| 1 | namespace dnnl { |
|---|---|
| 2 | namespace impl { |
| 3 | namespace gpu { |
| 4 | namespace ocl { |
| 5 | const char *xe_hp_systolic_gemm_copy_kernel = R"==(/******************************************************************************* )==" "\n" |
| 6 | R"==(* Copyright 2019-2022 Intel Corporation )==" "\n" |
| 7 | R"==(* )==" "\n" |
| 8 | R"==(* Licensed under the Apache License, Version 2.0 (the "License"); )==" "\n" |
| 9 | R"==(* you may not use this file except in compliance with the License. )==" "\n" |
| 10 | R"==(* You may obtain a copy of the License at )==" "\n" |
| 11 | R"==(* )==" "\n" |
| 12 | R"==(* http: )==" "\n" |
| 13 | R"==(* )==" "\n" |
| 14 | R"==(* Unless required by applicable law or agreed to in writing, software )==" "\n" |
| 15 | R"==(* distributed under the License is distributed on an "AS IS" BASIS, )==" "\n" |
| 16 | R"==(* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. )==" "\n" |
| 17 | R"==(* See the License for the specific language governing permissions and )==" "\n" |
| 18 | R"==(* limitations under the License. )==" "\n" |
| 19 | R"==(*******************************************************************************/ )==" "\n" |
| 20 | R"==(#include "gpu/ocl/ocl_math_utils.h" )==" "\n" |
| 21 | R"==(#if ELEMENT_SIZE == 2 )==" "\n" |
| 22 | R"==(#pragma OPENCL EXTENSION cl_intel_subgroups_short : enable )==" "\n" |
| 23 | R"==(#define ELEMENT ushort )==" "\n" |
| 24 | R"==(#define ELEMENT2 ushort2 )==" "\n" |
| 25 | R"==(#define ELEMENT4 ushort4 )==" "\n" |
| 26 | R"==(#define ELEMENT8 ushort8 )==" "\n" |
| 27 | R"==(#define ELEMENT16 ushort16 )==" "\n" |
| 28 | R"==(#define ELEMENT_INT ushort2 )==" "\n" |
| 29 | R"==(#define ELEMENT_INT4 ushort8 )==" "\n" |
| 30 | R"==(#define VLOAD_ELEMENT_INT vload2 )==" "\n" |
| 31 | R"==(#define ELEMENTS_PER_INT 2 )==" "\n" |
| 32 | R"==(#define BLOCK_READ_ELEMENT2 intel_sub_group_block_read_us2 )==" "\n" |
| 33 | R"==(#define BLOCK_READ_ELEMENT4 intel_sub_group_block_read_us4 )==" "\n" |
| 34 | R"==(#define BLOCK_READ_ELEMENT_INT intel_sub_group_block_read_us2 )==" "\n" |
| 35 | R"==(#define MASKED_BLOCK_READ_ELEMENT_INT masked_block_read_element2 )==" "\n" |
| 36 | R"==(#define BLOCK_WRITE_ELEMENT_INT4 intel_sub_group_block_write_us8 )==" "\n" |
| 37 | R"==(#elif ELEMENT_SIZE == 1 )==" "\n" |
| 38 | R"==(#define ELEMENT uchar )==" "\n" |
| 39 | R"==(#define ELEMENT2 uchar2 )==" "\n" |
| 40 | R"==(#define ELEMENT4 uchar4 )==" "\n" |
| 41 | R"==(#define ELEMENT8 uchar8 )==" "\n" |
| 42 | R"==(#define ELEMENT16 uchar16 )==" "\n" |
| 43 | R"==(#define ELEMENT_INT uchar4 )==" "\n" |
| 44 | R"==(#define ELEMENT_INT4 uchar16 )==" "\n" |
| 45 | R"==(#define VLOAD_ELEMENT_INT vload4 )==" "\n" |
| 46 | R"==(#define BLOCK_READ_ELEMENT2 intel_sub_group_block_read_uc2 )==" "\n" |
| 47 | R"==(#define BLOCK_READ_ELEMENT4 intel_sub_group_block_read_uc4 )==" "\n" |
| 48 | R"==(#define BLOCK_READ_ELEMENT_INT intel_sub_group_block_read_uc4 )==" "\n" |
| 49 | R"==(#define MASKED_BLOCK_READ_ELEMENT_INT masked_block_read_element4 )==" "\n" |
| 50 | R"==(#define BLOCK_WRITE_ELEMENT_INT4 intel_sub_group_block_write_uc16 )==" "\n" |
| 51 | R"==(#define ELEMENTS_PER_INT 4 )==" "\n" |
| 52 | R"==(#define SUM_T int )==" "\n" |
| 53 | R"==(#define SUM_T4 int4 )==" "\n" |
| 54 | R"==(#define CONVERT_SUM_T convert_int )==" "\n" |
| 55 | R"==(#define CONVERT_SUM_T4 convert_int4 )==" "\n" |
| 56 | R"==(#if COPY_SIGNED )==" "\n" |
| 57 | R"==(#define AS_SIGNED_ELEMENT as_char )==" "\n" |
| 58 | R"==(#define AS_SIGNED_ELEMENT4 as_char4 )==" "\n" |
| 59 | R"==(#define AS_SIGNED_ELEMENT_INT as_char4 )==" "\n" |
| 60 | R"==(#define SIGNED_ELEMENT_INT char4 )==" "\n" |
| 61 | R"==(#else )==" "\n" |
| 62 | R"==(#define AS_SIGNED_ELEMENT as_uchar )==" "\n" |
| 63 | R"==(#define AS_SIGNED_ELEMENT4 as_uchar4 )==" "\n" |
| 64 | R"==(#define AS_SIGNED_ELEMENT_INT as_uchar4 )==" "\n" |
| 65 | R"==(#define SIGNED_ELEMENT_INT uchar4 )==" "\n" |
| 66 | R"==(#endif )==" "\n" |
| 67 | R"==(#else )==" "\n" |
| 68 | R"==(#error Unsupported element size. )==" "\n" |
| 69 | R"==(#endif )==" "\n" |
| 70 | R"==(#if !COPY_A && !COPY_B )==" "\n" |
| 71 | R"==(#error Source matrix not defined. )==" "\n" |
| 72 | R"==(#endif )==" "\n" |
| 73 | R"==(inline ELEMENT2 masked_block_read_element2(global ELEMENT *p, int rem) { )==" "\n" |
| 74 | R"==(ELEMENT2 v; )==" "\n" |
| 75 | R"==(int lid = get_sub_group_local_id(); )==" "\n" |
| 76 | R"==(int sg = get_sub_group_size(); )==" "\n" |
| 77 | R"==(v.s0 = (lid < rem) ? p[lid] : 0; )==" "\n" |
| 78 | R"==(v.s1 = (lid + sg < rem) ? p[lid + sg] : 0; )==" "\n" |
| 79 | R"==(return v; )==" "\n" |
| 80 | R"==(} )==" "\n" |
| 81 | R"==(inline ELEMENT4 masked_block_read_element4(global ELEMENT *p, int rem) { )==" "\n" |
| 82 | R"==(ELEMENT4 v; )==" "\n" |
| 83 | R"==(int lid = get_sub_group_local_id(); )==" "\n" |
| 84 | R"==(int sg = get_sub_group_size(); )==" "\n" |
| 85 | R"==(v.s0 = (lid < rem) ? p[lid] : 0; )==" "\n" |
| 86 | R"==(v.s1 = (lid + sg < rem) ? p[lid + sg] : 0; )==" "\n" |
| 87 | R"==(v.s2 = (lid + 2 * sg < rem) ? p[lid + 2 * sg] : 0; )==" "\n" |
| 88 | R"==(v.s3 = (lid + 3 * sg < rem) ? p[lid + 3 * sg] : 0; )==" "\n" |
| 89 | R"==(return v; )==" "\n" |
| 90 | R"==(} )==" "\n" |
| 91 | R"==(__attribute__((overloadable)) inline int sum(int v) { )==" "\n" |
| 92 | R"==(return sub_group_reduce_add(v); )==" "\n" |
| 93 | R"==(} )==" "\n" |
| 94 | R"==(__attribute__((overloadable)) inline int sum(int4 v) { )==" "\n" |
| 95 | R"==(return sub_group_reduce_add(v.s0) + sub_group_reduce_add(v.s1) )==" "\n" |
| 96 | R"==(+ sub_group_reduce_add(v.s2) + sub_group_reduce_add(v.s3); )==" "\n" |
| 97 | R"==(} )==" "\n" |
| 98 | R"==(void dummy_dpas() { )==" "\n" |
| 99 | R"==(if (get_sub_group_local_id() >= 16) { )==" "\n" |
| 100 | R"==(int __builtin_IB_sub_group_idpas_s8_s8_8_1(int, int, int8) )==" "\n" |
| 101 | R"==(__attribute__((const)); )==" "\n" |
| 102 | R"==(global volatile int *_; )==" "\n" |
| 103 | R"==(int z = __builtin_IB_sub_group_idpas_s8_s8_8_1(0, _[0], 1); )==" "\n" |
| 104 | R"==(for (int i = 0; i < z; i++) )==" "\n" |
| 105 | R"==((void)_[0]; )==" "\n" |
| 106 | R"==(} )==" "\n" |
| 107 | R"==(} )==" "\n" |
| 108 | R"==(#define DUMMY_DPAS dummy_dpas() )==" "\n" |
| 109 | R"==(#if ELEMENT_SIZE == 2 )==" "\n" |
| 110 | R"==(#define PARTIAL_LOAD(regs, rrem, crem, cc, p) \ )==" "\n" |
| 111 | R"==(if ((2 * cc + 1) < crem) { \ )==" "\n" |
| 112 | R"==(if (lid < rrem) regs[cc] = vload2(0, p); \ )==" "\n" |
| 113 | R"==(} else if ((2 * cc) < crem) { \ )==" "\n" |
| 114 | R"==(if (lid < rrem) regs[cc].s0 = *(p); \ )==" "\n" |
| 115 | R"==(} )==" "\n" |
| 116 | R"==(#elif ELEMENT_SIZE == 1 )==" "\n" |
| 117 | R"==(#define PARTIAL_LOAD(regs, rrem, crem, cc, p) \ )==" "\n" |
| 118 | R"==(if ((4 * cc + 3) < crem) { \ )==" "\n" |
| 119 | R"==(if (lid < rrem) regs[cc] = vload4(0, p); \ )==" "\n" |
| 120 | R"==(} else if ((4 * cc + 2) < crem) { \ )==" "\n" |
| 121 | R"==(if (lid < rrem) regs[cc].s012 = vload3(0, p); \ )==" "\n" |
| 122 | R"==(} else if ((4 * cc + 1) < crem) { \ )==" "\n" |
| 123 | R"==(if (lid < rrem) regs[cc].s01 = vload2(0, p); \ )==" "\n" |
| 124 | R"==(} else if (4 * cc < crem) { \ )==" "\n" |
| 125 | R"==(if (lid < rrem) regs[cc].s0 = *(p); \ )==" "\n" |
| 126 | R"==(} )==" "\n" |
| 127 | R"==(#endif )==" "\n" |
| 128 | R"==(#if COPY_A )==" "\n" |
| 129 | R"==(#define UNROLL_M 32 )==" "\n" |
| 130 | R"==(#define UNROLL_K (32 / ELEMENT_SIZE) )==" "\n" |
| 131 | R"==(#if COPY_SUM )==" "\n" |
| 132 | R"==(#define GET_A_SUM_ADDRESS \ )==" "\n" |
| 133 | R"==(global int *a_sum = (global int *)(a_packed + offseta_packed \ )==" "\n" |
| 134 | R"==(+ (m0 + UNROLL_M) * lda_packed - UNROLL_M * sizeof(int)); )==" "\n" |
| 135 | R"==(#else )==" "\n" |
| 136 | R"==(#define GET_A_SUM_ADDRESS )==" "\n" |
| 137 | R"==(#endif )==" "\n" |
| 138 | R"==(#if COPY_CLEAR_SUM )==" "\n" |
| 139 | R"==(__attribute__((intel_reqd_sub_group_size(8))) kernel void )==" "\n" |
| 140 | R"==(xe_hp_systolic_gemm_copy(long m, long k, global ELEMENT *a_packed, )==" "\n" |
| 141 | R"==(int offseta_packed, int lda_packed) { )==" "\n" |
| 142 | R"==(uint m0 = (sub_group_broadcast(get_global_id(0), 0) / 8) * UNROLL_M; )==" "\n" |
| 143 | R"==(GET_A_SUM_ADDRESS; )==" "\n" |
| 144 | R"==(uint4 zero = 0; )==" "\n" |
| 145 | R"==(intel_sub_group_block_write4(a_sum, zero); )==" "\n" |
| 146 | R"==(} )==" "\n" |
| 147 | R"==(#elif !COPY_TRANS )==" "\n" |
| 148 | R"==(#if ELEMENT_SIZE == 2 )==" "\n" |
| 149 | R"==(#define REPACK_REG(rr, cc) \ )==" "\n" |
| 150 | R"==(blk_r[rr].s##cc = (((uint)c[2 * cc + 1].s##rr) << 16) | c[2 * cc].s##rr )==" "\n" |
| 151 | R"==(#elif ELEMENT_SIZE == 1 )==" "\n" |
| 152 | R"==(#define REPACK_REG(rr, cc) \ )==" "\n" |
| 153 | R"==(blk_r[rr].s##cc = (((uint)c[4 * cc + 3].s##rr) << 24) \ )==" "\n" |
| 154 | R"==(| (((uint)c[4 * cc + 2].s##rr) << 16) \ )==" "\n" |
| 155 | R"==(| (((uint)c[4 * cc + 1].s##rr) << 8) | c[4 * cc].s##rr )==" "\n" |
| 156 | R"==(#endif )==" "\n" |
| 157 | R"==(#define REPACK_CC(cc) \ )==" "\n" |
| 158 | R"==(REPACK_REG(0, cc); \ )==" "\n" |
| 159 | R"==(REPACK_REG(1, cc); \ )==" "\n" |
| 160 | R"==(REPACK_REG(2, cc); \ )==" "\n" |
| 161 | R"==(REPACK_REG(3, cc) )==" "\n" |
| 162 | R"==(#define REPACK \ )==" "\n" |
| 163 | R"==(REPACK_CC(0); \ )==" "\n" |
| 164 | R"==(REPACK_CC(1); \ )==" "\n" |
| 165 | R"==(REPACK_CC(2); \ )==" "\n" |
| 166 | R"==(REPACK_CC(3); \ )==" "\n" |
| 167 | R"==(REPACK_CC(4); \ )==" "\n" |
| 168 | R"==(REPACK_CC(5); \ )==" "\n" |
| 169 | R"==(REPACK_CC(6); \ )==" "\n" |
| 170 | R"==(REPACK_CC(7) )==" "\n" |
| 171 | R"==(__attribute__((intel_reqd_sub_group_size(8))) kernel void )==" "\n" |
| 172 | R"==(xe_hp_systolic_gemm_copy(long m, long k, global ELEMENT *a, long offseta, )==" "\n" |
| 173 | R"==(long lda, global ELEMENT *a_packed, int offseta_packed, )==" "\n" |
| 174 | R"==(int lda_packed) { )==" "\n" |
| 175 | R"==(int lid = get_sub_group_local_id(); )==" "\n" |
| 176 | R"==(uint m0 = (sub_group_broadcast(get_global_id(0), 0) / 8) * UNROLL_M; )==" "\n" |
| 177 | R"==(uint k0 = get_global_id(1) * UNROLL_K; )==" "\n" |
| 178 | R"==(int mrem = m - m0; )==" "\n" |
| 179 | R"==(int krem = k - k0; )==" "\n" |
| 180 | R"==(bool aligned = ((as_long(a) | lda | offseta) & (ELEMENTS_PER_INT - 1)) == 0; )==" "\n" |
| 181 | R"==(if (mrem <= 0 || krem <= 0) return; )==" "\n" |
| 182 | R"==(GET_A_SUM_ADDRESS; )==" "\n" |
| 183 | R"==(a += offseta + m0 + k0 * lda; )==" "\n" |
| 184 | R"==(a_packed += offseta_packed + m0 * lda_packed + k0 * UNROLL_M; )==" "\n" |
| 185 | R"==(ELEMENT4 c[UNROLL_K]; )==" "\n" |
| 186 | R"==(if (mrem >= UNROLL_M && krem >= UNROLL_K && aligned) { )==" "\n" |
| 187 | R"==(for (int h = 0; h < UNROLL_K; h++) )==" "\n" |
| 188 | R"==(c[h] = BLOCK_READ_ELEMENT4(a + h * lda); )==" "\n" |
| 189 | R"==(} else { )==" "\n" |
| 190 | R"==(for (int h = 0; h < UNROLL_K; h++) )==" "\n" |
| 191 | R"==(if (h < krem) )==" "\n" |
| 192 | R"==(c[h] = masked_block_read_element4(a + h * lda, mrem); )==" "\n" |
| 193 | R"==(else )==" "\n" |
| 194 | R"==(c[h] = 0; )==" "\n" |
| 195 | R"==(} )==" "\n" |
| 196 | R"==(uint8 blk_r[UNROLL_M / 8]; )==" "\n" |
| 197 | R"==(REPACK; )==" "\n" |
| 198 | R"==(for (int rr = 0; rr < UNROLL_M / 8; rr++) )==" "\n" |
| 199 | R"==(intel_sub_group_block_write8( )==" "\n" |
| 200 | R"==((global uint *)(a_packed + rr * UNROLL_K * 8), blk_r[rr]); )==" "\n" |
| 201 | R"==(#if COPY_SUM )==" "\n" |
| 202 | R"==(SUM_T4 sum = 0; )==" "\n" |
| 203 | R"==(for (int h = 0; h < UNROLL_K; h++) )==" "\n" |
| 204 | R"==(sum += CONVERT_SUM_T4(AS_SIGNED_ELEMENT4(c[h])); )==" "\n" |
| 205 | R"==(atomic_add(a_sum + lid, sum.s0); )==" "\n" |
| 206 | R"==(atomic_add(a_sum + lid + 8, sum.s1); )==" "\n" |
| 207 | R"==(atomic_add(a_sum + lid + 16, sum.s2); )==" "\n" |
| 208 | R"==(atomic_add(a_sum + lid + 24, sum.s3); )==" "\n" |
| 209 | R"==(#endif )==" "\n" |
| 210 | R"==(DUMMY_DPAS; )==" "\n" |
| 211 | R"==(} )==" "\n" |
| 212 | R"==(#else /* COPY_TRANS */ )==" "\n" |
| 213 | R"==(__attribute__((intel_reqd_workgroup_walk_order(1, 0))) )==" "\n" |
| 214 | R"==(__attribute__((intel_reqd_sub_group_size(8))) kernel void )==" "\n" |
| 215 | R"==(xe_hp_systolic_gemm_copy(long m, long k, global ELEMENT *a, long offseta, )==" "\n" |
| 216 | R"==(long lda, global ELEMENT *a_packed, int offseta_packed, )==" "\n" |
| 217 | R"==(int lda_packed) { )==" "\n" |
| 218 | R"==(int lid = get_sub_group_local_id(); )==" "\n" |
| 219 | R"==(uint m0 = (sub_group_broadcast(get_global_id(0), 0) / 8) * UNROLL_M; )==" "\n" |
| 220 | R"==(uint k0 = get_global_id(1) * UNROLL_K; )==" "\n" |
| 221 | R"==(int mrem = m - m0; )==" "\n" |
| 222 | R"==(int krem = k - k0; )==" "\n" |
| 223 | R"==(if (mrem <= 0 || krem <= 0) return; )==" "\n" |
| 224 | R"==(GET_A_SUM_ADDRESS; )==" "\n" |
| 225 | R"==(a += offseta + m0 * lda + k0; )==" "\n" |
| 226 | R"==(a_packed += offseta_packed + m0 * lda_packed + k0 * UNROLL_M; )==" "\n" |
| 227 | R"==(#if COPY_SUM )==" "\n" |
| 228 | R"==(SUM_T sum[UNROLL_M / 8] = {0}; )==" "\n" |
| 229 | R"==(#endif )==" "\n" |
| 230 | R"==(for (int rr = 0; rr < UNROLL_M / 8; rr++, mrem -= 8) { )==" "\n" |
| 231 | R"==(ELEMENT_INT regs[8]; )==" "\n" |
| 232 | R"==(if (mrem >= UNROLL_M && krem >= UNROLL_K) { )==" "\n" |
| 233 | R"==(for (int cc = 0; cc < UNROLL_K / ELEMENTS_PER_INT; cc++) )==" "\n" |
| 234 | R"==(regs[cc] = VLOAD_ELEMENT_INT(0, )==" "\n" |
| 235 | R"==(a + ((rr * 8) + lid) * lda + (cc * ELEMENTS_PER_INT)); )==" "\n" |
| 236 | R"==(} else { )==" "\n" |
| 237 | R"==(for (int cc = 0; cc < UNROLL_K / ELEMENTS_PER_INT; cc++) { )==" "\n" |
| 238 | R"==(regs[cc] = 0; )==" "\n" |
| 239 | R"==(PARTIAL_LOAD(regs, mrem, krem, cc, )==" "\n" |
| 240 | R"==(a + ((rr * 8) + lid) * lda + (cc * ELEMENTS_PER_INT)); )==" "\n" |
| 241 | R"==(} )==" "\n" |
| 242 | R"==(} )==" "\n" |
| 243 | R"==(uint8 blk_r; )==" "\n" |
| 244 | R"==(blk_r.s0 = as_uint(regs[0]); )==" "\n" |
| 245 | R"==(blk_r.s1 = as_uint(regs[1]); )==" "\n" |
| 246 | R"==(blk_r.s2 = as_uint(regs[2]); )==" "\n" |
| 247 | R"==(blk_r.s3 = as_uint(regs[3]); )==" "\n" |
| 248 | R"==(blk_r.s4 = as_uint(regs[4]); )==" "\n" |
| 249 | R"==(blk_r.s5 = as_uint(regs[5]); )==" "\n" |
| 250 | R"==(blk_r.s6 = as_uint(regs[6]); )==" "\n" |
| 251 | R"==(blk_r.s7 = as_uint(regs[7]); )==" "\n" |
| 252 | R"==(#if COPY_SUM )==" "\n" |
| 253 | R"==(for (int cc = 0; cc < UNROLL_K / ELEMENTS_PER_INT; cc++) { )==" "\n" |
| 254 | R"==(sum[rr] += CONVERT_SUM_T(AS_SIGNED_ELEMENT(regs[cc].s0)); )==" "\n" |
| 255 | R"==(sum[rr] += CONVERT_SUM_T(AS_SIGNED_ELEMENT(regs[cc].s1)); )==" "\n" |
| 256 | R"==(sum[rr] += CONVERT_SUM_T(AS_SIGNED_ELEMENT(regs[cc].s2)); )==" "\n" |
| 257 | R"==(sum[rr] += CONVERT_SUM_T(AS_SIGNED_ELEMENT(regs[cc].s3)); )==" "\n" |
| 258 | R"==(} )==" "\n" |
| 259 | R"==(#endif )==" "\n" |
| 260 | R"==(intel_sub_group_block_write8( )==" "\n" |
| 261 | R"==((global uint *)(a_packed + rr * UNROLL_K * 8), blk_r); )==" "\n" |
| 262 | R"==(} )==" "\n" |
| 263 | R"==(#if COPY_SUM )==" "\n" |
| 264 | R"==(atomic_add(a_sum + lid, sum[0]); )==" "\n" |
| 265 | R"==(atomic_add(a_sum + lid + 8, sum[1]); )==" "\n" |
| 266 | R"==(atomic_add(a_sum + lid + 16, sum[2]); )==" "\n" |
| 267 | R"==(atomic_add(a_sum + lid + 24, sum[3]); )==" "\n" |
| 268 | R"==(#endif )==" "\n" |
| 269 | R"==(DUMMY_DPAS; )==" "\n" |
| 270 | R"==(} )==" "\n" |
| 271 | R"==(#endif /* !COPY_TRANS */ )==" "\n" |
| 272 | R"==(#endif /* COPY_A */ )==" "\n" |
| 273 | R"==(#if COPY_B )==" "\n" |
| 274 | R"==(#define UNROLL_K (32 / ELEMENT_SIZE) )==" "\n" |
| 275 | R"==(#if ELEMENT_SIZE == 2 )==" "\n" |
| 276 | R"==(#define REPACK_CC(cc) \ )==" "\n" |
| 277 | R"==(do { \ )==" "\n" |
| 278 | R"==(colgroups[cc].s01 = cols[cc * 4]; \ )==" "\n" |
| 279 | R"==(colgroups[cc].s23 = cols[cc * 4 + 1]; \ )==" "\n" |
| 280 | R"==(colgroups[cc].s45 = cols[cc * 4 + 2]; \ )==" "\n" |
| 281 | R"==(colgroups[cc].s67 = cols[cc * 4 + 3]; \ )==" "\n" |
| 282 | R"==(} while (false) )==" "\n" |
| 283 | R"==(#define REPACK_CC2(cc) \ )==" "\n" |
| 284 | R"==(do { \ )==" "\n" |
| 285 | R"==(colgroups[cc].s02 = cols[cc * 2]; \ )==" "\n" |
| 286 | R"==(colgroups[cc].s13 = cols2[cc * 2]; \ )==" "\n" |
| 287 | R"==(colgroups[cc].s46 = cols[cc * 2 + 1]; \ )==" "\n" |
| 288 | R"==(colgroups[cc].s57 = cols2[cc * 2 + 1]; \ )==" "\n" |
| 289 | R"==(} while (false) )==" "\n" |
| 290 | R"==(#elif ELEMENT_SIZE == 1 )==" "\n" |
| 291 | R"==(#define REPACK_CC(cc) \ )==" "\n" |
| 292 | R"==(do { \ )==" "\n" |
| 293 | R"==(colgroups[cc].s0123 = cols[cc * 4]; \ )==" "\n" |
| 294 | R"==(colgroups[cc].s4567 = cols[cc * 4 + 1]; \ )==" "\n" |
| 295 | R"==(colgroups[cc].s89ab = cols[cc * 4 + 2]; \ )==" "\n" |
| 296 | R"==(colgroups[cc].scdef = cols[cc * 4 + 3]; \ )==" "\n" |
| 297 | R"==(} while (false) )==" "\n" |
| 298 | R"==(#define REPACK_CC4(cc) \ )==" "\n" |
| 299 | R"==(do { \ )==" "\n" |
| 300 | R"==(colgroups[cc].s048c = cols[cc]; \ )==" "\n" |
| 301 | R"==(colgroups[cc].s159d = cols2[cc]; \ )==" "\n" |
| 302 | R"==(colgroups[cc].s26ae = cols3[cc]; \ )==" "\n" |
| 303 | R"==(colgroups[cc].s37bf = cols4[cc]; \ )==" "\n" |
| 304 | R"==(} while (false) )==" "\n" |
| 305 | R"==(#endif )==" "\n" |
| 306 | R"==(#if COPY_SUM )==" "\n" |
| 307 | R"==(#define GET_B_SUM_ADDRESS \ )==" "\n" |
| 308 | R"==(global int *b_sum = (global int *)(b_packed + offsetb_packed \ )==" "\n" |
| 309 | R"==(+ (n0 + UNROLL_N) * ldb_packed - UNROLL_N * sizeof(int)); )==" "\n" |
| 310 | R"==(#else )==" "\n" |
| 311 | R"==(#define GET_B_SUM_ADDRESS )==" "\n" |
| 312 | R"==(#endif )==" "\n" |
| 313 | R"==(#if COPY_CLEAR_SUM )==" "\n" |
| 314 | R"==(__attribute__((intel_reqd_sub_group_size(8))) kernel void )==" "\n" |
| 315 | R"==(xe_hp_systolic_gemm_copy(long k, long n, global ELEMENT *b_packed, )==" "\n" |
| 316 | R"==(int offsetb_packed, int ldb_packed) { )==" "\n" |
| 317 | R"==(uint n0 = (sub_group_broadcast(get_global_id(0), 0) / 8) * UNROLL_N; )==" "\n" |
| 318 | R"==(GET_B_SUM_ADDRESS; )==" "\n" |
| 319 | R"==(uint4 zero = 0; )==" "\n" |
| 320 | R"==(intel_sub_group_block_write4(b_sum, zero); )==" "\n" |
| 321 | R"==(#if UNROLL_N > 32 )==" "\n" |
| 322 | R"==(intel_sub_group_block_write2(b_sum + 32, zero.s01); )==" "\n" |
| 323 | R"==(#endif )==" "\n" |
| 324 | R"==(} )==" "\n" |
| 325 | R"==(#elif !COPY_TRANS )==" "\n" |
| 326 | R"==(__attribute__((intel_reqd_sub_group_size(8))) kernel void )==" "\n" |
| 327 | R"==(xe_hp_systolic_gemm_copy(long k, long n, global ELEMENT *b, long offsetb, )==" "\n" |
| 328 | R"==(long ldb, global ELEMENT *b_packed, int offsetb_packed, )==" "\n" |
| 329 | R"==(int ldb_packed) { )==" "\n" |
| 330 | R"==(int lid = get_sub_group_local_id(); )==" "\n" |
| 331 | R"==(uint k0 = (sub_group_broadcast(get_global_id(0), 0) / 8) * UNROLL_K; )==" "\n" |
| 332 | R"==(uint n0 = get_global_id(1) * UNROLL_N; )==" "\n" |
| 333 | R"==(int krem = k - k0; )==" "\n" |
| 334 | R"==(int nrem = n - n0; )==" "\n" |
| 335 | R"==(bool aligned = ((as_long(b) | ldb | offsetb) & (ELEMENTS_PER_INT - 1)) == 0; )==" "\n" |
| 336 | R"==(if (nrem <= 0 || krem <= 0) return; )==" "\n" |
| 337 | R"==(GET_B_SUM_ADDRESS; )==" "\n" |
| 338 | R"==(b += offsetb + k0 + n0 * ldb; )==" "\n" |
| 339 | R"==(b_packed += offsetb_packed + n0 * ldb_packed + k0 * UNROLL_N; )==" "\n" |
| 340 | R"==(#define UNROLL_N_CHUNK (UNROLL_N / 2) )==" "\n" |
| 341 | R"==(#if COPY_SUM )==" "\n" |
| 342 | R"==(SUM_T sums[UNROLL_N]; )==" "\n" |
| 343 | R"==(#endif )==" "\n" |
| 344 | R"==(ELEMENT_INT cols[UNROLL_N / 2]; )==" "\n" |
| 345 | R"==(for (int c0 = 0; c0 < UNROLL_N; )==" "\n" |
| 346 | R"==(c0 += UNROLL_N_CHUNK, nrem -= UNROLL_N_CHUNK) { )==" "\n" |
| 347 | R"==(if (krem >= UNROLL_K && nrem >= UNROLL_N_CHUNK && aligned) { )==" "\n" |
| 348 | R"==(for (int c = 0; c < UNROLL_N_CHUNK; c++) )==" "\n" |
| 349 | R"==(cols[c] = BLOCK_READ_ELEMENT_INT(b + (c + c0) * ldb); )==" "\n" |
| 350 | R"==(} else { )==" "\n" |
| 351 | R"==(for (int c = 0; c < UNROLL_N_CHUNK; c++) )==" "\n" |
| 352 | R"==(if (c < nrem) )==" "\n" |
| 353 | R"==(cols[c] = MASKED_BLOCK_READ_ELEMENT_INT( )==" "\n" |
| 354 | R"==(b + (c + c0) * ldb, krem); )==" "\n" |
| 355 | R"==(else )==" "\n" |
| 356 | R"==(cols[c] = 0; )==" "\n" |
| 357 | R"==(} )==" "\n" |
| 358 | R"==(ELEMENT_INT4 colgroups[UNROLL_N_CHUNK / 4]; )==" "\n" |
| 359 | R"==(for (int cc = 0; cc < UNROLL_N_CHUNK / 4; cc++) )==" "\n" |
| 360 | R"==(REPACK_CC(cc); )==" "\n" |
| 361 | R"==(for (int cc = 0; cc < UNROLL_N_CHUNK / 4; cc++) )==" "\n" |
| 362 | R"==(BLOCK_WRITE_ELEMENT_INT4( )==" "\n" |
| 363 | R"==(b_packed + (cc * 4 + c0) * UNROLL_K, colgroups[cc]); )==" "\n" |
| 364 | R"==(#if COPY_SUM )==" "\n" |
| 365 | R"==(for (int c = 0; c < UNROLL_N_CHUNK; c++) )==" "\n" |
| 366 | R"==(sums[c + c0] = sum(CONVERT_SUM_T4(AS_SIGNED_ELEMENT_INT(cols[c]))); )==" "\n" |
| 367 | R"==(#endif )==" "\n" |
| 368 | R"==(} )==" "\n" |
| 369 | R"==(#if COPY_SUM )==" "\n" |
| 370 | R"==(for (int c0 = 0; c0 < UNROLL_N; c0 += get_sub_group_size()) )==" "\n" |
| 371 | R"==(atomic_add(b_sum + c0 + lid, sums[c0 + lid]); )==" "\n" |
| 372 | R"==(#endif )==" "\n" |
| 373 | R"==(DUMMY_DPAS; )==" "\n" |
| 374 | R"==(} )==" "\n" |
| 375 | R"==(#else /* COPY_TRANS */ )==" "\n" |
| 376 | R"==(#define ADD_SUM(coln) \ )==" "\n" |
| 377 | R"==(for (int cc = 0; cc < UNROLL_N / 4; cc++) { \ )==" "\n" |
| 378 | R"==(sums[4 * cc + 0] \ )==" "\n" |
| 379 | R"==(+= sum(CONVERT_SUM_T(AS_SIGNED_ELEMENT(coln[cc].s0))); \ )==" "\n" |
| 380 | R"==(sums[4 * cc + 1] \ )==" "\n" |
| 381 | R"==(+= sum(CONVERT_SUM_T(AS_SIGNED_ELEMENT(coln[cc].s1))); \ )==" "\n" |
| 382 | R"==(sums[4 * cc + 2] \ )==" "\n" |
| 383 | R"==(+= sum(CONVERT_SUM_T(AS_SIGNED_ELEMENT(coln[cc].s2))); \ )==" "\n" |
| 384 | R"==(sums[4 * cc + 3] \ )==" "\n" |
| 385 | R"==(+= sum(CONVERT_SUM_T(AS_SIGNED_ELEMENT(coln[cc].s3))); \ )==" "\n" |
| 386 | R"==(} )==" "\n" |
| 387 | R"==(__attribute__((intel_reqd_workgroup_walk_order(1, 0))) )==" "\n" |
| 388 | R"==(__attribute__((intel_reqd_sub_group_size(8))) kernel void )==" "\n" |
| 389 | R"==(xe_hp_systolic_gemm_copy(long k, long n, global ELEMENT *b, long offsetb, )==" "\n" |
| 390 | R"==(long ldb, global ELEMENT *b_packed, int offsetb_packed, )==" "\n" |
| 391 | R"==(int ldb_packed) { )==" "\n" |
| 392 | R"==(int lid = get_sub_group_local_id(); )==" "\n" |
| 393 | R"==(uint k0 = (sub_group_broadcast(get_global_id(0), 0) / 8) * UNROLL_K; )==" "\n" |
| 394 | R"==(uint n0 = get_global_id(1) * UNROLL_N; )==" "\n" |
| 395 | R"==(int krem = k - k0; )==" "\n" |
| 396 | R"==(int nrem = n - n0; )==" "\n" |
| 397 | R"==(int sg = get_sub_group_size(); )==" "\n" |
| 398 | R"==(if (nrem <= 0 || krem <= 0) return; )==" "\n" |
| 399 | R"==(GET_B_SUM_ADDRESS; )==" "\n" |
| 400 | R"==(b += offsetb + n0 + k0 * ldb; )==" "\n" |
| 401 | R"==(b_packed += offsetb_packed + n0 * ldb_packed + k0 * UNROLL_N; )==" "\n" |
| 402 | R"==(ELEMENT_INT cols[UNROLL_N / ELEMENTS_PER_INT]; )==" "\n" |
| 403 | R"==(ELEMENT_INT cols2[UNROLL_N / ELEMENTS_PER_INT]; )==" "\n" |
| 404 | R"==(ELEMENT_INT4 colgroups[UNROLL_N / 4]; )==" "\n" |
| 405 | R"==(if (krem >= 2 * sg && nrem >= UNROLL_N) { )==" "\n" |
| 406 | R"==(for (int cc = 0; cc < UNROLL_N / ELEMENTS_PER_INT; cc++) { )==" "\n" |
| 407 | R"==(cols[cc] = VLOAD_ELEMENT_INT( )==" "\n" |
| 408 | R"==(0, b + cc * ELEMENTS_PER_INT + lid * ldb); )==" "\n" |
| 409 | R"==(cols2[cc] = VLOAD_ELEMENT_INT( )==" "\n" |
| 410 | R"==(0, b + cc * ELEMENTS_PER_INT + (lid + sg) * ldb); )==" "\n" |
| 411 | R"==(} )==" "\n" |
| 412 | R"==(} else { )==" "\n" |
| 413 | R"==(for (int cc = 0; cc < UNROLL_N / ELEMENTS_PER_INT; cc++) { )==" "\n" |
| 414 | R"==(cols[cc] = 0; )==" "\n" |
| 415 | R"==(cols2[cc] = 0; )==" "\n" |
| 416 | R"==(PARTIAL_LOAD(cols, krem, nrem, cc, )==" "\n" |
| 417 | R"==(b + cc * ELEMENTS_PER_INT + lid * ldb); )==" "\n" |
| 418 | R"==(PARTIAL_LOAD(cols2, krem - sg, nrem, cc, )==" "\n" |
| 419 | R"==(b + cc * ELEMENTS_PER_INT + (lid + sg) * ldb); )==" "\n" |
| 420 | R"==(} )==" "\n" |
| 421 | R"==(} )==" "\n" |
| 422 | R"==(#if ELEMENT_SIZE == 2 )==" "\n" |
| 423 | R"==(for (int cc = 0; cc < UNROLL_N / 4; cc++) )==" "\n" |
| 424 | R"==(REPACK_CC2(cc); )==" "\n" |
| 425 | R"==(#else )==" "\n" |
| 426 | R"==(ELEMENT_INT cols3[UNROLL_N / ELEMENTS_PER_INT]; )==" "\n" |
| 427 | R"==(ELEMENT_INT cols4[UNROLL_N / ELEMENTS_PER_INT]; )==" "\n" |
| 428 | R"==(krem -= 2 * sg; )==" "\n" |
| 429 | R"==(if (krem >= 2 * sg && nrem >= UNROLL_N) { )==" "\n" |
| 430 | R"==(for (int cc = 0; cc < UNROLL_N / ELEMENTS_PER_INT; cc++) { )==" "\n" |
| 431 | R"==(cols3[cc] = VLOAD_ELEMENT_INT( )==" "\n" |
| 432 | R"==(0, b + cc * ELEMENTS_PER_INT + (lid + 2 * sg) * ldb); )==" "\n" |
| 433 | R"==(cols4[cc] = VLOAD_ELEMENT_INT( )==" "\n" |
| 434 | R"==(0, b + cc * ELEMENTS_PER_INT + (lid + 3 * sg) * ldb); )==" "\n" |
| 435 | R"==(} )==" "\n" |
| 436 | R"==(} else { )==" "\n" |
| 437 | R"==(for (int cc = 0; cc < UNROLL_N / ELEMENTS_PER_INT; cc++) { )==" "\n" |
| 438 | R"==(cols3[cc] = 0; )==" "\n" |
| 439 | R"==(cols4[cc] = 0; )==" "\n" |
| 440 | R"==(PARTIAL_LOAD(cols3, krem, nrem, cc, )==" "\n" |
| 441 | R"==(b + cc * ELEMENTS_PER_INT + (lid + 2 * sg) * ldb); )==" "\n" |
| 442 | R"==(PARTIAL_LOAD(cols4, krem - sg, nrem, cc, )==" "\n" |
| 443 | R"==(b + cc * ELEMENTS_PER_INT + (lid + 3 * sg) * ldb); )==" "\n" |
| 444 | R"==(} )==" "\n" |
| 445 | R"==(} )==" "\n" |
| 446 | R"==(for (int cc = 0; cc < UNROLL_N / 4; cc++) )==" "\n" |
| 447 | R"==(REPACK_CC4(cc); )==" "\n" |
| 448 | R"==(#endif )==" "\n" |
| 449 | R"==(for (int cc = 0; cc < UNROLL_N / 4; cc++) )==" "\n" |
| 450 | R"==(BLOCK_WRITE_ELEMENT_INT4(b_packed + cc * 4 * UNROLL_K, colgroups[cc]); )==" "\n" |
| 451 | R"==(#if COPY_SUM )==" "\n" |
| 452 | R"==(SUM_T sums[UNROLL_N] = {0}; )==" "\n" |
| 453 | R"==(ADD_SUM(cols); )==" "\n" |
| 454 | R"==(ADD_SUM(cols2); )==" "\n" |
| 455 | R"==(ADD_SUM(cols3); )==" "\n" |
| 456 | R"==(ADD_SUM(cols4); )==" "\n" |
| 457 | R"==(for (int c0 = 0; c0 < UNROLL_N; c0 += get_sub_group_size()) )==" "\n" |
| 458 | R"==(atomic_add(b_sum + c0 + lid, sums[c0 + lid]); )==" "\n" |
| 459 | R"==(#endif )==" "\n" |
| 460 | R"==(DUMMY_DPAS; )==" "\n" |
| 461 | R"==(} )==" "\n" |
| 462 | R"==(#endif /* !COPY_TRANS */ )==" "\n" |
| 463 | R"==(#endif /* COPY_B */ )==" "\n" |
| 464 | R"==()=="; |
| 465 | } |
| 466 | } |
| 467 | } |
| 468 | } |
Generated on 2022-Nov-30 from project oneDNN revision 84bc949a1
Powered by 
Code Browser 2.1
Generator usage only permitted with license.