnp_sort_build,
4759 .copy = _ccv_cnnp_sort_copy,
4760};
4761
4762ccv_cnnp_model_t* ccv_cnnp_sort(const int along_axis, const int descending, const char* const name)
4763{
4764 ccv_cnnp_model_sort_t* const model_sort = (ccv_cnnp_model_sort_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_sort_t));
4765 model_sort->super.isa = &ccv_cnnp_sort_isa;
4766 model_sort->super.input_size = 0;
4767 model_sort->super.outputs = model_sort->outputs;
4768 model_sort->super.output_size = 2;
4769 model_sort->along_axis = along_axis;
4770 model_sort->descending = descending;
4771 ccv_cnnp_model_copy_name(&model_sort->super, name);
4772 return (ccv_cnnp_model_t*)model_sort;
4773}
4774
4775static ccv_cnnp_model_t* _ccv_cnnp_sort_copy(const ccv_cnnp_model_t* const super, void* const context)
4776{
4777 ccv_cnnp_model_sort_t* const self = (ccv_cnnp_model_sort_t*)super;
4778 return ccv_cnnp_sort(self->along_axis, self->descending, self->super.name);
4779}
4780
4781/// MARK - Partition layer.
4782
4783typedef struct {
4784 ccv_cnnp_model_t super;
4785 ccv_nnc_tensor_symbol_t outputs[2];
4786 int kth;
4787 int along_axis;
4788 int descending;
4789} ccv_cnnp_model_partition_t;
4790
4791static 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)
4792{
4793 ccv_cnnp_model_partition_t* const self = (ccv_cnnp_model_partition_t*)super;
4794 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);
4795 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4796 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", 4796, __extension__ __PRETTY_FUNCTION__
); }));
4797 if (self->kth > 0)
4798 params.dim[self->along_axis] = self->kth;
4799 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4800 params.datatype = CCV_32S;
4801 outputs[1] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4802 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");
4803}
4804
4805static ccv_cnnp_model_t* _ccv_cnnp_partition_copy(const ccv_cnnp_model_t* const self, void* const context);
4806
4807static const ccv_cnnp_model_vtab_t ccv_cnnp_partition_isa = {
4808 .build = _ccv_cnnp_partition_build,
4809 .copy = _ccv_cnnp_partition_copy,
4810};
4811
4812ccv_cnnp_model_t* ccv_cnnp_partition(const int kth, const int along_axis, const int descending, const char* const name)
4813{
4814 ccv_cnnp_model_partition_t* const model_partition = (ccv_cnnp_model_partition_t*)cccalloccalloc(1, sizeof(ccv_cnnp_model_partition_t));
4815 model_partition->super.isa = &ccv_cnnp_partition_isa;
4816 model_partition->super.input_size = 0;
4817 model_partition->super.outputs = model_partition->outputs;
4818 model_partition->super.output_size = 2;
4819 model_partition->kth = kth;
4820 model_partition->along_axis = along_axis;
4821 model_partition->descending = descending;
4822 ccv_cnnp_model_copy_name(&model_partition->super, name);
4823 return (ccv_cnnp_model_t*)model_partition;
4824}
4825
4826static ccv_cnnp_model_t* _ccv_cnnp_partition_copy(const ccv_cnnp_model_t* const super, void* const context)
4827{
4828 ccv_cnnp_model_partition_t* const self = (ccv_cnnp_model_partition_t*)super;
4829 return ccv_cnnp_partition(self->kth, self->along_axis, self->descending, self->super.name);
4830}
4831
4832/// MARK - Unique consecutive layer.
4833
4834typedef struct {
4835 ccv_cnnp_model_t super;
4836 ccv_nnc_tensor_symbol_t outputs[2];
4837 int bincount;
4838} ccv_cnnp_model_unique_consecutive_t;
4839
4840static 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)
4841{
4842 ccv_cnnp_model_unique_consecutive_t* const self = (ccv_cnnp_model_unique_consecutive_t*)super;
4843 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);
4844 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4845 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", 4845, __extension__ __PRETTY_FUNCTION__
); }));
4846 if (self->bincount > 0)
4847 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; }
);
4848 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4849 params.datatype = CCV_32S;
4850 outputs[1] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4851 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");
4852}
4853
4854static ccv_cnnp_model_t* _ccv_cnnp_unique_consecutive_copy(const ccv_cnnp_model_t* const self, void* const context);
4855
4856static const ccv_cnnp_model_vtab_t ccv_cnnp_unique_consecutive_isa = {
4857 .build = _ccv_cnnp_unique_consecutive_build,
4858 .copy = _ccv_cnnp_unique_consecutive_copy,
4859};
4860
4861ccv_cnnp_model_t* ccv_cnnp_unique_consecutive(const int bincount, const char* const name)
4862{
4863 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));
4864 model_unique_consecutive->super.isa = &ccv_cnnp_unique_consecutive_isa;
4865 model_unique_consecutive->super.input_size = 0;
4866 model_unique_consecutive->super.outputs = model_unique_consecutive->outputs;
4867 model_unique_consecutive->super.output_size = 2;
4868 model_unique_consecutive->bincount = bincount;
4869 ccv_cnnp_model_copy_name(&model_unique_consecutive->super, name);
4870 return (ccv_cnnp_model_t*)model_unique_consecutive;
4871}
4872
4873static ccv_cnnp_model_t* _ccv_cnnp_unique_consecutive_copy(const ccv_cnnp_model_t* const super, void* const context)
4874{
4875 ccv_cnnp_model_unique_consecutive_t* const self = (ccv_cnnp_model_unique_consecutive_t*)super;
4876 return ccv_cnnp_unique_consecutive(self->bincount, self->super.name);
4877}
4878
4879/// MARK - Scatter add layer.
4880
4881typedef struct {
4882 ccv_cnnp_model_t super;
4883 ccv_nnc_tensor_symbol_t output;
4884 int bincount;
4885} ccv_cnnp_model_scatter_add_t;
4886
4887static 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)
4888{
4889 ccv_cnnp_model_scatter_add_t* const self = (ccv_cnnp_model_scatter_add_t*)super;
4890 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);
4891 ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4892 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 4892, __extension__ __PRETTY_FUNCTION__
); }));
4893 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", 4893, __extension__ __PRETTY_FUNCTION__
); }));
4894 params.dim[0] = self->bincount;
4895 outputs[0] = ccv_nnc_tensor_symbol_new(graph, params, 0);
4896 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");
4897}
4898
4899static ccv_cnnp_model_t* _ccv_cnnp_scatter_add_copy(const ccv_cnnp_model_t* const self, void* const context);
4900
4901static const ccv_cnnp_model_vtab_t ccv_cnnp_scatter_add_isa = {
4902 .build = _ccv_cnnp_scatter_add_build,
4903 .copy = _ccv_cnnp_scatter_add_copy,
4904};
4905
4906ccv_cnnp_model_t* ccv_cnnp_scatter_add(const int bincount, const char* const name)
4907{
4908 assert(bincount > 0)((void) sizeof ((bincount > 0) ? 1 : 0), __extension__ ({ if
(bincount > 0) ; else __assert_fail ("bincount > 0", "ccv_cnnp_model_addons.c"
, 4908, __extension__ __PRETTY_FUNCTION__); }));
4909 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));
4910 model_scatter_add->super.isa = &ccv_cnnp_scatter_add_isa;
4911 model_scatter_add->super.input_size = 0;
4912 model_scatter_add->super.outputs = &model_scatter_add->output;
4913 model_scatter_add->super.output_size = 1;
4914 model_scatter_add->bincount = bincount;
4915 ccv_cnnp_model_copy_name(&model_scatter_add->super, name);
4916 return (ccv_cnnp_model_t*)model_scatter_add;
4917}
4918
4919static ccv_cnnp_model_t* _ccv_cnnp_scatter_add_copy(const ccv_cnnp_model_t* const super, void* const context)
4920{
4921 ccv_cnnp_model_scatter_add_t* const self = (ccv_cnnp_model_scatter_add_t*)super;
4922 return ccv_cnnp_scatter_add(self->bincount, self->super.name);
4923}
4924
4925// MARK - Segmented Dense Layer
4926
4927typedef struct {
4928 ccv_cnnp_model_t super;
4929 ccv_nnc_tensor_symbol_t output;
4930 ccv_nnc_tensor_symbol_t weights;
4931 ccv_nnc_tensor_symbol_t bias;
4932 int segments;
4933 int count;
4934 int no_bias;
4935 int flags;
4936} ccv_cnnp_model_segmented_dense_t;
4937
4938static 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)
4939{
4940 ccv_cnnp_model_segmented_dense_t* const self = (ccv_cnnp_model_segmented_dense_t*)super;
4941 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);
4942 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"
, 4942, __extension__ __PRETTY_FUNCTION__); }));
4943 assert(output_size == 1)((void) sizeof ((output_size == 1) ? 1 : 0), __extension__ ({
if (output_size == 1) ; else __assert_fail ("output_size == 1"
, "ccv_cnnp_model_addons.c", 4943, __extension__ __PRETTY_FUNCTION__
); }));
4944 const ccv_nnc_tensor_param_t params = ccv_nnc_tensor_symbol_params(graph, inputs[0]);
4945 const ccv_nnc_tensor_param_t indices_params = ccv_nnc_tensor_symbol_params(graph, inputs[1]);
4946 const ccv_nnc_tensor_param_t counts_params = ccv_nnc_tensor_symbol_params(graph, inputs[2]);
4947 ccv_nnc_tensor_param_t weights_params = params;
4948 memset(weights_params.dim, 0, sizeof(weights_params.dim));
4949 weights_params.dim[0] = self->segments;
4950 weights_params.dim[1] = self->count;
4951 weights_params.dim[2] = params.dim[ccv_nnc_tensor_nd(params.dim) - 1];
4952 if (!self->weights.graph)
4953 self->weights = ccv_nnc_tensor_symbol_new(graph, weights_params, "weights");
4954 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"
, 4954, __extension__ __PRETTY_FUNCTION__); }));
4955 const ccv_nnc_tensor_symbol_t weights = ccv_cnnp_model_get_symbol(super, self->weights);
4956 ccv_nnc_tensor_param_t bias_params = params;
4957 memset(bias_params.dim, 0, sizeof(bias_params.dim));
4958 bias_params.dim[0] = self->segments;
4959 bias_params.dim[1] = self->count;
4960 ccv_nnc_cmd_t cmd = {0};
4961 cmd.cmd = CCV_NNC_SEGMENTED_GEMM_FORWARD;
4962 cmd.info.blas.a[0] = 1;
4963 cmd.info.blas.a[1] = 1;
4964 cmd.info.blas.transpose_b[0] = 1;
4965 cmd.info.blas.transpose_b[1] = 2;
4966 cmd.info.blas.flags = self->flags;
4967 ccv_nnc_tensor_param_t output_params;
4968 ccv_nnc_hint_tensor_auto(cmd, (ccv_nnc_tensor_param_t []){
4969 params, indices_params, counts_params,
4970 weights_params,
4971 bias_params,
4972 }, 5, ccv_nnc_no_hint, &output_params, 1);
4973 const ccv_nnc_tensor_symbol_t output = ccv_nnc_tensor_symbol_new(graph, output_params, 0);
4974 if (self->no_bias)
4975 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");
4976 else {
4977 if (!self->bias.graph)
4978 self->bias = ccv_nnc_tensor_symbol_new(graph, bias_params, "bias");
4979 const ccv_nnc_tensor_symbol_t bias = ccv_cnnp_model_get_symbol(super, self->bias);
4980 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");
4981 }
4982 outputs[0] = output;
4983}
4984
4985static 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)
4986{
4987 ccv_cnnp_model_segmented_dense_t* const self = (ccv_cnnp_model_segmented_dense_t*)super;
4988 const ccv_nnc_tensor_param_t weight_params = ccv_nnc_tensor_symbol_params(graph, self->weights);
4989 const int c = weight_params.dim[1];
4990 const float std = sqrtf(2) / sqrtf(c);
4991 const float bound = sqrtf(3) * std;
4992 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);
4993 if (self->bias.graph)
4994 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);
4995}
4996
4997static 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)
4998{
4999 ccv_cnnp_model_segmented_dense_t* const self = (ccv_cnnp_model_segmented_dense_t*)super;
5000 add_to_array(parameters, self->weights, is_trainable);
5001 if (self->bias.graph)
5002 add_to_array(parameters, self->bias, is_trainable);
5003}
5004
5005static ccv_cnnp_model_t* _ccv_cnnp_segmented_dense_copy(const ccv_cnnp_model_t* const super, void* const context);
5006
5007static const ccv_cnnp_model_vtab_t ccv_cnnp_segmented_dense_isa = {
5008 .build = _ccv_cnnp_segmented_dense_build,
5009 .init_states = _ccv_cnnp_segmented_dense_init_states,
5010 .add_to_parameter = _ccv_cnnp_segmented_dense_add_to_parameter,
5011 .copy = _ccv_cnnp_segmented_dense_copy,
5012};
5013
5014ccv_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)
5015{
5016 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));
5017 model_segmented_dense->super.isa = &ccv_cnnp_segmented_dense_isa;
5018 model_segmented_dense->super.input_size = 3;
5019 model_segmented_dense->super.outputs = &model_segmented_dense->output;
5020 model_segmented_dense->super.output_size = 1;
5021 model_segmented_dense->super.is_trainable = is_trainable;
5022 ccv_cnnp_model_copy_name(&model_segmented_dense->super, name);
5023 model_segmented_dense->weights.d = CCV_NNC_NO_TENSOR_SYMBOL;
5024 model_segmented_dense->weights.graph = 0;
5025 model_segmented_dense->bias.d = CCV_NNC_NO_TENSOR_SYMBOL;
5026 model_segmented_dense->bias.graph = 0;
5027 model_segmented_dense->segments = segments;
5028 model_segmented_dense->count = count;
5029 model_segmented_dense->no_bias = no_bias;
5030 model_segmented_dense->flags = flags;
5031 return (ccv_cnnp_model_t*)model_segmented_dense;
5032}
5033
5034static ccv_cnnp_model_t* _ccv_cnnp_segmented_dense_copy(const ccv_cnnp_model_t* const super, void* const context)
5035{
5036 const ccv_cnnp_model_segmented_dense_t* const self = (const ccv_cnnp_model_segmented_dense_t*)super;
5037 return ccv_cnnp_segmented_dense(self->segments, self->count, self->no_bias, self->flags, self->super.is_trainable, self->super.name);
5038}