| 1 | namespace dnnl { |
|---|---|
| 2 | namespace impl { |
| 3 | namespace gpu { |
| 4 | namespace ocl { |
| 5 | const char *reorder_common_header = 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"==(#undef cl_future_bf16_cvt )==" "\n" |
| 21 | R"==(#define DT_UNDEF 1 )==" "\n" |
| 22 | R"==(#include "gpu/ocl/ocl_types.h" )==" "\n" |
| 23 | R"==(#include "gpu/ocl/ocl_math_utils.h" )==" "\n" |
| 24 | R"==(#if SRC_DT_F16 || DST_DT_F16 )==" "\n" |
| 25 | R"==(#pragma OPENCL EXTENSION cl_khr_fp16 : enable )==" "\n" |
| 26 | R"==(#endif )==" "\n" |
| 27 | R"==(#if SRC_DT_F64 || DST_DT_F64 )==" "\n" |
| 28 | R"==(#pragma OPENCL EXTENSION cl_khr_fp64 : enable )==" "\n" |
| 29 | R"==(#endif )==" "\n" |
| 30 | R"==(#undef SRC_OFF )==" "\n" |
| 31 | R"==(#undef DST_OFF )==" "\n" |
| 32 | R"==(#define SRC_OFF(x0, x1, x2, x3, x4, x5) \ )==" "\n" |
| 33 | R"==(OFF_MD(SRC, (x0), (x1), (x2), (x3), (x4), (x5)) )==" "\n" |
| 34 | R"==(#define DST_OFF(x0, x1, x2, x3, x4, x5) \ )==" "\n" |
| 35 | R"==(OFF_MD(DST, (x0), (x1), (x2), (x3), (x4), (x5)) )==" "\n" |
| 36 | R"==(#define SRC_OFF_G(gr, x0, x1, x2, x3, x4) \ )==" "\n" |
| 37 | R"==(OFF_MD(SRC, gr, (x0), (x1), (x2), (x3), (x4)) )==" "\n" |
| 38 | R"==(#define DST_OFF_G(gr, x0, x1, x2, x3, x4) \ )==" "\n" |
| 39 | R"==(OFF_MD(DST, gr, (x0), (x1), (x2), (x3), (x4)) )==" "\n" |
| 40 | R"==(#if SRC_DT_S8 )==" "\n" |
| 41 | R"==(#define SRC_BLOCK_READ(src) \ )==" "\n" |
| 42 | R"==(as_char(intel_sub_group_block_read_uc((const __global uchar *)(src))) )==" "\n" |
| 43 | R"==(#define SRC_BLOCK_READ8(src) \ )==" "\n" |
| 44 | R"==(as_char8(intel_sub_group_block_read_uc8((const __global uchar *)(src))) )==" "\n" |
| 45 | R"==(#define SRC_BLOCK_WRITE(dst, val) \ )==" "\n" |
| 46 | R"==(intel_sub_group_block_write_uc((__global uchar *)(dst), as_uchar(val)) )==" "\n" |
| 47 | R"==(#define SRC_BLOCK_WRITE8(dst, val) \ )==" "\n" |
| 48 | R"==(intel_sub_group_block_write_uc8((__global uchar *)(dst), as_uchar8(val)) )==" "\n" |
| 49 | R"==(#endif )==" "\n" |
| 50 | R"==(#if SRC_DT_U8 )==" "\n" |
| 51 | R"==(#define SRC_BLOCK_READ(src) \ )==" "\n" |
| 52 | R"==(as_uchar(intel_sub_group_block_read_uc((const __global uchar *)(src))) )==" "\n" |
| 53 | R"==(#define SRC_BLOCK_READ8(src) \ )==" "\n" |
| 54 | R"==(as_uchar8(intel_sub_group_block_read_uc8((const __global uchar *)(src))) )==" "\n" |
| 55 | R"==(#define SRC_BLOCK_WRITE(dst, val) \ )==" "\n" |
| 56 | R"==(intel_sub_group_block_write_uc((__global uchar *)(dst), as_uchar(val)) )==" "\n" |
| 57 | R"==(#define SRC_BLOCK_WRITE8(dst, val) \ )==" "\n" |
| 58 | R"==(intel_sub_group_block_write_uc8((__global uchar *)(dst), as_uchar8(val)) )==" "\n" |
| 59 | R"==(#endif )==" "\n" |
| 60 | R"==(#if SRC_DT_F16 )==" "\n" |
| 61 | R"==(#define SRC_BLOCK_READ(src) \ )==" "\n" |
| 62 | R"==(as_half(intel_sub_group_block_read_us((const __global ushort *)(src))) )==" "\n" |
| 63 | R"==(#define SRC_BLOCK_READ8(src) \ )==" "\n" |
| 64 | R"==(as_half8(intel_sub_group_block_read_us8((const __global ushort *)(src))) )==" "\n" |
| 65 | R"==(#define SRC_BLOCK_WRITE(dst, val) \ )==" "\n" |
| 66 | R"==(intel_sub_group_block_write_us((__global ushort *)(dst), as_ushort(val)) )==" "\n" |
| 67 | R"==(#define SRC_BLOCK_WRITE8(dst, val) \ )==" "\n" |
| 68 | R"==(intel_sub_group_block_write_us8((__global ushort *)(dst), as_ushort8(val)) )==" "\n" |
| 69 | R"==(#endif )==" "\n" |
| 70 | R"==(#if SRC_DT_S32 )==" "\n" |
| 71 | R"==(#define SRC_BLOCK_READ(src) \ )==" "\n" |
| 72 | R"==(as_int(intel_sub_group_block_read((const __global uint *)(src))) )==" "\n" |
| 73 | R"==(#define SRC_BLOCK_READ8(src) \ )==" "\n" |
| 74 | R"==(as_int8(intel_sub_group_block_read8((const __global uint *)(src))) )==" "\n" |
| 75 | R"==(#define SRC_BLOCK_WRITE(dst, val) \ )==" "\n" |
| 76 | R"==(intel_sub_group_block_write((__global uint *)(dst), as_uint(val)) )==" "\n" |
| 77 | R"==(#define SRC_BLOCK_WRITE8(dst, val) \ )==" "\n" |
| 78 | R"==(intel_sub_group_block_write8((__global uint *)(dst), as_uint8(val)) )==" "\n" |
| 79 | R"==(#endif )==" "\n" |
| 80 | R"==(#if SRC_DT_F32 )==" "\n" |
| 81 | R"==(#define SRC_BLOCK_READ(src) \ )==" "\n" |
| 82 | R"==(as_float(intel_sub_group_block_read((const __global uint *)(src))) )==" "\n" |
| 83 | R"==(#define SRC_BLOCK_READ8(src) \ )==" "\n" |
| 84 | R"==(as_float8(intel_sub_group_block_read8((const __global uint *)(src))) )==" "\n" |
| 85 | R"==(#define SRC_BLOCK_WRITE(dst, val) \ )==" "\n" |
| 86 | R"==(intel_sub_group_block_write((__global uint *)(dst), as_uint(val)) )==" "\n" |
| 87 | R"==(#define SRC_BLOCK_WRITE8(dst, val) \ )==" "\n" |
| 88 | R"==(intel_sub_group_block_write8((__global uint *)(dst), as_uint8(val)) )==" "\n" |
| 89 | R"==(#endif )==" "\n" |
| 90 | R"==(#if SRC_DT_F64 )==" "\n" |
| 91 | R"==(#define SRC_BLOCK_READ(src) \ )==" "\n" |
| 92 | R"==(as_double(intel_sub_group_block_read2((const __global uint *)(src))) )==" "\n" |
| 93 | R"==(#define SRC_BLOCK_READ8(src) \ )==" "\n" |
| 94 | R"==((double8)((as_double4(intel_sub_group_block_read8( \ )==" "\n" |
| 95 | R"==((const __global uint *)(src)))), \ )==" "\n" |
| 96 | R"==((as_double4(intel_sub_group_block_read8( \ )==" "\n" |
| 97 | R"==((const __global uint *)(src + 8))))) )==" "\n" |
| 98 | R"==(#define SRC_BLOCK_WRITE(dst, val) \ )==" "\n" |
| 99 | R"==(intel_sub_group_block_write2((__global uint *)(dst), as_uint2(val)) )==" "\n" |
| 100 | R"==(#define SRC_BLOCK_WRITE8(dst, val) \ )==" "\n" |
| 101 | R"==(do { \ )==" "\n" |
| 102 | R"==(intel_sub_group_block_write8( \ )==" "\n" |
| 103 | R"==((__global uint *)(dst), as_uint8(val.lo)); \ )==" "\n" |
| 104 | R"==(intel_sub_group_block_write8( \ )==" "\n" |
| 105 | R"==((__global uint *)(dst + 8), as_uint8(val.hi)); \ )==" "\n" |
| 106 | R"==(} while (0) )==" "\n" |
| 107 | R"==(#endif )==" "\n" |
| 108 | R"==(#if SRC_DT_BF16 )==" "\n" |
| 109 | R"==(#define SRC_BLOCK_READ(src) \ )==" "\n" |
| 110 | R"==(as_ushort(intel_sub_group_block_read_us((const __global ushort *)(src))) )==" "\n" |
| 111 | R"==(#define SRC_BLOCK_READ8(src) \ )==" "\n" |
| 112 | R"==(as_ushort8(intel_sub_group_block_read_us8((const __global ushort *)(src))) )==" "\n" |
| 113 | R"==(#define SRC_BLOCK_WRITE(dst, val) \ )==" "\n" |
| 114 | R"==(intel_sub_group_block_write_us((__global ushort *)(dst), as_ushort(val)) )==" "\n" |
| 115 | R"==(#define SRC_BLOCK_WRITE8(dst, val) \ )==" "\n" |
| 116 | R"==(intel_sub_group_block_write_us8((__global ushort *)(dst), as_ushort8(val)) )==" "\n" |
| 117 | R"==(#endif )==" "\n" |
| 118 | R"==(#if DST_DT_S8 )==" "\n" |
| 119 | R"==(#define DST_BLOCK_READ(src) \ )==" "\n" |
| 120 | R"==(as_char(intel_sub_group_block_read_uc((const __global uchar *)(src))) )==" "\n" |
| 121 | R"==(#define DST_BLOCK_READ8(src) \ )==" "\n" |
| 122 | R"==(as_char8(intel_sub_group_block_read_uc8((const __global uchar *)(src))) )==" "\n" |
| 123 | R"==(#define DST_BLOCK_WRITE(dst, val) \ )==" "\n" |
| 124 | R"==(intel_sub_group_block_write_uc((__global uchar *)(dst), as_uchar(val)) )==" "\n" |
| 125 | R"==(#define DST_BLOCK_WRITE8(dst, val) \ )==" "\n" |
| 126 | R"==(intel_sub_group_block_write_uc8((__global uchar *)(dst), as_uchar8(val)) )==" "\n" |
| 127 | R"==(#endif )==" "\n" |
| 128 | R"==(#if DST_DT_U8 )==" "\n" |
| 129 | R"==(#define DST_BLOCK_READ(src) \ )==" "\n" |
| 130 | R"==(as_uchar(intel_sub_group_block_read_uc((const __global uchar *)(src))) )==" "\n" |
| 131 | R"==(#define DST_BLOCK_READ8(src) \ )==" "\n" |
| 132 | R"==(as_uchar8(intel_sub_group_block_read_uc8((const __global uchar *)(src))) )==" "\n" |
| 133 | R"==(#define DST_BLOCK_WRITE(dst, val) \ )==" "\n" |
| 134 | R"==(intel_sub_group_block_write_uc((__global uchar *)(dst), as_uchar(val)) )==" "\n" |
| 135 | R"==(#define DST_BLOCK_WRITE8(dst, val) \ )==" "\n" |
| 136 | R"==(intel_sub_group_block_write_uc8((__global uchar *)(dst), as_uchar8(val)) )==" "\n" |
| 137 | R"==(#endif )==" "\n" |
| 138 | R"==(#if DST_DT_F16 )==" "\n" |
| 139 | R"==(#define DST_BLOCK_READ(src) \ )==" "\n" |
| 140 | R"==(as_half(intel_sub_group_block_read_us((const __global ushort *)(src))) )==" "\n" |
| 141 | R"==(#define DST_BLOCK_READ8(src) \ )==" "\n" |
| 142 | R"==(as_half8(intel_sub_group_block_read_us8((const __global ushort *)(src))) )==" "\n" |
| 143 | R"==(#define DST_BLOCK_WRITE(dst, val) \ )==" "\n" |
| 144 | R"==(intel_sub_group_block_write_us((__global ushort *)(dst), as_ushort(val)) )==" "\n" |
| 145 | R"==(#define DST_BLOCK_WRITE8(dst, val) \ )==" "\n" |
| 146 | R"==(intel_sub_group_block_write_us8((__global ushort *)(dst), as_ushort8(val)) )==" "\n" |
| 147 | R"==(#endif )==" "\n" |
| 148 | R"==(#if DST_DT_S32 )==" "\n" |
| 149 | R"==(#define DST_BLOCK_READ(src) \ )==" "\n" |
| 150 | R"==(as_int(intel_sub_group_block_read((const __global uint *)(src))) )==" "\n" |
| 151 | R"==(#define DST_BLOCK_READ8(src) \ )==" "\n" |
| 152 | R"==(as_int8(intel_sub_group_block_read8((const __global uint *)(src))) )==" "\n" |
| 153 | R"==(#define DST_BLOCK_WRITE(dst, val) \ )==" "\n" |
| 154 | R"==(intel_sub_group_block_write((__global uint *)(dst), as_uint(val)) )==" "\n" |
| 155 | R"==(#define DST_BLOCK_WRITE8(dst, val) \ )==" "\n" |
| 156 | R"==(intel_sub_group_block_write8((__global uint *)(dst), as_uint8(val)) )==" "\n" |
| 157 | R"==(#endif )==" "\n" |
| 158 | R"==(#if DST_DT_F32 )==" "\n" |
| 159 | R"==(#define DST_BLOCK_READ(src) \ )==" "\n" |
| 160 | R"==(as_float(intel_sub_group_block_read((const __global uint *)(src))) )==" "\n" |
| 161 | R"==(#define DST_BLOCK_READ8(src) \ )==" "\n" |
| 162 | R"==(as_float8(intel_sub_group_block_read8((const __global uint *)(src))) )==" "\n" |
| 163 | R"==(#define DST_BLOCK_WRITE(dst, val) \ )==" "\n" |
| 164 | R"==(intel_sub_group_block_write((__global uint *)(dst), as_uint(val)) )==" "\n" |
| 165 | R"==(#define DST_BLOCK_WRITE8(dst, val) \ )==" "\n" |
| 166 | R"==(intel_sub_group_block_write8((__global uint *)(dst), as_uint8(val)) )==" "\n" |
| 167 | R"==(#endif )==" "\n" |
| 168 | R"==(#if DST_DT_F64 )==" "\n" |
| 169 | R"==(#define DST_BLOCK_READ(src) \ )==" "\n" |
| 170 | R"==(as_double(intel_sub_group_block_read2((const __global uint *)(src))) )==" "\n" |
| 171 | R"==(#define DST_BLOCK_READ8(src) \ )==" "\n" |
| 172 | R"==((double8)((as_double4(intel_sub_group_block_read8( \ )==" "\n" |
| 173 | R"==((const __global uint *)(src)))), \ )==" "\n" |
| 174 | R"==((as_double4(intel_sub_group_block_read8( \ )==" "\n" |
| 175 | R"==((const __global uint *)(src))))) )==" "\n" |
| 176 | R"==(#define DST_BLOCK_WRITE(dst, val) \ )==" "\n" |
| 177 | R"==(intel_sub_group_block_write2((__global uint *)(dst), as_uint2(val)) )==" "\n" |
| 178 | R"==(#define DST_BLOCK_WRITE8(dst, val) \ )==" "\n" |
| 179 | R"==(do { \ )==" "\n" |
| 180 | R"==(intel_sub_group_block_write8( \ )==" "\n" |
| 181 | R"==((__global uint *)(dst), as_uint8(val.lo)); \ )==" "\n" |
| 182 | R"==(intel_sub_group_block_write8( \ )==" "\n" |
| 183 | R"==((__global uint *)(dst + 8), as_uint8(val.hi)); \ )==" "\n" |
| 184 | R"==(} while (0) )==" "\n" |
| 185 | R"==(#endif )==" "\n" |
| 186 | R"==(#if DST_DT_BF16 )==" "\n" |
| 187 | R"==(#define DST_BLOCK_READ(src) \ )==" "\n" |
| 188 | R"==(as_ushort(intel_sub_group_block_read_us((const __global ushort *)(src))) )==" "\n" |
| 189 | R"==(#define DST_BLOCK_READ8(src) \ )==" "\n" |
| 190 | R"==(as_ushort8(intel_sub_group_block_read_us8((const __global ushort *)(src))) )==" "\n" |
| 191 | R"==(#define DST_BLOCK_WRITE(dst, val) \ )==" "\n" |
| 192 | R"==(intel_sub_group_block_write_us((__global ushort *)(dst), as_ushort(val)) )==" "\n" |
| 193 | R"==(#define DST_BLOCK_WRITE8(dst, val) \ )==" "\n" |
| 194 | R"==(intel_sub_group_block_write_us8((__global ushort *)(dst), as_ushort8(val)) )==" "\n" |
| 195 | R"==(#endif )==" "\n" |
| 196 | R"==(#if (SRC_DT_S8 && DST_DT_S8) || (SRC_DT_U8 && DST_DT_U8) \ )==" "\n" |
| 197 | R"==(|| (SRC_DT_BF16 && DST_DT_BF16) || (SRC_DT_F16 && DST_DT_F16) \ )==" "\n" |
| 198 | R"==(|| (SRC_DT_F32 && DST_DT_F32) || (SRC_DT_S32 && DST_DT_S32) \ )==" "\n" |
| 199 | R"==(|| (SRC_DT_F64 && DST_DT_F64) )==" "\n" |
| 200 | R"==(#define SRC_TO_DST(x) (x) )==" "\n" |
| 201 | R"==(#define SRC_TO_DST8(x) (x) )==" "\n" |
| 202 | R"==(#else )==" "\n" |
| 203 | R"==(#define SRC_TO_DST(x) TO_DST(SRC_TO_REF(x)) )==" "\n" |
| 204 | R"==(#define SRC_TO_DST8(x) TO_DST8(SRC_TO_REF8(x)) )==" "\n" |
| 205 | R"==(#endif )==" "\n" |
| 206 | R"==(#define SCALE_MUL(x, s) (x) * (s) )==" "\n" |
| 207 | R"==(#define SCALE_DIV(x, s) (x) / (s) )==" "\n" |
| 208 | R"==(#define SCALE_NONE(x, s) (x) )==" "\n" |
| 209 | R"==(#if WITH_SRC_SCALE )==" "\n" |
| 210 | R"==(#define SRC_SCALE SCALE_MUL )==" "\n" |
| 211 | R"==(#else )==" "\n" |
| 212 | R"==(#define SRC_SCALE SCALE_NONE )==" "\n" |
| 213 | R"==(#endif )==" "\n" |
| 214 | R"==(#if WITH_DST_SCALE )==" "\n" |
| 215 | R"==(#define DST_SCALE SCALE_DIV )==" "\n" |
| 216 | R"==(#else )==" "\n" |
| 217 | R"==(#define DST_SCALE SCALE_NONE )==" "\n" |
| 218 | R"==(#endif )==" "\n" |
| 219 | R"==(#if WITH_SUM_SCALE )==" "\n" |
| 220 | R"==(#define SUM_SCALE SCALE_MUL )==" "\n" |
| 221 | R"==(#else )==" "\n" |
| 222 | R"==(#define SUM_SCALE SCALE_NONE )==" "\n" |
| 223 | R"==(#endif )==" "\n" |
| 224 | R"==(#define ZP_SHIFT(x, x0) (x) - (float)(x0) )==" "\n" |
| 225 | R"==(#define ZP_UNSHIFT(x, x0) (x) + (float)(x0) )==" "\n" |
| 226 | R"==(#define ZP_NO_SHIFT(x, x0) (x) )==" "\n" |
| 227 | R"==(#define ZP_READ_VAL(x) x )==" "\n" |
| 228 | R"==(#define ZP_READ_PTR(x) x[0] )==" "\n" |
| 229 | R"==(#define ZP_ZERO(x) 0 )==" "\n" |
| 230 | R"==(#if WITH_SRC_ZPOINT )==" "\n" |
| 231 | R"==(#define SRC_SHIFT ZP_SHIFT )==" "\n" |
| 232 | R"==(#define GET_SRC_ZP ZP_READ_PTR )==" "\n" |
| 233 | R"==(#else )==" "\n" |
| 234 | R"==(#define SRC_SHIFT ZP_NO_SHIFT )==" "\n" |
| 235 | R"==(#define GET_SRC_ZP ZP_ZERO )==" "\n" |
| 236 | R"==(#endif )==" "\n" |
| 237 | R"==(#if WITH_DST_ZPOINT )==" "\n" |
| 238 | R"==(#define DST_SHIFT ZP_UNSHIFT )==" "\n" |
| 239 | R"==(#define GET_DST_ZP ZP_READ_PTR )==" "\n" |
| 240 | R"==(#else )==" "\n" |
| 241 | R"==(#define DST_SHIFT ZP_NO_SHIFT )==" "\n" |
| 242 | R"==(#define GET_DST_ZP ZP_ZERO )==" "\n" |
| 243 | R"==(#endif )==" "\n" |
| 244 | R"==(#if WITH_SUM_ZPOINT )==" "\n" |
| 245 | R"==(#define SUM_SHIFT ZP_SHIFT )==" "\n" |
| 246 | R"==(#define GET_SUM_ZP ZP_READ_VAL )==" "\n" |
| 247 | R"==(#else )==" "\n" |
| 248 | R"==(#define SUM_SHIFT ZP_NO_SHIFT )==" "\n" |
| 249 | R"==(#define GET_SUM_ZP ZP_ZERO )==" "\n" |
| 250 | R"==(#endif )==" "\n" |
| 251 | R"==(#define SRC_AXPY(a, x, x0) SRC_SCALE(SRC_SHIFT(x, x0), a) )==" "\n" |
| 252 | R"==(#define DST_AXPY(a, x, x0) DST_SHIFT(DST_SCALE(x, a), x0) )==" "\n" |
| 253 | R"==(#define SUM_AXPY(a, x, x0) SUM_SCALE(SUM_SHIFT(x, x0), a) )==" "\n" |
| 254 | R"==(#define AXPY(src, dst, a, b, c, x0, y0, z0) \ )==" "\n" |
| 255 | R"==(DST_AXPY(b, (SRC_AXPY(a, src, x0) + SUM_AXPY(c, dst, y0 + z0)), y0) )==" "\n" |
| 256 | R"==(#if WITH_SRC_SCALE || WITH_SRC_ZPOINT )==" "\n" |
| 257 | R"==(#define WITH_SRC_MOD 1 )==" "\n" |
| 258 | R"==(#endif )==" "\n" |
| 259 | R"==(#if WITH_DST_SCALE || WITH_DST_ZPOINT )==" "\n" |
| 260 | R"==(#define WITH_DST_MOD 1 )==" "\n" |
| 261 | R"==(#endif )==" "\n" |
| 262 | R"==(#if WITH_SUM_SCALE || WITH_SUM_ZPOINT )==" "\n" |
| 263 | R"==(#define WITH_SUM_MOD 1 )==" "\n" |
| 264 | R"==(#endif )==" "\n" |
| 265 | R"==(#if WITH_SUM_MOD )==" "\n" |
| 266 | R"==(#define REORDER(_dst, _src, _a, _b, _c, _x0, _y0, _z0) \ )==" "\n" |
| 267 | R"==(do { \ )==" "\n" |
| 268 | R"==(const float _x = SRC_TO_REF(_src); \ )==" "\n" |
| 269 | R"==(const float _y = DST_TO_REF(_dst); \ )==" "\n" |
| 270 | R"==(const float _s = AXPY(_x, _y, _a, _b, _c, _x0, _y0, _z0); \ )==" "\n" |
| 271 | R"==(_dst = TO_DST(_s); \ )==" "\n" |
| 272 | R"==(} while (0) )==" "\n" |
| 273 | R"==(#define REORDER8(_dst, _src, _a, _b, _c, _x0, _y0, _z0) \ )==" "\n" |
| 274 | R"==(do { \ )==" "\n" |
| 275 | R"==(const float8 _x = convert_float8(SRC_TO_REF8(_src)); \ )==" "\n" |
| 276 | R"==(const float8 _y = convert_float8(DST_TO_REF8(_dst)); \ )==" "\n" |
| 277 | R"==(const float8 _s = AXPY(_x, _y, _a, _b, _c, _x0, _y0, _z0); \ )==" "\n" |
| 278 | R"==(_dst = TO_DST8(_s); \ )==" "\n" |
| 279 | R"==(} while (0) )==" "\n" |
| 280 | R"==(#elif WITH_SRC_MOD || WITH_DST_MOD )==" "\n" |
| 281 | R"==(#define REORDER(_dst, _src, _a, _b, _c, _x0, _y0, _z0) \ )==" "\n" |
| 282 | R"==(do { \ )==" "\n" |
| 283 | R"==(const float _x = SRC_TO_REF(_src); \ )==" "\n" |
| 284 | R"==(const float _s = AXPY(_x, 0.f, _a, _b, _c, _x0, _y0, _z0); \ )==" "\n" |
| 285 | R"==(_dst = TO_DST(_s); \ )==" "\n" |
| 286 | R"==(} while (0) )==" "\n" |
| 287 | R"==(#define REORDER8(_dst, _src, _a, _b, _c, _x0, _y0, _z0) \ )==" "\n" |
| 288 | R"==(do { \ )==" "\n" |
| 289 | R"==(const float8 _x = convert_float8(SRC_TO_REF8(_src)); \ )==" "\n" |
| 290 | R"==(const float8 _s = AXPY(_x, 0.f, _a, _b, _c, _x0, _y0, _z0); \ )==" "\n" |
| 291 | R"==(_dst = TO_DST8(_s); \ )==" "\n" |
| 292 | R"==(} while (0) )==" "\n" |
| 293 | R"==(#else )==" "\n" |
| 294 | R"==(#define REORDER(_dst, _src, _a, _b, _c, _x0, _y0, _z0) \ )==" "\n" |
| 295 | R"==(do { \ )==" "\n" |
| 296 | R"==(_dst = SRC_TO_DST(_src); \ )==" "\n" |
| 297 | R"==(} while (0) )==" "\n" |
| 298 | R"==(#define REORDER8(_dst, _src, _a, _b, _c, _x0, _y0, _z0) \ )==" "\n" |
| 299 | R"==(do { \ )==" "\n" |
| 300 | R"==(_dst = SRC_TO_DST8(_src); \ )==" "\n" |
| 301 | R"==(} while (0) )==" "\n" |
| 302 | R"==(#endif )==" "\n" |
| 303 | R"==(#if WITH_SRC_SCALE || WITH_DST_SCALE )==" "\n" |
| 304 | R"==(#define MASK_DIM(prefix, dim) ((CONCAT2(prefix, _SCALE_MASK) >> dim) & 1) )==" "\n" |
| 305 | R"==(#define SCALE_DIM(prefix, dim) \ )==" "\n" |
| 306 | R"==((MASK_DIM(prefix, dim) ? CONCAT3(prefix, _D, dim) : 1) )==" "\n" |
| 307 | R"==(#define SCALE_S5(prefix) (1) )==" "\n" |
| 308 | R"==(#define SCALE_S4(prefix) (SCALE_DIM(prefix, 5) * SCALE_S5(prefix)) )==" "\n" |
| 309 | R"==(#define SCALE_S3(prefix) (SCALE_DIM(prefix, 4) * SCALE_S4(prefix)) )==" "\n" |
| 310 | R"==(#define SCALE_S2(prefix) (SCALE_DIM(prefix, 3) * SCALE_S3(prefix)) )==" "\n" |
| 311 | R"==(#define SCALE_S1(prefix) (SCALE_DIM(prefix, 2) * SCALE_S2(prefix)) )==" "\n" |
| 312 | R"==(#define SCALE_S0(prefix) (SCALE_DIM(prefix, 1) * SCALE_S1(prefix)) )==" "\n" |
| 313 | R"==(#define SCALE_STRIDE(prefix, dim) \ )==" "\n" |
| 314 | R"==((CONCAT2(SCALE_S, dim)(prefix) * MASK_DIM(prefix, dim)) )==" "\n" |
| 315 | R"==(#define SCALE_OFF(prefix, x0, x1, x2, x3, x4, x5) \ )==" "\n" |
| 316 | R"==(((x0)*SCALE_STRIDE(prefix, 0) + (x1)*SCALE_STRIDE(prefix, 1) \ )==" "\n" |
| 317 | R"==(+ (x2)*SCALE_STRIDE(prefix, 2) + (x3)*SCALE_STRIDE(prefix, 3) \ )==" "\n" |
| 318 | R"==(+ (x4)*SCALE_STRIDE(prefix, 4) + (x5)*SCALE_STRIDE(prefix, 5)) )==" "\n" |
| 319 | R"==(#endif )==" "\n" |
| 320 | R"==()=="; |
| 321 | } |
| 322 | } |
| 323 | } |
| 324 | } |
Generated on 2022-Nov-30 from project oneDNN revision 84bc949a1
Powered by 
Code Browser 2.1
Generator usage only permitted with license.