| 1 | namespace dnnl { |
|---|---|
| 2 | namespace impl { |
| 3 | namespace gpu { |
| 4 | namespace ocl { |
| 5 | const char *ref_convolution_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"==(#include "gpu/ocl/ocl_post_ops.h" )==" "\n" |
| 21 | R"==(#include "gpu/ocl/ocl_types.h" )==" "\n" |
| 22 | R"==(#undef SRC_OFF )==" "\n" |
| 23 | R"==(#undef WEI_OFF )==" "\n" |
| 24 | R"==(#undef DST_OFF )==" "\n" |
| 25 | R"==(#define SRC_OFF CONV_SRC_OFF )==" "\n" |
| 26 | R"==(#define WEI_OFF CONV_WEI_OFF )==" "\n" |
| 27 | R"==(#define DST_OFF CONV_DST_OFF )==" "\n" |
| 28 | R"==(#if IS_FWD )==" "\n" |
| 29 | R"==(KERNEL_ATTR )==" "\n" |
| 30 | R"==(__kernel void ref_convolution_fwd(const __global SRC_DATA_T *src, )==" "\n" |
| 31 | R"==(const __global WEI_DATA_T *wei, const __global BIA_DATA_T *bias, )==" "\n" |
| 32 | R"==(__global DST_DATA_T *dst POST_OP_ARGS, const __global float *src_scales, )==" "\n" |
| 33 | R"==(const __global float *wei_scales, const __global float *dst_scales, )==" "\n" |
| 34 | R"==(const __global int *src_zpoints, const __global int *dst_zpoints) { )==" "\n" |
| 35 | R"==(src += SRC_OFFSET0; )==" "\n" |
| 36 | R"==(dst += DST_OFFSET0; )==" "\n" |
| 37 | R"==(const int n = GWS_GET_MB(); )==" "\n" |
| 38 | R"==(const int oc = GWS_GET_OC(); )==" "\n" |
| 39 | R"==(const int g = GWS_GET_G(); )==" "\n" |
| 40 | R"==(const int od = GWS_GET_OD(); )==" "\n" |
| 41 | R"==(const int oh = GWS_GET_OH(); )==" "\n" |
| 42 | R"==(const int ow = GWS_GET_OW(); )==" "\n" |
| 43 | R"==(ACC_DATA_T d = 0; )==" "\n" |
| 44 | R"==(for (int ic = 0; ic < IC; ++ic) )==" "\n" |
| 45 | R"==(for (int kd = 0; kd < KD; ++kd) )==" "\n" |
| 46 | R"==(for (int kh = 0; kh < KH; ++kh) )==" "\n" |
| 47 | R"==(for (int kw = 0; kw < KW; ++kw) { )==" "\n" |
| 48 | R"==(const int id = od * SD - PD + kd * (1 + DD); )==" "\n" |
| 49 | R"==(const int ih = oh * SH - PH + kh * (1 + DH); )==" "\n" |
| 50 | R"==(const int iw = ow * SW - PW + kw * (1 + DW); )==" "\n" |
| 51 | R"==(if (id < 0 || id >= ID || ih < 0 || ih >= IH || iw < 0 )==" "\n" |
| 52 | R"==(|| iw >= IW) )==" "\n" |
| 53 | R"==(continue; )==" "\n" |
| 54 | R"==(const uint src_off = SRC_OFF(n, g * IC + ic, id, ih, iw); )==" "\n" |
| 55 | R"==(const uint wei_off = WEI_OFF(g, oc, ic, kd, kh, kw); )==" "\n" |
| 56 | R"==(d += SRC_TO_REF(src[src_off]) * WEI_TO_REF(wei[wei_off]); )==" "\n" |
| 57 | R"==(#if WITH_SRC_ZPOINTS )==" "\n" |
| 58 | R"==(const int src_zp )==" "\n" |
| 59 | R"==(= src_zpoints[WITH_SRC_ZPOINTS_PER_IC ? g * IC + ic )==" "\n" |
| 60 | R"==(: 0]; )==" "\n" |
| 61 | R"==(d -= src_zp * WEI_TO_REF(wei[wei_off]); )==" "\n" |
| 62 | R"==(#endif )==" "\n" |
| 63 | R"==(} )==" "\n" |
| 64 | R"==(POST_OP_DATA_T tmp = d; )==" "\n" |
| 65 | R"==(#if WITH_SRC_SCALES )==" "\n" |
| 66 | R"==(tmp *= src_scales[0]; )==" "\n" |
| 67 | R"==(#endif )==" "\n" |
| 68 | R"==(#if WITH_WEI_SCALES )==" "\n" |
| 69 | R"==(#if WEI_SCALES_MASK == 0 )==" "\n" |
| 70 | R"==(tmp *= wei_scales[0]; )==" "\n" |
| 71 | R"==(#else )==" "\n" |
| 72 | R"==(tmp *= wei_scales[g * OC + oc]; )==" "\n" |
| 73 | R"==(#endif )==" "\n" |
| 74 | R"==(#endif )==" "\n" |
| 75 | R"==(#if WITH_BIAS )==" "\n" |
| 76 | R"==(tmp += (POST_OP_DATA_T)BIA_TO_REF(bias[g * OC + oc]); )==" "\n" |
| 77 | R"==(#endif )==" "\n" |
| 78 | R"==(POST_OP_DATA_T sum_src; )==" "\n" |
| 79 | R"==(#if WITH_SUM )==" "\n" |
| 80 | R"==(sum_src = (POST_OP_DATA_T)SUM_TO_REF( )==" "\n" |
| 81 | R"==(AS_SUM_DATA_T(dst[DST_OFF(n, g * OC + oc, od, oh, ow)])); )==" "\n" |
| 82 | R"==(#endif )==" "\n" |
| 83 | R"==(#if NDIMS == 3 )==" "\n" |
| 84 | R"==(const unsigned po_d2 = ow; )==" "\n" |
| 85 | R"==(const unsigned po_d3 = 0; )==" "\n" |
| 86 | R"==(const unsigned po_d4 = 0; )==" "\n" |
| 87 | R"==(#elif NDIMS == 4 )==" "\n" |
| 88 | R"==(const unsigned po_d2 = oh; )==" "\n" |
| 89 | R"==(const unsigned po_d3 = ow; )==" "\n" |
| 90 | R"==(const unsigned po_d4 = 0; )==" "\n" |
| 91 | R"==(#elif NDIMS == 5 )==" "\n" |
| 92 | R"==(const unsigned po_d2 = od; )==" "\n" |
| 93 | R"==(const unsigned po_d3 = oh; )==" "\n" |
| 94 | R"==(const unsigned po_d4 = ow; )==" "\n" |
| 95 | R"==(#else )==" "\n" |
| 96 | R"==(const unsigned po_d2 = 0; )==" "\n" |
| 97 | R"==(const unsigned po_d3 = 0; )==" "\n" |
| 98 | R"==(const unsigned po_d4 = 0; )==" "\n" |
| 99 | R"==(#endif )==" "\n" |
| 100 | R"==(APPLY_POST_OPS_SERIAL(tmp, POST_OP_DATA_T, sum_src, POST_OP_DATA_T, n, 1, )==" "\n" |
| 101 | R"==(g * OC + oc, 1, po_d2, 1, po_d3, 1, po_d4, 1, 0, 1); )==" "\n" |
| 102 | R"==(#if WITH_DST_SCALES )==" "\n" |
| 103 | R"==(tmp /= dst_scales[0]; )==" "\n" |
| 104 | R"==(#endif )==" "\n" |
| 105 | R"==(#if WITH_DST_ZPOINTS )==" "\n" |
| 106 | R"==(const int dst_zp = dst_zpoints[WITH_DST_ZPOINTS_PER_OC ? g * OC + oc : 0]; )==" "\n" |
| 107 | R"==(tmp += dst_zp; )==" "\n" |
| 108 | R"==(#endif )==" "\n" |
| 109 | R"==(dst[DST_OFF(n, g * OC + oc, od, oh, ow)] = TO_DST(tmp); )==" "\n" |
| 110 | R"==(} )==" "\n" |
| 111 | R"==(#endif )==" "\n" |
| 112 | R"==(#if IS_BWD_D )==" "\n" |
| 113 | R"==(KERNEL_ATTR )==" "\n" |
| 114 | R"==(__kernel void ref_convolution_bwd_data(__global SRC_DATA_T *diff_src, )==" "\n" |
| 115 | R"==(const __global WEI_DATA_T *wei, const __global DST_DATA_T *diff_dst, )==" "\n" |
| 116 | R"==(const __global BIA_DATA_T *bias POST_OP_ARGS, )==" "\n" |
| 117 | R"==(const __global float *src_scales, const __global float *wei_scales, )==" "\n" |
| 118 | R"==(const __global float *dst_scales, const __global int *src_zpoints, )==" "\n" |
| 119 | R"==(const __global int *dst_zpoints) { )==" "\n" |
| 120 | R"==(const int n = GWS_GET_MB(); )==" "\n" |
| 121 | R"==(const int ic = GWS_GET_IC(); )==" "\n" |
| 122 | R"==(const int g = GWS_GET_G(); )==" "\n" |
| 123 | R"==(const int id = GWS_GET_ID(); )==" "\n" |
| 124 | R"==(const int ih = GWS_GET_IH(); )==" "\n" |
| 125 | R"==(const int iw = GWS_GET_IW(); )==" "\n" |
| 126 | R"==(ACC_DATA_T d = WITH_BIAS ? BIA_TO_REF(bias[g * IC + ic]) : 0.0; )==" "\n" |
| 127 | R"==(for_(int oc = 0; oc < OC; ++oc) )==" "\n" |
| 128 | R"==(for_(int kd = 0; kd < KD; ++kd) )==" "\n" |
| 129 | R"==(for_(int kh = 0; kh < KH; ++kh) )==" "\n" |
| 130 | R"==(for (int kw = 0; kw < KW; ++kw) { )==" "\n" |
| 131 | R"==(if (iw + PW < kw * (1 + DW) || ih + PH < kh * (1 + DH) )==" "\n" |
| 132 | R"==(|| id + PD < kd * (1 + DD)) )==" "\n" |
| 133 | R"==(continue; )==" "\n" |
| 134 | R"==(int ow = iw - kw * (1 + DW) + PW; )==" "\n" |
| 135 | R"==(int oh = ih - kh * (1 + DH) + PH; )==" "\n" |
| 136 | R"==(int od = id - kd * (1 + DD) + PD; )==" "\n" |
| 137 | R"==(if (ow % SW != 0 || oh % SH != 0 || od % SD != 0) continue; )==" "\n" |
| 138 | R"==(ow /= SW; )==" "\n" |
| 139 | R"==(oh /= SH; )==" "\n" |
| 140 | R"==(od /= SD; )==" "\n" |
| 141 | R"==(if (oh < OH && ow < OW && od < OD) { )==" "\n" |
| 142 | R"==(const uint dst_off = DST_OFF(n, g * OC + oc, od, oh, ow); )==" "\n" |
| 143 | R"==(const uint wei_off = WEI_OFF(g, oc, ic, kd, kh, kw); )==" "\n" |
| 144 | R"==(d += DST_TO_REF(diff_dst[dst_off]) * WEI_TO_REF(wei[wei_off]); )==" "\n" |
| 145 | R"==(#if WITH_SRC_ZPOINTS )==" "\n" |
| 146 | R"==(const int src_zp )==" "\n" |
| 147 | R"==(= src_zpoints[WITH_SRC_ZPOINTS_PER_IC ? g * OC + oc : 0]; )==" "\n" |
| 148 | R"==(d -= src_zp * WEI_TO_REF(wei[wei_off]); )==" "\n" |
| 149 | R"==(#endif )==" "\n" |
| 150 | R"==(} )==" "\n" |
| 151 | R"==(} )==" "\n" |
| 152 | R"==(float sum_src; )==" "\n" |
| 153 | R"==(#if WITH_SUM )==" "\n" |
| 154 | R"==(sum_src = convert_float( )==" "\n" |
| 155 | R"==(SRC_TO_REF(diff_src[SRC_OFF(n, g * IC + ic, id, ih, iw)])); )==" "\n" |
| 156 | R"==(#endif )==" "\n" |
| 157 | R"==(float accumulator = convert_float(d); )==" "\n" |
| 158 | R"==(#if WITH_SRC_SCALES )==" "\n" |
| 159 | R"==(accumulator *= src_scales[0]; )==" "\n" |
| 160 | R"==(#endif )==" "\n" |
| 161 | R"==(#if WITH_WEI_SCALES )==" "\n" |
| 162 | R"==(#if WEI_SCALES_MASK == 0 )==" "\n" |
| 163 | R"==(accumulator *= wei_scales[0]; )==" "\n" |
| 164 | R"==(#else )==" "\n" |
| 165 | R"==(accumulator *= wei_scales[g * IC + ic]; )==" "\n" |
| 166 | R"==(#endif )==" "\n" |
| 167 | R"==(#endif )==" "\n" |
| 168 | R"==(#if NDIMS == 3 )==" "\n" |
| 169 | R"==(const unsigned po_d2 = iw; )==" "\n" |
| 170 | R"==(const unsigned po_d3 = 0; )==" "\n" |
| 171 | R"==(const unsigned po_d4 = 0; )==" "\n" |
| 172 | R"==(#elif NDIMS == 4 )==" "\n" |
| 173 | R"==(const unsigned po_d2 = ih; )==" "\n" |
| 174 | R"==(const unsigned po_d3 = iw; )==" "\n" |
| 175 | R"==(const unsigned po_d4 = 0; )==" "\n" |
| 176 | R"==(#elif NDIMS == 5 )==" "\n" |
| 177 | R"==(const unsigned po_d2 = id; )==" "\n" |
| 178 | R"==(const unsigned po_d3 = ih; )==" "\n" |
| 179 | R"==(const unsigned po_d4 = iw; )==" "\n" |
| 180 | R"==(#else )==" "\n" |
| 181 | R"==(const unsigned po_d2 = 0; )==" "\n" |
| 182 | R"==(const unsigned po_d3 = 0; )==" "\n" |
| 183 | R"==(const unsigned po_d4 = 0; )==" "\n" |
| 184 | R"==(#endif )==" "\n" |
| 185 | R"==(APPLY_POST_OPS_SERIAL(accumulator, float, sum_src, float, n, 1, g *IC + ic, )==" "\n" |
| 186 | R"==(1, po_d2, 1, po_d3, 1, po_d4, 1, 0, 1); )==" "\n" |
| 187 | R"==(#if WITH_DST_SCALES )==" "\n" |
| 188 | R"==(accumulator /= dst_scales[0]; )==" "\n" |
| 189 | R"==(#endif )==" "\n" |
| 190 | R"==(#if WITH_DST_ZPOINTS )==" "\n" |
| 191 | R"==(const int dst_zp = dst_zpoints[WITH_DST_ZPOINTS_PER_OC ? g * IC + ic : 0]; )==" "\n" |
| 192 | R"==(accumulator += dst_zp; )==" "\n" |
| 193 | R"==(#endif )==" "\n" |
| 194 | R"==(diff_src[SRC_OFF(n, g * IC + ic, id, ih, iw)] = TO_SRC(accumulator); )==" "\n" |
| 195 | R"==(} )==" "\n" |
| 196 | R"==(#endif )==" "\n" |
| 197 | R"==(#if IS_BWD_W )==" "\n" |
| 198 | R"==(KERNEL_ATTR )==" "\n" |
| 199 | R"==(__kernel void ref_convolution_bwd_weights(const __global SRC_DATA_T *src, )==" "\n" |
| 200 | R"==(__global WEI_DATA_T *diff_wei, __global BIA_DATA_T *diff_bias, )==" "\n" |
| 201 | R"==(const __global DST_DATA_T *diff_dst) { )==" "\n" |
| 202 | R"==(const int g = GWS_GET_G(); )==" "\n" |
| 203 | R"==(const int ic = GWS_GET_IC(); )==" "\n" |
| 204 | R"==(const int oc = GWS_GET_OC(); )==" "\n" |
| 205 | R"==(const int kd = GWS_GET_KD(); )==" "\n" |
| 206 | R"==(const int kh = GWS_GET_KH(); )==" "\n" |
| 207 | R"==(const int kw = GWS_GET_KW(); )==" "\n" |
| 208 | R"==(#if WITH_BIAS )==" "\n" |
| 209 | R"==(if (ic == 0 && kh == 0 && kw == 0 & kd == 0) { )==" "\n" |
| 210 | R"==(ACC_DATA_T d = 0.0; )==" "\n" |
| 211 | R"==(for (int n = 0; n < MB; ++n) )==" "\n" |
| 212 | R"==(for (int od = 0; od < OD; ++od) )==" "\n" |
| 213 | R"==(for (int oh = 0; oh < OH; ++oh) )==" "\n" |
| 214 | R"==(for (int ow = 0; ow < OW; ++ow) { )==" "\n" |
| 215 | R"==(d += DST_TO_REF( )==" "\n" |
| 216 | R"==(diff_dst[DST_OFF(n, g * OC + oc, od, oh, ow)]); )==" "\n" |
| 217 | R"==(} )==" "\n" |
| 218 | R"==(diff_bias[g * OC + oc] = TO_BIA(d); )==" "\n" |
| 219 | R"==(} )==" "\n" |
| 220 | R"==(#endif )==" "\n" |
| 221 | R"==(ACC_DATA_T dw = 0.0; )==" "\n" |
| 222 | R"==(for (int n = 0; n < MB; ++n) )==" "\n" |
| 223 | R"==(for (int od = 0; od < OD; ++od) )==" "\n" |
| 224 | R"==(for (int oh = 0; oh < OH; ++oh) )==" "\n" |
| 225 | R"==(for (int ow = 0; ow < OW; ++ow) { )==" "\n" |
| 226 | R"==(if (ow * SW + kw * (1 + DW) < PW )==" "\n" |
| 227 | R"==(|| oh * SH + kh * (1 + DH) < PH )==" "\n" |
| 228 | R"==(|| od * SD + kd * (1 + DD) < PD )==" "\n" |
| 229 | R"==(|| ow * SW + kw * (1 + DW) >= IW + PW )==" "\n" |
| 230 | R"==(|| oh * SH + kh * (1 + DH) >= IH + PH )==" "\n" |
| 231 | R"==(|| od * SD + kd * (1 + DD) >= ID + PD) )==" "\n" |
| 232 | R"==(continue; )==" "\n" |
| 233 | R"==(int id = od * SD - PD + kd * (1 + DD); )==" "\n" |
| 234 | R"==(int ih = oh * SH - PH + kh * (1 + DH); )==" "\n" |
| 235 | R"==(int iw = ow * SW - PW + kw * (1 + DW); )==" "\n" |
| 236 | R"==(dw += DST_TO_REF( )==" "\n" |
| 237 | R"==(diff_dst[DST_OFF(n, g * OC + oc, od, oh, ow)]) )==" "\n" |
| 238 | R"==(* SRC_TO_REF( )==" "\n" |
| 239 | R"==(src[SRC_OFF(n, g * IC + ic, id, ih, iw)]); )==" "\n" |
| 240 | R"==(} )==" "\n" |
| 241 | R"==(diff_wei[WEI_OFF(g, oc, ic, kd, kh, kw)] = TO_WEI(dw); )==" "\n" |
| 242 | R"==(} )==" "\n" |
| 243 | R"==(#endif )==" "\n" |
| 244 | R"==()=="; |
| 245 | } |
| 246 | } |
| 247 | } |
| 248 | } |
Generated on 2022-Nov-30 from project oneDNN revision 84bc949a1
Powered by 
Code Browser 2.1
Generator usage only permitted with license.