| 1 | namespace dnnl { |
|---|---|
| 2 | namespace impl { |
| 3 | namespace gpu { |
| 4 | namespace ocl { |
| 5 | const char *xe_lp_conv_dw_fwd_data_ow_block_x8s8x_kernel = R"==(/******************************************************************************* )==" "\n" |
| 6 | R"==(* Copyright 2020-2022 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_math_utils.h" )==" "\n" |
| 21 | R"==(#include "gpu/ocl/ocl_post_ops.h" )==" "\n" |
| 22 | R"==(#include "gpu/ocl/ocl_types.h" )==" "\n" |
| 23 | R"==(inline ushort16 read_block16(const __global ushort *p) )==" "\n" |
| 24 | R"==(__attribute__((overloadable)) { )==" "\n" |
| 25 | R"==(ushort16 __builtin_IB_simd_block_read_16_global_h(const __global ushort *p) )==" "\n" |
| 26 | R"==(__attribute__((const)); )==" "\n" |
| 27 | R"==(return __builtin_IB_simd_block_read_16_global_h(p); )==" "\n" |
| 28 | R"==(} )==" "\n" |
| 29 | R"==(#define PWX (PW > 1 ? PW : 1) )==" "\n" |
| 30 | R"==(#if SCALES_PER_OC )==" "\n" |
| 31 | R"==(#define SCALE scales.s0101010101010101 )==" "\n" |
| 32 | R"==(#elif SCALES_COMMON )==" "\n" |
| 33 | R"==(#define SCALE runtime_scales[0] )==" "\n" |
| 34 | R"==(#else )==" "\n" |
| 35 | R"==(#define SCALE 1 )==" "\n" |
| 36 | R"==(#endif )==" "\n" |
| 37 | R"==(void block_read_dst(int n, DST_DATA_T *d, const __global DST_DATA_T *dst); )==" "\n" |
| 38 | R"==(void block_write_dst(int n, const DST_DATA_T *d, __global DST_DATA_T *dst); )==" "\n" |
| 39 | R"==(__attribute__((intel_reqd_sub_group_size(SUB_GROUP_SIZE))) )==" "\n" |
| 40 | R"==(__attribute__((reqd_work_group_size(LWS_0, LWS_1, LWS_2))) __kernel void )==" "\n" |
| 41 | R"==(conv_dw_fwd_ow_block_x8s8x(const __global uchar *src, const __global char *wei, )==" "\n" |
| 42 | R"==(const __global float *bias, __global DST_DATA_T *dst POST_OP_ARGS, )==" "\n" |
| 43 | R"==(const __global float *runtime_scales, )==" "\n" |
| 44 | R"==(const __global int *src_compensation, const __global int *src_zpoints, )==" "\n" |
| 45 | R"==(const __global int *dst_compensation) { )==" "\n" |
| 46 | R"==(const int osp = get_global_id(1); )==" "\n" |
| 47 | R"==(const int od = osp / (OWB * OH); )==" "\n" |
| 48 | R"==(const int ohw = osp % (OWB * OH); )==" "\n" |
| 49 | R"==(const int ow = (ohw % OWB) * OW_BLOCK; )==" "\n" |
| 50 | R"==(const int oh = ohw / OWB; )==" "\n" |
| 51 | R"==(const int g = get_group_id(0) * OC_BLOCK; )==" "\n" |
| 52 | R"==(const int mb = get_global_id(2) * MB_BLOCK; )==" "\n" |
| 53 | R"==(const int id = od * SD - PD; )==" "\n" |
| 54 | R"==(const int ih = oh * SH - PH; )==" "\n" |
| 55 | R"==(const int iw = ow * SW - PW; )==" "\n" |
| 56 | R"==(const __global uchar *src_original = src; )==" "\n" |
| 57 | R"==(dst += mb * G_PADDED * OD * OH * OW + g * OD * OH * OW * MB_BLOCK )==" "\n" |
| 58 | R"==(+ (od * OH * OW + oh * OW + ow) * MB_BLOCK * OC_BLOCK; )==" "\n" |
| 59 | R"==(src += mb * G_PADDED * ID * IH * IW + g * ID * IH * IW * MB_BLOCK )==" "\n" |
| 60 | R"==(+ (id * IH * IW + ih * IW + iw) * MB_BLOCK * IC_BLOCK; )==" "\n" |
| 61 | R"==(wei += g * KD * KH * KW; )==" "\n" |
| 62 | R"==(int16 S0 = 0; )==" "\n" |
| 63 | R"==(int16 S1 = 0; )==" "\n" |
| 64 | R"==(#if SCALES_PER_OC )==" "\n" |
| 65 | R"==(float2 scales = 1; )==" "\n" |
| 66 | R"==(if (g + 2 * get_sub_group_local_id() < G) )==" "\n" |
| 67 | R"==(scales.s0 = runtime_scales[g + 2 * get_sub_group_local_id()]; )==" "\n" |
| 68 | R"==(if (g + 2 * get_sub_group_local_id() + 1 < G) )==" "\n" |
| 69 | R"==(scales.s1 = runtime_scales[g + 2 * get_sub_group_local_id() + 1]; )==" "\n" |
| 70 | R"==(#endif )==" "\n" |
| 71 | R"==(for (int kd = 0; kd < KD; kd++) { )==" "\n" |
| 72 | R"==(if (kd * (1 + DD) + id < 0 || kd * (1 + DD) + id >= ID) { )==" "\n" |
| 73 | R"==(src += IC_BLOCK * MB_BLOCK * IH * IW * (1 + DD); )==" "\n" |
| 74 | R"==(wei += OC_BLOCK * KH * KW; )==" "\n" |
| 75 | R"==(continue; )==" "\n" |
| 76 | R"==(} )==" "\n" |
| 77 | R"==(__attribute__((opencl_unroll_hint)) for (int kh = 0; kh < KH; kh++) { )==" "\n" |
| 78 | R"==(if (kh * (1 + DH) + ih < 0 || kh * (1 + DH) + ih >= IH) { )==" "\n" |
| 79 | R"==(src += IC_BLOCK * MB_BLOCK * IW * (1 + DH); )==" "\n" |
| 80 | R"==(wei += OC_BLOCK * KW; )==" "\n" |
| 81 | R"==(continue; )==" "\n" |
| 82 | R"==(} )==" "\n" |
| 83 | R"==(#if SW == 2 )==" "\n" |
| 84 | R"==(ushort16 AAA = 0; )==" "\n" |
| 85 | R"==(#else )==" "\n" |
| 86 | R"==(ushort AAA; )==" "\n" |
| 87 | R"==(#endif )==" "\n" |
| 88 | R"==(ushort16 AA = 0; )==" "\n" |
| 89 | R"==(#if OW % (SW * OW_BLOCK) + KW - 1 \ )==" "\n" |
| 90 | R"==(>= SW - 1 + PW )==" "\n" |
| 91 | R"==(/* get main block */ )==" "\n" |
| 92 | R"==(if (iw + SW * (OW_BLOCK) + KW - 1 > IW) { )==" "\n" |
| 93 | R"==(if (iw >= 0) { )==" "\n" |
| 94 | R"==(AA.s0 = intel_sub_group_block_read_us( )==" "\n" |
| 95 | R"==((const __global ushort *)(&src[0 * IC_BLOCK])); )==" "\n" |
| 96 | R"==(#if PW >= 2 )==" "\n" |
| 97 | R"==(AA.s1 = intel_sub_group_block_read_us( )==" "\n" |
| 98 | R"==((const __global ushort *)(&src[1 * IC_BLOCK])); )==" "\n" |
| 99 | R"==(#endif )==" "\n" |
| 100 | R"==(#if PW >= 3 )==" "\n" |
| 101 | R"==(AA.s2 = intel_sub_group_block_read_us( )==" "\n" |
| 102 | R"==((const __global ushort *)(&src[2 * IC_BLOCK])); )==" "\n" |
| 103 | R"==(#endif )==" "\n" |
| 104 | R"==(} )==" "\n" |
| 105 | R"==(#if IW_TAIL > 16 )==" "\n" |
| 106 | R"==(#if PW == 2 )==" "\n" |
| 107 | R"==(AA.s23456789 = intel_sub_group_block_read_us8( )==" "\n" |
| 108 | R"==((const __global ushort *)(&src[PWX * IC_BLOCK])); )==" "\n" |
| 109 | R"==(AA.sabcd = intel_sub_group_block_read_us4( )==" "\n" |
| 110 | R"==((const __global ushort *)(&src[(PWX + 8) * IC_BLOCK])); )==" "\n" |
| 111 | R"==(AA.sef = intel_sub_group_block_read_us2( )==" "\n" |
| 112 | R"==((const __global ushort *)(&src[(PWX + 12) * IC_BLOCK])); )==" "\n" |
| 113 | R"==(#elif PW == 3 )==" "\n" |
| 114 | R"==(AA.s3456789a = intel_sub_group_block_read_us8( )==" "\n" |
| 115 | R"==((const __global ushort *)(&src[PWX * IC_BLOCK])); )==" "\n" |
| 116 | R"==(AA.sbcde = intel_sub_group_block_read_us4( )==" "\n" |
| 117 | R"==((const __global ushort *)(&src[(PWX + 8) * IC_BLOCK])); )==" "\n" |
| 118 | R"==(AA.sf = intel_sub_group_block_read_us( )==" "\n" |
| 119 | R"==((const __global ushort *)(&src[(PWX + 12) * IC_BLOCK])); )==" "\n" |
| 120 | R"==(#else )==" "\n" |
| 121 | R"==(AA.s12345678 = intel_sub_group_block_read_us8( )==" "\n" |
| 122 | R"==((const __global ushort *)(&src[PWX * IC_BLOCK])); )==" "\n" |
| 123 | R"==(AA.s9abc = intel_sub_group_block_read_us4( )==" "\n" |
| 124 | R"==((const __global ushort *)(&src[(PWX + 8) * IC_BLOCK])); )==" "\n" |
| 125 | R"==(AA.sde = intel_sub_group_block_read_us2( )==" "\n" |
| 126 | R"==((const __global ushort *)(&src[(PWX + 12) * IC_BLOCK])); )==" "\n" |
| 127 | R"==(AA.sf = intel_sub_group_block_read_us( )==" "\n" |
| 128 | R"==((const __global ushort *)(&src[(PWX + 14) * IC_BLOCK])); )==" "\n" |
| 129 | R"==(#endif )==" "\n" |
| 130 | R"==(#if ((IW_TAIL - 16) & 0b1000) == 0b1000 )==" "\n" |
| 131 | R"==(AAA.s01234567 = intel_sub_group_block_read_us8( )==" "\n" |
| 132 | R"==((const __global ushort *)(&src[(16) * IC_BLOCK])); )==" "\n" |
| 133 | R"==(#endif )==" "\n" |
| 134 | R"==(#if ((IW_TAIL - 16) & 0b100) == 0b100 )==" "\n" |
| 135 | R"==(*((ushort4 *)(((ushort *)&AAA) + ((IW_TAIL - 16) & 0b1000))) )==" "\n" |
| 136 | R"==(= intel_sub_group_block_read_us4((const __global ushort )==" "\n" |
| 137 | R"==(*)(&src[(16 + ((IW_TAIL - 16) & 0b1000)) )==" "\n" |
| 138 | R"==(* IC_BLOCK])); )==" "\n" |
| 139 | R"==(#endif )==" "\n" |
| 140 | R"==(#if ((IW_TAIL - 16) & 0b10) == 0b10 )==" "\n" |
| 141 | R"==(*((ushort2 *)(((ushort *)&AAA) + ((IW_TAIL - 16) & 0b1100))) )==" "\n" |
| 142 | R"==(= intel_sub_group_block_read_us2((const __global ushort )==" "\n" |
| 143 | R"==(*)(&src[(16 + ((IW_TAIL - 16) & 0b1100)) )==" "\n" |
| 144 | R"==(* IC_BLOCK])); )==" "\n" |
| 145 | R"==(#endif )==" "\n" |
| 146 | R"==(#if ((IW_TAIL - 16) & 0b1) == 0b1 )==" "\n" |
| 147 | R"==(*(((ushort *)&AAA) + ((IW_TAIL - 16) & 0b1110)) )==" "\n" |
| 148 | R"==(= intel_sub_group_block_read_us((const __global ushort )==" "\n" |
| 149 | R"==(*)(&src[(16 + ((IW_TAIL - 16) & 0b1110)) )==" "\n" |
| 150 | R"==(* IC_BLOCK])); )==" "\n" |
| 151 | R"==(#endif )==" "\n" |
| 152 | R"==(#else )==" "\n" |
| 153 | R"==(#if ((IW_TAIL - PWX) & 0b1000) == 0b1000 )==" "\n" |
| 154 | R"==(*((ushort8 *)(((ushort *)&AA) + PWX)) )==" "\n" |
| 155 | R"==(= intel_sub_group_block_read_us8(( )==" "\n" |
| 156 | R"==(const __global ushort *)(&src[PWX * IC_BLOCK])); )==" "\n" |
| 157 | R"==(#endif )==" "\n" |
| 158 | R"==(#if ((IW_TAIL - PWX) & 0b100) == 0b100 )==" "\n" |
| 159 | R"==(*((ushort4 *)(((ushort *)&AA) + PWX )==" "\n" |
| 160 | R"==(+ ((IW_TAIL - PWX) & 0b1000))) )==" "\n" |
| 161 | R"==(= intel_sub_group_block_read_us4((const __global ushort )==" "\n" |
| 162 | R"==(*)(&src[(PWX )==" "\n" |
| 163 | R"==(+ ((IW_TAIL - PWX) )==" "\n" |
| 164 | R"==(& 0b1000)) )==" "\n" |
| 165 | R"==(* IC_BLOCK])); )==" "\n" |
| 166 | R"==(#endif )==" "\n" |
| 167 | R"==(#if ((IW_TAIL - PWX) & 0b10) == 0b10 )==" "\n" |
| 168 | R"==(*((ushort2 *)(((ushort *)&AA) + PWX )==" "\n" |
| 169 | R"==(+ ((IW_TAIL - PWX) & 0b1100))) )==" "\n" |
| 170 | R"==(= intel_sub_group_block_read_us2((const __global ushort )==" "\n" |
| 171 | R"==(*)(&src[(PWX )==" "\n" |
| 172 | R"==(+ ((IW_TAIL - PWX) )==" "\n" |
| 173 | R"==(& 0b1100)) )==" "\n" |
| 174 | R"==(* IC_BLOCK])); )==" "\n" |
| 175 | R"==(#endif )==" "\n" |
| 176 | R"==(#if ((IW_TAIL - PWX) & 0b1) == 0b1 )==" "\n" |
| 177 | R"==(*(((ushort *)&AA) + PWX + ((IW_TAIL - PWX) & 0b1110)) )==" "\n" |
| 178 | R"==(= intel_sub_group_block_read_us((const __global ushort )==" "\n" |
| 179 | R"==(*)(&src[(PWX )==" "\n" |
| 180 | R"==(+ ((IW_TAIL - PWX) )==" "\n" |
| 181 | R"==(& 0b1110)) )==" "\n" |
| 182 | R"==(* IC_BLOCK])); )==" "\n" |
| 183 | R"==(#endif )==" "\n" |
| 184 | R"==(#endif )==" "\n" |
| 185 | R"==(} else { )==" "\n" |
| 186 | R"==(#endif )==" "\n" |
| 187 | R"==(#if SW == 1 )==" "\n" |
| 188 | R"==(#define READ_BLOCK (OW_BLOCK + KW - 1 - PWX) )==" "\n" |
| 189 | R"==(if (iw >= 0) { )==" "\n" |
| 190 | R"==(AA.s0 = intel_sub_group_block_read_us( )==" "\n" |
| 191 | R"==((const __global ushort *)(&src[0 * IC_BLOCK])); )==" "\n" |
| 192 | R"==(#if PW >= 2 )==" "\n" |
| 193 | R"==(AA.s1 = intel_sub_group_block_read_us( )==" "\n" |
| 194 | R"==((const __global ushort *)(&src[1 * IC_BLOCK])); )==" "\n" |
| 195 | R"==(#endif )==" "\n" |
| 196 | R"==(#if PW >= 3 )==" "\n" |
| 197 | R"==(AA.s2 = intel_sub_group_block_read_us( )==" "\n" |
| 198 | R"==((const __global ushort *)(&src[2 * IC_BLOCK])); )==" "\n" |
| 199 | R"==(#endif )==" "\n" |
| 200 | R"==(} )==" "\n" |
| 201 | R"==(#if (READ_BLOCK & 0b1000) == 0b1000 )==" "\n" |
| 202 | R"==(*((ushort8 *)(((ushort *)&AA) + PWX)) )==" "\n" |
| 203 | R"==(= intel_sub_group_block_read_us8(( )==" "\n" |
| 204 | R"==(const __global ushort *)(&src[(PWX)*IC_BLOCK])); )==" "\n" |
| 205 | R"==(#endif )==" "\n" |
| 206 | R"==(#if (READ_BLOCK & 0b100) == 0b100 )==" "\n" |
| 207 | R"==(*((ushort4 *)(((ushort *)&AA) + PWX + (READ_BLOCK & 0b1000))) )==" "\n" |
| 208 | R"==(= intel_sub_group_block_read_us4((const __global ushort )==" "\n" |
| 209 | R"==(*)(&src[(PWX + (READ_BLOCK & 0b1000)) )==" "\n" |
| 210 | R"==(* IC_BLOCK])); )==" "\n" |
| 211 | R"==(#endif )==" "\n" |
| 212 | R"==(#if (READ_BLOCK & 0b10) == 0b10 )==" "\n" |
| 213 | R"==(*((ushort2 *)(((ushort *)&AA) + PWX + (READ_BLOCK & 0b1100))) )==" "\n" |
| 214 | R"==(= intel_sub_group_block_read_us2((const __global ushort )==" "\n" |
| 215 | R"==(*)(&src[(PWX + (READ_BLOCK & 0b1100)) )==" "\n" |
| 216 | R"==(* IC_BLOCK])); )==" "\n" |
| 217 | R"==(#endif )==" "\n" |
| 218 | R"==(#if (READ_BLOCK & 0b1) == 0b1 )==" "\n" |
| 219 | R"==(*(((ushort *)&AA) + PWX + (READ_BLOCK & 0b1110)) )==" "\n" |
| 220 | R"==(= intel_sub_group_block_read_us((const __global ushort )==" "\n" |
| 221 | R"==(*)(&src[(PWX + (READ_BLOCK & 0b1110)) )==" "\n" |
| 222 | R"==(* IC_BLOCK])); )==" "\n" |
| 223 | R"==(#endif )==" "\n" |
| 224 | R"==(#elif SW == 2 )==" "\n" |
| 225 | R"==(#if OW_BLOCK + KW - 1 >= 8 )==" "\n" |
| 226 | R"==(#define READ_BLOCK (2 * (OW_BLOCK) + KW - 1) )==" "\n" |
| 227 | R"==(if (iw >= 0) { )==" "\n" |
| 228 | R"==(AA = read_block16( )==" "\n" |
| 229 | R"==((const __global ushort *)(&src[0 * IC_BLOCK])); )==" "\n" |
| 230 | R"==(} else { )==" "\n" |
| 231 | R"==(#if PW == 0 )==" "\n" |
| 232 | R"==(AA = read_block16( )==" "\n" |
| 233 | R"==((const __global ushort *)(&src[0 * IC_BLOCK])); )==" "\n" |
| 234 | R"==(#elif PW == 2 )==" "\n" |
| 235 | R"==(AA.s23456789 = intel_sub_group_block_read_us8( )==" "\n" |
| 236 | R"==((const __global ushort *)(&src[PW * IC_BLOCK])); )==" "\n" |
| 237 | R"==(AA.sabcd = intel_sub_group_block_read_us4( )==" "\n" |
| 238 | R"==((const __global ushort *)(&src[(PW + 8) * IC_BLOCK])); )==" "\n" |
| 239 | R"==(AA.sef = intel_sub_group_block_read_us2( )==" "\n" |
| 240 | R"==((const __global ushort *)(&src[(PW + 12) * IC_BLOCK])); )==" "\n" |
| 241 | R"==(#elif PW == 3 )==" "\n" |
| 242 | R"==(AA.s3456789a = intel_sub_group_block_read_us8( )==" "\n" |
| 243 | R"==((const __global ushort *)(&src[PW * IC_BLOCK])); )==" "\n" |
| 244 | R"==(AA.sbcde = intel_sub_group_block_read_us4( )==" "\n" |
| 245 | R"==((const __global ushort *)(&src[(PW + 8) * IC_BLOCK])); )==" "\n" |
| 246 | R"==(AA.sf = intel_sub_group_block_read_us( )==" "\n" |
| 247 | R"==((const __global ushort *)(&src[(PW + 12) * IC_BLOCK])); )==" "\n" |
| 248 | R"==(#else )==" "\n" |
| 249 | R"==(AA.s12345678 = intel_sub_group_block_read_us8( )==" "\n" |
| 250 | R"==((const __global ushort *)(&src[PW * IC_BLOCK])); )==" "\n" |
| 251 | R"==(AA.s9abc = intel_sub_group_block_read_us4( )==" "\n" |
| 252 | R"==((const __global ushort *)(&src[(PW + 8) * IC_BLOCK])); )==" "\n" |
| 253 | R"==(AA.sde = intel_sub_group_block_read_us2( )==" "\n" |
| 254 | R"==((const __global ushort *)(&src[(PW + 12) * IC_BLOCK])); )==" "\n" |
| 255 | R"==(AA.sf = intel_sub_group_block_read_us( )==" "\n" |
| 256 | R"==((const __global ushort *)(&src[(PW + 14) * IC_BLOCK])); )==" "\n" |
| 257 | R"==(#endif )==" "\n" |
| 258 | R"==(} )==" "\n" |
| 259 | R"==(#if ((READ_BLOCK - 16) & 0b1000) == 0b1000 )==" "\n" |
| 260 | R"==(AAA.s01234567 = intel_sub_group_block_read_us8( )==" "\n" |
| 261 | R"==((const __global ushort *)(&src[(16) * IC_BLOCK])); )==" "\n" |
| 262 | R"==(#endif )==" "\n" |
| 263 | R"==(#if ((READ_BLOCK - 16) & 0b100) == 0b100 )==" "\n" |
| 264 | R"==(*((ushort4 *)(((ushort *)&AAA) + ((READ_BLOCK - 16) & 0b1000))) )==" "\n" |
| 265 | R"==(= intel_sub_group_block_read_us4((const __global ushort )==" "\n" |
| 266 | R"==(*)(&src[(16 + ((READ_BLOCK - 16) & 0b1000)) )==" "\n" |
| 267 | R"==(* IC_BLOCK])); )==" "\n" |
| 268 | R"==(#endif )==" "\n" |
| 269 | R"==(#if ((READ_BLOCK - 16) & 0b10) == 0b10 )==" "\n" |
| 270 | R"==(*((ushort2 *)(((ushort *)&AAA) + ((READ_BLOCK - 16) & 0b1100))) )==" "\n" |
| 271 | R"==(= intel_sub_group_block_read_us2((const __global ushort )==" "\n" |
| 272 | R"==(*)(&src[(16 + ((READ_BLOCK - 16) & 0b1100)) )==" "\n" |
| 273 | R"==(* IC_BLOCK])); )==" "\n" |
| 274 | R"==(#endif )==" "\n" |
| 275 | R"==(#if ((READ_BLOCK - 16) & 0b1) == 0b1 )==" "\n" |
| 276 | R"==(*(((ushort *)&AAA) + ((READ_BLOCK - 16) & 0b1110)) )==" "\n" |
| 277 | R"==(= intel_sub_group_block_read_us((const __global ushort )==" "\n" |
| 278 | R"==(*)(&src[(16 + ((READ_BLOCK - 16) & 0b1110)) )==" "\n" |
| 279 | R"==(* IC_BLOCK])); )==" "\n" |
| 280 | R"==(#endif )==" "\n" |
| 281 | R"==(#else )==" "\n" |
| 282 | R"==(#define READ_BLOCK (2 * (OW_BLOCK) + KW - 1 - PWX) )==" "\n" |
| 283 | R"==(if (iw >= 0) { )==" "\n" |
| 284 | R"==(AA.s0 = intel_sub_group_block_read_us( )==" "\n" |
| 285 | R"==((const __global ushort *)(&src[0 * IC_BLOCK])); )==" "\n" |
| 286 | R"==(#if PW >= 2 )==" "\n" |
| 287 | R"==(AA.s1 = intel_sub_group_block_read_us( )==" "\n" |
| 288 | R"==((const __global ushort *)(&src[1 * IC_BLOCK])); )==" "\n" |
| 289 | R"==(#endif )==" "\n" |
| 290 | R"==(#if PW >= 3 )==" "\n" |
| 291 | R"==(AA.s2 = intel_sub_group_block_read_us( )==" "\n" |
| 292 | R"==((const __global ushort *)(&src[2 * IC_BLOCK])); )==" "\n" |
| 293 | R"==(#endif )==" "\n" |
| 294 | R"==(} )==" "\n" |
| 295 | R"==(#if (READ_BLOCK & 0b1000) == 0b1000 )==" "\n" |
| 296 | R"==(*((ushort8 *)(((ushort *)&AA) + PWX)) )==" "\n" |
| 297 | R"==(= intel_sub_group_block_read_us8( )==" "\n" |
| 298 | R"==((const __global ushort *)(&src[(PWX)*IC_BLOCK])); )==" "\n" |
| 299 | R"==(#endif )==" "\n" |
| 300 | R"==(#if (READ_BLOCK & 0b100) == 0b100 )==" "\n" |
| 301 | R"==(*((ushort4 *)(((ushort *)&AA) + PWX + (READ_BLOCK & 0b1000))) )==" "\n" |
| 302 | R"==(= intel_sub_group_block_read_us4((const __global ushort )==" "\n" |
| 303 | R"==(*)(&src[(PWX + (READ_BLOCK & 0b1000)) )==" "\n" |
| 304 | R"==(* IC_BLOCK])); )==" "\n" |
| 305 | R"==(#endif )==" "\n" |
| 306 | R"==(#if (READ_BLOCK & 0b10) == 0b10 )==" "\n" |
| 307 | R"==(*((ushort2 *)(((ushort *)&AA) + PWX + (READ_BLOCK & 0b1100))) )==" "\n" |
| 308 | R"==(= intel_sub_group_block_read_us2((const __global ushort )==" "\n" |
| 309 | R"==(*)(&src[(PWX + (READ_BLOCK & 0b1100)) )==" "\n" |
| 310 | R"==(* IC_BLOCK])); )==" "\n" |
| 311 | R"==(#endif )==" "\n" |
| 312 | R"==(#if (READ_BLOCK & 0b1) == 0b1 )==" "\n" |
| 313 | R"==(*(((ushort *)&AA) + PWX + (READ_BLOCK & 0b1110)) )==" "\n" |
| 314 | R"==(= intel_sub_group_block_read_us((const __global ushort )==" "\n" |
| 315 | R"==(*)(&src[(PWX + (READ_BLOCK & 0b1110)) )==" "\n" |
| 316 | R"==(* IC_BLOCK])); )==" "\n" |
| 317 | R"==(#endif )==" "\n" |
| 318 | R"==(#endif )==" "\n" |
| 319 | R"==(#endif )==" "\n" |
| 320 | R"==(#if OW % (SW * OW_BLOCK) + KW - 1 >= SW - 1 + PW )==" "\n" |
| 321 | R"==(} )==" "\n" |
| 322 | R"==(#endif )==" "\n" |
| 323 | R"==(#if OW > OWX )==" "\n" |
| 324 | R"==(if (iw + READ_BLOCK > IW) { )==" "\n" |
| 325 | R"==(if (iw < IW) { )==" "\n" |
| 326 | R"==(for (int i = IW - iw; i < READ_BLOCK; i++) { )==" "\n" |
| 327 | R"==(if (i < 16) { )==" "\n" |
| 328 | R"==(AA[i] = 0; )==" "\n" |
| 329 | R"==(} )==" "\n" |
| 330 | R"==(#if SW == 2 )==" "\n" |
| 331 | R"==(else { )==" "\n" |
| 332 | R"==(AAA[i - 16] = 0; )==" "\n" |
| 333 | R"==(} )==" "\n" |
| 334 | R"==(#endif )==" "\n" |
| 335 | R"==(} )==" "\n" |
| 336 | R"==(} else { )==" "\n" |
| 337 | R"==(AA = 0; )==" "\n" |
| 338 | R"==(#if SW == 2 )==" "\n" |
| 339 | R"==(AAA = 0; )==" "\n" |
| 340 | R"==(#endif )==" "\n" |
| 341 | R"==(} )==" "\n" |
| 342 | R"==(} )==" "\n" |
| 343 | R"==(#endif )==" "\n" |
| 344 | R"==(ushort4 WW = 0; )==" "\n" |
| 345 | R"==(#if KW == 4 )==" "\n" |
| 346 | R"==(WW = intel_sub_group_block_read_us4((const __global ushort *)wei); )==" "\n" |
| 347 | R"==(#endif )==" "\n" |
| 348 | R"==(#if KW == 3 )==" "\n" |
| 349 | R"==(WW.s01 = intel_sub_group_block_read_us2( )==" "\n" |
| 350 | R"==((const __global ushort *)wei); )==" "\n" |
| 351 | R"==(WW.s2 = intel_sub_group_block_read_us( )==" "\n" |
| 352 | R"==((const __global ushort *)wei + OC_BLOCK); )==" "\n" |
| 353 | R"==(#endif )==" "\n" |
| 354 | R"==(#if KW == 2 )==" "\n" |
| 355 | R"==(WW.s01 = intel_sub_group_block_read_us2( )==" "\n" |
| 356 | R"==((const __global ushort *)wei); )==" "\n" |
| 357 | R"==(#endif )==" "\n" |
| 358 | R"==(#if KW == 1 )==" "\n" |
| 359 | R"==(WW.s0 = intel_sub_group_block_read_us((const __global ushort *)wei); )==" "\n" |
| 360 | R"==(#endif )==" "\n" |
| 361 | R"==(SRC_DATA16_T A0 = 0, A1 = 0; )==" "\n" |
| 362 | R"==(A0.s01234567 = AS_SRC_DATA16_T(AA.s01234567).s02468ace; )==" "\n" |
| 363 | R"==(A0.s89abcdef = AS_SRC_DATA16_T(AA.s89abcdef).s02468ace; )==" "\n" |
| 364 | R"==(A1.s01234567 = AS_SRC_DATA16_T(AA.s01234567).s13579bdf; )==" "\n" |
| 365 | R"==(A1.s89abcdef = AS_SRC_DATA16_T(AA.s89abcdef).s13579bdf; )==" "\n" |
| 366 | R"==(#if SW == 2 )==" "\n" |
| 367 | R"==(SRC_DATA16_T right0, right1; )==" "\n" |
| 368 | R"==(right0.s01234567 = AS_SRC_DATA16_T(AAA.s01234567).s02468ace; )==" "\n" |
| 369 | R"==(right0.s89abcdef = AS_SRC_DATA16_T(AAA.s89abcdef).s02468ace; )==" "\n" |
| 370 | R"==(right1.s01234567 = AS_SRC_DATA16_T(AAA.s01234567).s13579bdf; )==" "\n" |
| 371 | R"==(right1.s89abcdef = AS_SRC_DATA16_T(AAA.s89abcdef).s13579bdf; )==" "\n" |
| 372 | R"==(#else )==" "\n" |
| 373 | R"==(#if OW_BLOCK >= 14 )==" "\n" |
| 374 | R"==(SRC_DATA_T right0, right1; )==" "\n" |
| 375 | R"==(right0 = AS_SRC_DATA2_T(AAA).s0; )==" "\n" |
| 376 | R"==(right1 = AS_SRC_DATA2_T(AAA).s1; )==" "\n" |
| 377 | R"==(#endif )==" "\n" |
| 378 | R"==(#endif )==" "\n" |
| 379 | R"==(char8 W = as_char8(WW); )==" "\n" |
| 380 | R"==(#if SW == 1 )==" "\n" |
| 381 | R"==(S0.s0 = idot4(A0.s0123, W.s0246, S0.s0); )==" "\n" |
| 382 | R"==(#if OW_BLOCK >= 2 )==" "\n" |
| 383 | R"==(S0.s2 = idot4(A0.s1234, W.s0246, S0.s2); )==" "\n" |
| 384 | R"==(#endif )==" "\n" |
| 385 | R"==(#if OW_BLOCK >= 3 )==" "\n" |
| 386 | R"==(S0.s4 = idot4(A0.s2345, W.s0246, S0.s4); )==" "\n" |
| 387 | R"==(#endif )==" "\n" |
| 388 | R"==(#if OW_BLOCK >= 4 )==" "\n" |
| 389 | R"==(S0.s6 = idot4(A0.s3456, W.s0246, S0.s6); )==" "\n" |
| 390 | R"==(#endif )==" "\n" |
| 391 | R"==(#if OW_BLOCK >= 5 )==" "\n" |
| 392 | R"==(S0.s8 = idot4(A0.s4567, W.s0246, S0.s8); )==" "\n" |
| 393 | R"==(#endif )==" "\n" |
| 394 | R"==(#if OC_BLOCK >= 6 )==" "\n" |
| 395 | R"==(S0.sa = idot4(A0.s5678, W.s0246, S0.sa); )==" "\n" |
| 396 | R"==(#endif )==" "\n" |
| 397 | R"==(#if OW_BLOCK >= 7 )==" "\n" |
| 398 | R"==(S0.sc = idot4(A0.s6789, W.s0246, S0.sc); )==" "\n" |
| 399 | R"==(#endif )==" "\n" |
| 400 | R"==(#if OW_BLOCK >= 8 )==" "\n" |
| 401 | R"==(S0.se = idot4(A0.s789a, W.s0246, S0.se); )==" "\n" |
| 402 | R"==(#endif )==" "\n" |
| 403 | R"==(#if OW_BLOCK >= 9 )==" "\n" |
| 404 | R"==(S1.s0 = idot4(A0.s89ab, W.s0246, S1.s0); )==" "\n" |
| 405 | R"==(#endif )==" "\n" |
| 406 | R"==(#if OW_BLOCK >= 10 )==" "\n" |
| 407 | R"==(S1.s2 = idot4(A0.s9abc, W.s0246, S1.s2); )==" "\n" |
| 408 | R"==(#endif )==" "\n" |
| 409 | R"==(#if OW_BLOCK >= 11 )==" "\n" |
| 410 | R"==(S1.s4 = idot4(A0.sabcd, W.s0246, S1.s4); )==" "\n" |
| 411 | R"==(#endif )==" "\n" |
| 412 | R"==(#if OW_BLOCK >= 12 )==" "\n" |
| 413 | R"==(S1.s6 = idot4(A0.sbcde, W.s0246, S1.s6); )==" "\n" |
| 414 | R"==(#endif )==" "\n" |
| 415 | R"==(#if OW_BLOCK >= 13 )==" "\n" |
| 416 | R"==(S1.s8 = idot4(A0.scdef, W.s0246, S1.s8); )==" "\n" |
| 417 | R"==(#endif )==" "\n" |
| 418 | R"==(#if OW_BLOCK >= 14 )==" "\n" |
| 419 | R"==(S1.sa = idot4((SRC_DATA4_T)(A0.sde, A0.sf, right0), W.s0246, S1.sa); )==" "\n" |
| 420 | R"==(#endif )==" "\n" |
| 421 | R"==(#if OW_BLOCK >= 15 )==" "\n" |
| 422 | R"==(S1.sc = idot4( )==" "\n" |
| 423 | R"==((SRC_DATA4_T)(A0.sef, right0, right1), W.s0246, S1.sc); )==" "\n" |
| 424 | R"==(#endif )==" "\n" |
| 425 | R"==(S0.s1 = idot4(A1.s0123, W.s1357, S0.s1); )==" "\n" |
| 426 | R"==(#if OW_BLOCK >= 2 )==" "\n" |
| 427 | R"==(S0.s3 = idot4(A1.s1234, W.s1357, S0.s3); )==" "\n" |
| 428 | R"==(#endif )==" "\n" |
| 429 | R"==(#if OW_BLOCK >= 3 )==" "\n" |
| 430 | R"==(S0.s5 = idot4(A1.s2345, W.s1357, S0.s5); )==" "\n" |
| 431 | R"==(#endif )==" "\n" |
| 432 | R"==(#if OW_BLOCK >= 4 )==" "\n" |
| 433 | R"==(S0.s7 = idot4(A1.s3456, W.s1357, S0.s7); )==" "\n" |
| 434 | R"==(#endif )==" "\n" |
| 435 | R"==(#if OW_BLOCK >= 5 )==" "\n" |
| 436 | R"==(S0.s9 = idot4(A1.s4567, W.s1357, S0.s9); )==" "\n" |
| 437 | R"==(#endif )==" "\n" |
| 438 | R"==(#if OW_BLOCK >= 6 )==" "\n" |
| 439 | R"==(S0.sb = idot4(A1.s5678, W.s1357, S0.sb); )==" "\n" |
| 440 | R"==(#endif )==" "\n" |
| 441 | R"==(#if OW_BLOCK >= 7 )==" "\n" |
| 442 | R"==(S0.sd = idot4(A1.s6789, W.s1357, S0.sd); )==" "\n" |
| 443 | R"==(#endif )==" "\n" |
| 444 | R"==(#if OW_BLOCK >= 8 )==" "\n" |
| 445 | R"==(S0.sf = idot4(A1.s789a, W.s1357, S0.sf); )==" "\n" |
| 446 | R"==(#endif )==" "\n" |
| 447 | R"==(#if OW_BLOCK >= 9 )==" "\n" |
| 448 | R"==(S1.s1 = idot4(A1.s89ab, W.s1357, S1.s1); )==" "\n" |
| 449 | R"==(#endif )==" "\n" |
| 450 | R"==(#if OW_BLOCK >= 10 )==" "\n" |
| 451 | R"==(S1.s3 = idot4(A1.s9abc, W.s1357, S1.s3); )==" "\n" |
| 452 | R"==(#endif )==" "\n" |
| 453 | R"==(#if OW_BLOCK >= 11 )==" "\n" |
| 454 | R"==(S1.s5 = idot4(A1.sabcd, W.s1357, S1.s5); )==" "\n" |
| 455 | R"==(#endif )==" "\n" |
| 456 | R"==(#if OW_BLOCK >= 12 )==" "\n" |
| 457 | R"==(S1.s7 = idot4(A1.sbcde, W.s1357, S1.s7); )==" "\n" |
| 458 | R"==(#endif )==" "\n" |
| 459 | R"==(#if OW_BLOCK >= 13 )==" "\n" |
| 460 | R"==(S1.s9 = idot4(A1.scdef, W.s1357, S1.s9); )==" "\n" |
| 461 | R"==(#endif )==" "\n" |
| 462 | R"==(#if OW_BLOCK >= 14 )==" "\n" |
| 463 | R"==(S1.sb = idot4((SRC_DATA4_T)(A1.sde, A1.sf, 0), W.s1357, S1.sb); )==" "\n" |
| 464 | R"==(#endif )==" "\n" |
| 465 | R"==(#if OW_BLOCK >= 15 )==" "\n" |
| 466 | R"==(S1.sd = idot4((SRC_DATA4_T)(A1.sef, right1, 0), W.s1357, S1.sd); )==" "\n" |
| 467 | R"==(#endif )==" "\n" |
| 468 | R"==(#elif SW == 2 )==" "\n" |
| 469 | R"==(S0.s0 = idot4(A0.s0123, W.s0246, S0.s0); )==" "\n" |
| 470 | R"==(#if OW_BLOCK >= 2 )==" "\n" |
| 471 | R"==(S0.s2 = idot4(A0.s2345, W.s0246, S0.s2); )==" "\n" |
| 472 | R"==(#endif )==" "\n" |
| 473 | R"==(#if OW_BLOCK >= 3 )==" "\n" |
| 474 | R"==(S0.s4 = idot4(A0.s4567, W.s0246, S0.s4); )==" "\n" |
| 475 | R"==(#endif )==" "\n" |
| 476 | R"==(#if OW_BLOCK >= 4 )==" "\n" |
| 477 | R"==(S0.s6 = idot4(A0.s6789, W.s0246, S0.s6); )==" "\n" |
| 478 | R"==(#endif )==" "\n" |
| 479 | R"==(#if OW_BLOCK >= 5 )==" "\n" |
| 480 | R"==(S0.s8 = idot4(A0.s89ab, W.s0246, S0.s8); )==" "\n" |
| 481 | R"==(#endif )==" "\n" |
| 482 | R"==(#if OW_BLOCK >= 6 )==" "\n" |
| 483 | R"==(S0.sa = idot4(A0.sabcd, W.s0246, S0.sa); )==" "\n" |
| 484 | R"==(#endif )==" "\n" |
| 485 | R"==(#if OW_BLOCK >= 7 )==" "\n" |
| 486 | R"==(S0.sc = idot4(A0.scdef, W.s0246, S0.sc); )==" "\n" |
| 487 | R"==(#endif )==" "\n" |
| 488 | R"==(#if OW_BLOCK >= 8 )==" "\n" |
| 489 | R"==(S0.se = idot4((SRC_DATA4_T)(A0.sef, right0.s0, right0.s1), W.s0246, )==" "\n" |
| 490 | R"==(S0.se); )==" "\n" |
| 491 | R"==(#endif )==" "\n" |
| 492 | R"==(#if OW_BLOCK >= 9 )==" "\n" |
| 493 | R"==(S1.s0 = idot4(right0.s0123, W.s0246, S1.s0); )==" "\n" |
| 494 | R"==(#endif )==" "\n" |
| 495 | R"==(#if OW_BLOCK >= 10 )==" "\n" |
| 496 | R"==(S1.s2 = idot4(right0.s2345, W.s0246, S1.s2); )==" "\n" |
| 497 | R"==(#endif )==" "\n" |
| 498 | R"==(#if OW_BLOCK >= 11 )==" "\n" |
| 499 | R"==(S1.s4 = idot4(right0.s4567, W.s0246, S1.s4); )==" "\n" |
| 500 | R"==(#endif )==" "\n" |
| 501 | R"==(#if OW_BLOCK >= 12 )==" "\n" |
| 502 | R"==(S1.s6 = idot4(right0.s6789, W.s0246, S1.s6); )==" "\n" |
| 503 | R"==(#endif )==" "\n" |
| 504 | R"==(#if OW_BLOCK >= 13 )==" "\n" |
| 505 | R"==(S1.s8 = idot4(right0.s89ab, W.s0246, S1.s8); )==" "\n" |
| 506 | R"==(#endif )==" "\n" |
| 507 | R"==(#if OW_BLOCK >= 14 )==" "\n" |
| 508 | R"==(S1.sa = idot4(right0.sabcd, W.s0246, S1.sa); )==" "\n" |
| 509 | R"==(#endif )==" "\n" |
| 510 | R"==(#if OW_BLOCK >= 15 )==" "\n" |
| 511 | R"==(S1.sc = idot4(right0.scdef, W.s0246, S1.sc); )==" "\n" |
| 512 | R"==(#endif )==" "\n" |
| 513 | R"==(S0.s1 = idot4(A1.s0123, W.s1357, S0.s1); )==" "\n" |
| 514 | R"==(#if OW_BLOCK >= 2 )==" "\n" |
| 515 | R"==(S0.s3 = idot4(A1.s2345, W.s1357, S0.s3); )==" "\n" |
| 516 | R"==(#endif )==" "\n" |
| 517 | R"==(#if OW_BLOCK >= 3 )==" "\n" |
| 518 | R"==(S0.s5 = idot4(A1.s4567, W.s1357, S0.s5); )==" "\n" |
| 519 | R"==(#endif )==" "\n" |
| 520 | R"==(#if OW_BLOCK >= 4 )==" "\n" |
| 521 | R"==(S0.s7 = idot4(A1.s6789, W.s1357, S0.s7); )==" "\n" |
| 522 | R"==(#endif )==" "\n" |
| 523 | R"==(#if OW_BLOCK >= 5 )==" "\n" |
| 524 | R"==(S0.s9 = idot4(A1.s89ab, W.s1357, S0.s9); )==" "\n" |
| 525 | R"==(#endif )==" "\n" |
| 526 | R"==(#if OW_BLOCK >= 6 )==" "\n" |
| 527 | R"==(S0.sb = idot4(A1.sabcd, W.s1357, S0.sb); )==" "\n" |
| 528 | R"==(#endif )==" "\n" |
| 529 | R"==(#if OW_BLOCK >= 7 )==" "\n" |
| 530 | R"==(S0.sd = idot4(A1.scdef, W.s1357, S0.sd); )==" "\n" |
| 531 | R"==(#endif )==" "\n" |
| 532 | R"==(#if OW_BLOCK >= 8 )==" "\n" |
| 533 | R"==(S0.sf = idot4((SRC_DATA4_T)(A1.sef, right1.s0, right1.s1), W.s1357, )==" "\n" |
| 534 | R"==(S0.sf); )==" "\n" |
| 535 | R"==(#endif )==" "\n" |
| 536 | R"==(#if OW_BLOCK >= 9 )==" "\n" |
| 537 | R"==(S1.s1 = idot4(right1.s0123, W.s1357, S1.s1); )==" "\n" |
| 538 | R"==(#endif )==" "\n" |
| 539 | R"==(#if OW_BLOCK >= 10 )==" "\n" |
| 540 | R"==(S1.s3 = idot4(right1.s2345, W.s1357, S1.s3); )==" "\n" |
| 541 | R"==(#endif )==" "\n" |
| 542 | R"==(#if OW_BLOCK >= 11 )==" "\n" |
| 543 | R"==(S1.s5 = idot4(right1.s4567, W.s1357, S1.s5); )==" "\n" |
| 544 | R"==(#endif )==" "\n" |
| 545 | R"==(#if OW_BLOCK >= 12 )==" "\n" |
| 546 | R"==(S1.s7 = idot4(right1.s6789, W.s1357, S1.s7); )==" "\n" |
| 547 | R"==(#endif )==" "\n" |
| 548 | R"==(#if OW_BLOCK >= 13 )==" "\n" |
| 549 | R"==(S1.s9 = idot4(right1.s89ab, W.s1357, S1.s9); )==" "\n" |
| 550 | R"==(#endif )==" "\n" |
| 551 | R"==(#if OW_BLOCK >= 14 )==" "\n" |
| 552 | R"==(S1.sb = idot4(right1.sabcd, W.s1357, S1.sb); )==" "\n" |
| 553 | R"==(#endif )==" "\n" |
| 554 | R"==(#if OW_BLOCK >= 15 )==" "\n" |
| 555 | R"==(S1.sd = idot4(right1.scdef, W.s1357, S1.sd); )==" "\n" |
| 556 | R"==(#endif )==" "\n" |
| 557 | R"==(#else )==" "\n" |
| 558 | R"==(#error )==" "\n" |
| 559 | R"==(#endif )==" "\n" |
| 560 | R"==(src += IC_BLOCK * MB_BLOCK * IW * (1 + DH); )==" "\n" |
| 561 | R"==(wei += OC_BLOCK * KW; )==" "\n" |
| 562 | R"==(} )==" "\n" |
| 563 | R"==(src += IC_BLOCK * MB_BLOCK * IW * (IH * (1 + DD) - KH * (1 + DH)); )==" "\n" |
| 564 | R"==(} )==" "\n" |
| 565 | R"==(#if WITH_BIAS || (WITH_POST_OP && !SUM_SCALE1 || SCALES_PER_OC || SCALES_COMMON) )==" "\n" |
| 566 | R"==(float16 tmp00 = convert_float16(S0); )==" "\n" |
| 567 | R"==(float16 tmp01 = convert_float16(S1); )==" "\n" |
| 568 | R"==(#if WITH_BIAS )==" "\n" |
| 569 | R"==(float2 B; )==" "\n" |
| 570 | R"==(int g_off = g + 2 * get_sub_group_local_id(); )==" "\n" |
| 571 | R"==(B.s0 = g_off >= G ? 0 : bias[g_off]; )==" "\n" |
| 572 | R"==(B.s1 = g_off + 1 >= G ? 0 : bias[g_off + 1]; )==" "\n" |
| 573 | R"==(tmp00 += convert_float16(B.s0101010101010101); )==" "\n" |
| 574 | R"==(tmp01 += convert_float16(B.s0101010101010101); )==" "\n" |
| 575 | R"==(#endif )==" "\n" |
| 576 | R"==(tmp00 *= SCALE; )==" "\n" |
| 577 | R"==(tmp01 *= SCALE; )==" "\n" |
| 578 | R"==(#define ACC_DATA_TYPE float )==" "\n" |
| 579 | R"==(#define ACC0 tmp00 )==" "\n" |
| 580 | R"==(#define ACC1 tmp01 )==" "\n" |
| 581 | R"==(#else )==" "\n" |
| 582 | R"==(#define ACC_DATA_TYPE int )==" "\n" |
| 583 | R"==(#define ACC0 S0 )==" "\n" |
| 584 | R"==(#define ACC1 S1 )==" "\n" |
| 585 | R"==(#endif )==" "\n" |
| 586 | R"==(DST_DATA16_T D0 = 0; )==" "\n" |
| 587 | R"==(DST_DATA16_T D1 = 0; )==" "\n" |
| 588 | R"==(#if WITH_SUM )==" "\n" |
| 589 | R"==(if (OW_TAIL != 0 && ow + OW_BLOCK >= OW) { )==" "\n" |
| 590 | R"==(block_read_dst(min(8, OW_TAIL), &D0, dst); )==" "\n" |
| 591 | R"==(block_read_dst(OW_TAIL - 8, &D1, dst + 8 * OC_BLOCK); )==" "\n" |
| 592 | R"==(} else { )==" "\n" |
| 593 | R"==(block_read_dst(min(8, OW_BLOCK), &D0, dst); )==" "\n" |
| 594 | R"==(block_read_dst(OW_BLOCK - 8, &D1, dst + 8 * OC_BLOCK); )==" "\n" |
| 595 | R"==(} )==" "\n" |
| 596 | R"==(#endif )==" "\n" |
| 597 | R"==(SUM_DATA16_T D0_sdt = AS_SUM_DATA16_T(D0); )==" "\n" |
| 598 | R"==(SUM_DATA16_T D1_sdt = AS_SUM_DATA16_T(D1); )==" "\n" |
| 599 | R"==(#define APPLY_POST_OPS_COMMON(accumulator, sum, offset) \ )==" "\n" |
| 600 | R"==({ \ )==" "\n" |
| 601 | R"==(/*This kernel is using serial post op processing due to non-trivial \ )==" "\n" |
| 602 | R"==(data ordering in accumulator. */ \ )==" "\n" |
| 603 | R"==(for (int didx = 0; didx < 16; ++didx) { \ )==" "\n" |
| 604 | R"==(int po_mb = mb; \ )==" "\n" |
| 605 | R"==(int po_oc = g * OC + 2 * get_sub_group_local_id() + (didx % 2); \ )==" "\n" |
| 606 | R"==(ACC_DATA_TYPE accum = accumulator[didx]; \ )==" "\n" |
| 607 | R"==(SUM_DATA_T sum_di = sum[didx]; \ )==" "\n" |
| 608 | R"==(APPLY_POST_OPS_SERIAL_BINARY_2D(accum, ACC_DATA_TYPE, sum_di, \ )==" "\n" |
| 609 | R"==(SUM_DATA_T, po_mb, 1, po_oc, 1); \ )==" "\n" |
| 610 | R"==(accumulator[didx] = accum; \ )==" "\n" |
| 611 | R"==(} \ )==" "\n" |
| 612 | R"==(} )==" "\n" |
| 613 | R"==(APPLY_POST_OPS_COMMON(ACC0, D0_sdt, 0); )==" "\n" |
| 614 | R"==(APPLY_POST_OPS_COMMON(ACC1, D1_sdt, 8); )==" "\n" |
| 615 | R"==(DST_DATA16_T R0 = CONVERT_DST_DATA16_T(ACC0); )==" "\n" |
| 616 | R"==(DST_DATA16_T R1 = CONVERT_DST_DATA16_T(ACC1); )==" "\n" |
| 617 | R"==(if (OW_TAIL != 0 && ow + OW_BLOCK > OW) { )==" "\n" |
| 618 | R"==(block_write_dst(min(8, OW_TAIL), &R0, dst); )==" "\n" |
| 619 | R"==(block_write_dst(OW_TAIL - 8, &R1, dst + 8 * OC_BLOCK); )==" "\n" |
| 620 | R"==(} else { )==" "\n" |
| 621 | R"==(block_write_dst(min(8, OW_BLOCK), &R0, dst); )==" "\n" |
| 622 | R"==(block_write_dst(OW_BLOCK - 8, &R1, dst + 8 * OC_BLOCK); )==" "\n" |
| 623 | R"==(} )==" "\n" |
| 624 | R"==(} )==" "\n" |
| 625 | R"==(void block_read_dst(int n, DST_DATA_T *d, const __global DST_DATA_T *dst) { )==" "\n" |
| 626 | R"==(#if DST_DT_S8 || DST_DT_U8 )==" "\n" |
| 627 | R"==(__attribute__((opencl_unroll_hint)) )==" "\n" |
| 628 | R"==(for (int i = 0; i < n / 8 * 8; i += 8) { )==" "\n" |
| 629 | R"==(ushort8 block = intel_sub_group_block_read_us8( )==" "\n" |
| 630 | R"==((const __global ushort *)(dst + i * SUB_GROUP_SIZE * 2)); )==" "\n" |
| 631 | R"==(*(DST_DATA16_T *)(&d[i * 2]) = AS_DST_DATA16_T(block); )==" "\n" |
| 632 | R"==(} )==" "\n" |
| 633 | R"==(__attribute__((opencl_unroll_hint)) )==" "\n" |
| 634 | R"==(for (int i = n / 8 * 8; i < n; i++) { )==" "\n" |
| 635 | R"==(ushort block = intel_sub_group_block_read_us( )==" "\n" |
| 636 | R"==((const __global ushort *)(dst + i * SUB_GROUP_SIZE * 2)); )==" "\n" |
| 637 | R"==(*(DST_DATA2_T *)(&d[i * 2]) = AS_DST_DATA2_T(block); )==" "\n" |
| 638 | R"==(} )==" "\n" |
| 639 | R"==(return; )==" "\n" |
| 640 | R"==(#elif DST_DT_S32 || DST_DT_F32 )==" "\n" |
| 641 | R"==(int sglid = get_sub_group_local_id(); )==" "\n" |
| 642 | R"==(__attribute__((opencl_unroll_hint)) )==" "\n" |
| 643 | R"==(for (int i = 0; i < n; i++) { )==" "\n" |
| 644 | R"==(uint block0 = intel_sub_group_block_read( )==" "\n" |
| 645 | R"==((const __global uint *)(dst + i * SUB_GROUP_SIZE * 2)); )==" "\n" |
| 646 | R"==(uint block1 = intel_sub_group_block_read((const __global uint *)(dst )==" "\n" |
| 647 | R"==(+ i * SUB_GROUP_SIZE * 2 + SUB_GROUP_SIZE)); )==" "\n" |
| 648 | R"==(int from00 = min(2 * sglid, SUB_GROUP_SIZE - 1); )==" "\n" |
| 649 | R"==(int from01 = max(2 * sglid - 16, 0); )==" "\n" |
| 650 | R"==(int from10 = min(2 * sglid + 1, SUB_GROUP_SIZE - 1); )==" "\n" |
| 651 | R"==(int from11 = max(2 * sglid + 1 - 16, 0); )==" "\n" |
| 652 | R"==(uint block00 = intel_sub_group_shuffle(block0, from00); )==" "\n" |
| 653 | R"==(uint block01 = intel_sub_group_shuffle(block1, from01); )==" "\n" |
| 654 | R"==(uint block10 = intel_sub_group_shuffle(block0, from10); )==" "\n" |
| 655 | R"==(uint block11 = intel_sub_group_shuffle(block1, from11); )==" "\n" |
| 656 | R"==(block0 = (2 * sglid < SUB_GROUP_SIZE) ? block00 : block01; )==" "\n" |
| 657 | R"==(block1 = (2 * sglid + 1 < SUB_GROUP_SIZE) ? block10 : block11; )==" "\n" |
| 658 | R"==(*(DST_DATA2_T *)(&d[i * 2]) = AS_DST_DATA2_T((uint2)(block0, block1)); )==" "\n" |
| 659 | R"==(} )==" "\n" |
| 660 | R"==(return; )==" "\n" |
| 661 | R"==(#else )==" "\n" |
| 662 | R"==(#error "Not expected" )==" "\n" |
| 663 | R"==(#endif )==" "\n" |
| 664 | R"==(} )==" "\n" |
| 665 | R"==(void block_write_dst(int n, const DST_DATA_T *d, __global DST_DATA_T *dst) { )==" "\n" |
| 666 | R"==(#if DST_DT_S8 || DST_DT_U8 )==" "\n" |
| 667 | R"==(__attribute__((opencl_unroll_hint)) )==" "\n" |
| 668 | R"==(for (int i = 0; i < n / 8 * 8; i += 8) { )==" "\n" |
| 669 | R"==(intel_sub_group_block_write_us8( )==" "\n" |
| 670 | R"==((__global ushort *)(dst + i * SUB_GROUP_SIZE * 2), )==" "\n" |
| 671 | R"==(as_ushort8(*(DST_DATA16_T *)(&d[i * 2]))); )==" "\n" |
| 672 | R"==(} )==" "\n" |
| 673 | R"==(__attribute__((opencl_unroll_hint)) )==" "\n" |
| 674 | R"==(for (int i = n / 8 * 8; i < n; i++) { )==" "\n" |
| 675 | R"==(intel_sub_group_block_write_us( )==" "\n" |
| 676 | R"==((__global ushort *)(dst + i * SUB_GROUP_SIZE * 2), )==" "\n" |
| 677 | R"==(as_ushort(*(DST_DATA2_T *)(&d[i * 2]))); )==" "\n" |
| 678 | R"==(} )==" "\n" |
| 679 | R"==(return; )==" "\n" |
| 680 | R"==(#elif DST_DT_S32 || DST_DT_F32 )==" "\n" |
| 681 | R"==(int sglid = get_sub_group_local_id(); )==" "\n" |
| 682 | R"==(__attribute__((opencl_unroll_hint)) )==" "\n" |
| 683 | R"==(for (int i = 0; i < n; i++) { )==" "\n" |
| 684 | R"==(DST_DATA2_T block = AS_DST_DATA2_T(*(DST_DATA2_T *)(&d[i * 2])); )==" "\n" |
| 685 | R"==(dst[i * SUB_GROUP_SIZE * 2 + 2 * sglid] = block.S0; )==" "\n" |
| 686 | R"==(dst[i * SUB_GROUP_SIZE * 2 + 1 + 2 * sglid] = block.S1; )==" "\n" |
| 687 | R"==(} )==" "\n" |
| 688 | R"==(return; )==" "\n" |
| 689 | R"==(#else )==" "\n" |
| 690 | R"==(#error "Not expected" )==" "\n" |
| 691 | R"==(#endif )==" "\n" |
| 692 | R"==(} )==" "\n" |
| 693 | R"==()=="; |
| 694 | } |
| 695 | } |
| 696 | } |
| 697 | } |
Generated on 2022-Nov-30 from project oneDNN revision 84bc949a1
Powered by 
Code Browser 2.1
Generator usage only permitted with license.