1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
|
/*
* Copyright (c) 2017-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to
* deal in the Software without restriction, including without limitation the
* rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
* sell copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#include "helpers.h"
#define ADD_OP(a, b) ((a) + (b))
#define SUB_OP(a, b) ((a) - (b))
#define MUL_OP(a, b) ((a) * (b))
#define INVSQRT_OP(a) rsqrt((a))
#define SQCVT_SAT(a) (a)
#if defined(VEC_SIZE) && defined(DATA_TYPE) && defined(ACTIVATION_TYPE)
#include "activation_float_helpers.h"
/** Apply batch normalization.
*
* @note It is possible to select the activation function to apply using -DACTIVATION_TYPE e.g. -DACTIVATION_TYPE=relu
* @note A, B variables required by some activation functions are set using -DA_VAL= and -DB_VAL= respectively
*
* @param[in] input_ptr Pointer to the first source tensor. Supported data types: F16/F32
* @param[in] input_stride_x Stride of the first source tensor in X dimension (in bytes)
* @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] input_stride_y Stride of the first source tensor in Y dimension (in bytes)
* @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] input_stride_z Stride of the first source tensor in Z dimension (in bytes)
* @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] input_offset_first_element_in_bytes The offset of the first element in the first source tensor
* @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr
* @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes)
* @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes)
* @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] output_stride_z Stride of the destination tensor in Z dimension (in bytes)
* @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor
* @param[in] mean_ptr Pointer to the mean source tensor. Supported data types: same as @p input_ptr
* @param[in] mean_stride_x Stride of the mean source tensor in X dimension (in bytes)
* @param[in] mean_step_x mean_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] mean_offset_first_element_in_bytes The offset of the first element in the mean source tensor
* @param[in] var_ptr Pointer to the var tensor. Supported data types: same as @p input_ptr
* @param[in] var_stride_x Stride of the var tensor in X dimension (in bytes)
* @param[in] var_step_x var_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] var_offset_first_element_in_bytes The offset of the first element in the var source tensor
* @param[in] beta_ptr Pointer to the beta source tensor. Supported data types: same as @p input_ptr
* @param[in] beta_stride_x Stride of the beta source tensor in X dimension (in bytes)
* @param[in] beta_step_x beta_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] beta_offset_first_element_in_bytes The offset of the first element in the beta source tensor
* @param[in] gamma_ptr Pointer to the gamma source tensor. Supported data types: same as @p input_ptr
* @param[in] gamma_stride_x Stride of the gamma source tensor in X dimension (in bytes)
* @param[in] gamma_step_x gamma_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] gamma_offset_first_element_in_bytes The offset of the first element in the gamma source tensor
* @param[in] epsilon Epsilon parameter in the batch normalization equation
*/
__kernel void batchnormalization_layer_nchw(TENSOR3D_DECLARATION(input),
#ifndef IN_PLACE
TENSOR3D_DECLARATION(output),
#endif /* not IN_PLACE */
VECTOR_DECLARATION(mean),
VECTOR_DECLARATION(var),
#ifndef USE_DEFAULT_BETA
VECTOR_DECLARATION(beta),
#endif /* USE_DEFAULT_BETA */
#ifndef USE_DEFAULT_GAMMA
VECTOR_DECLARATION(gamma),
#endif /* USE_DEFAULT_GAMMA */
float epsilon)
{
Tensor3D in = CONVERT_TO_TENSOR3D_STRUCT(input);
#ifdef IN_PLACE
Tensor3D out = in;
#else /* IN_PLACE */
Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(output);
#endif /* IN_PLACE */
Vector mean = CONVERT_TO_VECTOR_STRUCT(mean);
Vector var = CONVERT_TO_VECTOR_STRUCT(var);
#ifndef USE_DEFAULT_BETA
Vector beta = CONVERT_TO_VECTOR_STRUCT(beta);
#endif /* USE_DEFAULT_BETA */
#ifndef USE_DEFAULT_GAMMA
Vector gamma = CONVERT_TO_VECTOR_STRUCT(gamma);
#endif /* USE_DEFAULT_GAMMA */
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
data = 0;
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
denominator = 0;
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
numerator = 0;
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
x_bar = 0;
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
res = 0;
const int current_slice = get_global_id(2);
data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)in.ptr);
denominator = *((__global DATA_TYPE *)(var.ptr + current_slice * var.stride_x));
denominator = INVSQRT_OP(ADD_OP(denominator, ((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))SQCVT_SAT(epsilon))));
// Calculate x bar and store results
numerator = *((__global DATA_TYPE *)(mean.ptr + current_slice * mean.stride_x));
numerator = SUB_OP(data, numerator);
x_bar = MUL_OP(numerator, denominator);
#ifndef USE_DEFAULT_GAMMA
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
gamma_vec = *((__global DATA_TYPE *)(gamma.ptr + current_slice * gamma.stride_x));
res = MUL_OP(gamma_vec, x_bar);
#else /* USE_DEFAULT_GAMMA */
// gamma is equal to 1, no need to perform multiplications
res = x_bar;
#endif /* USE_DEFAULT_GAMMA */
#ifndef USE_DEFAULT_BETA
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
beta_vec = *((__global DATA_TYPE *)(beta.ptr + current_slice * beta.stride_x));
// beta is not zero, hence we need to perform the addition
res = ADD_OP(res, beta_vec);
#endif /* USE_DEFAULT_BETA */
res = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, res, A_VAL, B_VAL);
VSTORE(VEC_SIZE)
(res, 0, (__global DATA_TYPE *)out.ptr);
}
/** Apply batch normalization on tensors with NHWC format.
*
* @note It is possible to select the activation function to apply using -DACTIVATION_TYPE e.g. -DACTIVATION_TYPE=relu
* @note A, B variables required by some activation functions are set using -DA_VAL= and -DB_VAL= respectively
*
* @param[in] input_ptr Pointer to the first source tensor. Supported data types: F16/F32
* @param[in] input_stride_x Stride of the first source tensor in X dimension (in bytes)
* @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] input_stride_y Stride of the first source tensor in Y dimension (in bytes)
* @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] input_stride_z Stride of the first source tensor in Z dimension (in bytes)
* @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] input_offset_first_element_in_bytes The offset of the first element in the first source tensor
* @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr
* @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes)
* @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes)
* @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] output_stride_z Stride of the destination tensor in Z dimension (in bytes)
* @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor
* @param[in] mean_ptr Pointer to the mean source tensor. Supported data types: same as @p input_ptr
* @param[in] mean_stride_x Stride of the mean source tensor in X dimension (in bytes)
* @param[in] mean_step_x mean_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] mean_offset_first_element_in_bytes The offset of the first element in the mean source tensor
* @param[in] var_ptr Pointer to the var tensor. Supported data types: same as @p input_ptr
* @param[in] var_stride_x Stride of the var tensor in X dimension (in bytes)
* @param[in] var_step_x var_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] var_offset_first_element_in_bytes The offset of the first element in the var source tensor
* @param[in] beta_ptr Pointer to the beta source tensor. Supported data types: same as @p input_ptr
* @param[in] beta_stride_x Stride of the beta source tensor in X dimension (in bytes)
* @param[in] beta_step_x beta_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] beta_offset_first_element_in_bytes The offset of the first element in the beta source tensor
* @param[in] gamma_ptr Pointer to the gamma source tensor. Supported data types: same as @p input_ptr
* @param[in] gamma_stride_x Stride of the gamma source tensor in X dimension (in bytes)
* @param[in] gamma_step_x gamma_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] gamma_offset_first_element_in_bytes The offset of the first element in the gamma source tensor
* @param[in] epsilon Epsilon parameter in the batch normalization equation
*/
__kernel void batchnormalization_layer_nhwc(TENSOR3D_DECLARATION(input),
#ifndef IN_PLACE
TENSOR3D_DECLARATION(output),
#endif /* not IN_PLACE */
VECTOR_DECLARATION(mean),
VECTOR_DECLARATION(var),
#ifndef USE_DEFAULT_BETA
VECTOR_DECLARATION(beta),
#endif /* USE_DEFAULT_BETA */
#ifndef USE_DEFAULT_GAMMA
VECTOR_DECLARATION(gamma),
#endif /* USE_DEFAULT_GAMMA */
float epsilon)
{
Tensor3D in = CONVERT_TO_TENSOR3D_STRUCT(input);
#ifdef IN_PLACE
Tensor3D out = in;
#else /* IN_PLACE */
Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(output);
#endif /* IN_PLACE */
Vector mean = CONVERT_TO_VECTOR_STRUCT(mean);
Vector var = CONVERT_TO_VECTOR_STRUCT(var);
#ifndef USE_DEFAULT_BETA
Vector beta = CONVERT_TO_VECTOR_STRUCT(beta);
#endif /* USE_DEFAULT_BETA */
#ifndef USE_DEFAULT_GAMMA
Vector gamma = CONVERT_TO_VECTOR_STRUCT(gamma);
#endif /* USE_DEFAULT_GAMMA */
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
data = 0;
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
denominator = 0;
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
numerator = 0;
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
x_bar = 0;
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
res = 0;
const int current_slice = get_global_id(0);
data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)in.ptr);
denominator = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(var.ptr + current_slice * VEC_SIZE * var.stride_x));
denominator = INVSQRT_OP(ADD_OP(denominator, ((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))SQCVT_SAT(epsilon))));
// Calculate x bar and store results
numerator = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(mean.ptr + current_slice * VEC_SIZE * mean.stride_x));
numerator = SUB_OP(data, numerator);
x_bar = MUL_OP(numerator, denominator);
#ifndef USE_DEFAULT_GAMMA
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
gamma_vec = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(gamma.ptr + current_slice * VEC_SIZE * gamma.stride_x));
res = MUL_OP(gamma_vec, x_bar);
#else /* USE_DEFAULT_GAMMA */
// gamma is equal to 1, no need to perform multiplications
res = x_bar;
#endif /* USE_DEFAULT_GAMMA */
#ifndef USE_DEFAULT_BETA
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
beta_vec = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(beta.ptr + current_slice * VEC_SIZE * beta.stride_x));
// beta is not zero, hence we need to perform the addition
res = ADD_OP(res, beta_vec);
#endif /* USE_DEFAULT_BETA */
res = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, res, A_VAL, B_VAL);
VSTORE(VEC_SIZE)
(res, 0, (__global DATA_TYPE *)out.ptr);
}
#endif /* defined(VEC_SIZE) && defined(DATA_TYPE) && defined(DATA_TYPE)*/
#if defined(NUM_CHANNELS) && defined(DATA_TYPE) && defined(EPSILON)
/** Fuse batchnorm parameters to convolution layer parameters
*
* @attention Data type should be passed using the -DDATA_TYPE compile flag, e.g. -DDATA_TYPE=float
* @attention Input tensor depth should be given as a preprocessor argument using -DNUM_CHANNELS=size. e.g. -DNUM_CHANNELS=16
* @attention Batch normalization epsilon parameter should be given as a preprocessor argument with -DEPSILON=value. e.g. -DEPSILON=0.001f
*
* @param[in] conv_w_ptr Pointer to the source tensor. Supported data types: F16/F32
* @param[in] conv_w_stride_x Stride of the source tensor in X dimension (in bytes)
* @param[in] conv_w_step_x input_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] conv_w_stride_y Stride of the source tensor in Y dimension (in bytes)
* @param[in] conv_w_step_y input_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] conv_w_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] conv_w_step_z input_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] conv_w_stride_w Stride of the source tensor in W dimension (in bytes)
* @param[in] conv_w_step_w input_stride_w * number of elements along W processed per workitem(in bytes)
* @param[in] conv_w_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[in] bn_mean_ptr Pointer to the mean source tensor. Supported data types: same as @p input_ptr
* @param[in] bn_mean_stride_x Stride of the mean source tensor in X dimension (in bytes)
* @param[in] bn_mean_step_x bn_mean_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] bn_mean_offset_first_element_in_bytes The offset of the first element in the mean source tensor
* @param[in] bn_var_ptr Pointer to the var tensor. Supported data types: same as @p input_ptr
* @param[in] bn_var_stride_x Stride of the var tensor in X dimension (in bytes)
* @param[in] bn_var_step_x bn_var_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] bn_var_offset_first_element_in_bytes The offset of the first element in the var source tensor
* @param[out] fused_w_ptr Pointer to the destination weights tensors. Supported data types: same as @p input_ptr
* @param[in] fused_w_stride_x Stride of the destination tensor in X dimension (in bytes)
* @param[in] fused_w_step_x fused_w_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] fused_w_stride_y Stride of the destination tensor in Y dimension (in bytes)
* @param[in] fused_w_step_y fused_w_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] fused_w_stride_z Stride of the destination tensor in Z dimension (in bytes)
* @param[in] fused_w_step_z fused_w_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] fused_w_stride_w Stride of the destination tensor in W dimension (in bytes)
* @param[in] fused_w_step_w fused_w_stride_w * number of elements along W processed per workitem(in bytes)
* @param[in] fused_w_offset_first_element_in_bytes The offset of the first element in the destination tensor
* @param[in] fused_b_ptr Pointer to the destination bias tensor. Supported data types: same as @p input_ptr
* @param[in] fused_b_stride_x Stride of the bias source tensor in X dimension (in bytes)
* @param[in] fused_b_step_x fused_b_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] fused_b_offset_first_element_in_bytes The offset of the first element in the destination tensor
* @param[in] conv_b_ptr Pointer to the source bias tensor. Supported data types: same as @p input_ptr
* @param[in] conv_b_stride_x Stride of the beta source tensor in X dimension (in bytes)
* @param[in] conv_b_step_x conv_b_beta_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] conv_b_offset_first_element_in_bytes The offset of the first element in the source bias tensor
* @param[in] bn_beta_ptr Pointer to the beta source tensor. Supported data types: same as @p input_ptr
* @param[in] bn_beta_stride_x Stride of the beta source tensor in X dimension (in bytes)
* @param[in] bn_beta_step_x bn_beta_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] bn_beta_offset_first_element_in_bytes The offset of the first element in the beta source tensor
* @param[in] bn_gamma_ptr Pointer to the gamma source tensor. Supported data types: same as @p input_ptr
* @param[in] bn_gamma_stride_x Stride of the gamma source tensor in X dimension (in bytes)
* @param[in] bn_gamma_step_x bn_gamma_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] bn_gamma_offset_first_element_in_bytes The offset of the first element in the gamma source tensor
* @param[in] epsilon Epsilon parameter in the batch normalization equation
*/
__kernel void fuse_batchnormalization_layer(TENSOR4D_DECLARATION(conv_w),
VECTOR_DECLARATION(bn_mean),
VECTOR_DECLARATION(bn_var)
#ifndef IN_PLACE_W
,
TENSOR4D_DECLARATION(fused_w)
#endif /* not IN_PLACE_W */
#ifndef IN_PLACE_B
,
VECTOR_DECLARATION(fused_b)
#endif /* not IN_PLACE_B */
#ifdef HAS_BIAS
,
VECTOR_DECLARATION(conv_b)
#endif /* HAS_BIAS */
#ifndef USE_DEFAULT_BETA
,
VECTOR_DECLARATION(bn_beta)
#endif /* USE_DEFAULT_BETA */
#ifndef USE_DEFAULT_GAMMA
,
VECTOR_DECLARATION(bn_gamma)
#endif /* USE_DEFAULT_GAMMA */
)
{
Tensor4D conv_w = CONVERT_TO_TENSOR4D_STRUCT(conv_w, NUM_CHANNELS);
Vector bn_mean = CONVERT_TO_VECTOR_STRUCT_NO_STEP(bn_mean);
Vector bn_var = CONVERT_TO_VECTOR_STRUCT_NO_STEP(bn_var);
// Conditional ops
#ifdef HAS_BIAS
Vector conv_b = CONVERT_TO_VECTOR_STRUCT_NO_STEP(conv_b);
#endif /* HAS_BIAS */
#ifndef USE_DEFAULT_BETA
Vector bn_beta = CONVERT_TO_VECTOR_STRUCT_NO_STEP(bn_beta);
#endif /* USE_DEFAULT_BETA */
#ifndef USE_DEFAULT_GAMMA
Vector bn_gamma = CONVERT_TO_VECTOR_STRUCT_NO_STEP(bn_gamma);
#endif /* USE_DEFAULT_GAMMA */
// In-place ops
#ifdef IN_PLACE_W
Tensor4D fused_w = conv_w;
uint fused_w_stride_x = conv_w_stride_x;
#else /* IN_PLACE_W */
Tensor4D fused_w = CONVERT_TO_TENSOR4D_STRUCT(fused_w, NUM_CHANNELS);
#endif /* IN_PLACE_W */
#ifdef IN_PLACE_B
Vector fused_b = conv_b;
#else /* IN_PLACE_B */
Vector fused_b = CONVERT_TO_VECTOR_STRUCT_NO_STEP(fused_b);
#endif /* IN_PLACE_B */
const int current_slice = get_global_id(2) / NUM_CHANNELS;
#if defined(VEC_SIZE) && defined(LAST_ACCESSED_X)
// Check if access on width gets out of bounds
// If it does shift access vector to access elements within bounds
const int xi = (int)(get_global_id(0) * VEC_SIZE);
conv_w.ptr -= max(xi - (int)LAST_ACCESSED_X, 0) * conv_w_stride_x;
fused_w.ptr -= max(xi - (int)LAST_ACCESSED_X, 0) * fused_w_stride_x;
// Load W
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
wn = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)conv_w.ptr);
#else // !defined(VEC_SIZE) || !defined(LAST_ACCESSED_X)
DATA_TYPE wn = *((__global DATA_TYPE *)(conv_w.ptr));
#endif // defined(VEC_SIZE) && defined(LAST_ACCESSED_X)
// rvar = 1 / sqrt(var + epsilon)
const DATA_TYPE var = *((__global DATA_TYPE *)(bn_var.ptr + current_slice * bn_var.stride_x));
const DATA_TYPE rvar = INVSQRT_OP(ADD_OP(var, SQCVT_SAT((float)EPSILON)));
wn *= rvar;
// Load b
const DATA_TYPE mean = *((__global DATA_TYPE *)(bn_mean.ptr + current_slice * bn_mean.stride_x));
DATA_TYPE bn = 0;
#ifdef HAS_BIAS
bn = *((__global DATA_TYPE *)(conv_b.ptr + current_slice * conv_b.stride_x));
#endif /* HAS_BIAS */
bn = (bn - mean) * rvar;
#ifndef USE_DEFAULT_GAMMA
const DATA_TYPE gamma_scalar = *((__global DATA_TYPE *)(bn_gamma.ptr + current_slice * bn_gamma.stride_x));
wn *= gamma_scalar;
bn *= gamma_scalar;
#endif /* USE_DEFAULT_GAMMA */
#ifndef USE_DEFAULT_BETA
const DATA_TYPE beta_scalar = *((__global DATA_TYPE *)(bn_beta.ptr + current_slice * bn_beta.stride_x));
bn += beta_scalar;
#endif /* USE_DEFAULT_BETA */
#if defined(VEC_SIZE) && defined(LAST_ACCESSED_X)
// Store updated weights
VSTORE(VEC_SIZE)
(wn, 0, (__global DATA_TYPE *)fused_w.ptr);
#else // !defined(VEC_SIZE) || !defined(LAST_ACCESSED_X)
*((__global DATA_TYPE *)(fused_w.ptr)) = wn;
#endif // defined(VEC_SIZE) && defined(LAST_ACCESSED_X)
// Store updated bias
*((__global DATA_TYPE *)(fused_b.ptr + current_slice * fused_b.stride_x)) = bn;
}
#endif /* defined(NUM_CHANNELS) && defined(DATA_TYPE) && defined(EPSILON) */
|
