| 1 | namespace dnnl { |
|---|---|
| 2 | namespace impl { |
| 3 | namespace gpu { |
| 4 | namespace ocl { |
| 5 | const char *xe_lp_conv_nhwc_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_types.h" )==" "\n" |
| 23 | R"==(#if IC % IC_BLOCK != 0 )==" "\n" |
| 24 | R"==(#define IC_NBLOCKS_TAIL ((IC % IC_BLOCK + 3) / 4) )==" "\n" |
| 25 | R"==(#else )==" "\n" |
| 26 | R"==(#define IC_NBLOCKS_TAIL 8 )==" "\n" |
| 27 | R"==(#endif )==" "\n" |
| 28 | R"==(#if OW_BLOCK == 4 )==" "\n" |
| 29 | R"==(#define BLOCK 4 )==" "\n" |
| 30 | R"==(#define ACC_DATA_BLOCK int4 )==" "\n" |
| 31 | R"==(#define SRC_DATA_BLOCK_T SRC_MMAD_DATA4_T )==" "\n" |
| 32 | R"==(#define READ_BLOCK intel_sub_group_block_read4 )==" "\n" |
| 33 | R"==(#define WRITE_LOCAL block_write4 )==" "\n" |
| 34 | R"==(DECLARE_MMAD_EMU(mmad_tail, idot4, IC_NBLOCKS_TAIL, 4, SRC_DATA_BLOCK_T, int8, )==" "\n" |
| 35 | R"==(ACC_DATA_BLOCK) )==" "\n" |
| 36 | R"==(#define MMAD_FULL mmad8x4 )==" "\n" |
| 37 | R"==(#define MMAD_TAIL mmad_tail )==" "\n" |
| 38 | R"==(#elif OW_BLOCK == 8 )==" "\n" |
| 39 | R"==(#define BLOCK 8 )==" "\n" |
| 40 | R"==(#define ACC_DATA_BLOCK int8 )==" "\n" |
| 41 | R"==(#define SRC_DATA_BLOCK_T SRC_MMAD_DATA8_T )==" "\n" |
| 42 | R"==(#define READ_BLOCK intel_sub_group_block_read8 )==" "\n" |
| 43 | R"==(#define WRITE_LOCAL block_write8 )==" "\n" |
| 44 | R"==(DECLARE_MMAD_EMU(mmad_tail, idot4, IC_NBLOCKS_TAIL, 8, SRC_DATA_BLOCK_T, int8, )==" "\n" |
| 45 | R"==(ACC_DATA_BLOCK) )==" "\n" |
| 46 | R"==(#define MMAD_FULL mmad8x8 )==" "\n" |
| 47 | R"==(#define MMAD_TAIL mmad_tail )==" "\n" |
| 48 | R"==(#else )==" "\n" |
| 49 | R"==(#error "Wrong OW_BLOCK" )==" "\n" |
| 50 | R"==(#endif )==" "\n" |
| 51 | R"==(#define BLOCK_READ_SRC(data, idx) \ )==" "\n" |
| 52 | R"==(data = intel_sub_group_block_read8((__global uint *)&src[idx]); )==" "\n" |
| 53 | R"==(#define BLOCK_READ_WHT_1x32(data, idx) \ )==" "\n" |
| 54 | R"==(data = as_int(intel_sub_group_block_read((__global uint *)&wei[idx])); )==" "\n" |
| 55 | R"==(#define BLOCK_READ_WHT_8x32(data, idx) \ )==" "\n" |
| 56 | R"==(data = as_int8(intel_sub_group_block_read8((__global uint *)&wei[idx])); )==" "\n" |
| 57 | R"==(#define BLOCK_READ_BIA(data, idx) \ )==" "\n" |
| 58 | R"==(data = as_float4(intel_sub_group_block_read4((__global uint *)&bias[idx])); )==" "\n" |
| 59 | R"==(#if SCALES_PER_OC )==" "\n" |
| 60 | R"==(#define SCALE scales )==" "\n" |
| 61 | R"==(#elif SCALES_COMMON )==" "\n" |
| 62 | R"==(#define SCALE runtime_scales[0] )==" "\n" |
| 63 | R"==(#else )==" "\n" |
| 64 | R"==(#define SCALE 1 )==" "\n" |
| 65 | R"==(#endif )==" "\n" |
| 66 | R"==(#define BLOCK_READ_BOUND 1 )==" "\n" |
| 67 | R"==(#define BLOCK_WRITE_BOUND 4 )==" "\n" |
| 68 | R"==(inline float4 read_bias_scale_block4(const __global float *dst, int off) { )==" "\n" |
| 69 | R"==(const int local_id = get_sub_group_local_id(); )==" "\n" |
| 70 | R"==(#if OC % OC_BLOCK != 0 )==" "\n" |
| 71 | R"==(int tail = OC - off; )==" "\n" |
| 72 | R"==(if (tail < OC_BLOCK) { )==" "\n" |
| 73 | R"==(return (float4)(local_id < tail - 8 * 0 ? dst[0 * 8 + local_id] : 0, )==" "\n" |
| 74 | R"==(local_id < tail - 8 * 1 ? dst[1 * 8 + local_id] : 0, )==" "\n" |
| 75 | R"==(local_id < tail - 8 * 2 ? dst[2 * 8 + local_id] : 0, )==" "\n" |
| 76 | R"==(local_id < tail - 8 * 3 ? dst[3 * 8 + local_id] : 0); )==" "\n" |
| 77 | R"==(} )==" "\n" |
| 78 | R"==(#endif )==" "\n" |
| 79 | R"==(#if OC % BLOCK_READ_BOUND != 0 )==" "\n" |
| 80 | R"==(return (float4)(dst[0 * 8 + local_id], dst[1 * 8 + local_id], )==" "\n" |
| 81 | R"==(dst[2 * 8 + local_id], dst[3 * 8 + local_id]); )==" "\n" |
| 82 | R"==(#else )==" "\n" |
| 83 | R"==(return as_float4(intel_sub_group_block_read4((__global uint *)dst)); )==" "\n" |
| 84 | R"==(#endif )==" "\n" |
| 85 | R"==(} )==" "\n" |
| 86 | R"==(inline DST_DATA4_T read_oc_block4(const __global DATA_T *dst, int off) { )==" "\n" |
| 87 | R"==(const int local_id = get_sub_group_local_id(); )==" "\n" |
| 88 | R"==(#if OC % OC_BLOCK != 0 )==" "\n" |
| 89 | R"==(int tail = OC - off; )==" "\n" |
| 90 | R"==(if (tail < OC_BLOCK) { )==" "\n" |
| 91 | R"==(return (DST_DATA4_T)( )==" "\n" |
| 92 | R"==(local_id < tail - 8 * 0 ? dst[0 * 8 + local_id] : 0, )==" "\n" |
| 93 | R"==(local_id < tail - 8 * 1 ? dst[1 * 8 + local_id] : 0, )==" "\n" |
| 94 | R"==(local_id < tail - 8 * 2 ? dst[2 * 8 + local_id] : 0, )==" "\n" |
| 95 | R"==(local_id < tail - 8 * 3 ? dst[3 * 8 + local_id] : 0); )==" "\n" |
| 96 | R"==(} )==" "\n" |
| 97 | R"==(#endif )==" "\n" |
| 98 | R"==(#if OC % BLOCK_READ_BOUND != 0 )==" "\n" |
| 99 | R"==(return (DST_DATA4_T)(dst[0 * 8 + local_id], dst[1 * 8 + local_id], )==" "\n" |
| 100 | R"==(dst[2 * 8 + local_id], dst[3 * 8 + local_id]); )==" "\n" |
| 101 | R"==(#else )==" "\n" |
| 102 | R"==(return BLOCK_READ_DST4(dst); )==" "\n" |
| 103 | R"==(#endif )==" "\n" |
| 104 | R"==(} )==" "\n" |
| 105 | R"==(inline void write_oc_block4(__global DATA_T *dst, int off, DATA4_T value) { )==" "\n" |
| 106 | R"==(const int local_id = get_sub_group_local_id(); )==" "\n" |
| 107 | R"==(#if OC % OC_BLOCK != 0 )==" "\n" |
| 108 | R"==(int tail = OC - off; )==" "\n" |
| 109 | R"==(if (tail < OC_BLOCK) { )==" "\n" |
| 110 | R"==(if (local_id < tail) dst[0 * 8 + local_id] = value.s0; )==" "\n" |
| 111 | R"==(if (local_id < tail - 8 * 1) dst[1 * 8 + local_id] = value.s1; )==" "\n" |
| 112 | R"==(if (local_id < tail - 8 * 2) dst[2 * 8 + local_id] = value.s2; )==" "\n" |
| 113 | R"==(if (local_id < tail - 8 * 3) dst[3 * 8 + local_id] = value.s3; )==" "\n" |
| 114 | R"==(return; )==" "\n" |
| 115 | R"==(} )==" "\n" |
| 116 | R"==(#endif )==" "\n" |
| 117 | R"==(#if OC % BLOCK_WRITE_BOUND != 0 )==" "\n" |
| 118 | R"==(dst[0 * 8 + local_id] = value.s0; )==" "\n" |
| 119 | R"==(dst[1 * 8 + local_id] = value.s1; )==" "\n" |
| 120 | R"==(dst[2 * 8 + local_id] = value.s2; )==" "\n" |
| 121 | R"==(dst[3 * 8 + local_id] = value.s3; )==" "\n" |
| 122 | R"==(return; )==" "\n" |
| 123 | R"==(#else )==" "\n" |
| 124 | R"==(BLOCK_WRITE_DST4(dst, value); )==" "\n" |
| 125 | R"==(return; )==" "\n" |
| 126 | R"==(#endif )==" "\n" |
| 127 | R"==(} )==" "\n" |
| 128 | R"==(inline void write_local_1(__local uint *S, __global SRC_DATA_T *src1) { )==" "\n" |
| 129 | R"==(const int local_id = get_sub_group_local_id(); )==" "\n" |
| 130 | R"==(#if IC % 4 != 0 )==" "\n" |
| 131 | R"==(__local SRC_DATA_T *S1 = S + local_id; )==" "\n" |
| 132 | R"==(__global SRC_DATA_T *src2 = (const __global uint *)(src1) + local_id; )==" "\n" |
| 133 | R"==(S1[0] = src2[0]; )==" "\n" |
| 134 | R"==(S1[1] = src2[1]; )==" "\n" |
| 135 | R"==(S1[2] = src2[2]; )==" "\n" |
| 136 | R"==(S1[3] = src2[3]; )==" "\n" |
| 137 | R"==(#else )==" "\n" |
| 138 | R"==(block_write(S, intel_sub_group_block_read((const __global uint *)(src1))); )==" "\n" |
| 139 | R"==(#endif )==" "\n" |
| 140 | R"==(return; )==" "\n" |
| 141 | R"==(} )==" "\n" |
| 142 | R"==(__attribute__((intel_reqd_sub_group_size(SUB_GROUP_SIZE))) )==" "\n" |
| 143 | R"==(__attribute__((reqd_work_group_size(LWS_0, LWS_1, LWS_2))) )==" "\n" |
| 144 | R"==(__kernel void )==" "\n" |
| 145 | R"==(conv_nhwc_fwd_x8s8x(const __global SRC_DATA_T *src, const __global char *wei, )==" "\n" |
| 146 | R"==(const __global float *bias, __global DATA_T *dst POST_OP_ARGS, )==" "\n" |
| 147 | R"==(const __global float *runtime_scales, )==" "\n" |
| 148 | R"==(const __global int *src_compensation, const __global int *src_zpoints, )==" "\n" |
| 149 | R"==(const __global int *dst_compensation) { )==" "\n" |
| 150 | R"==(const int group_oc = get_group_id(0) * OC_GROUP; )==" "\n" |
| 151 | R"==(const int group_mb = get_group_id(2) * MB_GROUP; )==" "\n" |
| 152 | R"==(const int group_sp = get_group_id(1) * SP_GROUP; )==" "\n" |
| 153 | R"==(const int sub_group_id = get_sub_group_id(); )==" "\n" |
| 154 | R"==(const int oc = (sub_group_id % OC_GROUP); )==" "\n" |
| 155 | R"==(const int sp = (sub_group_id / OC_GROUP); )==" "\n" |
| 156 | R"==(const int g = (group_oc + oc) / OC_NCHUNK; )==" "\n" |
| 157 | R"==(const int group_ic = IC_NCHUNK * g; )==" "\n" |
| 158 | R"==(const int god = group_sp / (OW_PADDED * OH); )==" "\n" |
| 159 | R"==(const int gohw = group_sp % (OW_PADDED * OH); )==" "\n" |
| 160 | R"==(const int goh = gohw / OW_PADDED; )==" "\n" |
| 161 | R"==(const int gow = OW_BLOCK * (gohw % OW_PADDED); )==" "\n" |
| 162 | R"==(const int gid = god * SD; )==" "\n" |
| 163 | R"==(const int gih = goh * SH; )==" "\n" |
| 164 | R"==(const int giw = gow * SW; )==" "\n" |
| 165 | R"==(const int local_ow = OW_BLOCK * sp; )==" "\n" |
| 166 | R"==(const int local_iw = local_ow * SW; )==" "\n" |
| 167 | R"==(const int od = god; )==" "\n" |
| 168 | R"==(const int ow = gow + local_ow; )==" "\n" |
| 169 | R"==(const int oh = goh; )==" "\n" |
| 170 | R"==(const int id = gid - PD; )==" "\n" |
| 171 | R"==(const int iw = giw + local_iw - PW; )==" "\n" |
| 172 | R"==(const int ih = gih - PH; )==" "\n" |
| 173 | R"==(const int local_id = get_sub_group_local_id(); )==" "\n" |
| 174 | R"==(__local uint S_slice[SRC_SLM_SIZE]; )==" "\n" |
| 175 | R"==(__local uint *S_part = S_slice + IC_BLOCK / 4 * (sp * SW * OW_BLOCK + PW); )==" "\n" |
| 176 | R"==(__local uint *S_work = S_slice + IC_BLOCK / 4 * (sp * SW * OW_BLOCK); )==" "\n" |
| 177 | R"==(const bool left_tail = iw < 0; )==" "\n" |
| 178 | R"==(const bool left_nozero_tail = sub_group_id == 0 && iw > -PW; )==" "\n" |
| 179 | R"==(const bool right_tail = (iw + PW + SLM_TAIL >= IW) && (iw + PW < IW); )==" "\n" |
| 180 | R"==(const bool right_nozero_tail )==" "\n" |
| 181 | R"==(= sp == (LWS_1 - 1) && (iw + PW + SLM_TAIL < IW); )==" "\n" |
| 182 | R"==(const bool empty = (iw + PW >= IW); )==" "\n" |
| 183 | R"==(dst += group_mb * MB_BLOCK * OD * OH * OW * G * OC; )==" "\n" |
| 184 | R"==(dst += (OW * OH * od + OW * oh + ow) * G * OC; )==" "\n" |
| 185 | R"==(dst += OC_BLOCK * (group_oc + oc); )==" "\n" |
| 186 | R"==(src += group_mb * MB_BLOCK * ID * IH * IW * G * IC; )==" "\n" |
| 187 | R"==(src += (IW * IH * id + IW * ih + iw + PW) * G * IC; )==" "\n" |
| 188 | R"==(src += group_ic * IC_BLOCK; )==" "\n" |
| 189 | R"==(wei += IC_BLOCK * KD * KH * KW * OC_BLOCK * (group_oc + oc) * IC_NCHUNK; )==" "\n" |
| 190 | R"==(/* Prepare S_slice tails */ )==" "\n" |
| 191 | R"==(#if PW > 0 )==" "\n" |
| 192 | R"==(if (left_tail) { )==" "\n" |
| 193 | R"==(for (int i = 0; i < PW; i++) { )==" "\n" |
| 194 | R"==(block_write(S_slice + i * 8, 0); )==" "\n" |
| 195 | R"==(} )==" "\n" |
| 196 | R"==(} )==" "\n" |
| 197 | R"==(#endif )==" "\n" |
| 198 | R"==(#if ZERO_TAIL > 0 )==" "\n" |
| 199 | R"==(if (right_tail) { )==" "\n" |
| 200 | R"==(for (int i = SLM_TAIL; i < SW * OW_BLOCK + (KW - 1) * (1 + DW) - PW; )==" "\n" |
| 201 | R"==(i++) { )==" "\n" |
| 202 | R"==(block_write(S_part + i * 8, 0); )==" "\n" |
| 203 | R"==(} )==" "\n" |
| 204 | R"==(} )==" "\n" |
| 205 | R"==(#if SLM_NCHUNK < OW_NCHUNK )==" "\n" |
| 206 | R"==(if (empty) { )==" "\n" |
| 207 | R"==(for (int i = -PW; i < SW * OW_BLOCK + (KW - 1) * (1 + DW) - PW; i++) { )==" "\n" |
| 208 | R"==(block_write(S_part + i * 8, 0); )==" "\n" |
| 209 | R"==(} )==" "\n" |
| 210 | R"==(} )==" "\n" |
| 211 | R"==(#endif )==" "\n" |
| 212 | R"==(#endif )==" "\n" |
| 213 | R"==(ACC_DATA_BLOCK C00 = 0, C01 = 0, C02 = 0, C03 = 0; )==" "\n" |
| 214 | R"==(__attribute__((opencl_unroll_hint(1))) )==" "\n" |
| 215 | R"==(for (int ic_chunk = 0; ic_chunk < IC_NCHUNK; ic_chunk++) { )==" "\n" |
| 216 | R"==(SRC_DATA_BLOCK_T S0; )==" "\n" |
| 217 | R"==(__attribute__((opencl_unroll_hint(1))) )==" "\n" |
| 218 | R"==(for (int kd = 0; kd < KD; kd++) { )==" "\n" |
| 219 | R"==(if (kd * (1 + DD) + id < 0 || kd * (1 + DD) + id >= ID) { )==" "\n" |
| 220 | R"==(wei += IC_BLOCK * OC_BLOCK * KH * KW; )==" "\n" |
| 221 | R"==(continue; )==" "\n" |
| 222 | R"==(} )==" "\n" |
| 223 | R"==(__attribute__((opencl_unroll_hint(1))) )==" "\n" |
| 224 | R"==(for (int kh = 0; kh < KH; kh++) { )==" "\n" |
| 225 | R"==(if (kh * (1 + DH) + ih < 0 || kh * (1 + DH) + ih >= IH) { )==" "\n" |
| 226 | R"==(wei += IC_BLOCK * OC_BLOCK * KW; )==" "\n" |
| 227 | R"==(continue; )==" "\n" |
| 228 | R"==(} )==" "\n" |
| 229 | R"==(barrier(CLK_LOCAL_MEM_FENCE); )==" "\n" |
| 230 | R"==(const __global SRC_DATA_T *src1 = src )==" "\n" |
| 231 | R"==(+ kd * (1 + DD) * IH * IW * G * IC )==" "\n" |
| 232 | R"==(+ kh * (1 + DH) * IW * G * IC; )==" "\n" |
| 233 | R"==(#if SLM_NCHUNK < OW_NCHUNK )==" "\n" |
| 234 | R"==(if (iw + PW < IW) { )==" "\n" |
| 235 | R"==(#endif )==" "\n" |
| 236 | R"==(#if OW_NCHUNK > LWS_1 )==" "\n" |
| 237 | R"==(/* Copy tails in case of multigroups */ )==" "\n" |
| 238 | R"==(if (ow < OW) { )==" "\n" |
| 239 | R"==(#if PW > 0 )==" "\n" |
| 240 | R"==(if (left_nozero_tail) { )==" "\n" |
| 241 | R"==(for (int i = -PW - min(iw, 0); i < 0; i++) { )==" "\n" |
| 242 | R"==(write_local_1( )==" "\n" |
| 243 | R"==(S_part + i * 8, src1 + i * G * IC); )==" "\n" |
| 244 | R"==(} )==" "\n" |
| 245 | R"==(} )==" "\n" |
| 246 | R"==(#endif )==" "\n" |
| 247 | R"==(if (right_nozero_tail) { )==" "\n" |
| 248 | R"==(int buffer_last = (KW - 1) * (1 + DW) - PW; )==" "\n" |
| 249 | R"==(int src_last = IW - iw - SW * OW_BLOCK - PW; )==" "\n" |
| 250 | R"==(for (int i = SW * OW_BLOCK; i < SW * OW_BLOCK )==" "\n" |
| 251 | R"==(+ min(buffer_last, src_last); )==" "\n" |
| 252 | R"==(i++) { )==" "\n" |
| 253 | R"==(write_local_1( )==" "\n" |
| 254 | R"==(S_part + i * 8, src1 + i * G * IC); )==" "\n" |
| 255 | R"==(} )==" "\n" |
| 256 | R"==(for (int i = SW * OW_BLOCK )==" "\n" |
| 257 | R"==(+ min(buffer_last, src_last); )==" "\n" |
| 258 | R"==(i < SW * OW_BLOCK + buffer_last; i++) { )==" "\n" |
| 259 | R"==(block_write(S_part + i * 8, 0); )==" "\n" |
| 260 | R"==(} )==" "\n" |
| 261 | R"==(} )==" "\n" |
| 262 | R"==(#endif )==" "\n" |
| 263 | R"==(#if SLM_TAIL != OW_BLOCK * SW )==" "\n" |
| 264 | R"==(/* Copy last block to SLM */ )==" "\n" |
| 265 | R"==(if (right_tail) { )==" "\n" |
| 266 | R"==(__attribute__(( )==" "\n" |
| 267 | R"==(opencl_unroll_hint)) )==" "\n" |
| 268 | R"==(for (int i = 0; i < SLM_TAIL; i++) { )==" "\n" |
| 269 | R"==(write_local_1( )==" "\n" |
| 270 | R"==(S_part + i * 8, src1 + i * G * IC); )==" "\n" |
| 271 | R"==(} )==" "\n" |
| 272 | R"==(} else { )==" "\n" |
| 273 | R"==(#endif )==" "\n" |
| 274 | R"==(__attribute__(( )==" "\n" |
| 275 | R"==(opencl_unroll_hint)) )==" "\n" |
| 276 | R"==(for (int i = 0; i < SW * OW_BLOCK; i++) { )==" "\n" |
| 277 | R"==(write_local_1( )==" "\n" |
| 278 | R"==(S_part + i * 8, src1 + i * G * IC); )==" "\n" |
| 279 | R"==(} )==" "\n" |
| 280 | R"==(#if SLM_TAIL != OW_BLOCK * SW )==" "\n" |
| 281 | R"==(} )==" "\n" |
| 282 | R"==(#endif )==" "\n" |
| 283 | R"==(#if OW_NCHUNK > LWS_1 )==" "\n" |
| 284 | R"==(} )==" "\n" |
| 285 | R"==(#endif )==" "\n" |
| 286 | R"==(#if SLM_NCHUNK < OW_NCHUNK )==" "\n" |
| 287 | R"==(} )==" "\n" |
| 288 | R"==(#endif )==" "\n" |
| 289 | R"==(barrier(CLK_LOCAL_MEM_FENCE); )==" "\n" |
| 290 | R"==(__attribute__((opencl_unroll_hint)) )==" "\n" |
| 291 | R"==(for (int kw = 0; kw < KW; kw++) { )==" "\n" |
| 292 | R"==(__attribute__(( )==" "\n" |
| 293 | R"==(opencl_unroll_hint(OW_BLOCK))) )==" "\n" |
| 294 | R"==(for (int i = 0; i < OW_BLOCK; i++) { )==" "\n" |
| 295 | R"==(S0[i] = block_read( )==" "\n" |
| 296 | R"==(S_work + (kw * (1 + DW) + SW * i) * 8); )==" "\n" |
| 297 | R"==(} )==" "\n" |
| 298 | R"==(int8 W0 = 0, W1 = 0, W2 = 0, W3 = 0; )==" "\n" |
| 299 | R"==(#if IC % IC_BLOCK != 0 )==" "\n" |
| 300 | R"==(if (ic_chunk == IC_NCHUNK - 1) { )==" "\n" |
| 301 | R"==(unroll_for(int i = 0; i < IC_NBLOCKS_TAIL; ++i) )==" "\n" |
| 302 | R"==(BLOCK_READ_WHT_1x32(W0[i], (i + 0) * IC_BLOCK); )==" "\n" |
| 303 | R"==(if (OC > 8) )==" "\n" |
| 304 | R"==(unroll_for(int i = 0; i < IC_NBLOCKS_TAIL; ++i) )==" "\n" |
| 305 | R"==(BLOCK_READ_WHT_1x32( )==" "\n" |
| 306 | R"==(W1[i], (i + 8) * IC_BLOCK); )==" "\n" |
| 307 | R"==(if (OC > 16) )==" "\n" |
| 308 | R"==(unroll_for(int i = 0; i < IC_NBLOCKS_TAIL; ++i) )==" "\n" |
| 309 | R"==(BLOCK_READ_WHT_1x32( )==" "\n" |
| 310 | R"==(W2[i], (i + 16) * IC_BLOCK); )==" "\n" |
| 311 | R"==(if (OC > 24) )==" "\n" |
| 312 | R"==(unroll_for(int i = 0; i < IC_NBLOCKS_TAIL; ++i) )==" "\n" |
| 313 | R"==(BLOCK_READ_WHT_1x32( )==" "\n" |
| 314 | R"==(W3[i], (i + 24) * IC_BLOCK); )==" "\n" |
| 315 | R"==(C00 = MMAD_TAIL(S0, W0, C00); )==" "\n" |
| 316 | R"==(if (OC > 8) C01 = MMAD_TAIL(S0, W1, C01); )==" "\n" |
| 317 | R"==(if (OC > 16) C02 = MMAD_TAIL(S0, W2, C02); )==" "\n" |
| 318 | R"==(if (OC > 24) C03 = MMAD_TAIL(S0, W3, C03); )==" "\n" |
| 319 | R"==(} else )==" "\n" |
| 320 | R"==(#endif )==" "\n" |
| 321 | R"==({ )==" "\n" |
| 322 | R"==(BLOCK_READ_WHT_8x32(W0, 0); )==" "\n" |
| 323 | R"==(if (OC > 8) BLOCK_READ_WHT_8x32(W1, 8 * IC_BLOCK); )==" "\n" |
| 324 | R"==(if (OC > 16) BLOCK_READ_WHT_8x32(W2, 16 * IC_BLOCK); )==" "\n" |
| 325 | R"==(if (OC > 24) BLOCK_READ_WHT_8x32(W3, 24 * IC_BLOCK); )==" "\n" |
| 326 | R"==(C00 = MMAD_FULL(S0, W0, C00); )==" "\n" |
| 327 | R"==(if (OC > 8) C01 = MMAD_FULL(S0, W1, C01); )==" "\n" |
| 328 | R"==(if (OC > 16) C02 = MMAD_FULL(S0, W2, C02); )==" "\n" |
| 329 | R"==(if (OC > 24) C03 = MMAD_FULL(S0, W3, C03); )==" "\n" |
| 330 | R"==(} )==" "\n" |
| 331 | R"==(wei += IC_BLOCK * OC_BLOCK; )==" "\n" |
| 332 | R"==(} )==" "\n" |
| 333 | R"==(} )==" "\n" |
| 334 | R"==(} )==" "\n" |
| 335 | R"==(src += IC_BLOCK; )==" "\n" |
| 336 | R"==(} )==" "\n" |
| 337 | R"==(if (ow < OW) { )==" "\n" |
| 338 | R"==(float4 tmp = {0, 0, 0, 0}; )==" "\n" |
| 339 | R"==(DST_DATA4_T dst_pack[BLOCK] = {0}; )==" "\n" |
| 340 | R"==(DST_DATA4_T D0[BLOCK] = {0}; )==" "\n" |
| 341 | R"==(#if SCALES_PER_OC )==" "\n" |
| 342 | R"==(float4 scales = {0, 0, 0, 0}; )==" "\n" |
| 343 | R"==(scales = read_bias_scale_block4( )==" "\n" |
| 344 | R"==(runtime_scales + (group_oc + oc) * OC_BLOCK, )==" "\n" |
| 345 | R"==((group_oc + oc) * OC_BLOCK); )==" "\n" |
| 346 | R"==(#endif )==" "\n" |
| 347 | R"==(#if WITH_BIAS )==" "\n" |
| 348 | R"==(float4 bia = {0, 0, 0, 0}; )==" "\n" |
| 349 | R"==(bia = read_bias_scale_block4( )==" "\n" |
| 350 | R"==(bias + (group_oc + oc) * OC_BLOCK, (group_oc + oc) * OC_BLOCK); )==" "\n" |
| 351 | R"==(#define QUANTIZE_ADD_BIAS() tmp = SCALE * fma(tmp, (float4)1, bia); )==" "\n" |
| 352 | R"==(#else )==" "\n" |
| 353 | R"==(#define QUANTIZE_ADD_BIAS() tmp *= SCALE; )==" "\n" |
| 354 | R"==(#endif )==" "\n" |
| 355 | R"==(#if WITH_SUM )==" "\n" |
| 356 | R"==(#if OW_BLOCK >= 4 )==" "\n" |
| 357 | R"==(*(DST_DATA16_T *)D0 = (DST_DATA16_T)( )==" "\n" |
| 358 | R"==(read_oc_block4(dst + G * OC * 0, (group_oc + oc) * OC_BLOCK), )==" "\n" |
| 359 | R"==(read_oc_block4(dst + G * OC * 1, (group_oc + oc) * OC_BLOCK), )==" "\n" |
| 360 | R"==(read_oc_block4(dst + G * OC * 2, (group_oc + oc) * OC_BLOCK), )==" "\n" |
| 361 | R"==(read_oc_block4(dst + G * OC * 3, (group_oc + oc) * OC_BLOCK)); )==" "\n" |
| 362 | R"==(#endif )==" "\n" |
| 363 | R"==(#if OW_BLOCK == 8 )==" "\n" |
| 364 | R"==(*(DST_DATA16_T *)(D0 + 4) = (DST_DATA16_T)( )==" "\n" |
| 365 | R"==(read_oc_block4(dst + G * OC * 4, (group_oc + oc) * OC_BLOCK), )==" "\n" |
| 366 | R"==(read_oc_block4(dst + G * OC * 5, (group_oc + oc) * OC_BLOCK), )==" "\n" |
| 367 | R"==(read_oc_block4(dst + G * OC * 6, (group_oc + oc) * OC_BLOCK), )==" "\n" |
| 368 | R"==(read_oc_block4(dst + G * OC * 7, (group_oc + oc) * OC_BLOCK)); )==" "\n" |
| 369 | R"==(#endif )==" "\n" |
| 370 | R"==(#endif )==" "\n" |
| 371 | R"==(#define PACK(C0, C1, C2, C3, idx) \ )==" "\n" |
| 372 | R"==(do { \ )==" "\n" |
| 373 | R"==(tmp[0] = C0[idx]; \ )==" "\n" |
| 374 | R"==(tmp[1] = C1[idx]; \ )==" "\n" |
| 375 | R"==(tmp[2] = C2[idx]; \ )==" "\n" |
| 376 | R"==(tmp[3] = C3[idx]; \ )==" "\n" |
| 377 | R"==(} while (0) )==" "\n" |
| 378 | R"==(#define CONVERT_PACK(idx) \ )==" "\n" |
| 379 | R"==(do { \ )==" "\n" |
| 380 | R"==(dst_pack[idx] = CONVERT_DST_DATA4_T(tmp); \ )==" "\n" |
| 381 | R"==(} while (0) )==" "\n" |
| 382 | R"==(#define PACK_DST(C0, C1, C2, C3, D) \ )==" "\n" |
| 383 | R"==(do { \ )==" "\n" |
| 384 | R"==(for (int n_i = 0; n_i < OW_BLOCK; n_i++) { \ )==" "\n" |
| 385 | R"==(PACK(C0, C1, C2, C3, n_i); \ )==" "\n" |
| 386 | R"==(QUANTIZE_ADD_BIAS(); \ )==" "\n" |
| 387 | R"==(for (int didx = 0; didx < 4; ++didx) { \ )==" "\n" |
| 388 | R"==(float tmp_i = tmp[didx]; \ )==" "\n" |
| 389 | R"==(float dni_i = convert_float(AS_SUM_DATA_T(D[n_i][didx])); \ )==" "\n" |
| 390 | R"==(int po_mb; \ )==" "\n" |
| 391 | R"==(po_mb = group_mb % MB; \ )==" "\n" |
| 392 | R"==(const int po_oc = ((group_oc + oc) * OC_BLOCK + local_id \ )==" "\n" |
| 393 | R"==(+ didx * SUB_GROUP_SIZE) \ )==" "\n" |
| 394 | R"==(% (OC * G); \ )==" "\n" |
| 395 | R"==(APPLY_POST_OPS_SERIAL_BINARY_2D( \ )==" "\n" |
| 396 | R"==(tmp_i, float, dni_i, float, po_mb, 1, po_oc, 1); \ )==" "\n" |
| 397 | R"==(tmp[didx] = tmp_i; \ )==" "\n" |
| 398 | R"==(} \ )==" "\n" |
| 399 | R"==(CONVERT_PACK(n_i); \ )==" "\n" |
| 400 | R"==(} \ )==" "\n" |
| 401 | R"==(} while (0) )==" "\n" |
| 402 | R"==(PACK_DST(C00, C01, C02, C03, D0); )==" "\n" |
| 403 | R"==(#if OW_TAIL )==" "\n" |
| 404 | R"==(if (ow + OW_BLOCK > OW) { )==" "\n" |
| 405 | R"==(__attribute__((opencl_unroll_hint(OW_TAIL))) )==" "\n" |
| 406 | R"==(for (int i = 0; i < OW_TAIL; i++) { )==" "\n" |
| 407 | R"==(write_oc_block4(dst + i * G * OC, (group_oc + oc) * OC_BLOCK, )==" "\n" |
| 408 | R"==(dst_pack[i]); )==" "\n" |
| 409 | R"==(} )==" "\n" |
| 410 | R"==(} else { )==" "\n" |
| 411 | R"==(#endif )==" "\n" |
| 412 | R"==(#if OW_BLOCK >= 4 )==" "\n" |
| 413 | R"==(write_oc_block4( )==" "\n" |
| 414 | R"==(dst + G * OC * 0, (group_oc + oc) * OC_BLOCK, dst_pack[0]); )==" "\n" |
| 415 | R"==(write_oc_block4( )==" "\n" |
| 416 | R"==(dst + G * OC * 1, (group_oc + oc) * OC_BLOCK, dst_pack[1]); )==" "\n" |
| 417 | R"==(write_oc_block4( )==" "\n" |
| 418 | R"==(dst + G * OC * 2, (group_oc + oc) * OC_BLOCK, dst_pack[2]); )==" "\n" |
| 419 | R"==(write_oc_block4( )==" "\n" |
| 420 | R"==(dst + G * OC * 3, (group_oc + oc) * OC_BLOCK, dst_pack[3]); )==" "\n" |
| 421 | R"==(#endif )==" "\n" |
| 422 | R"==(#if OW_BLOCK == 8 )==" "\n" |
| 423 | R"==(write_oc_block4( )==" "\n" |
| 424 | R"==(dst + G * OC * 4, (group_oc + oc) * OC_BLOCK, dst_pack[4]); )==" "\n" |
| 425 | R"==(write_oc_block4( )==" "\n" |
| 426 | R"==(dst + G * OC * 5, (group_oc + oc) * OC_BLOCK, dst_pack[5]); )==" "\n" |
| 427 | R"==(write_oc_block4( )==" "\n" |
| 428 | R"==(dst + G * OC * 6, (group_oc + oc) * OC_BLOCK, dst_pack[6]); )==" "\n" |
| 429 | R"==(write_oc_block4( )==" "\n" |
| 430 | R"==(dst + G * OC * 7, (group_oc + oc) * OC_BLOCK, dst_pack[7]); )==" "\n" |
| 431 | R"==(#endif )==" "\n" |
| 432 | R"==(#if OW_TAIL )==" "\n" |
| 433 | R"==(} )==" "\n" |
| 434 | R"==(#endif )==" "\n" |
| 435 | R"==(} )==" "\n" |
| 436 | R"==(} )==" "\n" |
| 437 | R"==()=="; |
| 438 | } |
| 439 | } |
| 440 | } |
| 441 | } |
Generated on 2022-Nov-30 from project oneDNN revision 84bc949a1
Powered by 
Code Browser 2.1
Generator usage only permitted with license.