| 1 | namespace dnnl { |
|---|---|
| 2 | namespace impl { |
| 3 | namespace gpu { |
| 4 | namespace ocl { |
| 5 | const char *xe_lp_nhwc_1x1_conv_fwd_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_scales.h" )==" "\n" |
| 23 | R"==(#include "gpu/ocl/ocl_types.h" )==" "\n" |
| 24 | R"==(#include "gpu/ocl/ocl_zero_points.h" )==" "\n" |
| 25 | R"==(#if IC % IC_BLOCK != 0 )==" "\n" |
| 26 | R"==(#define IC_NBLOCKS_TAIL ((IC % IC_BLOCK + 3) / 4) )==" "\n" |
| 27 | R"==(#else )==" "\n" |
| 28 | R"==(#define IC_NBLOCKS_TAIL 8 )==" "\n" |
| 29 | R"==(#endif )==" "\n" |
| 30 | R"==(#define IC_TAIL (IC % IC_BLOCK) )==" "\n" |
| 31 | R"==(#define HAS_PADDING (PW > 0 || PH > 0 || PD > 0) )==" "\n" |
| 32 | R"==(#define USE_SP_BLOCK ((SW == 1 && SH == 1 && SW == 1) || HAS_PADDING) )==" "\n" |
| 33 | R"==(#define SRC_SP (IW * IH * ID) )==" "\n" |
| 34 | R"==(#define SRC_SP_STRIDE (G * IC) )==" "\n" |
| 35 | R"==(#define SRC_ICB_STRIDE IC_BLOCK )==" "\n" |
| 36 | R"==(#define DST_SP (OW * OH * OD) )==" "\n" |
| 37 | R"==(#define DST_SP_STRIDE (G * OC) )==" "\n" |
| 38 | R"==(#define DST_OCB_STRIDE OC_BLOCK )==" "\n" |
| 39 | R"==(#define WEI_BLOCK_STRIDE (4 * 8 * 8 * 4) )==" "\n" |
| 40 | R"==(#if DST_DT_S8 || DST_DT_U8 )==" "\n" |
| 41 | R"==(#define OC_BLOCK_READ_BOUND 4 )==" "\n" |
| 42 | R"==(#define OC_BLOCK_WRITE_BOUND 16 )==" "\n" |
| 43 | R"==(#else )==" "\n" |
| 44 | R"==(#define OC_BLOCK_READ_BOUND 1 )==" "\n" |
| 45 | R"==(#define OC_BLOCK_WRITE_BOUND 4 )==" "\n" |
| 46 | R"==(#endif )==" "\n" |
| 47 | R"==(#define IC_BLOCK_READ_BOUND 4 )==" "\n" |
| 48 | R"==(#if IC % IC_BLOCK == 0 )==" "\n" |
| 49 | R"==(#define BLOCK_READ_SRC_Xx32(start, end, d_idx, data, idx, sp_off) \ )==" "\n" |
| 50 | R"==(do { \ )==" "\n" |
| 51 | R"==(uint *d = (uint *)&data; \ )==" "\n" |
| 52 | R"==(unroll_for(uint _i = (start); _i < (end); ++_i) { \ )==" "\n" |
| 53 | R"==(if (HAS_PADDING) { \ )==" "\n" |
| 54 | R"==(PAD_BLOCK_READ(d[_i + d_idx], src, sp, _i + sp_off, 0, 0) \ )==" "\n" |
| 55 | R"==(} else { \ )==" "\n" |
| 56 | R"==(d[_i + d_idx] = AS_SRC_MMAD_DATA_T(intel_sub_group_block_read( \ )==" "\n" |
| 57 | R"==((__global uint *)&src[idx + _i * SW * G * IC])); \ )==" "\n" |
| 58 | R"==(} \ )==" "\n" |
| 59 | R"==(} \ )==" "\n" |
| 60 | R"==(} while (0); )==" "\n" |
| 61 | R"==(#elif IC % 4 == 0 )==" "\n" |
| 62 | R"==(#define BLOCK_READ_SRC_Xx32(start, end, d_idx, data, idx, sp_off) \ )==" "\n" |
| 63 | R"==(do { \ )==" "\n" |
| 64 | R"==(uint *d = (uint *)&data; \ )==" "\n" |
| 65 | R"==(unroll_for(uint _i = (start); _i < (end); ++_i) { \ )==" "\n" |
| 66 | R"==(__global uchar *s = &src[idx + _i * SW * G * IC]; \ )==" "\n" |
| 67 | R"==(if (ic_block_id < IC_NCHUNK - 1 \ )==" "\n" |
| 68 | R"==(|| sg_local_id * 4 < IC_TAIL - IC_TAIL % 4) { \ )==" "\n" |
| 69 | R"==(if (HAS_PADDING) { \ )==" "\n" |
| 70 | R"==(PAD_BLOCK_READ(d[_i + d_idx], src, sp, _i + sp_off, 1, 0) \ )==" "\n" |
| 71 | R"==(} else { \ )==" "\n" |
| 72 | R"==(d[_i + d_idx] = *((__global uint *)&s[sg_local_id * 4]); \ )==" "\n" |
| 73 | R"==(} \ )==" "\n" |
| 74 | R"==(} \ )==" "\n" |
| 75 | R"==(} \ )==" "\n" |
| 76 | R"==(} while (0); )==" "\n" |
| 77 | R"==(#else )==" "\n" |
| 78 | R"==(#define BLOCK_READ_SRC_Xx32(start, end, d_idx, data, idx, sp_off) \ )==" "\n" |
| 79 | R"==(do { \ )==" "\n" |
| 80 | R"==(uint *d = (uint *)&data; \ )==" "\n" |
| 81 | R"==(unroll_for(uint _i = (start); _i < (end); ++_i) { \ )==" "\n" |
| 82 | R"==(__global uchar *s = &src[idx + _i * SW * G * IC]; \ )==" "\n" |
| 83 | R"==(uint _j_max = (sg_local_id * 4 < IC_TAIL - IC_TAIL % 4) \ )==" "\n" |
| 84 | R"==(|| (ic_block_id < IC_NCHUNK - 1) \ )==" "\n" |
| 85 | R"==(? 4 \ )==" "\n" |
| 86 | R"==(: (sg_local_id * 4 == IC_TAIL - IC_TAIL % 4 ? IC_TAIL % 4 \ )==" "\n" |
| 87 | R"==(: 0); \ )==" "\n" |
| 88 | R"==(unroll_for(uint _j = 0; _j < _j_max; ++_j) { \ )==" "\n" |
| 89 | R"==(if (HAS_PADDING) { \ )==" "\n" |
| 90 | R"==(PAD_BLOCK_READ(d[_i + d_idx], src, sp, _i + sp_off, 2, _j) \ )==" "\n" |
| 91 | R"==(} else { \ )==" "\n" |
| 92 | R"==(*((uchar *)&d[_i + d_idx] + _j) = s[sg_local_id * 4 + _j]; \ )==" "\n" |
| 93 | R"==(} \ )==" "\n" |
| 94 | R"==(} \ )==" "\n" |
| 95 | R"==(} \ )==" "\n" |
| 96 | R"==(} while (0); )==" "\n" |
| 97 | R"==(#endif )==" "\n" |
| 98 | R"==(#if SP_BLOCK == 4 )==" "\n" |
| 99 | R"==(#define BLOCK0 4 )==" "\n" |
| 100 | R"==(#define ACC_DATA_BLOCK int4 )==" "\n" |
| 101 | R"==(#define SRC_DATA_BLOCK_T SRC_MMAD_DATA4_T )==" "\n" |
| 102 | R"==(#define BLOCK_READ_SRC BLOCK_READ_SRC_4x32 )==" "\n" |
| 103 | R"==(DECLARE_MMAD_EMU(mmad_tail0, idot4, IC_NBLOCKS_TAIL, 4, SRC_DATA_BLOCK_T, int8, )==" "\n" |
| 104 | R"==(ACC_DATA_BLOCK) )==" "\n" |
| 105 | R"==(#define MMAD_FULL0 mmad8x4 )==" "\n" |
| 106 | R"==(#define MMAD_TAIL0 mmad_tail0 )==" "\n" |
| 107 | R"==(#else )==" "\n" |
| 108 | R"==(#define BLOCK0 8 )==" "\n" |
| 109 | R"==(#define ACC_DATA_BLOCK int8 )==" "\n" |
| 110 | R"==(#define SRC_DATA_BLOCK_T SRC_MMAD_DATA8_T )==" "\n" |
| 111 | R"==(#define BLOCK_READ_SRC BLOCK_READ_SRC_8x32 )==" "\n" |
| 112 | R"==(DECLARE_MMAD_EMU(mmad_tail0, idot4, IC_NBLOCKS_TAIL, 8, SRC_DATA_BLOCK_T, int8, )==" "\n" |
| 113 | R"==(ACC_DATA_BLOCK) )==" "\n" |
| 114 | R"==(#define MMAD_FULL0 mmad8x8 )==" "\n" |
| 115 | R"==(#define MMAD_TAIL0 mmad_tail0 )==" "\n" |
| 116 | R"==(#endif )==" "\n" |
| 117 | R"==(#if SP_BLOCK == 12 )==" "\n" |
| 118 | R"==(#define BLOCK1 4 )==" "\n" |
| 119 | R"==(#define ACC_DATA_BLOCK1 int4 )==" "\n" |
| 120 | R"==(#define SRC_DATA_BLOCK_T1 SRC_MMAD_DATA4_T )==" "\n" |
| 121 | R"==(#define DST_DATA_BLOCK_T1 uint4 )==" "\n" |
| 122 | R"==(#define BLOCK_READ_SRC1 BLOCK_READ_SRC_4x32 )==" "\n" |
| 123 | R"==(DECLARE_MMAD_EMU(mmad_tail1, idot4, IC_NBLOCKS_TAIL, 4, SRC_DATA_BLOCK_T1, int8, )==" "\n" |
| 124 | R"==(ACC_DATA_BLOCK1) )==" "\n" |
| 125 | R"==(#define MMAD_FULL1 mmad8x4 )==" "\n" |
| 126 | R"==(#define MMAD_TAIL1 mmad_tail1 )==" "\n" |
| 127 | R"==(#else )==" "\n" |
| 128 | R"==(#define BLOCK1 8 )==" "\n" |
| 129 | R"==(#define ACC_DATA_BLOCK1 int8 )==" "\n" |
| 130 | R"==(#define SRC_DATA_BLOCK_T1 SRC_MMAD_DATA8_T )==" "\n" |
| 131 | R"==(#define DST_DATA_BLOCK_T1 uint8 )==" "\n" |
| 132 | R"==(#define BLOCK_READ_SRC1 BLOCK_READ_SRC_8x32 )==" "\n" |
| 133 | R"==(DECLARE_MMAD_EMU(mmad_tail1, idot4, IC_NBLOCKS_TAIL, 8, SRC_DATA_BLOCK_T1, int8, )==" "\n" |
| 134 | R"==(ACC_DATA_BLOCK1) )==" "\n" |
| 135 | R"==(#define MMAD_FULL1 mmad8x8 )==" "\n" |
| 136 | R"==(#define MMAD_TAIL1 mmad_tail1 )==" "\n" |
| 137 | R"==(#endif )==" "\n" |
| 138 | R"==(#if INT8_WEI_SLM )==" "\n" |
| 139 | R"==(#define BLOCK_READ_WHT_1x32(data, idx) \ )==" "\n" |
| 140 | R"==(data = as_int(block_read((__local uint *)&wei_tmp[idx])); )==" "\n" |
| 141 | R"==(#define BLOCK_READ_WHT_8x32(data, idx) \ )==" "\n" |
| 142 | R"==(data = as_int8(block_read8((__local uint *)&wei_tmp[idx])); )==" "\n" |
| 143 | R"==(#else )==" "\n" |
| 144 | R"==(#define BLOCK_READ_WHT_1x32(data, idx) \ )==" "\n" |
| 145 | R"==(data = as_int(intel_sub_group_block_read((__global uint *)&wei[idx])); )==" "\n" |
| 146 | R"==(#define BLOCK_READ_WHT_8x32(data, idx) \ )==" "\n" |
| 147 | R"==(data = as_int8(intel_sub_group_block_read8((__global uint *)&wei[idx])); )==" "\n" |
| 148 | R"==(#endif )==" "\n" |
| 149 | R"==(#if OC % OC_BLOCK == 0 )==" "\n" |
| 150 | R"==(#define BLOCK_READ_BIA(data, idx) \ )==" "\n" |
| 151 | R"==(data = as_float4(intel_sub_group_block_read4((__global uint *)&bias[idx])); )==" "\n" |
| 152 | R"==(#else )==" "\n" |
| 153 | R"==(#define BLOCK_READ_BIA(data, idx) \ )==" "\n" |
| 154 | R"==(data = (float4)0; \ )==" "\n" |
| 155 | R"==(int i; \ )==" "\n" |
| 156 | R"==(for (i = idx; i < idx + OC_BLOCK && i < OC - (OC % SUB_GROUP_SIZE); \ )==" "\n" |
| 157 | R"==(i += SUB_GROUP_SIZE) { \ )==" "\n" |
| 158 | R"==(data[(i - idx) / SUB_GROUP_SIZE] = as_float( \ )==" "\n" |
| 159 | R"==(intel_sub_group_block_read((__global uint *)&bias[i])); \ )==" "\n" |
| 160 | R"==(} \ )==" "\n" |
| 161 | R"==(if ((get_sub_group_local_id() < OC % SUB_GROUP_SIZE) && i < OC \ )==" "\n" |
| 162 | R"==(&& (i - idx) / SUB_GROUP_SIZE < 4) { \ )==" "\n" |
| 163 | R"==(data[(i - idx) / SUB_GROUP_SIZE] \ )==" "\n" |
| 164 | R"==(= as_float(bias[i + get_sub_group_local_id()]); \ )==" "\n" |
| 165 | R"==(} )==" "\n" |
| 166 | R"==(#endif )==" "\n" |
| 167 | R"==(#define PAD_BLOCK_READ(data, src, sp, i, read_kind, l_off) \ )==" "\n" |
| 168 | R"==(do { \ )==" "\n" |
| 169 | R"==(const int od = (sp + i) / (OW * OH); \ )==" "\n" |
| 170 | R"==(const int ohw = (sp + i) % (OW * OH); \ )==" "\n" |
| 171 | R"==(const int oh = ohw / OW; \ )==" "\n" |
| 172 | R"==(const int ow = (ohw % OW); \ )==" "\n" |
| 173 | R"==(const int id = SD * od - PD; \ )==" "\n" |
| 174 | R"==(const int ih = SH * oh - PH; \ )==" "\n" |
| 175 | R"==(const int iw = SW * ow - PW; \ )==" "\n" |
| 176 | R"==(bool pad = ((PW > 0 || PH > 0 || PD > 0) \ )==" "\n" |
| 177 | R"==(&& (iw < 0 || ih < 0 || id < 0 || iw >= IW || ih >= IH \ )==" "\n" |
| 178 | R"==(|| id >= ID)); \ )==" "\n" |
| 179 | R"==(int off = id * IH * IW + ih * IW + iw; \ )==" "\n" |
| 180 | R"==(if (read_kind == 0) { \ )==" "\n" |
| 181 | R"==(data = pad ? 0 \ )==" "\n" |
| 182 | R"==(: AS_SRC_MMAD_DATA_T(intel_sub_group_block_read( \ )==" "\n" |
| 183 | R"==((global uint *)&src[off * SRC_SP_STRIDE])); \ )==" "\n" |
| 184 | R"==(} else if (read_kind == 1) { \ )==" "\n" |
| 185 | R"==(data = pad ? 0 \ )==" "\n" |
| 186 | R"==(: *((__global uint *)&src[off * SRC_SP_STRIDE \ )==" "\n" |
| 187 | R"==(+ sg_local_id * 4]); \ )==" "\n" |
| 188 | R"==(} else if (read_kind == 2) { \ )==" "\n" |
| 189 | R"==(*((uchar *)&data + l_off) = pad \ )==" "\n" |
| 190 | R"==(? 0 \ )==" "\n" |
| 191 | R"==(: src[off * SRC_SP_STRIDE + sg_local_id * 4 + l_off]; \ )==" "\n" |
| 192 | R"==(} \ )==" "\n" |
| 193 | R"==(} while (0); )==" "\n" |
| 194 | R"==(#if SCALES_PER_OC )==" "\n" |
| 195 | R"==(#define SCALE scales )==" "\n" |
| 196 | R"==(#elif SCALES_COMMON )==" "\n" |
| 197 | R"==(#define SCALE runtime_scales[0] )==" "\n" |
| 198 | R"==(#else )==" "\n" |
| 199 | R"==(#define SCALE 1 )==" "\n" |
| 200 | R"==(#endif )==" "\n" |
| 201 | R"==(void block_read_dst( )==" "\n" |
| 202 | R"==(int n, DST_DATA_T *d, const __global DST_DATA_T *dst, int oc_tail); )==" "\n" |
| 203 | R"==(void block_write_dst( )==" "\n" |
| 204 | R"==(int n, const DST_DATA_T *d, __global DST_DATA_T *dst, int oc_tail); )==" "\n" |
| 205 | R"==(__attribute__((intel_reqd_sub_group_size(SUB_GROUP_SIZE))) )==" "\n" |
| 206 | R"==(__attribute__((reqd_work_group_size(LWS_0, LWS_1, LWS_2))) __kernel void )==" "\n" |
| 207 | R"==(xe_lp_nhwc_1x1_conv_fwd_x8s8x(const __global SRC_DATA_T *src, )==" "\n" |
| 208 | R"==(const __global char *wei, const __global float *bias, )==" "\n" |
| 209 | R"==(__global DST_DATA_T *dst POST_OP_ARGS, )==" "\n" |
| 210 | R"==(const __global float *runtime_scales, )==" "\n" |
| 211 | R"==(const __global int *src_compensation, )==" "\n" |
| 212 | R"==(const __global int *dst_compensation) { )==" "\n" |
| 213 | R"==(const uint oc_group_id = get_group_id(0); )==" "\n" |
| 214 | R"==(const uint sp_group_id = get_group_id(1); )==" "\n" |
| 215 | R"==(const uint mb_group_id = get_group_id(2); )==" "\n" |
| 216 | R"==(const uint ic_group_id = oc_group_id / OC_NCHUNK * IC_NCHUNK; )==" "\n" |
| 217 | R"==(const uint sg_local_id = get_sub_group_local_id(); )==" "\n" |
| 218 | R"==(const uint sg_id = get_sub_group_id(); )==" "\n" |
| 219 | R"==(#define OWB ((OW + SP_BLOCK - 1) / SP_BLOCK) )==" "\n" |
| 220 | R"==(#if USE_SP_BLOCK )==" "\n" |
| 221 | R"==(const uint sp = get_global_id(1) * SP_BLOCK; )==" "\n" |
| 222 | R"==(const int sp_local_id = get_local_id(1); )==" "\n" |
| 223 | R"==(const uint od = sp / (OH * OW); )==" "\n" |
| 224 | R"==(const uint ohw = sp % (OH * OW); )==" "\n" |
| 225 | R"==(const uint oh = ohw / OW; )==" "\n" |
| 226 | R"==(const uint ow = (ohw % OW); )==" "\n" |
| 227 | R"==(#else )==" "\n" |
| 228 | R"==(const uint sp = get_global_id(1); )==" "\n" |
| 229 | R"==(const int sp_local_id = get_local_id(1); )==" "\n" |
| 230 | R"==(const uint od = sp / (OWB * OH); )==" "\n" |
| 231 | R"==(const uint ohw = sp % (OWB * OH); )==" "\n" |
| 232 | R"==(const uint oh = ohw / OWB; )==" "\n" |
| 233 | R"==(const uint ow = (ohw % OWB) * SP_BLOCK; )==" "\n" |
| 234 | R"==(#endif )==" "\n" |
| 235 | R"==(const uint id = SD * od; )==" "\n" |
| 236 | R"==(const uint ih = SH * oh; )==" "\n" |
| 237 | R"==(const uint iw = SW * ow; )==" "\n" |
| 238 | R"==(src += mb_group_id * SRC_SP * SRC_SP_STRIDE; )==" "\n" |
| 239 | R"==(#if !HAS_PADDING )==" "\n" |
| 240 | R"==(src += (id * IH * IW + ih * IW + iw) * SRC_SP_STRIDE; )==" "\n" |
| 241 | R"==(#endif )==" "\n" |
| 242 | R"==(src += ic_group_id * SRC_ICB_STRIDE; )==" "\n" |
| 243 | R"==(dst += mb_group_id * DST_SP * DST_SP_STRIDE; )==" "\n" |
| 244 | R"==(dst += (od * OH * OW + oh * OW + ow) * DST_SP_STRIDE; )==" "\n" |
| 245 | R"==(dst += oc_group_id * DST_OCB_STRIDE; )==" "\n" |
| 246 | R"==(wei += oc_group_id * WEI_BLOCK_STRIDE * IC_NCHUNK; )==" "\n" |
| 247 | R"==(ACC_DATA_BLOCK C00 = 0, C01 = 0, C02 = 0, C03 = 0; )==" "\n" |
| 248 | R"==(ACC_DATA_BLOCK1 C10 = 0, C11 = 0, C12 = 0, C13 = 0; )==" "\n" |
| 249 | R"==(const int oc_tail )==" "\n" |
| 250 | R"==(= (oc_group_id + 1) * OC_BLOCK > G * OC ? OC % OC_BLOCK : OC_BLOCK; )==" "\n" |
| 251 | R"==(#if INT8_WEI_SLM )==" "\n" |
| 252 | R"==(#define READ_SLM() \ )==" "\n" |
| 253 | R"==(barrier(CLK_LOCAL_MEM_FENCE); \ )==" "\n" |
| 254 | R"==(const __global char *wei_copy_from \ )==" "\n" |
| 255 | R"==(= wei + sp_local_id * WEI_BLOCK_STRIDE / LWS_1; \ )==" "\n" |
| 256 | R"==(__local char *wei_copy_to \ )==" "\n" |
| 257 | R"==(= wei_slm + sp_local_id * WEI_BLOCK_STRIDE / LWS_1; \ )==" "\n" |
| 258 | R"==(block_write4((__local uint *)wei_copy_to, \ )==" "\n" |
| 259 | R"==(intel_sub_group_block_read4((__global uint *)wei_copy_from)); \ )==" "\n" |
| 260 | R"==(__local char *wei_tmp = wei_slm; \ )==" "\n" |
| 261 | R"==(barrier(CLK_LOCAL_MEM_FENCE); )==" "\n" |
| 262 | R"==(__local char wei_slm[WEI_BLOCK_STRIDE]; )==" "\n" |
| 263 | R"==(#endif )==" "\n" |
| 264 | R"==(for (uint ic_block_id = 0; ic_block_id < IC_NCHUNK; ++ic_block_id) { )==" "\n" |
| 265 | R"==(#if INT8_WEI_SLM )==" "\n" |
| 266 | R"==(READ_SLM() )==" "\n" |
| 267 | R"==(#if SP_TAIL )==" "\n" |
| 268 | R"==(if (ow < OW) )==" "\n" |
| 269 | R"==(#endif )==" "\n" |
| 270 | R"==(#endif )==" "\n" |
| 271 | R"==({ )==" "\n" |
| 272 | R"==(SRC_DATA_BLOCK_T S0; )==" "\n" |
| 273 | R"==(SRC_DATA_BLOCK_T1 S1; )==" "\n" |
| 274 | R"==(#if OUT_SP_TAIL )==" "\n" |
| 275 | R"==(#if USE_SP_BLOCK )==" "\n" |
| 276 | R"==(if (od * OH * OW + oh * OW + ow + SP_BLOCK > DST_SP) { )==" "\n" |
| 277 | R"==(#else )==" "\n" |
| 278 | R"==(if (ow + SP_BLOCK > OW) { )==" "\n" |
| 279 | R"==(#endif )==" "\n" |
| 280 | R"==(#if OUT_SP_TAIL < 8 )==" "\n" |
| 281 | R"==(S0 = 0; )==" "\n" |
| 282 | R"==(BLOCK_READ_SRC_Xx32(0, OUT_SP_TAIL, 0, S0, 0 * IC, 0); )==" "\n" |
| 283 | R"==(#else )==" "\n" |
| 284 | R"==(BLOCK_READ_SRC_Xx32(0, BLOCK0, 0, S0, 0, 0); )==" "\n" |
| 285 | R"==(S1 = 0; )==" "\n" |
| 286 | R"==(BLOCK_READ_SRC_Xx32(8, OUT_SP_TAIL, -8, S1, 0, BLOCK0); )==" "\n" |
| 287 | R"==(#endif )==" "\n" |
| 288 | R"==(} else )==" "\n" |
| 289 | R"==(#endif )==" "\n" |
| 290 | R"==({ )==" "\n" |
| 291 | R"==(BLOCK_READ_SRC_Xx32(0, BLOCK0, 0, S0, 0 * IC, 0); )==" "\n" |
| 292 | R"==(#if SP_BLOCK > 8 )==" "\n" |
| 293 | R"==(BLOCK_READ_SRC_Xx32(0, BLOCK1, 0, S1, 8 * SW * G * IC, BLOCK0); )==" "\n" |
| 294 | R"==(#endif )==" "\n" |
| 295 | R"==(} )==" "\n" |
| 296 | R"==(int8 W0 = 0, W1 = 0, W2 = 0, W3 = 0; )==" "\n" |
| 297 | R"==(#if IC % IC_BLOCK != 0 )==" "\n" |
| 298 | R"==(if (ic_block_id == IC_NCHUNK - 1) { )==" "\n" |
| 299 | R"==(unroll_for(int i = 0; i < IC_NBLOCKS_TAIL; ++i) )==" "\n" |
| 300 | R"==(BLOCK_READ_WHT_1x32(W0[i], (i + 0) * IC_BLOCK); )==" "\n" |
| 301 | R"==(if (OC > 8) )==" "\n" |
| 302 | R"==(unroll_for(int i = 0; i < IC_NBLOCKS_TAIL; ++i) )==" "\n" |
| 303 | R"==(BLOCK_READ_WHT_1x32(W1[i], (i + 8) * IC_BLOCK); )==" "\n" |
| 304 | R"==(if (OC > 16) )==" "\n" |
| 305 | R"==(unroll_for(int i = 0; i < IC_NBLOCKS_TAIL; ++i) )==" "\n" |
| 306 | R"==(BLOCK_READ_WHT_1x32(W2[i], (i + 16) * IC_BLOCK); )==" "\n" |
| 307 | R"==(if (OC > 24) )==" "\n" |
| 308 | R"==(unroll_for(int i = 0; i < IC_NBLOCKS_TAIL; ++i) )==" "\n" |
| 309 | R"==(BLOCK_READ_WHT_1x32(W3[i], (i + 24) * IC_BLOCK); )==" "\n" |
| 310 | R"==(C00 = MMAD_TAIL0(S0, W0, C00); )==" "\n" |
| 311 | R"==(if (OC > 8) C01 = MMAD_TAIL0(S0, W1, C01); )==" "\n" |
| 312 | R"==(if (OC > 16) C02 = MMAD_TAIL0(S0, W2, C02); )==" "\n" |
| 313 | R"==(if (OC > 24) C03 = MMAD_TAIL0(S0, W3, C03); )==" "\n" |
| 314 | R"==(#if SP_BLOCK > 8 )==" "\n" |
| 315 | R"==(C10 = MMAD_TAIL1(S1, W0, C10); )==" "\n" |
| 316 | R"==(if (OC > 8) C11 = MMAD_TAIL1(S1, W1, C11); )==" "\n" |
| 317 | R"==(if (OC > 16) C12 = MMAD_TAIL1(S1, W2, C12); )==" "\n" |
| 318 | R"==(if (OC > 24) C13 = MMAD_TAIL1(S1, W3, C13); )==" "\n" |
| 319 | R"==(#endif )==" "\n" |
| 320 | R"==(} else )==" "\n" |
| 321 | R"==(#endif )==" "\n" |
| 322 | R"==({ )==" "\n" |
| 323 | R"==(BLOCK_READ_WHT_8x32(W0, 0); )==" "\n" |
| 324 | R"==(if (OC > 8) BLOCK_READ_WHT_8x32(W1, 8 * IC_BLOCK); )==" "\n" |
| 325 | R"==(if (OC > 16) BLOCK_READ_WHT_8x32(W2, 16 * IC_BLOCK); )==" "\n" |
| 326 | R"==(if (OC > 24) BLOCK_READ_WHT_8x32(W3, 24 * IC_BLOCK); )==" "\n" |
| 327 | R"==(C00 = MMAD_FULL0(S0, W0, C00); )==" "\n" |
| 328 | R"==(if (OC > 8) C01 = MMAD_FULL0(S0, W1, C01); )==" "\n" |
| 329 | R"==(if (OC > 16) C02 = MMAD_FULL0(S0, W2, C02); )==" "\n" |
| 330 | R"==(if (OC > 24) C03 = MMAD_FULL0(S0, W3, C03); )==" "\n" |
| 331 | R"==(#if SP_BLOCK > 8 )==" "\n" |
| 332 | R"==(C10 = MMAD_FULL1(S1, W0, C10); )==" "\n" |
| 333 | R"==(if (OC > 8) C11 = MMAD_FULL1(S1, W1, C11); )==" "\n" |
| 334 | R"==(if (OC > 16) C12 = MMAD_FULL1(S1, W2, C12); )==" "\n" |
| 335 | R"==(if (OC > 24) C13 = MMAD_FULL1(S1, W3, C13); )==" "\n" |
| 336 | R"==(#endif )==" "\n" |
| 337 | R"==(} )==" "\n" |
| 338 | R"==(} )==" "\n" |
| 339 | R"==(src += SRC_ICB_STRIDE; )==" "\n" |
| 340 | R"==(wei += WEI_BLOCK_STRIDE; )==" "\n" |
| 341 | R"==(} )==" "\n" |
| 342 | R"==(#if WITH_SRC_ZPOINTS )==" "\n" |
| 343 | R"==(int4 src_comp = as_int4(intel_sub_group_block_read4( )==" "\n" |
| 344 | R"==((__global uint *)(&src_compensation[oc_group_id * OC_BLOCK]))); )==" "\n" |
| 345 | R"==(C00 -= src_comp.s0; )==" "\n" |
| 346 | R"==(C01 -= src_comp.s1; )==" "\n" |
| 347 | R"==(C02 -= src_comp.s2; )==" "\n" |
| 348 | R"==(C03 -= src_comp.s3; )==" "\n" |
| 349 | R"==(#if SP_BLOCK > 8 )==" "\n" |
| 350 | R"==(C10 -= src_comp.s0; )==" "\n" |
| 351 | R"==(C11 -= src_comp.s1; )==" "\n" |
| 352 | R"==(C12 -= src_comp.s2; )==" "\n" |
| 353 | R"==(C13 -= src_comp.s3; )==" "\n" |
| 354 | R"==(#endif )==" "\n" |
| 355 | R"==(#endif )==" "\n" |
| 356 | R"==(float4 tmp; )==" "\n" |
| 357 | R"==(DST_DATA4_T dst_pack[8]; )==" "\n" |
| 358 | R"==(DST_DATA4_T D0[BLOCK0] = {0}; )==" "\n" |
| 359 | R"==(DST_DATA4_T D1[BLOCK1] = {0}; )==" "\n" |
| 360 | R"==(#if SCALES_PER_OC )==" "\n" |
| 361 | R"==(float4 scales; )==" "\n" |
| 362 | R"==(block_read_scales( )==" "\n" |
| 363 | R"==(&scales, oc_group_id * OC_BLOCK, sg_local_id, runtime_scales); )==" "\n" |
| 364 | R"==(#endif )==" "\n" |
| 365 | R"==(#if WITH_BIAS )==" "\n" |
| 366 | R"==(float4 bia; )==" "\n" |
| 367 | R"==(BLOCK_READ_BIA(bia, oc_group_id * OC_BLOCK); )==" "\n" |
| 368 | R"==(#define QUANTIZE_ADD_BIAS() tmp = SCALE * fma(tmp, (float4)1, bia); )==" "\n" |
| 369 | R"==(#else )==" "\n" |
| 370 | R"==(#define QUANTIZE_ADD_BIAS() tmp *= SCALE; )==" "\n" |
| 371 | R"==(#endif )==" "\n" |
| 372 | R"==(#if WITH_SUM )==" "\n" |
| 373 | R"==(#if USE_SP_BLOCK )==" "\n" |
| 374 | R"==(if (OUT_SP_TAIL && od * OH * OW + oh * OW + ow + SP_BLOCK > DST_SP) { )==" "\n" |
| 375 | R"==(#else )==" "\n" |
| 376 | R"==(if (OUT_SP_TAIL && ow + SP_BLOCK > OW) { )==" "\n" |
| 377 | R"==(#endif )==" "\n" |
| 378 | R"==(#if OUT_SP_TAIL < 8 )==" "\n" |
| 379 | R"==(block_read_dst(OUT_SP_TAIL, D0, dst, oc_tail); )==" "\n" |
| 380 | R"==(#else )==" "\n" |
| 381 | R"==(block_read_dst(BLOCK0, D0, dst, oc_tail); )==" "\n" |
| 382 | R"==(block_read_dst(OUT_SP_TAIL - 8, D1, dst + 8 * G * OC, oc_tail); )==" "\n" |
| 383 | R"==(#endif )==" "\n" |
| 384 | R"==(} else { )==" "\n" |
| 385 | R"==(block_read_dst(BLOCK0, D0, dst, oc_tail); )==" "\n" |
| 386 | R"==(if (SP_BLOCK > 8) { )==" "\n" |
| 387 | R"==(block_read_dst(BLOCK1, D1, dst + 8 * G * OC, oc_tail); )==" "\n" |
| 388 | R"==(} )==" "\n" |
| 389 | R"==(} )==" "\n" |
| 390 | R"==(#endif )==" "\n" |
| 391 | R"==(#if WITH_DST_ZPOINTS )==" "\n" |
| 392 | R"==(int4 dst_comp = read_dst_zero_points_32c( )==" "\n" |
| 393 | R"==(dst_compensation, oc_group_id * OC_BLOCK); )==" "\n" |
| 394 | R"==(#define ADD_DST_COMPENSATION() tmp += convert_float4(dst_comp); )==" "\n" |
| 395 | R"==(#else )==" "\n" |
| 396 | R"==(#define ADD_DST_COMPENSATION() )==" "\n" |
| 397 | R"==(#endif )==" "\n" |
| 398 | R"==(#define PACK(C0, C1, C2, C3, idx) \ )==" "\n" |
| 399 | R"==(do { \ )==" "\n" |
| 400 | R"==(tmp[0] = C0[idx]; \ )==" "\n" |
| 401 | R"==(tmp[1] = C1[idx]; \ )==" "\n" |
| 402 | R"==(tmp[2] = C2[idx]; \ )==" "\n" |
| 403 | R"==(tmp[3] = C3[idx]; \ )==" "\n" |
| 404 | R"==(} while (0) )==" "\n" |
| 405 | R"==(#define CONVERT_PACK(idx) \ )==" "\n" |
| 406 | R"==(do { \ )==" "\n" |
| 407 | R"==(dst_pack[idx] = CONVERT_DST_DATA4_T(tmp); \ )==" "\n" |
| 408 | R"==(} while (0) )==" "\n" |
| 409 | R"==(#define STORE_DST(n, C0, C1, C2, C3, D, dst_ptr, mb_stride) \ )==" "\n" |
| 410 | R"==(do { \ )==" "\n" |
| 411 | R"==(for (int n_i = 0; n_i < n; n_i++) { \ )==" "\n" |
| 412 | R"==(PACK(C0, C1, C2, C3, n_i); \ )==" "\n" |
| 413 | R"==(QUANTIZE_ADD_BIAS(); \ )==" "\n" |
| 414 | R"==(for (int didx = 0; didx < 4; ++didx) { \ )==" "\n" |
| 415 | R"==(float tmp_i = tmp[didx]; \ )==" "\n" |
| 416 | R"==(float dni_i = convert_float(AS_SUM_DATA_T(D[n_i][didx])); \ )==" "\n" |
| 417 | R"==(int po_mb; \ )==" "\n" |
| 418 | R"==(if (MB_BLOCK == 32) \ )==" "\n" |
| 419 | R"==(po_mb = (mb_group_id * MB_BLOCK / 2 + mb_stride * 8 + n_i) \ )==" "\n" |
| 420 | R"==(% MB; \ )==" "\n" |
| 421 | R"==(else \ )==" "\n" |
| 422 | R"==(po_mb = mb_group_id % MB; \ )==" "\n" |
| 423 | R"==(const int po_oc = (oc_group_id * OC_BLOCK + sg_local_id \ )==" "\n" |
| 424 | R"==(+ didx * SUB_GROUP_SIZE) \ )==" "\n" |
| 425 | R"==(% (OC * G); \ )==" "\n" |
| 426 | R"==(APPLY_POST_OPS_SERIAL_BINARY_2D( \ )==" "\n" |
| 427 | R"==(tmp_i, float, dni_i, float, po_mb, 1, po_oc, 1); \ )==" "\n" |
| 428 | R"==(tmp[didx] = tmp_i; \ )==" "\n" |
| 429 | R"==(} \ )==" "\n" |
| 430 | R"==(ADD_DST_COMPENSATION(); \ )==" "\n" |
| 431 | R"==(CONVERT_PACK(n_i); \ )==" "\n" |
| 432 | R"==(} \ )==" "\n" |
| 433 | R"==(block_write_dst(n, dst_pack, dst_ptr, oc_tail); \ )==" "\n" |
| 434 | R"==(} while (0) )==" "\n" |
| 435 | R"==(#if INT8_WEI_SLM && SP_TAIL )==" "\n" |
| 436 | R"==(if (ow < OW) )==" "\n" |
| 437 | R"==(#endif )==" "\n" |
| 438 | R"==({ )==" "\n" |
| 439 | R"==(#if USE_SP_BLOCK )==" "\n" |
| 440 | R"==(if (OUT_SP_TAIL && od * OH * OW + oh * OW + ow + SP_BLOCK > DST_SP) { )==" "\n" |
| 441 | R"==(#else )==" "\n" |
| 442 | R"==(if (OUT_SP_TAIL && ow + SP_BLOCK > OW) { )==" "\n" |
| 443 | R"==(#endif )==" "\n" |
| 444 | R"==(STORE_DST(min(BLOCK0, OUT_SP_TAIL), C00, C01, C02, C03, D0, dst, 0); )==" "\n" |
| 445 | R"==(STORE_DST(OUT_SP_TAIL - 8, C10, C11, C12, C13, D1, dst + 8 * G * OC, )==" "\n" |
| 446 | R"==(1); )==" "\n" |
| 447 | R"==(} else { )==" "\n" |
| 448 | R"==(STORE_DST(BLOCK0, C00, C01, C02, C03, D0, dst, 0); )==" "\n" |
| 449 | R"==(if (SP_BLOCK > 8) { )==" "\n" |
| 450 | R"==(STORE_DST(BLOCK1, C10, C11, C12, C13, D1, dst + 8 * G * OC, 1); )==" "\n" |
| 451 | R"==(} )==" "\n" |
| 452 | R"==(} )==" "\n" |
| 453 | R"==(} )==" "\n" |
| 454 | R"==(} )==" "\n" |
| 455 | R"==(void block_read_dst( )==" "\n" |
| 456 | R"==(int n, DST_DATA_T *d, const __global DST_DATA_T *dst, int oc_tail) { )==" "\n" |
| 457 | R"==(const int local_id = get_sub_group_local_id(); )==" "\n" |
| 458 | R"==(int nelems = n * 4; )==" "\n" |
| 459 | R"==(__attribute__((opencl_unroll_hint)) )==" "\n" |
| 460 | R"==(for (int i = 0; i < nelems; i += 4) { )==" "\n" |
| 461 | R"==(if (OC % OC_BLOCK_WRITE_BOUND == 0 && oc_tail == OC_BLOCK) { )==" "\n" |
| 462 | R"==(*((DST_DATA4_T *)&d[i]) = BLOCK_READ_DST4(dst + (i / 4) * G * OC); )==" "\n" |
| 463 | R"==(} else { )==" "\n" |
| 464 | R"==(unroll_for(int idx = 0; idx < 4; idx++) { )==" "\n" |
| 465 | R"==(if (local_id + 8 * idx < oc_tail) { )==" "\n" |
| 466 | R"==(d[i + idx] = dst[(i / 4) * G * OC + idx * 8 + local_id]; )==" "\n" |
| 467 | R"==(} )==" "\n" |
| 468 | R"==(} )==" "\n" |
| 469 | R"==(} )==" "\n" |
| 470 | R"==(} )==" "\n" |
| 471 | R"==(} )==" "\n" |
| 472 | R"==(void block_write_dst( )==" "\n" |
| 473 | R"==(int n, const DST_DATA_T *d, __global DST_DATA_T *dst, int oc_tail) { )==" "\n" |
| 474 | R"==(const int local_id = get_sub_group_local_id(); )==" "\n" |
| 475 | R"==(int nelems = n * 4; )==" "\n" |
| 476 | R"==(__attribute__((opencl_unroll_hint)) )==" "\n" |
| 477 | R"==(for (int i = 0; i < nelems; i += 4) { )==" "\n" |
| 478 | R"==(if (OC % OC_BLOCK_WRITE_BOUND == 0 && oc_tail == OC_BLOCK) { )==" "\n" |
| 479 | R"==(BLOCK_WRITE_DST4(dst + (i / 4) * G * OC, *((DST_DATA4_T *)&d[i])); )==" "\n" |
| 480 | R"==(} else { )==" "\n" |
| 481 | R"==(unroll_for(int idx = 0; idx < 4; idx++) { )==" "\n" |
| 482 | R"==(if (local_id + 8 * idx < oc_tail) { )==" "\n" |
| 483 | R"==(dst[(i / 4) * G * OC + idx * 8 + local_id] = d[i + idx]; )==" "\n" |
| 484 | R"==(} )==" "\n" |
| 485 | R"==(} )==" "\n" |
| 486 | R"==(} )==" "\n" |
| 487 | R"==(} )==" "\n" |
| 488 | R"==(} )==" "\n" |
| 489 | R"==()=="; |
| 490 | } |
| 491 | } |
| 492 | } |
| 493 | } |
Generated on 2022-Nov-30 from project oneDNN revision 84bc949a1
Powered by 
Code Browser 2.1
Generator usage only permitted with license.