| 1 | namespace dnnl { |
|---|---|
| 2 | namespace impl { |
| 3 | namespace gpu { |
| 4 | namespace ocl { |
| 5 | const char *ref_bnorm_kernel = R"==(/******************************************************************************* )==" "\n" |
| 6 | R"==(* Copyright 2019-2022 Intel Corporation )==" "\n" |
| 7 | R"==(* )==" "\n" |
| 8 | R"==(* Licensed under the Apache License, Version 2.0 (the "License"); )==" "\n" |
| 9 | R"==(* you may not use this file except in compliance with the License. )==" "\n" |
| 10 | R"==(* You may obtain a copy of the License at )==" "\n" |
| 11 | R"==(* )==" "\n" |
| 12 | R"==(* http: )==" "\n" |
| 13 | R"==(* )==" "\n" |
| 14 | R"==(* Unless required by applicable law or agreed to in writing, software )==" "\n" |
| 15 | R"==(* distributed under the License is distributed on an "AS IS" BASIS, )==" "\n" |
| 16 | R"==(* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. )==" "\n" |
| 17 | R"==(* See the License for the specific language governing permissions and )==" "\n" |
| 18 | R"==(* limitations under the License. )==" "\n" |
| 19 | R"==(*******************************************************************************/ )==" "\n" |
| 20 | R"==(#if MB_BLOCK == 16 )==" "\n" |
| 21 | R"==(#define MB16 )==" "\n" |
| 22 | R"==(#define VECT_DT_N 8 )==" "\n" |
| 23 | R"==(#elif VECTORIZE_CALC_STATS == 1 )==" "\n" |
| 24 | R"==(#define VECT_DT_N VECT_SIZE )==" "\n" |
| 25 | R"==(#else )==" "\n" |
| 26 | R"==(#define VECT_DT_N 1 )==" "\n" |
| 27 | R"==(#endif )==" "\n" |
| 28 | R"==(#if VECT_DT_N == 1 )==" "\n" |
| 29 | R"==(#define VECT_CHAR_TO_INT convert_int )==" "\n" |
| 30 | R"==(#else )==" "\n" |
| 31 | R"==(#define VECT_CHAR_TO_INT CONCAT2(convert_int, VECT_DT_N) )==" "\n" |
| 32 | R"==(#endif )==" "\n" |
| 33 | R"==(#if USE_16MB_UNROLL == 0 && (CALCULATE_STATS == 1 || IS_BWD == 1) )==" "\n" |
| 34 | R"==(int reduce_index(int x[5]) { )==" "\n" |
| 35 | R"==(int dim[5] = {MB, IC, ID, IH, IW}; )==" "\n" |
| 36 | R"==(dim[REDUCE_DIM_IDX] = 1; )==" "\n" |
| 37 | R"==(return x[0] * (dim[2] * dim[3] * dim[4]) + x[2] * (dim[3] * dim[4]) )==" "\n" |
| 38 | R"==(+ x[3] * dim[4] + x[4]; )==" "\n" |
| 39 | R"==(} )==" "\n" |
| 40 | R"==(#endif )==" "\n" |
| 41 | R"==(#include "gpu/ocl/ocl_types.h" )==" "\n" |
| 42 | R"==(#if IS_FWD == 1 )==" "\n" |
| 43 | R"==(#if USE_16MB_UNROLL == 0 && CALCULATE_STATS == 1 )==" "\n" |
| 44 | R"==(#if VECTORIZE_CALC_STATS == 1 )==" "\n" |
| 45 | R"==(NAMED_KERNEL_ATTR(CALC) )==" "\n" |
| 46 | R"==(__kernel void calculate_mean(__global DATA_T *src, __global float *mean) { )==" "\n" |
| 47 | R"==(int x[5]; )==" "\n" |
| 48 | R"==(x[0] = GWS_GET_STAT_MB(); )==" "\n" |
| 49 | R"==(x[1] = GWS_GET_STAT_IC(); )==" "\n" |
| 50 | R"==(x[2] = GWS_GET_STAT_ID(); )==" "\n" |
| 51 | R"==(x[3] = GWS_GET_STAT_IH(); )==" "\n" |
| 52 | R"==(x[4] = GWS_GET_STAT_IW(); )==" "\n" |
| 53 | R"==(VECT_FLOAT_T vect_sum = 0; )==" "\n" |
| 54 | R"==(for (int i = 0; i < REDUCE_DIM; i += SUB_GROUP_SIZE * VECT_DT_N) { )==" "\n" |
| 55 | R"==(x[REDUCE_DIM_IDX] = i; )==" "\n" |
| 56 | R"==(int src_off = SRC_OFF(x[0], x[1], x[2], x[3], x[4]); )==" "\n" |
| 57 | R"==(vect_sum += CONVERT_VECT_FLOAT_T(AS_VECT_DATA_T( )==" "\n" |
| 58 | R"==(VECT_BLOCK_READ((const __global BLOCK_DATA_T *)&src[src_off]))); )==" "\n" |
| 59 | R"==(} )==" "\n" |
| 60 | R"==(#if VECT_DT_N == 1 )==" "\n" |
| 61 | R"==(float sum = vect_sum; )==" "\n" |
| 62 | R"==(#else )==" "\n" |
| 63 | R"==(float sum = 0; )==" "\n" |
| 64 | R"==(for (int i = 0; i < VECT_DT_N; ++i) { )==" "\n" |
| 65 | R"==(sum += vect_sum[i]; )==" "\n" |
| 66 | R"==(} )==" "\n" |
| 67 | R"==(#endif )==" "\n" |
| 68 | R"==(x[REDUCE_DIM_IDX] = 0; )==" "\n" |
| 69 | R"==(int reduce_idx = reduce_index(x); )==" "\n" |
| 70 | R"==(float total_sum = sub_group_reduce_add(sum); )==" "\n" |
| 71 | R"==(int local_id = get_sub_group_local_id(); )==" "\n" |
| 72 | R"==(if (local_id == 0) { mean[reduce_idx * IC + x[1]] = total_sum; } )==" "\n" |
| 73 | R"==(} )==" "\n" |
| 74 | R"==(NAMED_KERNEL_ATTR(CALC) )==" "\n" |
| 75 | R"==(__kernel void calculate_variance( )==" "\n" |
| 76 | R"==(__global DATA_T *src, __global float *mean, __global float *variance) { )==" "\n" |
| 77 | R"==(int x[5]; )==" "\n" |
| 78 | R"==(x[0] = GWS_GET_STAT_MB(); )==" "\n" |
| 79 | R"==(x[1] = GWS_GET_STAT_IC(); )==" "\n" |
| 80 | R"==(x[2] = GWS_GET_STAT_ID(); )==" "\n" |
| 81 | R"==(x[3] = GWS_GET_STAT_IH(); )==" "\n" |
| 82 | R"==(x[4] = GWS_GET_STAT_IW(); )==" "\n" |
| 83 | R"==(VECT_FLOAT_T mean_tmp = mean[x[1]]; )==" "\n" |
| 84 | R"==(VECT_FLOAT_T vect_sum = 0; )==" "\n" |
| 85 | R"==(for (int i = 0; i < REDUCE_DIM; i += SUB_GROUP_SIZE * VECT_DT_N) { )==" "\n" |
| 86 | R"==(x[REDUCE_DIM_IDX] = i; )==" "\n" |
| 87 | R"==(int src_off = SRC_OFF(x[0], x[1], x[2], x[3], x[4]); )==" "\n" |
| 88 | R"==(VECT_FLOAT_T v0 = CONVERT_VECT_FLOAT_T(AS_VECT_DATA_T( )==" "\n" |
| 89 | R"==(VECT_BLOCK_READ((const __global BLOCK_DATA_T *)&src[src_off]))); )==" "\n" |
| 90 | R"==(v0 -= mean_tmp; )==" "\n" |
| 91 | R"==(vect_sum += v0 * v0; )==" "\n" |
| 92 | R"==(} )==" "\n" |
| 93 | R"==(#if VECT_DT_N == 1 )==" "\n" |
| 94 | R"==(float sum = vect_sum; )==" "\n" |
| 95 | R"==(#else )==" "\n" |
| 96 | R"==(float sum = 0; )==" "\n" |
| 97 | R"==(for (int i = 0; i < VECT_DT_N; ++i) { )==" "\n" |
| 98 | R"==(sum += vect_sum[i]; )==" "\n" |
| 99 | R"==(} )==" "\n" |
| 100 | R"==(#endif )==" "\n" |
| 101 | R"==(x[REDUCE_DIM_IDX] = 0; )==" "\n" |
| 102 | R"==(int reduce_idx = reduce_index(x); )==" "\n" |
| 103 | R"==(float total_sum = sub_group_reduce_add(sum); )==" "\n" |
| 104 | R"==(int local_id = get_sub_group_local_id(); )==" "\n" |
| 105 | R"==(if (local_id == 0) { )==" "\n" |
| 106 | R"==(variance += MB * ID * IH * IW * IC / REDUCE_DIM; )==" "\n" |
| 107 | R"==(variance[reduce_idx * IC + x[1]] = total_sum; )==" "\n" |
| 108 | R"==(} )==" "\n" |
| 109 | R"==(} )==" "\n" |
| 110 | R"==(NAMED_KERNEL_ATTR(CALC) )==" "\n" |
| 111 | R"==(__kernel void calculate_mean_variance( )==" "\n" |
| 112 | R"==(__global DATA_T *src, __global float *mean, __global float *variance) { )==" "\n" |
| 113 | R"==(#if SKIP_REDUCE_STATS == 1 )==" "\n" |
| 114 | R"==(int x[5]; )==" "\n" |
| 115 | R"==(x[0] = GWS_GET_STAT_MB(); )==" "\n" |
| 116 | R"==(x[1] = GWS_GET_STAT_IC(); )==" "\n" |
| 117 | R"==(x[2] = GWS_GET_STAT_ID(); )==" "\n" |
| 118 | R"==(x[3] = GWS_GET_STAT_IH(); )==" "\n" |
| 119 | R"==(x[4] = GWS_GET_STAT_IW(); )==" "\n" |
| 120 | R"==(VECT_FLOAT_T src_sum = 0; )==" "\n" |
| 121 | R"==(VECT_FLOAT_T src_pow_sum = 0; )==" "\n" |
| 122 | R"==(for (int i = 0; i < REDUCE_DIM; i += SUB_GROUP_SIZE * VECT_DT_N) { )==" "\n" |
| 123 | R"==(x[REDUCE_DIM_IDX] = i; )==" "\n" |
| 124 | R"==(int src_off = SRC_OFF(x[0], x[1], x[2], x[3], x[4]); )==" "\n" |
| 125 | R"==(VECT_FLOAT_T src_vect = CONVERT_VECT_FLOAT_T(AS_VECT_DATA_T( )==" "\n" |
| 126 | R"==(VECT_BLOCK_READ((const __global BLOCK_DATA_T *)&src[src_off]))); )==" "\n" |
| 127 | R"==(src_sum += src_vect; )==" "\n" |
| 128 | R"==(src_pow_sum += src_vect * src_vect; )==" "\n" |
| 129 | R"==(} )==" "\n" |
| 130 | R"==(#if VECT_DT_N == 1 )==" "\n" |
| 131 | R"==(float sum = src_sum; )==" "\n" |
| 132 | R"==(float pow_sum = src_pow_sum; )==" "\n" |
| 133 | R"==(#else )==" "\n" |
| 134 | R"==(float sum = 0; )==" "\n" |
| 135 | R"==(float pow_sum = 0; )==" "\n" |
| 136 | R"==(for (int i = 0; i < VECT_DT_N; ++i) { )==" "\n" |
| 137 | R"==(sum += src_sum[i]; )==" "\n" |
| 138 | R"==(pow_sum += src_pow_sum[i]; )==" "\n" |
| 139 | R"==(} )==" "\n" |
| 140 | R"==(#endif )==" "\n" |
| 141 | R"==(x[REDUCE_DIM_IDX] = 0; )==" "\n" |
| 142 | R"==(int reduce_idx = reduce_index(x); )==" "\n" |
| 143 | R"==(float total_sum = sub_group_reduce_add(sum); )==" "\n" |
| 144 | R"==(float total_pow_sum = sub_group_reduce_add(pow_sum); )==" "\n" |
| 145 | R"==(int local_id = get_sub_group_local_id(); )==" "\n" |
| 146 | R"==(if (local_id == 0) { )==" "\n" |
| 147 | R"==(float calc_mean = total_sum / (MB * ID * IH * IW); )==" "\n" |
| 148 | R"==(float calc_variance )==" "\n" |
| 149 | R"==(= total_pow_sum / (MB * ID * IH * IW) - calc_mean * calc_mean; )==" "\n" |
| 150 | R"==(mean[x[1]] = calc_mean; )==" "\n" |
| 151 | R"==(variance[x[1]] = calc_variance < 0 ? 0 : calc_variance; )==" "\n" |
| 152 | R"==(} )==" "\n" |
| 153 | R"==(#endif )==" "\n" |
| 154 | R"==(} )==" "\n" |
| 155 | R"==(#else )==" "\n" |
| 156 | R"==(NAMED_KERNEL_ATTR(CALC) )==" "\n" |
| 157 | R"==(__kernel void calculate_mean(__global DATA_T *src, __global float *mean) { )==" "\n" |
| 158 | R"==(int x[5]; )==" "\n" |
| 159 | R"==(x[0] = GWS_GET_STAT_MB(); )==" "\n" |
| 160 | R"==(x[1] = GWS_GET_STAT_IC(); )==" "\n" |
| 161 | R"==(x[2] = GWS_GET_STAT_ID(); )==" "\n" |
| 162 | R"==(x[3] = GWS_GET_STAT_IH(); )==" "\n" |
| 163 | R"==(x[4] = GWS_GET_STAT_IW(); )==" "\n" |
| 164 | R"==(float sum = 0; )==" "\n" |
| 165 | R"==(for (int i = 0; i < REDUCE_DIM; i++) { )==" "\n" |
| 166 | R"==(x[REDUCE_DIM_IDX] = i; )==" "\n" |
| 167 | R"==(sum += TO_DEF_ACC_DATA_T(src[SRC_OFF(x[0], x[1], x[2], x[3], x[4])]); )==" "\n" |
| 168 | R"==(} )==" "\n" |
| 169 | R"==(x[REDUCE_DIM_IDX] = 0; )==" "\n" |
| 170 | R"==(int reduce_idx = reduce_index(x); )==" "\n" |
| 171 | R"==(mean[reduce_idx * IC + x[1]] = sum; )==" "\n" |
| 172 | R"==(} )==" "\n" |
| 173 | R"==(NAMED_KERNEL_ATTR(CALC) )==" "\n" |
| 174 | R"==(__kernel void calculate_variance( )==" "\n" |
| 175 | R"==(__global DATA_T *src, __global float *mean, __global float *variance) { )==" "\n" |
| 176 | R"==(int x[5]; )==" "\n" |
| 177 | R"==(x[0] = GWS_GET_STAT_MB(); )==" "\n" |
| 178 | R"==(x[1] = GWS_GET_STAT_IC(); )==" "\n" |
| 179 | R"==(x[2] = GWS_GET_STAT_ID(); )==" "\n" |
| 180 | R"==(x[3] = GWS_GET_STAT_IH(); )==" "\n" |
| 181 | R"==(x[4] = GWS_GET_STAT_IW(); )==" "\n" |
| 182 | R"==(float sum = 0; )==" "\n" |
| 183 | R"==(for (int i = 0; i < REDUCE_DIM; i++) { )==" "\n" |
| 184 | R"==(x[REDUCE_DIM_IDX] = i; )==" "\n" |
| 185 | R"==(DEF_ACC_DATA_T v0 )==" "\n" |
| 186 | R"==(= TO_DEF_ACC_DATA_T(src[SRC_OFF(x[0], x[1], x[2], x[3], x[4])]) )==" "\n" |
| 187 | R"==(- mean[x[1]]; )==" "\n" |
| 188 | R"==(sum += v0 * v0; )==" "\n" |
| 189 | R"==(} )==" "\n" |
| 190 | R"==(variance += MB * ID * IH * IW * IC / REDUCE_DIM; )==" "\n" |
| 191 | R"==(x[REDUCE_DIM_IDX] = 0; )==" "\n" |
| 192 | R"==(int reduce_idx = reduce_index(x); )==" "\n" |
| 193 | R"==(variance[reduce_idx * IC + x[1]] = sum; )==" "\n" |
| 194 | R"==(} )==" "\n" |
| 195 | R"==(#endif )==" "\n" |
| 196 | R"==(NAMED_KERNEL_ATTR(REDUCE) )==" "\n" |
| 197 | R"==(__kernel void reduce_mean(__global float *reduce_temp, __global float *mean) { )==" "\n" |
| 198 | R"==(const int c = GWS_GET_REDUCE_STAT_IC(); )==" "\n" |
| 199 | R"==(reduce_temp += c; )==" "\n" |
| 200 | R"==(float sum = 0.0f; )==" "\n" |
| 201 | R"==(int reduce_size = MB * ID * IH * IW / REDUCE_DIM; )==" "\n" |
| 202 | R"==(for (int i = 0; i < reduce_size; i++) { )==" "\n" |
| 203 | R"==(sum += reduce_temp[i * IC]; )==" "\n" |
| 204 | R"==(} )==" "\n" |
| 205 | R"==(mean[c] = sum / (MB * ID * IH * IW); )==" "\n" |
| 206 | R"==(} )==" "\n" |
| 207 | R"==(NAMED_KERNEL_ATTR(REDUCE) )==" "\n" |
| 208 | R"==(__kernel void reduce_variance( )==" "\n" |
| 209 | R"==(__global float *reduce_temp, __global float *variance) { )==" "\n" |
| 210 | R"==(const int c = GWS_GET_REDUCE_STAT_IC(); )==" "\n" |
| 211 | R"==(#if SAVE_STATS == 0 )==" "\n" |
| 212 | R"==(variance += IC; )==" "\n" |
| 213 | R"==(#endif )==" "\n" |
| 214 | R"==(float sum = 0.0f; )==" "\n" |
| 215 | R"==(int reduce_size = MB * ID * IH * IW / REDUCE_DIM; )==" "\n" |
| 216 | R"==(reduce_temp += reduce_size * IC + c; )==" "\n" |
| 217 | R"==(for (int i = 0; i < reduce_size; i++) )==" "\n" |
| 218 | R"==(sum += reduce_temp[i * IC]; )==" "\n" |
| 219 | R"==(variance[c] = sum / (MB * ID * IH * IW); )==" "\n" |
| 220 | R"==(} )==" "\n" |
| 221 | R"==(#endif )==" "\n" |
| 222 | R"==(KERNEL_ATTR )==" "\n" |
| 223 | R"==(__kernel void ref_bnorm_fwd(__global DATA_T *src, __global float *mean, )==" "\n" |
| 224 | R"==(__global float *variance, __global DATA_T *dst, __global float *scale, )==" "\n" |
| 225 | R"==(__global float *shift, __global char *ws, float eps, )==" "\n" |
| 226 | R"==(__global DATA_T *src_add) { )==" "\n" |
| 227 | R"==(const int n = GWS_GET_MB(); )==" "\n" |
| 228 | R"==(const int c = GWS_GET_IC(); )==" "\n" |
| 229 | R"==(const int d = GWS_GET_ID(); )==" "\n" |
| 230 | R"==(const int h = GWS_GET_IH(); )==" "\n" |
| 231 | R"==(const int w = GWS_GET_IW(); )==" "\n" |
| 232 | R"==(#if USE_SCALE == 1 )==" "\n" |
| 233 | R"==(float sm = scale[c]; )==" "\n" |
| 234 | R"==(#else )==" "\n" |
| 235 | R"==(float sm = 1; )==" "\n" |
| 236 | R"==(#endif )==" "\n" |
| 237 | R"==(#if USE_SHIFT == 1 )==" "\n" |
| 238 | R"==(float sv = shift[c]; )==" "\n" |
| 239 | R"==(#else )==" "\n" |
| 240 | R"==(float sv = 0; )==" "\n" |
| 241 | R"==(#endif )==" "\n" |
| 242 | R"==(#if SAVE_STATS == 0 && CALCULATE_STATS == 1 )==" "\n" |
| 243 | R"==(variance += IC; )==" "\n" |
| 244 | R"==(#endif )==" "\n" |
| 245 | R"==(float v_mean = mean[c]; )==" "\n" |
| 246 | R"==(float v_variance = variance[c]; )==" "\n" |
| 247 | R"==(const int off = SRC_OFF(n, c, d, h, w); )==" "\n" |
| 248 | R"==(float v0 = TO_DEF_ACC_DATA_T(src[off]); )==" "\n" |
| 249 | R"==(float sqrt_variance = 1.0f / sqrt(v_variance + eps); )==" "\n" |
| 250 | R"==(float bn_res = sm * (v0 - v_mean) * sqrt_variance + sv; )==" "\n" |
| 251 | R"==(#if FUSE_BN_ADD_RELU == 1 )==" "\n" |
| 252 | R"==(bn_res += TO_DEF_ACC_DATA_T(src_add[off]); )==" "\n" |
| 253 | R"==(#endif )==" "\n" |
| 254 | R"==(#if FUSE_BN_RELU == 1 )==" "\n" |
| 255 | R"==(if (bn_res <= 0) { )==" "\n" |
| 256 | R"==(bn_res = 0; )==" "\n" |
| 257 | R"==(#if IS_TRAINING == 1 )==" "\n" |
| 258 | R"==(ws[off] = 0; )==" "\n" |
| 259 | R"==(} else { )==" "\n" |
| 260 | R"==(ws[off] = -1; )==" "\n" |
| 261 | R"==(#endif )==" "\n" |
| 262 | R"==(} )==" "\n" |
| 263 | R"==(#endif )==" "\n" |
| 264 | R"==(#if WITH_RELU )==" "\n" |
| 265 | R"==(bn_res = max(bn_res, 0.0f); )==" "\n" |
| 266 | R"==(#endif )==" "\n" |
| 267 | R"==(dst[off] = TO_DATA_T(bn_res); )==" "\n" |
| 268 | R"==(} )==" "\n" |
| 269 | R"==(#endif )==" "\n" |
| 270 | R"==(#if IS_BWD == 1 )==" "\n" |
| 271 | R"==(#if USE_16MB_UNROLL == 1 )==" "\n" |
| 272 | R"==(NAMED_KERNEL_ATTR(CALC) )==" "\n" |
| 273 | R"==(__kernel void calculate_stats(__global DATA_T *src, __global float *mean, )==" "\n" |
| 274 | R"==(__global DATA_T *diff_dst, __global char *ws, )==" "\n" |
| 275 | R"==(__global float *reduce_temp) { )==" "\n" |
| 276 | R"==(const int mb = GWS_GET_STAT_MB(); )==" "\n" |
| 277 | R"==(const int stat_mb_block_idx = mb / MB_BLOCK; )==" "\n" |
| 278 | R"==(const int c = GWS_GET_STAT_IC(); )==" "\n" |
| 279 | R"==(const int sp_beg = GWS_GET_STAT_SP(); )==" "\n" |
| 280 | R"==(const int stat_sp_block = GWS_GET_STAT_SP_BLOCK(); )==" "\n" |
| 281 | R"==(const int stat_sp_nblocks = ID * IH * IW / stat_sp_block; )==" "\n" |
| 282 | R"==(const int stat_sp_block_idx = sp_beg / stat_sp_block; )==" "\n" |
| 283 | R"==(const int mb_sp_idx )==" "\n" |
| 284 | R"==(= stat_mb_block_idx * stat_sp_nblocks + stat_sp_block_idx; )==" "\n" |
| 285 | R"==(const int s_off = c * ID * IH * IW * MB_BLOCK + mb * IC * ID * IH * IW )==" "\n" |
| 286 | R"==(+ sp_beg * MB_BLOCK * IC_BLOCK; )==" "\n" |
| 287 | R"==(src += s_off; )==" "\n" |
| 288 | R"==(diff_dst += s_off; )==" "\n" |
| 289 | R"==(#if FUSE_BN_RELU == 1 )==" "\n" |
| 290 | R"==(ws += s_off; )==" "\n" |
| 291 | R"==(#endif )==" "\n" |
| 292 | R"==(VECT_FLOAT_T diff_gamma0 = 0.0f, diff_beta0 = 0.0f; )==" "\n" |
| 293 | R"==(VECT_FLOAT_T diff_gamma1 = 0.0f, diff_beta1 = 0.0f; )==" "\n" |
| 294 | R"==(float v_mean = as_float( )==" "\n" |
| 295 | R"==(intel_sub_group_block_read((const __global uint *)&mean[c])); )==" "\n" |
| 296 | R"==(for (int sp = sp_beg; sp < sp_beg + stat_sp_block; sp++) { )==" "\n" |
| 297 | R"==(VECT_FLOAT_T dd0 = CONVERT_VECT_FLOAT_T(AS_VECT_DATA_T( )==" "\n" |
| 298 | R"==(VECT_BLOCK_READ((const __global BLOCK_DATA_T *)&diff_dst[0]))); )==" "\n" |
| 299 | R"==(VECT_FLOAT_T ss0 = CONVERT_VECT_FLOAT_T(AS_VECT_DATA_T( )==" "\n" |
| 300 | R"==(VECT_BLOCK_READ((const __global BLOCK_DATA_T *)&src[0]))); )==" "\n" |
| 301 | R"==(#ifdef MB16 )==" "\n" |
| 302 | R"==(VECT_FLOAT_T dd1 = CONVERT_VECT_FLOAT_T(AS_VECT_DATA_T(VECT_BLOCK_READ( )==" "\n" |
| 303 | R"==((const __global BLOCK_DATA_T *)&diff_dst[8 * 16]))); )==" "\n" |
| 304 | R"==(VECT_FLOAT_T ss1 = CONVERT_VECT_FLOAT_T(AS_VECT_DATA_T( )==" "\n" |
| 305 | R"==(VECT_BLOCK_READ((const __global BLOCK_DATA_T *)&src[8 * 16]))); )==" "\n" |
| 306 | R"==(#endif )==" "\n" |
| 307 | R"==(#if FUSE_BN_RELU == 1 )==" "\n" |
| 308 | R"==(VECT_INT_T ws0 = VECT_CHAR_TO_INT(AS_VECT_CHAR_T( )==" "\n" |
| 309 | R"==(VECT_UCHAR_READ((const __global uchar *)&ws[0]))); )==" "\n" |
| 310 | R"==(dd0 = select((VECT_FLOAT_T)0.0f, dd0, ws0); )==" "\n" |
| 311 | R"==(#ifdef MB16 )==" "\n" |
| 312 | R"==(VECT_INT_T ws1 = VECT_CHAR_TO_INT(AS_VECT_CHAR_T( )==" "\n" |
| 313 | R"==(VECT_UCHAR_READ((const __global uchar *)&ws[8 * 16]))); )==" "\n" |
| 314 | R"==(dd1 = select((VECT_FLOAT_T)0.0f, dd1, ws1); )==" "\n" |
| 315 | R"==(#endif )==" "\n" |
| 316 | R"==(ws += MB_BLOCK * IC_BLOCK; )==" "\n" |
| 317 | R"==(#endif )==" "\n" |
| 318 | R"==(diff_gamma0 = fma((ss0 - (VECT_FLOAT_T)v_mean), dd0, diff_gamma0); )==" "\n" |
| 319 | R"==(diff_beta0 += dd0; )==" "\n" |
| 320 | R"==(#ifdef MB16 )==" "\n" |
| 321 | R"==(diff_gamma1 = fma((ss1 - (VECT_FLOAT_T)v_mean), dd1, diff_gamma1); )==" "\n" |
| 322 | R"==(diff_beta1 += dd1; )==" "\n" |
| 323 | R"==(#endif )==" "\n" |
| 324 | R"==(src += MB_BLOCK * IC_BLOCK; )==" "\n" |
| 325 | R"==(diff_dst += MB_BLOCK * IC_BLOCK; )==" "\n" |
| 326 | R"==(} )==" "\n" |
| 327 | R"==(#ifdef MB16 )==" "\n" |
| 328 | R"==(float v_diff_gamma = 0.0f, v_diff_beta = 0.0; )==" "\n" |
| 329 | R"==(for (int i = 0; i < 8; i++) { )==" "\n" |
| 330 | R"==(v_diff_gamma += diff_gamma0[i] + diff_gamma1[i]; )==" "\n" |
| 331 | R"==(v_diff_beta += diff_beta0[i] + diff_beta1[i]; )==" "\n" |
| 332 | R"==(} )==" "\n" |
| 333 | R"==(#else )==" "\n" |
| 334 | R"==(float v_diff_gamma = diff_gamma0, v_diff_beta = diff_beta0; )==" "\n" |
| 335 | R"==(#endif )==" "\n" |
| 336 | R"==(intel_sub_group_block_write( )==" "\n" |
| 337 | R"==((__global uint *)&reduce_temp[mb_sp_idx * IC + c], )==" "\n" |
| 338 | R"==(as_uint(v_diff_gamma)); )==" "\n" |
| 339 | R"==(intel_sub_group_block_write( )==" "\n" |
| 340 | R"==((__global uint *)&reduce_temp[REDUCE_STAT_NBLOCKS * IC )==" "\n" |
| 341 | R"==(+ mb_sp_idx * IC + c], )==" "\n" |
| 342 | R"==(as_uint(v_diff_beta)); )==" "\n" |
| 343 | R"==(} )==" "\n" |
| 344 | R"==(NAMED_KERNEL_ATTR(REDUCE) )==" "\n" |
| 345 | R"==(__kernel void reduce_stats(__global float *reduce_temp, )==" "\n" |
| 346 | R"==(__global float *diff_scale, __global float *diff_shift, )==" "\n" |
| 347 | R"==(__global float *variance, float eps) { )==" "\n" |
| 348 | R"==(const int c = GWS_GET_REDUCE_STAT_IC(); )==" "\n" |
| 349 | R"==(reduce_temp += c; )==" "\n" |
| 350 | R"==(float diff_gamma = 0.0f, diff_beta = 0.0f; )==" "\n" |
| 351 | R"==(for (int i = 0; i < REDUCE_STAT_NBLOCKS; i++) { )==" "\n" |
| 352 | R"==(diff_gamma += reduce_temp[i * IC]; )==" "\n" |
| 353 | R"==(diff_beta += reduce_temp[REDUCE_STAT_NBLOCKS * IC + i * IC]; )==" "\n" |
| 354 | R"==(} )==" "\n" |
| 355 | R"==(float sqrt_variance = 1.0f / sqrt(variance[c] + eps); )==" "\n" |
| 356 | R"==(diff_scale[c] = diff_gamma * sqrt_variance; )==" "\n" |
| 357 | R"==(#if DIFF_SHIFT == 1 )==" "\n" |
| 358 | R"==(diff_shift[c] = diff_beta; )==" "\n" |
| 359 | R"==(#else )==" "\n" |
| 360 | R"==(diff_shift[REDUCE_STAT_NBLOCKS * IC + c] = diff_beta; )==" "\n" |
| 361 | R"==(#endif )==" "\n" |
| 362 | R"==(} )==" "\n" |
| 363 | R"==(#else )==" "\n" |
| 364 | R"==(NAMED_KERNEL_ATTR(CALC) )==" "\n" |
| 365 | R"==(__kernel void calculate_stats(__global DATA_T *src, __global float *mean, )==" "\n" |
| 366 | R"==(__global DATA_T *diff_dst, __global char *ws, )==" "\n" |
| 367 | R"==(__global float *reduce_temp) { )==" "\n" |
| 368 | R"==(float diff_gamma = 0; )==" "\n" |
| 369 | R"==(float diff_beta = 0; )==" "\n" |
| 370 | R"==(int x[5]; )==" "\n" |
| 371 | R"==(x[0] = GWS_GET_STAT_MB(); )==" "\n" |
| 372 | R"==(x[1] = GWS_GET_STAT_IC(); )==" "\n" |
| 373 | R"==(x[2] = GWS_GET_STAT_ID(); )==" "\n" |
| 374 | R"==(x[3] = GWS_GET_STAT_IH(); )==" "\n" |
| 375 | R"==(x[4] = GWS_GET_STAT_IW(); )==" "\n" |
| 376 | R"==(for (int i = 0; i < REDUCE_DIM; i++) { )==" "\n" |
| 377 | R"==(x[REDUCE_DIM_IDX] = i; )==" "\n" |
| 378 | R"==(int off = SRC_OFF(x[0], x[1], x[2], x[3], x[4]); )==" "\n" |
| 379 | R"==(float dd = CONVERT_FLOAT_T(diff_dst[off]); )==" "\n" |
| 380 | R"==(#if FUSE_BN_RELU == 1 )==" "\n" |
| 381 | R"==(if (!ws[off]) dd = 0; )==" "\n" |
| 382 | R"==(#endif )==" "\n" |
| 383 | R"==(diff_gamma += (CONVERT_FLOAT_T(src[off]) - mean[x[1]]) * dd; )==" "\n" |
| 384 | R"==(diff_beta += dd; )==" "\n" |
| 385 | R"==(} )==" "\n" |
| 386 | R"==(int ss_off = MB * ID * IH * IW * IC / REDUCE_DIM; )==" "\n" |
| 387 | R"==(x[REDUCE_DIM_IDX] = 0; )==" "\n" |
| 388 | R"==(int reduce_idx = reduce_index(x); )==" "\n" |
| 389 | R"==(reduce_temp[reduce_idx * IC + x[1]] = diff_gamma; )==" "\n" |
| 390 | R"==(reduce_temp[ss_off + reduce_idx * IC + x[1]] = diff_beta; )==" "\n" |
| 391 | R"==(} )==" "\n" |
| 392 | R"==(NAMED_KERNEL_ATTR(REDUCE) )==" "\n" |
| 393 | R"==(__kernel void reduce_stats(__global float *reduce_temp, )==" "\n" |
| 394 | R"==(__global float *diff_scale, __global float *diff_shift, )==" "\n" |
| 395 | R"==(__global float *variance, float eps) { )==" "\n" |
| 396 | R"==(const int c = GWS_GET_REDUCE_STAT_IC(); )==" "\n" |
| 397 | R"==(float diff_gamma = 0.0f; )==" "\n" |
| 398 | R"==(float diff_beta = 0.0f; )==" "\n" |
| 399 | R"==(int reduce_size = MB * ID * IH * IW / REDUCE_DIM; )==" "\n" |
| 400 | R"==(for (int i = 0; i < reduce_size; i++) { )==" "\n" |
| 401 | R"==(diff_gamma += reduce_temp[c + i * IC]; )==" "\n" |
| 402 | R"==(diff_beta += reduce_temp[IC * reduce_size + c + i * IC]; )==" "\n" |
| 403 | R"==(} )==" "\n" |
| 404 | R"==(float sqrt_variance = 1.0f / sqrt(variance[c] + eps); )==" "\n" |
| 405 | R"==(diff_scale[c] = diff_gamma * sqrt_variance; )==" "\n" |
| 406 | R"==(#if DIFF_SHIFT == 1 )==" "\n" |
| 407 | R"==(diff_shift[c] = diff_beta; )==" "\n" |
| 408 | R"==(#else )==" "\n" |
| 409 | R"==(diff_shift[IC * reduce_size + c] = diff_beta; )==" "\n" |
| 410 | R"==(#endif )==" "\n" |
| 411 | R"==(} )==" "\n" |
| 412 | R"==(#endif )==" "\n" |
| 413 | R"==(KERNEL_ATTR )==" "\n" |
| 414 | R"==(__kernel void ref_bnorm_bwd(__global DATA_T *src, __global float *mean, )==" "\n" |
| 415 | R"==(__global float *variance, __global DATA_T *diff_dst, )==" "\n" |
| 416 | R"==(__global float *scale, __global char *ws, __global DATA_T *diff_src, )==" "\n" |
| 417 | R"==(__global float *diff_scale, __global float *diff_shift, float eps, )==" "\n" |
| 418 | R"==(__global DATA_T *diff_src_add) { )==" "\n" |
| 419 | R"==(#if USE_16MB_UNROLL == 1 )==" "\n" |
| 420 | R"==(const int n = GWS_GET_MB(); )==" "\n" |
| 421 | R"==(const int c = GWS_GET_IC(); )==" "\n" |
| 422 | R"==(const int d = GWS_GET_ID(); )==" "\n" |
| 423 | R"==(const int h = GWS_GET_IH(); )==" "\n" |
| 424 | R"==(const int w = GWS_GET_IW(); )==" "\n" |
| 425 | R"==(#if USE_SCALE == 1 )==" "\n" |
| 426 | R"==(float gamma = as_float( )==" "\n" |
| 427 | R"==(intel_sub_group_block_read((const __global uint *)&scale[c])); )==" "\n" |
| 428 | R"==(#else )==" "\n" |
| 429 | R"==(float gamma = 1.0f; )==" "\n" |
| 430 | R"==(#endif )==" "\n" |
| 431 | R"==(float v_variance = as_float( )==" "\n" |
| 432 | R"==(intel_sub_group_block_read((const __global uint *)&variance[c])); )==" "\n" |
| 433 | R"==(float sqrt_variance = 1.0f / sqrt(v_variance + eps); )==" "\n" |
| 434 | R"==(#if CALCULATE_DIFF_STATS == 1 )==" "\n" |
| 435 | R"==(float v_mean = as_float( )==" "\n" |
| 436 | R"==(intel_sub_group_block_read((const __global uint *)&mean[c])); )==" "\n" |
| 437 | R"==(float diff_gamma = as_float( )==" "\n" |
| 438 | R"==(intel_sub_group_block_read((const __global uint *)&diff_scale[c])); )==" "\n" |
| 439 | R"==(#if DIFF_SHIFT == 1 )==" "\n" |
| 440 | R"==(float diff_beta = as_float( )==" "\n" |
| 441 | R"==(intel_sub_group_block_read((const __global uint *)&diff_shift[c])); )==" "\n" |
| 442 | R"==(#else )==" "\n" |
| 443 | R"==(float diff_beta = as_float(intel_sub_group_block_read( )==" "\n" |
| 444 | R"==((const __global uint *)&diff_shift[REDUCE_STAT_NBLOCKS * IC + c])); )==" "\n" |
| 445 | R"==(#endif )==" "\n" |
| 446 | R"==(#endif )==" "\n" |
| 447 | R"==(const uint d_off = SRC_OFF(n, c, d, h, w); )==" "\n" |
| 448 | R"==(diff_src += d_off; )==" "\n" |
| 449 | R"==(#if FUSE_BN_ADD_RELU == 1 )==" "\n" |
| 450 | R"==(diff_src_add += d_off; )==" "\n" |
| 451 | R"==(#endif )==" "\n" |
| 452 | R"==(diff_dst += d_off; )==" "\n" |
| 453 | R"==(src += d_off; )==" "\n" |
| 454 | R"==(VECT_FLOAT_T blockD0 = CONVERT_VECT_FLOAT_T(AS_VECT_DATA_T( )==" "\n" |
| 455 | R"==(VECT_BLOCK_READ((const __global BLOCK_DATA_T *)&diff_dst[0]))); )==" "\n" |
| 456 | R"==(#ifdef MB16 )==" "\n" |
| 457 | R"==(VECT_FLOAT_T blockD1 = CONVERT_VECT_FLOAT_T(AS_VECT_DATA_T(VECT_BLOCK_READ( )==" "\n" |
| 458 | R"==((const __global BLOCK_DATA_T *)&diff_dst[8 * IC_BLOCK]))); )==" "\n" |
| 459 | R"==(#endif )==" "\n" |
| 460 | R"==(#if FUSE_BN_RELU == 1 )==" "\n" |
| 461 | R"==(ws += d_off; )==" "\n" |
| 462 | R"==(VECT_INT_T blockWS0 = VECT_CHAR_TO_INT( )==" "\n" |
| 463 | R"==(AS_VECT_CHAR_T(VECT_UCHAR_READ((const __global uchar *)&ws[0]))); )==" "\n" |
| 464 | R"==(blockD0 = select((VECT_FLOAT_T)0.0f, blockD0, blockWS0); )==" "\n" |
| 465 | R"==(#if FUSE_BN_ADD_RELU == 1 )==" "\n" |
| 466 | R"==(VECT_BLOCK_WRITE((__global BLOCK_DATA_T *)&diff_src_add[0], )==" "\n" |
| 467 | R"==(AS_VECT_BLOCK_DATA_T(CONVERT_VECTOR_DATA_T(blockD0))); )==" "\n" |
| 468 | R"==(#endif )==" "\n" |
| 469 | R"==(#ifdef MB16 )==" "\n" |
| 470 | R"==(VECT_INT_T blockWS1 = VECT_CHAR_TO_INT(AS_VECT_CHAR_T( )==" "\n" |
| 471 | R"==(VECT_UCHAR_READ((const __global uchar *)&ws[8 * IC_BLOCK]))); )==" "\n" |
| 472 | R"==(blockD1 = select((VECT_FLOAT_T)0.0f, blockD1, blockWS1); )==" "\n" |
| 473 | R"==(#if FUSE_BN_ADD_RELU == 1 )==" "\n" |
| 474 | R"==(VECT_BLOCK_WRITE((__global BLOCK_DATA_T *)&diff_src_add[8 * 16], )==" "\n" |
| 475 | R"==(AS_VECT_BLOCK_DATA_T(CONVERT_VECTOR_DATA_T(blockD1))); )==" "\n" |
| 476 | R"==(#endif )==" "\n" |
| 477 | R"==(#endif )==" "\n" |
| 478 | R"==(#endif )==" "\n" |
| 479 | R"==(gamma *= sqrt_variance; )==" "\n" |
| 480 | R"==(#if CALCULATE_DIFF_STATS == 1 )==" "\n" |
| 481 | R"==(diff_gamma *= sqrt_variance; )==" "\n" |
| 482 | R"==(diff_gamma /= (MB * ID * IH * IW); )==" "\n" |
| 483 | R"==(diff_beta /= (MB * ID * IH * IW); )==" "\n" |
| 484 | R"==(VECT_FLOAT_T blockS0 = CONVERT_VECT_FLOAT_T(AS_VECT_DATA_T( )==" "\n" |
| 485 | R"==(VECT_BLOCK_READ((const __global BLOCK_DATA_T *)&src[0]))); )==" "\n" |
| 486 | R"==(blockD0 -= fma((VECT_FLOAT_T)diff_gamma, (blockS0 - (VECT_FLOAT_T)v_mean), )==" "\n" |
| 487 | R"==((VECT_FLOAT_T)diff_beta); )==" "\n" |
| 488 | R"==(#ifdef MB16 )==" "\n" |
| 489 | R"==(VECT_FLOAT_T blockS1 = CONVERT_VECT_FLOAT_T(AS_VECT_DATA_T(VECT_BLOCK_READ( )==" "\n" |
| 490 | R"==((const __global BLOCK_DATA_T *)&src[8 * IC_BLOCK]))); )==" "\n" |
| 491 | R"==(blockD1 -= fma((VECT_FLOAT_T)diff_gamma, (blockS1 - (VECT_FLOAT_T)v_mean), )==" "\n" |
| 492 | R"==((VECT_FLOAT_T)diff_beta); )==" "\n" |
| 493 | R"==(#endif )==" "\n" |
| 494 | R"==(#endif )==" "\n" |
| 495 | R"==(blockD0 *= gamma; )==" "\n" |
| 496 | R"==(VECT_BLOCK_WRITE((__global BLOCK_DATA_T *)&diff_src[0], )==" "\n" |
| 497 | R"==(AS_VECT_BLOCK_DATA_T(CONVERT_VECTOR_DATA_T(blockD0))); )==" "\n" |
| 498 | R"==(#ifdef MB16 )==" "\n" |
| 499 | R"==(blockD1 *= gamma; )==" "\n" |
| 500 | R"==(VECT_BLOCK_WRITE((__global BLOCK_DATA_T *)&diff_src[8 * 16], )==" "\n" |
| 501 | R"==(AS_VECT_BLOCK_DATA_T(CONVERT_VECTOR_DATA_T(blockD1))); )==" "\n" |
| 502 | R"==(#endif )==" "\n" |
| 503 | R"==(#else )==" "\n" |
| 504 | R"==(const int n = GWS_GET_MB(); )==" "\n" |
| 505 | R"==(const int c = GWS_GET_IC(); )==" "\n" |
| 506 | R"==(const int d = GWS_GET_ID(); )==" "\n" |
| 507 | R"==(const int h = GWS_GET_IH(); )==" "\n" |
| 508 | R"==(const int w = GWS_GET_IW(); )==" "\n" |
| 509 | R"==(float v_variance = variance[c]; )==" "\n" |
| 510 | R"==(float sqrt_variance = 1.0f / sqrt(v_variance + eps); )==" "\n" |
| 511 | R"==(#if USE_SCALE == 1 )==" "\n" |
| 512 | R"==(float gamma = scale[c]; )==" "\n" |
| 513 | R"==(#else )==" "\n" |
| 514 | R"==(float gamma = 1; )==" "\n" |
| 515 | R"==(#endif )==" "\n" |
| 516 | R"==(#if CALCULATE_DIFF_STATS == 1 )==" "\n" |
| 517 | R"==(float v_mean = mean[c]; )==" "\n" |
| 518 | R"==(float diff_gamma = diff_scale[c]; )==" "\n" |
| 519 | R"==(#if DIFF_SHIFT == 1 )==" "\n" |
| 520 | R"==(float diff_beta = diff_shift[c]; )==" "\n" |
| 521 | R"==(#else )==" "\n" |
| 522 | R"==(int reduce_size = MB * ID * IH * IW / REDUCE_DIM; )==" "\n" |
| 523 | R"==(float diff_beta = diff_shift[reduce_size * IC + c]; )==" "\n" |
| 524 | R"==(#endif )==" "\n" |
| 525 | R"==(#endif )==" "\n" |
| 526 | R"==(const int off = SRC_OFF(n, c, d, h, w); )==" "\n" |
| 527 | R"==(float dd = TO_DEF_ACC_DATA_T(diff_dst[off]); )==" "\n" |
| 528 | R"==(#if FUSE_BN_RELU == 1 )==" "\n" |
| 529 | R"==(if (!ws[off]) dd = 0; )==" "\n" |
| 530 | R"==(#if FUSE_BN_ADD_RELU == 1 )==" "\n" |
| 531 | R"==(diff_src_add[off] = TO_DATA_T(dd); )==" "\n" |
| 532 | R"==(#endif )==" "\n" |
| 533 | R"==(#endif )==" "\n" |
| 534 | R"==(float v_diff_src = dd; )==" "\n" |
| 535 | R"==(#if CALCULATE_DIFF_STATS == 1 )==" "\n" |
| 536 | R"==(v_diff_src -= diff_beta / (MB * ID * IH * IW) )==" "\n" |
| 537 | R"==(+ (CONVERT_FLOAT_T(src[off]) - v_mean) * diff_gamma * sqrt_variance )==" "\n" |
| 538 | R"==(/ (MB * ID * IH * IW); )==" "\n" |
| 539 | R"==(#endif )==" "\n" |
| 540 | R"==(v_diff_src *= gamma * sqrt_variance; )==" "\n" |
| 541 | R"==(diff_src[off] = TO_DATA_T(v_diff_src); )==" "\n" |
| 542 | R"==(#endif )==" "\n" |
| 543 | R"==(} )==" "\n" |
| 544 | R"==(#endif )==" "\n" |
| 545 | R"==()=="; |
| 546 | } |
| 547 | } |
| 548 | } |
| 549 | } |
Generated on 2022-Nov-30 from project oneDNN revision 84bc949a1
Powered by 
Code Browser 2.1
Generator usage only permitted with license.