Coverage Report

Created: 2024-08-19 11:27

/home/liu/actions-runner/_work/ccv/ccv/test/int/nnc/gelu.tests.c
Line
Count
Source (jump to first uncovered line)
1
#include "case.h"
2
#include "ccv_case.h"
3
#include "ccv_nnc_case.h"
4
#include <ccv.h>
5
#include <nnc/ccv_nnc.h>
6
#include <nnc/ccv_nnc_easy.h>
7
#include <3rdparty/dsfmt/dSFMT.h>
8
9
TEST_SETUP()
10
{
11
  ccv_nnc_init();
12
}
13
14
TEST_CASE("gelu in float")
15
{
16
  GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GELU_FORWARD, CCV_NNC_BACKEND_GPU_REF) || ccv_nnc_cmd_ok(CCV_NNC_GELU_FORWARD, CCV_NNC_BACKEND_MPS));
17
  ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new();
18
  ccv_nnc_tensor_symbol_t a = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NCHW(000, 32F, 20, 10), "a");
19
  ccv_nnc_tensor_symbol_t b = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NCHW(000, 32F, 20, 10), "b");
20
  ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_GELU_FORWARD(0), TENSOR_SYMBOL_LIST(a), TENSOR_SYMBOL_LIST(b), "gelu");
21
  ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
22
  SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH);
23
  ccv_nnc_graph_t* graph = 0;
24
  ccv_nnc_tensor_arena_t* tensor_arena = 0;
25
  ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0;
26
  ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena);
27
  GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH);
28
  ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 20, 10), 0);
29
  dsfmt_t dsfmt;
30
  dsfmt_init_gen_rand(&dsfmt, 0);
31
  int i;
32
  for (i = 0; i < 20 * 10; i++)
33
    x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt);
34
  ccv_nnc_tensor_t* const a_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, a);
35
  ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(a_tensor), 0);
36
  ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0);
37
  ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 20, 10), 0);
38
  ccv_nnc_tensor_t* const b_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, b);
39
  ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(b_tensor), TENSOR_LIST(y_tensor), 0);
40
  ccv_nnc_tensor_t* const ty = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 20, 10), 0);
41
  ccv_nnc_cmd_exec(CMD_GELU_FORWARD(0), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(ty), 0);
42
  REQUIRE_TENSOR_EQ(ty, y_tensor, "gelu from cudnn should match from CPU");
43
  ccv_nnc_tensor_free(x_tensor);
44
  ccv_nnc_tensor_free(y_tensor);
45
  ccv_nnc_tensor_free(ty);
46
  ccv_nnc_graph_free(graph);
47
  ccv_nnc_tensor_arena_free(tensor_arena);
48
  ccv_nnc_graph_exec_arena_free(graph_exec_arena);
49
  ccv_nnc_symbolic_graph_free(symbolic_graph);
50
}
51
52
TEST_CASE("gelu in half precision")
53
1
{
54
1
  GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GELU_FORWARD, CCV_NNC_BACKEND_GPU_REF) || ccv_nnc_cmd_ok(CCV_NNC_GELU_FORWARD, CCV_NNC_BACKEND_MPS));
55
1
  ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new();
56
1
  ccv_nnc_tensor_symbol_t a = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NCHW(000, 16F, 20, 10), "a");
57
1
  ccv_nnc_tensor_symbol_t b = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NCHW(000, 16F, 20, 10), "b");
58
1
  ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_GELU_FORWARD(0), TENSOR_SYMBOL_LIST(a), TENSOR_SYMBOL_LIST(b), "gelu");
59
1
  ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
60
1
  SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH);
61
1
  ccv_nnc_graph_t* graph = 0;
62
1
  ccv_nnc_tensor_arena_t* tensor_arena = 0;
63
1
  ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0;
64
1
  ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena);
65
1
  GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH);
66
1
  ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 20, 10), 0);
67
1
  dsfmt_t dsfmt;
68
1
  dsfmt_init_gen_rand(&dsfmt, 0);
69
1
  int i;
70
201
  for (i = 0; i < 20 * 10; 
i++200
)
71
200
    x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt);
72
1
  ccv_nnc_tensor_t* const a_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, a);
73
1
  ccv_nnc_tensor_t* const x16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(16F, 20, 10), 0);
74
1
  ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(x16_tensor), 0);
75
1
  ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x16_tensor), TENSOR_LIST(a_tensor), 0);
76
1
  ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0);
77
1
  ccv_nnc_tensor_t* const y16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(16F, 20, 10), 0);
78
1
  ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 20, 10), 0);
79
1
  ccv_nnc_tensor_t* const b_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, b);
80
1
  ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(b_tensor), TENSOR_LIST(y16_tensor), 0);
81
1
  ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(y16_tensor), TENSOR_LIST(y_tensor), 0);
82
1
  ccv_nnc_tensor_t* const ty = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 20, 10), 0);
83
1
  ccv_nnc_cmd_exec(CMD_GELU_FORWARD(0), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(ty), 0);
84
1
  REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, ty->data.f32, y_tensor->data.f32, 20 * 10, 1e-3, "gelu from cudnn should match from CPU");
85
1
  ccv_nnc_tensor_free(x_tensor);
86
1
  ccv_nnc_tensor_free(x16_tensor);
87
1
  ccv_nnc_tensor_free(y16_tensor);
88
1
  ccv_nnc_tensor_free(y_tensor);
89
1
  ccv_nnc_tensor_free(ty);
90
1
  ccv_nnc_graph_free(graph);
91
1
  ccv_nnc_tensor_arena_free(tensor_arena);
92
1
  ccv_nnc_graph_exec_arena_free(graph_exec_arena);
93
1
  ccv_nnc_symbolic_graph_free(symbolic_graph);
94
1
}
95
96
TEST_CASE("gelu gradient in float")
97
1
{
98
1
  GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GELU_FORWARD, CCV_NNC_BACKEND_GPU_REF) &&
99
1
    ccv_nnc_cmd_ok(CCV_NNC_GELU_BACKWARD, CCV_NNC_BACKEND_GPU_REF));
100
1
  ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new();
101
1
  ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10, 100), "x");
102
1
  ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10, 100), "y");
103
1
  ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_GELU_FORWARD(0), TENSOR_SYMBOL_LIST(x), TENSOR_SYMBOL_LIST(y), "gelu");
104
1
  ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
105
1
  ccv_nnc_symbolic_graph_backward(symbolic_graph, TENSOR_SYMBOL_LIST(y), TENSOR_SYMBOL_LIST(x), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph));
106
1
  ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
107
1
  SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH);
108
1
  ccv_nnc_tensor_symbol_t dy = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, y);
109
1
  ccv_nnc_tensor_symbol_t dx = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, x);
110
1
  ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0);
111
1
  dsfmt_t dsfmt;
112
1
  dsfmt_init_gen_rand(&dsfmt, 0);
113
1
  int i;
114
1.00k
  for (i = 0; i < 10 * 100; 
i++1.00k
)
115
1.00k
    x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt);
116
1
  ccv_nnc_tensor_t* const dy_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0);
117
1.00k
  for (i = 0; i < 10 * 100; 
i++1.00k
)
118
1.00k
    dy_tensor->data.f32[i] = 0;
119
11
  for (i = 0; i < 10; 
i++10
)
120
10
    dy_tensor->data.f32[i * 100 + i] = 1;
121
1
  ccv_nnc_tensor_t* const dyt = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 100), 0);
122
1
  ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor), TENSOR_LIST(dyt), 0);
123
1
  ccv_nnc_graph_t* graph = 0;
124
1
  ccv_nnc_tensor_arena_t* tensor_arena = 0;
125
1
  ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0;
126
1
  ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, TENSOR_BIND_MAP(KV(dy, dyt)), TENSOR_SYMBOL_LIST(y), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena);
127
1
  GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH);
128
1
  ccv_nnc_tensor_t* const xt = ccv_nnc_tensor_from_symbol(tensor_arena, x);
129
1
  ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(xt), 0);
130
1
  ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0);
131
1
  ccv_nnc_tensor_t* const dx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0);
132
1
  ccv_nnc_tensor_t* const dxt = ccv_nnc_tensor_from_symbol(tensor_arena, dx);
133
1
  ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0);
134
1
  ccv_nnc_tensor_t* const yt = ccv_nnc_tensor_from_symbol(tensor_arena, y);
135
1
  ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dxt), TENSOR_LIST(dx_tensor), 0);
136
1
  ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(yt), TENSOR_LIST(y_tensor), 0);
137
1
  ccv_nnc_tensor_t* const ty_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0);
138
1
  ccv_nnc_cmd_exec(CMD_GELU_FORWARD(0), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(ty_tensor), 0);
139
1
  REQUIRE_TENSOR_EQ(ty_tensor, y_tensor, "forward pass should match");
140
1
  ccv_nnc_tensor_t* const tdx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0);
141
1
  ccv_nnc_cmd_exec(CMD_GELU_BACKWARD(0), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor, x_tensor, 0), TENSOR_LIST(tdx_tensor), 0);
142
1
  REQUIRE_TENSOR_EQ(tdx_tensor, dx_tensor, "backward pass should match");
143
1
  ccv_nnc_tensor_free(x_tensor);
144
1
  ccv_nnc_tensor_free(y_tensor);
145
1
  ccv_nnc_tensor_free(dx_tensor);
146
1
  ccv_nnc_tensor_free(dy_tensor);
147
1
  ccv_nnc_tensor_free(ty_tensor);
148
1
  ccv_nnc_tensor_free(tdx_tensor);
149
1
  ccv_nnc_tensor_free(dyt);
150
1
  ccv_nnc_graph_free(graph);
151
1
  ccv_nnc_tensor_arena_free(tensor_arena);
152
1
  ccv_nnc_graph_exec_arena_free(graph_exec_arena);
153
1
  ccv_nnc_symbolic_graph_free(symbolic_graph);
154
1
}
155
156
TEST_CASE("mps gelu gradient in float")
157
1
{
158
1
  GUARD_ELSE_RETURN((ccv_nnc_cmd_ok(CCV_NNC_GELU_FORWARD, CCV_NNC_BACKEND_MPS) &&
159
1
    ccv_nnc_cmd_ok(CCV_NNC_GELU_BACKWARD, CCV_NNC_BACKEND_MPS)));
160
0
  ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new();
161
0
  ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10, 100), "x");
162
0
  ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10, 100), "y");
163
0
  ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_GELU_FORWARD(0), TENSOR_SYMBOL_LIST(x), TENSOR_SYMBOL_LIST(y), "gelu");
164
0
  ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
165
0
  ccv_nnc_symbolic_graph_backward(symbolic_graph, TENSOR_SYMBOL_LIST(y), TENSOR_SYMBOL_LIST(x), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph));
166
0
  ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
167
0
  SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH);
168
0
  ccv_nnc_tensor_symbol_t dy = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, y);
169
0
  ccv_nnc_tensor_symbol_t dx = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, x);
170
0
  ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0);
171
0
  dsfmt_t dsfmt;
172
0
  dsfmt_init_gen_rand(&dsfmt, 0);
173
0
  int i;
174
0
  for (i = 0; i < 10 * 100; i++)
175
0
    x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt);
176
0
  ccv_nnc_tensor_t* const dy_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0);
177
0
  for (i = 0; i < 10 * 100; i++)
178
0
    dy_tensor->data.f32[i] = 0;
179
0
  for (i = 0; i < 10; i++)
180
0
    dy_tensor->data.f32[i * 100 + i] = 1;
181
0
  ccv_nnc_tensor_t* const dyt = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 100), 0);
182
0
  ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor), TENSOR_LIST(dyt), 0);
183
0
  ccv_nnc_graph_t* graph = 0;
184
0
  ccv_nnc_tensor_arena_t* tensor_arena = 0;
185
0
  ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0;
186
0
  ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, TENSOR_BIND_MAP(KV(dy, dyt)), TENSOR_SYMBOL_LIST(y), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena);
187
0
  GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH);
188
0
  ccv_nnc_tensor_t* const xt = ccv_nnc_tensor_from_symbol(tensor_arena, x);
189
0
  ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(xt), 0);
190
0
  ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0);
191
0
  ccv_nnc_tensor_t* const dx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0);
192
0
  ccv_nnc_tensor_t* const dxt = ccv_nnc_tensor_from_symbol(tensor_arena, dx);
193
0
  ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0);
194
0
  ccv_nnc_tensor_t* const yt = ccv_nnc_tensor_from_symbol(tensor_arena, y);
195
0
  ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dxt), TENSOR_LIST(dx_tensor), 0);
196
0
  ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(yt), TENSOR_LIST(y_tensor), 0);
197
0
  ccv_nnc_tensor_t* const ty_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0);
198
0
  ccv_nnc_cmd_exec(CMD_GELU_FORWARD(0), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(ty_tensor), 0);
199
0
  REQUIRE_TENSOR_EQ(ty_tensor, y_tensor, "forward pass should match");
200
0
  ccv_nnc_tensor_t* const tdx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0);
201
0
  ccv_nnc_cmd_exec(CMD_GELU_BACKWARD(0), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor, x_tensor, 0), TENSOR_LIST(tdx_tensor), 0);
202
0
  REQUIRE_TENSOR_EQ(tdx_tensor, dx_tensor, "backward pass should match");
203
0
  ccv_nnc_tensor_free(x_tensor);
204
0
  ccv_nnc_tensor_free(y_tensor);
205
0
  ccv_nnc_tensor_free(dx_tensor);
206
0
  ccv_nnc_tensor_free(dy_tensor);
207
0
  ccv_nnc_tensor_free(ty_tensor);
208
0
  ccv_nnc_tensor_free(tdx_tensor);
209
0
  ccv_nnc_tensor_free(dyt);
210
0
  ccv_nnc_graph_free(graph);
211
0
  ccv_nnc_tensor_arena_free(tensor_arena);
212
0
  ccv_nnc_graph_exec_arena_free(graph_exec_arena);
213
0
  ccv_nnc_symbolic_graph_free(symbolic_graph);
214
0
}
215
216
TEST_CASE("gelu gradient in half precision")
217
1
{
218
1
  GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GELU_FORWARD, CCV_NNC_BACKEND_GPU_REF) &&
219
1
    ccv_nnc_cmd_ok(CCV_NNC_GELU_BACKWARD, CCV_NNC_BACKEND_GPU_REF));
220
1
  ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new();
221
1
  ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 16F, 10, 100), "x");
222
1
  ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 16F, 10, 100), "y");
223
1
  ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_GELU_FORWARD(0), TENSOR_SYMBOL_LIST(x), TENSOR_SYMBOL_LIST(y), "gelu");
224
1
  ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
225
1
  ccv_nnc_symbolic_graph_backward(symbolic_graph, TENSOR_SYMBOL_LIST(y), TENSOR_SYMBOL_LIST(x), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph));
226
1
  ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
227
1
  SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH);
228
1
  ccv_nnc_tensor_symbol_t dy = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, y);
229
1
  ccv_nnc_tensor_symbol_t dx = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, x);
230
1
  ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0);
231
1
  dsfmt_t dsfmt;
232
1
  dsfmt_init_gen_rand(&dsfmt, 0);
233
1
  int i;
234
1.00k
  for (i = 0; i < 10 * 100; 
i++1.00k
)
235
1.00k
    x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt);
236
1
  ccv_nnc_tensor_t* const dy_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0);
237
1.00k
  for (i = 0; i < 10 * 100; 
i++1.00k
)
238
1.00k
    dy_tensor->data.f32[i] = 0;
239
11
  for (i = 0; i < 10; 
i++10
)
240
10
    dy_tensor->data.f32[i * 100 + i] = 1;
241
1
  ccv_nnc_tensor_t* const dy16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 100), 0);
242
1
  ccv_nnc_tensor_t* const dyt = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 10, 100), 0);
243
1
  ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor), TENSOR_LIST(dy16_tensor), 0);
244
1
  ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy16_tensor), TENSOR_LIST(dyt), 0);
245
1
  ccv_nnc_graph_t* graph = 0;
246
1
  ccv_nnc_tensor_arena_t* tensor_arena = 0;
247
1
  ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0;
248
1
  ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, TENSOR_BIND_MAP(KV(dy, dyt)), TENSOR_SYMBOL_LIST(y), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena);
249
1
  GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH);
250
1
  ccv_nnc_tensor_t* const xt = ccv_nnc_tensor_from_symbol(tensor_arena, x);
251
1
  ccv_nnc_tensor_t* const x16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 100), 0);
252
1
  ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(x16_tensor), 0);
253
1
  ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x16_tensor), TENSOR_LIST(xt), 0);
254
1
  ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0);
255
1
  ccv_nnc_tensor_t* const dx16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 100), 0);
256
1
  ccv_nnc_tensor_t* const dx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0);
257
1
  ccv_nnc_tensor_t* const dxt = ccv_nnc_tensor_from_symbol(tensor_arena, dx);
258
1
  ccv_nnc_tensor_t* const y16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 100), 0);
259
1
  ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0);
260
1
  ccv_nnc_tensor_t* const yt = ccv_nnc_tensor_from_symbol(tensor_arena, y);
261
1
  ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dxt), TENSOR_LIST(dx16_tensor), 0);
262
1
  ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dx16_tensor), TENSOR_LIST(dx_tensor), 0);
263
1
  ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(yt), TENSOR_LIST(y16_tensor), 0);
264
1
  ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(y16_tensor), TENSOR_LIST(y_tensor), 0);
265
1
  ccv_nnc_tensor_t* const ty_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0);
266
1
  ccv_nnc_cmd_exec(CMD_GELU_FORWARD(0), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(ty_tensor), 0);
267
1
  REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, ty_tensor->data.f32, y_tensor->data.f32, 10 * 100, 1e-3, "forward pass should match");
268
1
  ccv_nnc_tensor_t* const tdx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0);
269
1
  ccv_nnc_cmd_exec(CMD_GELU_BACKWARD(0), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor, x_tensor, 0), TENSOR_LIST(tdx_tensor), 0);
270
1
  REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, tdx_tensor->data.f32, dx_tensor->data.f32, 10 * 100, 1e-3, "backward pass should match");
271
1
  ccv_nnc_tensor_free(x_tensor);
272
1
  ccv_nnc_tensor_free(x16_tensor);
273
1
  ccv_nnc_tensor_free(y_tensor);
274
1
  ccv_nnc_tensor_free(y16_tensor);
275
1
  ccv_nnc_tensor_free(dx_tensor);
276
1
  ccv_nnc_tensor_free(dx16_tensor);
277
1
  ccv_nnc_tensor_free(dy_tensor);
278
1
  ccv_nnc_tensor_free(dy16_tensor);
279
1
  ccv_nnc_tensor_free(ty_tensor);
280
1
  ccv_nnc_tensor_free(tdx_tensor);
281
1
  ccv_nnc_tensor_free(dyt);
282
1
  ccv_nnc_graph_free(graph);
283
1
  ccv_nnc_tensor_arena_free(tensor_arena);
284
1
  ccv_nnc_graph_exec_arena_free(graph_exec_arena);
285
1
  ccv_nnc_symbolic_graph_free(symbolic_graph);
286
1
}
287
288
TEST_CASE("mps gelu gradient in half precision")
289
1
{
290
1
  GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GELU_FORWARD, CCV_NNC_BACKEND_MPS) &&
291
1
    ccv_nnc_cmd_ok(CCV_NNC_GELU_BACKWARD, CCV_NNC_BACKEND_MPS));
292
0
  ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new();
293
0
  ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 16F, 10, 100), "x");
294
0
  ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 16F, 10, 100), "y");
295
0
  ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_GELU_FORWARD(0), TENSOR_SYMBOL_LIST(x), TENSOR_SYMBOL_LIST(y), "gelu");
296
0
  ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
297
0
  ccv_nnc_symbolic_graph_backward(symbolic_graph, TENSOR_SYMBOL_LIST(y), TENSOR_SYMBOL_LIST(x), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph));
298
0
  ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
299
0
  SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH);
300
0
  ccv_nnc_tensor_symbol_t dy = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, y);
301
0
  ccv_nnc_tensor_symbol_t dx = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, x);
302
0
  ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0);
303
0
  dsfmt_t dsfmt;
304
0
  dsfmt_init_gen_rand(&dsfmt, 0);
305
0
  int i;
306
0
  for (i = 0; i < 10 * 100; i++)
307
0
    x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt);
308
0
  ccv_nnc_tensor_t* const dy_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0);
309
0
  for (i = 0; i < 10 * 100; i++)
310
0
    dy_tensor->data.f32[i] = 0;
311
0
  for (i = 0; i < 10; i++)
312
0
    dy_tensor->data.f32[i * 100 + i] = 1;
313
0
  ccv_nnc_tensor_t* const dy16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 100), 0);
314
0
  ccv_nnc_tensor_t* const dyt = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 10, 100), 0);
315
0
  ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor), TENSOR_LIST(dy16_tensor), 0);
316
0
  ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy16_tensor), TENSOR_LIST(dyt), 0);
317
0
  ccv_nnc_graph_t* graph = 0;
318
0
  ccv_nnc_tensor_arena_t* tensor_arena = 0;
319
0
  ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0;
320
0
  ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, TENSOR_BIND_MAP(KV(dy, dyt)), TENSOR_SYMBOL_LIST(y), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena);
321
0
  GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH);
322
0
  ccv_nnc_tensor_t* const xt = ccv_nnc_tensor_from_symbol(tensor_arena, x);
323
0
  ccv_nnc_tensor_t* const x16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 100), 0);
324
0
  ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(x16_tensor), 0);
325
0
  ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x16_tensor), TENSOR_LIST(xt), 0);
326
0
  ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0);
327
0
  ccv_nnc_tensor_t* const dx16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 100), 0);
328
0
  ccv_nnc_tensor_t* const dx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0);
329
0
  ccv_nnc_tensor_t* const dxt = ccv_nnc_tensor_from_symbol(tensor_arena, dx);
330
0
  ccv_nnc_tensor_t* const y16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 100), 0);
331
0
  ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0);
332
0
  ccv_nnc_tensor_t* const yt = ccv_nnc_tensor_from_symbol(tensor_arena, y);
333
0
  ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dxt), TENSOR_LIST(dx16_tensor), 0);
334
0
  ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dx16_tensor), TENSOR_LIST(dx_tensor), 0);
335
0
  ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(yt), TENSOR_LIST(y16_tensor), 0);
336
0
  ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(y16_tensor), TENSOR_LIST(y_tensor), 0);
337
0
  ccv_nnc_tensor_t* const ty_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0);
338
0
  ccv_nnc_cmd_exec(CMD_GELU_FORWARD(0), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(ty_tensor), 0);
339
0
  REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, ty_tensor->data.f32, y_tensor->data.f32, 10 * 100, 1e-3, "forward pass should match");
340
0
  ccv_nnc_tensor_t* const tdx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0);
341
0
  ccv_nnc_cmd_exec(CMD_GELU_BACKWARD(0), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor, x_tensor, 0), TENSOR_LIST(tdx_tensor), 0);
342
0
  REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, tdx_tensor->data.f32, dx_tensor->data.f32, 10 * 100, 1e-3, "backward pass should match");
343
0
  ccv_nnc_tensor_free(x_tensor);
344
0
  ccv_nnc_tensor_free(x16_tensor);
345
0
  ccv_nnc_tensor_free(y_tensor);
346
0
  ccv_nnc_tensor_free(y16_tensor);
347
0
  ccv_nnc_tensor_free(dx_tensor);
348
0
  ccv_nnc_tensor_free(dx16_tensor);
349
0
  ccv_nnc_tensor_free(dy_tensor);
350
0
  ccv_nnc_tensor_free(dy16_tensor);
351
0
  ccv_nnc_tensor_free(ty_tensor);
352
0
  ccv_nnc_tensor_free(tdx_tensor);
353
0
  ccv_nnc_tensor_free(dyt);
354
0
  ccv_nnc_graph_free(graph);
355
0
  ccv_nnc_tensor_arena_free(tensor_arena);
356
0
  ccv_nnc_graph_exec_arena_free(graph_exec_arena);
357
0
  ccv_nnc_symbolic_graph_free(symbolic_graph);
358
0
}
359
360
TEST_CASE("fast gelu in float")
361
1
{
362
1
  GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GELU_FORWARD, CCV_NNC_BACKEND_GPU_REF) || ccv_nnc_cmd_ok(CCV_NNC_GELU_FORWARD, CCV_NNC_BACKEND_MPS));
363
1
  ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new();
364
1
  ccv_nnc_tensor_symbol_t a = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NCHW(000, 32F, 20, 10), "a");
365
1
  ccv_nnc_tensor_symbol_t b = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NCHW(000, 32F, 20, 10), "b");
366
1
  ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_GELU_FORWARD(1), TENSOR_SYMBOL_LIST(a), TENSOR_SYMBOL_LIST(b), "gelu");
367
1
  ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
368
1
  SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH);
369
1
  ccv_nnc_graph_t* graph = 0;
370
1
  ccv_nnc_tensor_arena_t* tensor_arena = 0;
371
1
  ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0;
372
1
  ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena);
373
1
  GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH);
374
1
  ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 20, 10), 0);
375
1
  dsfmt_t dsfmt;
376
1
  dsfmt_init_gen_rand(&dsfmt, 0);
377
1
  int i;
378
201
  for (i = 0; i < 20 * 10; 
i++200
)
379
200
    x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt);
380
1
  ccv_nnc_tensor_t* const a_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, a);
381
1
  ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(a_tensor), 0);
382
1
  ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0);
383
1
  ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 20, 10), 0);
384
1
  ccv_nnc_tensor_t* const b_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, b);
385
1
  ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(b_tensor), TENSOR_LIST(y_tensor), 0);
386
1
  ccv_nnc_tensor_t* const ty = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 20, 10), 0);
387
1
  ccv_nnc_cmd_exec(CMD_GELU_FORWARD(1), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(ty), 0);
388
1
  REQUIRE_TENSOR_EQ(ty, y_tensor, "gelu from cudnn should match from CPU");
389
1
  ccv_nnc_tensor_free(x_tensor);
390
1
  ccv_nnc_tensor_free(y_tensor);
391
1
  ccv_nnc_tensor_free(ty);
392
1
  ccv_nnc_graph_free(graph);
393
1
  ccv_nnc_tensor_arena_free(tensor_arena);
394
1
  ccv_nnc_graph_exec_arena_free(graph_exec_arena);
395
1
  ccv_nnc_symbolic_graph_free(symbolic_graph);
396
1
}
397
398
TEST_CASE("fast gelu in half precision")
399
1
{
400
1
  GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GELU_FORWARD, CCV_NNC_BACKEND_GPU_REF) || ccv_nnc_cmd_ok(CCV_NNC_GELU_FORWARD, CCV_NNC_BACKEND_MPS));
401
1
  ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new();
402
1
  ccv_nnc_tensor_symbol_t a = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NCHW(000, 16F, 20, 10), "a");
403
1
  ccv_nnc_tensor_symbol_t b = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NCHW(000, 16F, 20, 10), "b");
404
1
  ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_GELU_FORWARD(1), TENSOR_SYMBOL_LIST(a), TENSOR_SYMBOL_LIST(b), "gelu");
405
1
  ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
406
1
  SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH);
407
1
  ccv_nnc_graph_t* graph = 0;
408
1
  ccv_nnc_tensor_arena_t* tensor_arena = 0;
409
1
  ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0;
410
1
  ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, 0, 0, 0, 0, SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena);
411
1
  GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH);
412
1
  ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 20, 10), 0);
413
1
  dsfmt_t dsfmt;
414
1
  dsfmt_init_gen_rand(&dsfmt, 0);
415
1
  int i;
416
201
  for (i = 0; i < 20 * 10; 
i++200
)
417
200
    x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt);
418
1
  ccv_nnc_tensor_t* const a_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, a);
419
1
  ccv_nnc_tensor_t* const x16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(16F, 20, 10), 0);
420
1
  ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(x16_tensor), 0);
421
1
  ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x16_tensor), TENSOR_LIST(a_tensor), 0);
422
1
  ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0);
423
1
  ccv_nnc_tensor_t* const y16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(16F, 20, 10), 0);
424
1
  ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 20, 10), 0);
425
1
  ccv_nnc_tensor_t* const b_tensor = ccv_nnc_tensor_from_symbol(tensor_arena, b);
426
1
  ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(b_tensor), TENSOR_LIST(y16_tensor), 0);
427
1
  ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(y16_tensor), TENSOR_LIST(y_tensor), 0);
428
1
  ccv_nnc_tensor_t* const ty = ccv_nnc_tensor_new(0, CPU_TENSOR_NCHW(32F, 20, 10), 0);
429
1
  ccv_nnc_cmd_exec(CMD_GELU_FORWARD(1), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(ty), 0);
430
1
  REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, ty->data.f32, y_tensor->data.f32, 20 * 10, 1e-3, "gelu from cudnn should match from CPU");
431
1
  ccv_nnc_tensor_free(x_tensor);
432
1
  ccv_nnc_tensor_free(x16_tensor);
433
1
  ccv_nnc_tensor_free(y16_tensor);
434
1
  ccv_nnc_tensor_free(y_tensor);
435
1
  ccv_nnc_tensor_free(ty);
436
1
  ccv_nnc_graph_free(graph);
437
1
  ccv_nnc_tensor_arena_free(tensor_arena);
438
1
  ccv_nnc_graph_exec_arena_free(graph_exec_arena);
439
1
  ccv_nnc_symbolic_graph_free(symbolic_graph);
440
1
}
441
442
TEST_CASE("fast gelu gradient in float")
443
1
{
444
1
  GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GELU_FORWARD, CCV_NNC_BACKEND_GPU_REF) &&
445
1
    ccv_nnc_cmd_ok(CCV_NNC_GELU_BACKWARD, CCV_NNC_BACKEND_GPU_REF));
446
1
  ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new();
447
1
  ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10, 100), "x");
448
1
  ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 32F, 10, 100), "y");
449
1
  ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_GELU_FORWARD(1), TENSOR_SYMBOL_LIST(x), TENSOR_SYMBOL_LIST(y), "gelu");
450
1
  ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
451
1
  ccv_nnc_symbolic_graph_backward(symbolic_graph, TENSOR_SYMBOL_LIST(y), TENSOR_SYMBOL_LIST(x), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph));
452
1
  ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
453
1
  SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH);
454
1
  ccv_nnc_tensor_symbol_t dy = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, y);
455
1
  ccv_nnc_tensor_symbol_t dx = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, x);
456
1
  ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0);
457
1
  dsfmt_t dsfmt;
458
1
  dsfmt_init_gen_rand(&dsfmt, 0);
459
1
  int i;
460
1.00k
  for (i = 0; i < 10 * 100; 
i++1.00k
)
461
1.00k
    x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt);
462
1
  ccv_nnc_tensor_t* const dy_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0);
463
1.00k
  for (i = 0; i < 10 * 100; 
i++1.00k
)
464
1.00k
    dy_tensor->data.f32[i] = 0;
465
11
  for (i = 0; i < 10; 
i++10
)
466
10
    dy_tensor->data.f32[i * 100 + i] = 1;
467
1
  ccv_nnc_tensor_t* const dyt = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 32F, 10, 100), 0);
468
1
  ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor), TENSOR_LIST(dyt), 0);
469
1
  ccv_nnc_graph_t* graph = 0;
470
1
  ccv_nnc_tensor_arena_t* tensor_arena = 0;
471
1
  ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0;
472
1
  ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, TENSOR_BIND_MAP(KV(dy, dyt)), TENSOR_SYMBOL_LIST(y), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena);
473
1
  GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH);
474
1
  ccv_nnc_tensor_t* const xt = ccv_nnc_tensor_from_symbol(tensor_arena, x);
475
1
  ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(xt), 0);
476
1
  ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0);
477
1
  ccv_nnc_tensor_t* const dx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0);
478
1
  ccv_nnc_tensor_t* const dxt = ccv_nnc_tensor_from_symbol(tensor_arena, dx);
479
1
  ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0);
480
1
  ccv_nnc_tensor_t* const yt = ccv_nnc_tensor_from_symbol(tensor_arena, y);
481
1
  ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dxt), TENSOR_LIST(dx_tensor), 0);
482
1
  ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(yt), TENSOR_LIST(y_tensor), 0);
483
1
  ccv_nnc_tensor_t* const ty_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0);
484
1
  ccv_nnc_cmd_exec(CMD_GELU_FORWARD(1), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(ty_tensor), 0);
485
1
  REQUIRE_TENSOR_EQ(ty_tensor, y_tensor, "forward pass should match");
486
1
  ccv_nnc_tensor_t* const tdx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0);
487
1
  ccv_nnc_cmd_exec(CMD_GELU_BACKWARD(1), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor, x_tensor, 0), TENSOR_LIST(tdx_tensor), 0);
488
1
  REQUIRE_TENSOR_EQ(tdx_tensor, dx_tensor, "backward pass should match");
489
1
  ccv_nnc_tensor_free(x_tensor);
490
1
  ccv_nnc_tensor_free(y_tensor);
491
1
  ccv_nnc_tensor_free(dx_tensor);
492
1
  ccv_nnc_tensor_free(dy_tensor);
493
1
  ccv_nnc_tensor_free(ty_tensor);
494
1
  ccv_nnc_tensor_free(tdx_tensor);
495
1
  ccv_nnc_tensor_free(dyt);
496
1
  ccv_nnc_graph_free(graph);
497
1
  ccv_nnc_tensor_arena_free(tensor_arena);
498
1
  ccv_nnc_graph_exec_arena_free(graph_exec_arena);
499
1
  ccv_nnc_symbolic_graph_free(symbolic_graph);
500
1
}
501
502
TEST_CASE("fast gelu gradient in half precision")
503
1
{
504
1
  GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_GELU_FORWARD, CCV_NNC_BACKEND_GPU_REF) &&
505
1
    ccv_nnc_cmd_ok(CCV_NNC_GELU_BACKWARD, CCV_NNC_BACKEND_GPU_REF));
506
1
  ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new();
507
1
  ccv_nnc_tensor_symbol_t x = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 16F, 10, 100), "x");
508
1
  ccv_nnc_tensor_symbol_t y = ccv_nnc_tensor_symbol_new(symbolic_graph, GPU_TENSOR_NHWC(000, 16F, 10, 100), "y");
509
1
  ccv_nnc_graph_exec_symbol_new(symbolic_graph, CMD_GELU_FORWARD(1), TENSOR_SYMBOL_LIST(x), TENSOR_SYMBOL_LIST(y), "gelu");
510
1
  ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
511
1
  ccv_nnc_symbolic_graph_backward(symbolic_graph, TENSOR_SYMBOL_LIST(y), TENSOR_SYMBOL_LIST(x), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph));
512
1
  ccv_nnc_graph_exec_symbol_autogen(symbolic_graph, 0, 0, CCV_NNC_AUTOGEN_ALL_EXECS | CCV_NNC_AUTOGEN_SOURCES_AND_DESTINATIONS);
513
1
  SYMBOLIC_GRAPH_GEN(symbolic_graph, CCV_NNC_LONG_DOT_GRAPH);
514
1
  ccv_nnc_tensor_symbol_t dy = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, y);
515
1
  ccv_nnc_tensor_symbol_t dx = ccv_nnc_tensor_symbol_for_backward(symbolic_graph, x);
516
1
  ccv_nnc_tensor_t* const x_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0);
517
1
  dsfmt_t dsfmt;
518
1
  dsfmt_init_gen_rand(&dsfmt, 0);
519
1
  int i;
520
1.00k
  for (i = 0; i < 10 * 100; 
i++1.00k
)
521
1.00k
    x_tensor->data.f32[i] = dsfmt_genrand_open_close(&dsfmt);
522
1
  ccv_nnc_tensor_t* const dy_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0);
523
1.00k
  for (i = 0; i < 10 * 100; 
i++1.00k
)
524
1.00k
    dy_tensor->data.f32[i] = 0;
525
11
  for (i = 0; i < 10; 
i++10
)
526
10
    dy_tensor->data.f32[i * 100 + i] = 1;
527
1
  ccv_nnc_tensor_t* const dy16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 100), 0);
528
1
  ccv_nnc_tensor_t* const dyt = ccv_nnc_tensor_new(0, GPU_TENSOR_NHWC(000, 16F, 10, 100), 0);
529
1
  ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor), TENSOR_LIST(dy16_tensor), 0);
530
1
  ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dy16_tensor), TENSOR_LIST(dyt), 0);
531
1
  ccv_nnc_graph_t* graph = 0;
532
1
  ccv_nnc_tensor_arena_t* tensor_arena = 0;
533
1
  ccv_nnc_graph_exec_arena_t* graph_exec_arena = 0;
534
1
  ccv_nnc_symbolic_graph_compile(symbolic_graph, ccv_nnc_default_compile_params, TENSOR_BIND_MAP(KV(dy, dyt)), TENSOR_SYMBOL_LIST(y), SYMBOLIC_GRAPH_SOURCES(symbolic_graph), SYMBOLIC_GRAPH_DESTINATIONS(symbolic_graph), &graph, &tensor_arena, &graph_exec_arena);
535
1
  GRAPH_GEN(graph, CCV_NNC_LONG_DOT_GRAPH);
536
1
  ccv_nnc_tensor_t* const xt = ccv_nnc_tensor_from_symbol(tensor_arena, x);
537
1
  ccv_nnc_tensor_t* const x16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 100), 0);
538
1
  ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(x16_tensor), 0);
539
1
  ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(x16_tensor), TENSOR_LIST(xt), 0);
540
1
  ccv_nnc_graph_run(graph, 0, TRAVERSE_FULL, 0, 0);
541
1
  ccv_nnc_tensor_t* const dx16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 100), 0);
542
1
  ccv_nnc_tensor_t* const dx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0);
543
1
  ccv_nnc_tensor_t* const dxt = ccv_nnc_tensor_from_symbol(tensor_arena, dx);
544
1
  ccv_nnc_tensor_t* const y16_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(16F, 10, 100), 0);
545
1
  ccv_nnc_tensor_t* const y_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0);
546
1
  ccv_nnc_tensor_t* const yt = ccv_nnc_tensor_from_symbol(tensor_arena, y);
547
1
  ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dxt), TENSOR_LIST(dx16_tensor), 0);
548
1
  ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(dx16_tensor), TENSOR_LIST(dx_tensor), 0);
549
1
  ccv_nnc_cmd_exec(CMD_DATA_TRANSFER_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(yt), TENSOR_LIST(y16_tensor), 0);
550
1
  ccv_nnc_cmd_exec(CMD_DATATYPE_CONVERSION_FORWARD(), ccv_nnc_no_hint, 0, TENSOR_LIST(y16_tensor), TENSOR_LIST(y_tensor), 0);
551
1
  ccv_nnc_tensor_t* const ty_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0);
552
1
  ccv_nnc_cmd_exec(CMD_GELU_FORWARD(1), ccv_nnc_no_hint, 0, TENSOR_LIST(x_tensor), TENSOR_LIST(ty_tensor), 0);
553
1
  REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, ty_tensor->data.f32, y_tensor->data.f32, 10 * 100, 1e-3, "forward pass should match");
554
1
  ccv_nnc_tensor_t* const tdx_tensor = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, 10, 100), 0);
555
1
  ccv_nnc_cmd_exec(CMD_GELU_BACKWARD(1), ccv_nnc_no_hint, 0, TENSOR_LIST(dy_tensor, x_tensor, 0), TENSOR_LIST(tdx_tensor), 0);
556
1
  REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, tdx_tensor->data.f32, dx_tensor->data.f32, 10 * 100, 1e-3, "backward pass should match");
557
1
  ccv_nnc_tensor_free(x_tensor);
558
1
  ccv_nnc_tensor_free(x16_tensor);
559
1
  ccv_nnc_tensor_free(y_tensor);
560
1
  ccv_nnc_tensor_free(y16_tensor);
561
1
  ccv_nnc_tensor_free(dx_tensor);
562
1
  ccv_nnc_tensor_free(dx16_tensor);
563
1
  ccv_nnc_tensor_free(dy_tensor);
564
1
  ccv_nnc_tensor_free(dy16_tensor);
565
1
  ccv_nnc_tensor_free(ty_tensor);
566
1
  ccv_nnc_tensor_free(tdx_tensor);
567
1
  ccv_nnc_tensor_free(dyt);
568
1
  ccv_nnc_graph_free(graph);
569
1
  ccv_nnc_tensor_arena_free(tensor_arena);
570
1
  ccv_nnc_graph_exec_arena_free(graph_exec_arena);
571
1
  ccv_nnc_symbolic_graph_free(symbolic_graph);
572
1
}
573
574
#include "case_main.h"