1 | namespace dnnl { |
2 | namespace impl { |
3 | namespace gpu { |
4 | namespace ocl { |
5 | const char *gen9_softmax_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_types.h" )==" "\n" |
21 | R"==(#define LOAD_FLOAT8(prefix, ptr) \ )==" "\n" |
22 | R"==(DATA_TO_FLOAT8(prefix, \ )==" "\n" |
23 | R"==(BLOCK_TO_DATA8(prefix, \ )==" "\n" |
24 | R"==(READ_BLOCK8(prefix, \ )==" "\n" |
25 | R"==((__global BLOCK_T(ALIAS(prefix)) *)(ptr)))) )==" "\n" |
26 | R"==(#define STORE_FLOAT8(prefix, ptr, val) \ )==" "\n" |
27 | R"==(WRITE_BLOCK8(prefix, (__global BLOCK_T(ALIAS(prefix)) *)(ptr), \ )==" "\n" |
28 | R"==(DATA_TO_BLOCK8(prefix, FLOAT_TO_DATA8(prefix, val))) )==" "\n" |
29 | R"==(#define VECT_SIZE 8 )==" "\n" |
30 | R"==(#define NUM_BUF (SOFTMAX_AXIS_SIZE / SUB_GROUP_SIZE / VECT_SIZE) )==" "\n" |
31 | R"==(#if IS_FWD )==" "\n" |
32 | R"==(__attribute__((reqd_work_group_size(GROUP_SIZE, 1, 1))) )==" "\n" |
33 | R"==(__attribute__((intel_reqd_sub_group_size(SUB_GROUP_SIZE))) __kernel void )==" "\n" |
34 | R"==(gen9_softmax_fwd(__global SRC_DATA_T *src, __global DST_DATA_T *dst, )==" "\n" |
35 | R"==(__global float *src_scale, __global float *dst_scale) { )==" "\n" |
36 | R"==(float scale = 1.0f; )==" "\n" |
37 | R"==(#if WITH_SRC_SCALES )==" "\n" |
38 | R"==(scale *= src_scale[0]; )==" "\n" |
39 | R"==(#endif )==" "\n" |
40 | R"==(#if WITH_DST_SCALES )==" "\n" |
41 | R"==(scale /= dst_scale[0]; )==" "\n" |
42 | R"==(#endif )==" "\n" |
43 | R"==(#if IS_NHWC || IS_BLOCKED )==" "\n" |
44 | R"==(const int group = get_global_id(0) / GROUP_SIZE; )==" "\n" |
45 | R"==(const int mb = group / OC_PADDED; )==" "\n" |
46 | R"==(const int local_oc = group % OC_PADDED; )==" "\n" |
47 | R"==(const int local_id = get_local_id(0); )==" "\n" |
48 | R"==(const int axis_chunk = (local_id / SUB_GROUP_SIZE) * SOFTMAX_BUF; )==" "\n" |
49 | R"==(const int subgroup_id = get_sub_group_local_id(); )==" "\n" |
50 | R"==(const int oc_chunk = OC * VECT_SIZE * subgroup_id; )==" "\n" |
51 | R"==(#if IS_BLOCKED )==" "\n" |
52 | R"==(const int oc_block_id = local_oc / OC; )==" "\n" |
53 | R"==(const int oc_in_block = local_oc % OC; )==" "\n" |
54 | R"==(int data_off = (MB * oc_block_id + mb) * OC * SOFTMAX_AXIS_SIZE + oc_chunk )==" "\n" |
55 | R"==(+ oc_in_block; )==" "\n" |
56 | R"==(#else )==" "\n" |
57 | R"==(int data_off = mb * OC_PADDED * SOFTMAX_AXIS_SIZE + oc_chunk + local_oc; )==" "\n" |
58 | R"==(#endif )==" "\n" |
59 | R"==(float d[VECT_SIZE]; )==" "\n" |
60 | R"==(float max_ = -FLT_MAX; )==" "\n" |
61 | R"==(float denom_ = 0.f; )==" "\n" |
62 | R"==(src += data_off; )==" "\n" |
63 | R"==(for (int k = 0, axis_channel_id = OC * axis_chunk; k < VECT_SIZE; )==" "\n" |
64 | R"==(++k, axis_channel_id += OC) { )==" "\n" |
65 | R"==(d[k] = DATA_TO_FLOAT(SRC, src[axis_channel_id]); )==" "\n" |
66 | R"==(max_ = max(d[k], max_); )==" "\n" |
67 | R"==(} )==" "\n" |
68 | R"==(#if GROUP_SIZE == SUB_GROUP_SIZE )==" "\n" |
69 | R"==(max_ = sub_group_reduce_max(max_); )==" "\n" |
70 | R"==(#else )==" "\n" |
71 | R"==(max_ = work_group_reduce_max(max_); )==" "\n" |
72 | R"==(#endif )==" "\n" |
73 | R"==(for (int k = 0; k < VECT_SIZE; ++k) { )==" "\n" |
74 | R"==(#if LOGSOFTMAX )==" "\n" |
75 | R"==(denom_ += exp(d[k] - max_); )==" "\n" |
76 | R"==(#else )==" "\n" |
77 | R"==(d[k] = exp(d[k] - max_); )==" "\n" |
78 | R"==(denom_ += d[k]; )==" "\n" |
79 | R"==(#endif )==" "\n" |
80 | R"==(} )==" "\n" |
81 | R"==(#if GROUP_SIZE == SUB_GROUP_SIZE )==" "\n" |
82 | R"==(denom_ = sub_group_reduce_add(denom_); )==" "\n" |
83 | R"==(#else )==" "\n" |
84 | R"==(denom_ = work_group_reduce_add(denom_); )==" "\n" |
85 | R"==(#endif )==" "\n" |
86 | R"==(#if LOGSOFTMAX )==" "\n" |
87 | R"==(denom_ = log(denom_); )==" "\n" |
88 | R"==(#else )==" "\n" |
89 | R"==(denom_ = 1.0 / denom_; )==" "\n" |
90 | R"==(#endif )==" "\n" |
91 | R"==(dst += data_off; )==" "\n" |
92 | R"==(for (int k = 0, axis_channel_id = OC * axis_chunk; k < VECT_SIZE; )==" "\n" |
93 | R"==(++k, axis_channel_id += OC) { )==" "\n" |
94 | R"==(#if LOGSOFTMAX )==" "\n" |
95 | R"==(d[k] = d[k] - max_ - denom_; )==" "\n" |
96 | R"==(#else )==" "\n" |
97 | R"==(d[k] = d[k] * denom_; )==" "\n" |
98 | R"==(#endif )==" "\n" |
99 | R"==(dst[axis_channel_id] = FLOAT_TO_DATA(DST, d[k] * scale); )==" "\n" |
100 | R"==(} )==" "\n" |
101 | R"==(#else )==" "\n" |
102 | R"==(const int data_off = (get_global_id(0) / GROUP_SIZE) * SOFTMAX_AXIS_SIZE; )==" "\n" |
103 | R"==(float8 d[NUM_BUF]; )==" "\n" |
104 | R"==(float max_ = -FLT_MAX; )==" "\n" |
105 | R"==(float denom_ = 0.f; )==" "\n" |
106 | R"==(src += data_off; )==" "\n" |
107 | R"==(for (int k = 0; k < NUM_BUF; ++k) { )==" "\n" |
108 | R"==(d[k] = LOAD_FLOAT8(SRC, &src[k * VECT_SIZE * SUB_GROUP_SIZE]); )==" "\n" |
109 | R"==(for (int i = 0; i < VECT_SIZE; ++i) { )==" "\n" |
110 | R"==(max_ = max(d[k][i], max_); )==" "\n" |
111 | R"==(} )==" "\n" |
112 | R"==(} )==" "\n" |
113 | R"==(max_ = sub_group_reduce_max(max_); )==" "\n" |
114 | R"==(for (int k = 0; k < NUM_BUF; ++k) { )==" "\n" |
115 | R"==(#if LOGSOFTMAX )==" "\n" |
116 | R"==(for (int i = 0; i < VECT_SIZE; ++i) )==" "\n" |
117 | R"==(denom_ += exp(d[k][i] - max_); )==" "\n" |
118 | R"==(#else )==" "\n" |
119 | R"==(d[k] = exp(d[k] - max_); )==" "\n" |
120 | R"==(for (int i = 0; i < VECT_SIZE; ++i) )==" "\n" |
121 | R"==(denom_ += d[k][i]; )==" "\n" |
122 | R"==(#endif )==" "\n" |
123 | R"==(} )==" "\n" |
124 | R"==(denom_ = sub_group_reduce_add(denom_); )==" "\n" |
125 | R"==(#if LOGSOFTMAX )==" "\n" |
126 | R"==(denom_ = log(denom_); )==" "\n" |
127 | R"==(#else )==" "\n" |
128 | R"==(denom_ = 1.0 / denom_; )==" "\n" |
129 | R"==(#endif )==" "\n" |
130 | R"==(dst += data_off; )==" "\n" |
131 | R"==(for (int k = 0; k < NUM_BUF; ++k) { )==" "\n" |
132 | R"==(#if LOGSOFTMAX )==" "\n" |
133 | R"==(d[k] = d[k] - max_ - denom_; )==" "\n" |
134 | R"==(#else )==" "\n" |
135 | R"==(d[k] = d[k] * denom_; )==" "\n" |
136 | R"==(#endif )==" "\n" |
137 | R"==(STORE_FLOAT8(DST, &dst[k * VECT_SIZE * SUB_GROUP_SIZE], scale * d[k]); )==" "\n" |
138 | R"==(} )==" "\n" |
139 | R"==(#endif )==" "\n" |
140 | R"==(} )==" "\n" |
141 | R"==(#endif )==" "\n" |
142 | R"==(#if IS_BWD )==" "\n" |
143 | R"==(__attribute__((reqd_work_group_size(GROUP_SIZE, 1, 1))) )==" "\n" |
144 | R"==(__attribute__((intel_reqd_sub_group_size(SUB_GROUP_SIZE))) __kernel void )==" "\n" |
145 | R"==(gen9_softmax_bwd(__global DST_DATA_T *dst, __global SRC_DATA_T *diff_src, )==" "\n" |
146 | R"==(__global DST_DATA_T *diff_dst) { )==" "\n" |
147 | R"==(#if IS_NHWC || IS_16C )==" "\n" |
148 | R"==(const int groups = get_global_id(0) / GROUP_SIZE; )==" "\n" |
149 | R"==(const int batch = groups / IC_PADDED; )==" "\n" |
150 | R"==(const int ic = groups % IC_PADDED; )==" "\n" |
151 | R"==(const int sub_grp_id = get_sub_group_local_id(); )==" "\n" |
152 | R"==(const int local_id = get_local_id(0); )==" "\n" |
153 | R"==(const int slice = local_id / SUB_GROUP_SIZE; )==" "\n" |
154 | R"==(const int ic_buff = IC * VECT_SIZE; )==" "\n" |
155 | R"==(#if IS_16C )==" "\n" |
156 | R"==(const int ic_blk = ic / IC; )==" "\n" |
157 | R"==(const int ic_in_blk = ic % IC; )==" "\n" |
158 | R"==(int data_off = BATCH * IC * SOFTMAX_AXIS_SIZE * ic_blk )==" "\n" |
159 | R"==(+ batch * IC * SOFTMAX_AXIS_SIZE + ic_buff * sub_grp_id + ic_in_blk; )==" "\n" |
160 | R"==(#else )==" "\n" |
161 | R"==(int data_off = batch * IC * SOFTMAX_AXIS_SIZE + ic_buff * sub_grp_id + ic; )==" "\n" |
162 | R"==(#endif )==" "\n" |
163 | R"==(float sbr = 0.f; )==" "\n" |
164 | R"==(float diff_d[VECT_SIZE]; )==" "\n" |
165 | R"==(float dst_[VECT_SIZE]; )==" "\n" |
166 | R"==(diff_dst += data_off; )==" "\n" |
167 | R"==(dst += data_off; )==" "\n" |
168 | R"==(for (int i = 0, idx = IC * slice * SOFTMAX_BUF; i < VECT_SIZE; )==" "\n" |
169 | R"==(++i, idx += IC) { )==" "\n" |
170 | R"==(diff_d[i] = DATA_TO_FLOAT(DST, diff_dst[idx]); )==" "\n" |
171 | R"==(dst_[i] = DATA_TO_FLOAT(DST, dst[idx]); )==" "\n" |
172 | R"==(#if LOGSOFTMAX )==" "\n" |
173 | R"==(sbr += diff_d[i]; )==" "\n" |
174 | R"==(#else )==" "\n" |
175 | R"==(sbr += dst_[i] * diff_d[i]; )==" "\n" |
176 | R"==(#endif )==" "\n" |
177 | R"==(} )==" "\n" |
178 | R"==(#if GROUP_SIZE == SUB_GROUP_SIZE )==" "\n" |
179 | R"==(sbr = sub_group_reduce_add(sbr); )==" "\n" |
180 | R"==(#else )==" "\n" |
181 | R"==(sbr = work_group_reduce_add(sbr); )==" "\n" |
182 | R"==(#endif )==" "\n" |
183 | R"==(diff_src += data_off; )==" "\n" |
184 | R"==(for (int i = 0, idx = IC * slice * SOFTMAX_BUF; i < VECT_SIZE; )==" "\n" |
185 | R"==(++i, idx += IC) { )==" "\n" |
186 | R"==(#if LOGSOFTMAX )==" "\n" |
187 | R"==(diff_d[i] = diff_d[i] - exp(dst_[i]) * sbr; )==" "\n" |
188 | R"==(#else )==" "\n" |
189 | R"==(diff_d[i] = (diff_d[i] - sbr) * dst_[i]; )==" "\n" |
190 | R"==(#endif )==" "\n" |
191 | R"==(diff_src[idx] = FLOAT_TO_DATA(SRC, diff_d[i]); )==" "\n" |
192 | R"==(} )==" "\n" |
193 | R"==(#else )==" "\n" |
194 | R"==(const int data_off = (get_global_id(0) / GROUP_SIZE) * SOFTMAX_AXIS_SIZE; )==" "\n" |
195 | R"==(float sbr = 0.f; )==" "\n" |
196 | R"==(float8 diff_d[NUM_BUF]; )==" "\n" |
197 | R"==(float8 dst_[NUM_BUF]; )==" "\n" |
198 | R"==(diff_dst += data_off; )==" "\n" |
199 | R"==(dst += data_off; )==" "\n" |
200 | R"==(for (int k = 0; k < NUM_BUF; ++k) { )==" "\n" |
201 | R"==(diff_d[k] = LOAD_FLOAT8(DST, &diff_dst[k * VECT_SIZE * SUB_GROUP_SIZE]); )==" "\n" |
202 | R"==(dst_[k] = LOAD_FLOAT8(DST, &dst[k * VECT_SIZE * SUB_GROUP_SIZE]); )==" "\n" |
203 | R"==(for (int i = 0; i < VECT_SIZE; ++i) { )==" "\n" |
204 | R"==(#if LOGSOFTMAX )==" "\n" |
205 | R"==(sbr += diff_d[k][i]; )==" "\n" |
206 | R"==(#else )==" "\n" |
207 | R"==(sbr += dst_[k][i] * diff_d[k][i]; )==" "\n" |
208 | R"==(#endif )==" "\n" |
209 | R"==(} )==" "\n" |
210 | R"==(} )==" "\n" |
211 | R"==(sbr = sub_group_reduce_add(sbr); )==" "\n" |
212 | R"==(diff_src += data_off; )==" "\n" |
213 | R"==(for (int k = 0; k < NUM_BUF; ++k) { )==" "\n" |
214 | R"==(#if LOGSOFTMAX )==" "\n" |
215 | R"==(diff_d[k] = diff_d[k] - exp(dst_[k]) * sbr; )==" "\n" |
216 | R"==(#else )==" "\n" |
217 | R"==(diff_d[k] = (diff_d[k] - sbr) * dst_[k]; )==" "\n" |
218 | R"==(#endif )==" "\n" |
219 | R"==(STORE_FLOAT8(SRC, &diff_src[k * VECT_SIZE * SUB_GROUP_SIZE], diff_d[k]); )==" "\n" |
220 | R"==(} )==" "\n" |
221 | R"==(#endif )==" "\n" |
222 | R"==(} )==" "\n" |
223 | R"==(#endif )==" "\n" |
224 | R"==()==" ; |
225 | } |
226 | } |
227 | } |
228 | } |