| 1 | namespace dnnl { |
|---|---|
| 2 | namespace impl { |
| 3 | namespace gpu { |
| 4 | namespace ocl { |
| 5 | const char *ref_eltwise_kernel = R"==(/******************************************************************************* )==" "\n" |
| 6 | R"==(* Copyright 2019-2021 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_eltwise.h" )==" "\n" |
| 21 | R"==(#include "gpu/ocl/ocl_post_ops.h" )==" "\n" |
| 22 | R"==(#include "gpu/ocl/ocl_types.h" )==" "\n" |
| 23 | R"==(#define DATA_OFF(x0, x1, x2, x3, x4, x5) OFF_MD(DATA, x0, x1, x2, x3, x4, x5) )==" "\n" |
| 24 | R"==(#define DIFF_DATA_OFF(x0, x1, x2, x3, x4, x5) \ )==" "\n" |
| 25 | R"==(OFF_MD(DIFF_DATA, x0, x1, x2, x3, x4, x5) )==" "\n" |
| 26 | R"==(#define KERNEL_ATTR __attribute__((intel_reqd_sub_group_size(SUB_GROUP_SIZE))) )==" "\n" |
| 27 | R"==(#if IS_FWD )==" "\n" |
| 28 | R"==(KERNEL_ATTR )==" "\n" |
| 29 | R"==(__kernel void ref_eltwise_fwd(__global DATA_T *src, __global DATA_T *dst, )==" "\n" |
| 30 | R"==(float alpha, float beta POST_OP_ARGS) { )==" "\n" |
| 31 | R"==(#if USE_GWS_GET )==" "\n" |
| 32 | R"==(int d0 = GWS_GET_D0(); )==" "\n" |
| 33 | R"==(int d1 = GWS_GET_D1(); )==" "\n" |
| 34 | R"==(int d2 = GWS_GET_D2(); )==" "\n" |
| 35 | R"==(int d3 = GWS_GET_D3(); )==" "\n" |
| 36 | R"==(int d4 = GWS_GET_D4(); )==" "\n" |
| 37 | R"==(int d5 = GWS_GET_D5(); )==" "\n" |
| 38 | R"==(const size_t data_off = DATA_OFF(d0, d1, d2, d3, d4, d5); )==" "\n" |
| 39 | R"==(if (d0 >= DATA_D0 || d1 >= DATA_D1 || d2 >= DATA_D2 || d3 >= DATA_D3 )==" "\n" |
| 40 | R"==(|| d4 >= DATA_D4 || d5 >= DATA_D5) { )==" "\n" |
| 41 | R"==(dst[data_off] = CONVERT_DATA_T(0.f); )==" "\n" |
| 42 | R"==(return; )==" "\n" |
| 43 | R"==(} )==" "\n" |
| 44 | R"==(#else )==" "\n" |
| 45 | R"==(const size_t data_off = get_global_id(0) )==" "\n" |
| 46 | R"==(#if GWS1 > 1 )==" "\n" |
| 47 | R"==(+ get_global_id(1) * GWS0 )==" "\n" |
| 48 | R"==(#endif )==" "\n" |
| 49 | R"==(#if GWS2 > 1 )==" "\n" |
| 50 | R"==(+ get_global_id(2) * GWS0 * GWS1 )==" "\n" |
| 51 | R"==(#endif )==" "\n" |
| 52 | R"==(; )==" "\n" |
| 53 | R"==(const int d0 = 0; )==" "\n" |
| 54 | R"==(const int d1 = 0; )==" "\n" |
| 55 | R"==(const int d2 = 0; )==" "\n" |
| 56 | R"==(const int d3 = 0; )==" "\n" |
| 57 | R"==(const int d4 = 0; )==" "\n" |
| 58 | R"==(const int d5 = 0; )==" "\n" |
| 59 | R"==(#endif )==" "\n" |
| 60 | R"==(#if DT_F16 == 1 )==" "\n" |
| 61 | R"==(float tmp_s = CONVERT_FLOAT_T(src[data_off]); )==" "\n" |
| 62 | R"==(#else )==" "\n" |
| 63 | R"==(float tmp_s = DATA_TO_REF(src[data_off]); )==" "\n" |
| 64 | R"==(#endif )==" "\n" |
| 65 | R"==(tmp_s = fwd_eltwise(tmp_s, alpha, beta, 1.0f); )==" "\n" |
| 66 | R"==(float dst_data; )==" "\n" |
| 67 | R"==(#if WITH_SUM )==" "\n" |
| 68 | R"==(dst_data = convert_float(DATA_TO_REF(dst[data_off])); )==" "\n" |
| 69 | R"==(#endif )==" "\n" |
| 70 | R"==(APPLY_POST_OPS_SERIAL(tmp_s, float, dst_data, float, d0, 1, d1, 1, d2, 1, )==" "\n" |
| 71 | R"==(d3, 1, d4, 1, d5, 1); )==" "\n" |
| 72 | R"==(dst[data_off] = CONVERT_DATA_T(tmp_s); )==" "\n" |
| 73 | R"==(} )==" "\n" |
| 74 | R"==(#else )==" "\n" |
| 75 | R"==(#if DT_F32 == 1 || DT_BF16 == 1 )==" "\n" |
| 76 | R"==(KERNEL_ATTR )==" "\n" |
| 77 | R"==(__kernel void ref_eltwise_bwd(__global DATA_T *src, __global DATA_T *diff_src, )==" "\n" |
| 78 | R"==(__global DATA_T *diff_dst, float alpha, float beta) { )==" "\n" |
| 79 | R"==(int d0 = GWS_GET_D0(); )==" "\n" |
| 80 | R"==(int d1 = GWS_GET_D1(); )==" "\n" |
| 81 | R"==(int d2 = GWS_GET_D2(); )==" "\n" |
| 82 | R"==(int d3 = GWS_GET_D3(); )==" "\n" |
| 83 | R"==(int d4 = GWS_GET_D4(); )==" "\n" |
| 84 | R"==(int d5 = GWS_GET_D5(); )==" "\n" |
| 85 | R"==(const size_t data_off = DATA_OFF(d0, d1, d2, d3, d4, d5); )==" "\n" |
| 86 | R"==(const size_t diff_data_off = DIFF_DATA_OFF(d0, d1, d2, d3, d4, d5); )==" "\n" |
| 87 | R"==(if (d0 >= DATA_D0 || d1 >= DATA_D1 || d2 >= DATA_D2 || d3 >= DATA_D3 )==" "\n" |
| 88 | R"==(|| d4 >= DATA_D4 || d5 >= DATA_D5) { )==" "\n" |
| 89 | R"==(diff_src[diff_data_off] = CONVERT_DATA_T(0.f); )==" "\n" |
| 90 | R"==(return; )==" "\n" |
| 91 | R"==(} )==" "\n" |
| 92 | R"==(POST_OP_DATA_T tmp_dd = DATA_TO_REF(diff_dst[diff_data_off]); )==" "\n" |
| 93 | R"==(POST_OP_DATA_T tmp_s = DATA_TO_REF(src[data_off]); )==" "\n" |
| 94 | R"==(diff_src[diff_data_off] )==" "\n" |
| 95 | R"==(= CONVERT_DATA_T(bwd_eltwise(tmp_dd, tmp_s, alpha, beta)); )==" "\n" |
| 96 | R"==(} )==" "\n" |
| 97 | R"==(#endif )==" "\n" |
| 98 | R"==(#endif )==" "\n" |
| 99 | R"==()=="; |
| 100 | } |
| 101 | } |
| 102 | } |
| 103 | } |
Generated on 2022-Nov-30 from project oneDNN revision 84bc949a1
Powered by 
Code Browser 2.1
Generator usage only permitted with license.