1 | namespace dnnl { |
2 | namespace impl { |
3 | namespace gpu { |
4 | namespace ocl { |
5 | const char *xe_lp_conv_dw_fwd_data_mb_block_x8s8x_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"==(#include "gpu/ocl/ocl_post_ops.h" )==" "\n" |
22 | R"==(#include "gpu/ocl/ocl_types.h" )==" "\n" |
23 | R"==(#include "gpu/ocl/ocl_zero_points.h" )==" "\n" |
24 | R"==(#define KDHW_SIZE (KD * KH * KW) )==" "\n" |
25 | R"==(#if SCALES_PER_OC )==" "\n" |
26 | R"==(#define SCALE scales )==" "\n" |
27 | R"==(#define SCALE_VEC8 scales.s01010101 )==" "\n" |
28 | R"==(#elif SCALES_COMMON )==" "\n" |
29 | R"==(#define SCALE runtime_scales[0] )==" "\n" |
30 | R"==(#define SCALE_VEC8 runtime_scales[0] )==" "\n" |
31 | R"==(#else )==" "\n" |
32 | R"==(#define SCALE 1 )==" "\n" |
33 | R"==(#define SCALE_VEC8 1 )==" "\n" |
34 | R"==(#endif )==" "\n" |
35 | R"==(__attribute__((intel_reqd_sub_group_size(SUB_GROUP_SIZE))) )==" "\n" |
36 | R"==(__attribute__((reqd_work_group_size(LWS_0, LWS_1, LWS_2))) __kernel void )==" "\n" |
37 | R"==(conv_dw_fwd_mb_block_x8s8x(const __global uchar *src, const __global char *wei, )==" "\n" |
38 | R"==(const __global float *bias, __global DST_DATA_T *dst POST_OP_ARGS, )==" "\n" |
39 | R"==(const __global float *runtime_scales, )==" "\n" |
40 | R"==(const __global int *src_compensation, const __global int *src_zpoints, )==" "\n" |
41 | R"==(const __global int *dst_compensation) { )==" "\n" |
42 | R"==(const int osp = get_global_id(1); )==" "\n" |
43 | R"==(const int ocl_local_id = get_local_id(0); )==" "\n" |
44 | R"==(const int od = osp / (OW * OH); )==" "\n" |
45 | R"==(const int ohw = osp % (OW * OH); )==" "\n" |
46 | R"==(const int ow = (ohw % OW); )==" "\n" |
47 | R"==(const int oh = ohw / OW; )==" "\n" |
48 | R"==(const int g = get_group_id(0) * OC_BLOCK; )==" "\n" |
49 | R"==(const int mb = (get_global_id(2) / 4) * MB_BLOCK; )==" "\n" |
50 | R"==(const int mb_half = (get_global_id(2) % 4) * MB_BLOCK / 4; )==" "\n" |
51 | R"==(const int id = od * SD - PD; )==" "\n" |
52 | R"==(const int ih = oh * SH - PH; )==" "\n" |
53 | R"==(const int iw = ow * SW - PW; )==" "\n" |
54 | R"==(dst += mb * G_PADDED * OD * OH * OW + g * OD * OH * OW * MB_BLOCK )==" "\n" |
55 | R"==(+ (od * OH * OW + oh * OW + ow) * MB_BLOCK * OC_BLOCK )==" "\n" |
56 | R"==(+ mb_half * OC_BLOCK; )==" "\n" |
57 | R"==(src += mb * G_PADDED * ID * IH * IW + g * ID * IH * IW * MB_BLOCK )==" "\n" |
58 | R"==(+ (id * IH * IW + ih * IW + iw) * MB_BLOCK * IC_BLOCK )==" "\n" |
59 | R"==(+ mb_half * IC_BLOCK; )==" "\n" |
60 | R"==(wei += g * KDHW_SIZE; )==" "\n" |
61 | R"==(int8 S00 = 0; )==" "\n" |
62 | R"==(int8 S01 = 0; )==" "\n" |
63 | R"==(uchar16 A00, A10, A20, A30; )==" "\n" |
64 | R"==(#if WITH_SRC_ZPOINTS )==" "\n" |
65 | R"==(#if WITH_SRC_ZPOINTS_PER_IC )==" "\n" |
66 | R"==(const int2 z = read_src_zero_points_32g(src_zpoints, g); )==" "\n" |
67 | R"==(#else )==" "\n" |
68 | R"==(const int2 z = read_src_zero_point(src_zpoints); )==" "\n" |
69 | R"==(#endif )==" "\n" |
70 | R"==(#endif )==" "\n" |
71 | R"==(unroll_for(int sp = 0; sp < KDHW_SIZE - KDHW_SIZE % 4; sp += 4) { )==" "\n" |
72 | R"==(const int4 s = {sp, sp + 1, sp + 2, sp + 3}; )==" "\n" |
73 | R"==(const int4 kd = s / (KH * KW); )==" "\n" |
74 | R"==(const int4 kh = (s % (KH * KW)) / KW; )==" "\n" |
75 | R"==(const int4 kw = (s % (KH * KW)) % KW; )==" "\n" |
76 | R"==(const int4 src_index )==" "\n" |
77 | R"==(= (kd * (1 + DD) * IH * IW + kh * (1 + DH) * IW + kw * (1 + DW)) )==" "\n" |
78 | R"==(* MB_BLOCK * IC_BLOCK; )==" "\n" |
79 | R"==(const int4 index = id + kd * (1 + DD) < 0 || id + kd * (1 + DD) >= ID )==" "\n" |
80 | R"==(|| ih + kh * (1 + DH) < 0 || ih + kh * (1 + DH) >= IH )==" "\n" |
81 | R"==(|| iw + kw * (1 + DW) < 0 || iw + kw * (1 + DW) >= IW; )==" "\n" |
82 | R"==(if (index.s0) { )==" "\n" |
83 | R"==(A00 = 0; )==" "\n" |
84 | R"==(} else { )==" "\n" |
85 | R"==(A00 = (intel_sub_group_block_read_uc16( )==" "\n" |
86 | R"==((const __global uchar *)(&src[src_index.s0]))); )==" "\n" |
87 | R"==(} )==" "\n" |
88 | R"==(if (index.s1) { )==" "\n" |
89 | R"==(A10 = 0; )==" "\n" |
90 | R"==(} else { )==" "\n" |
91 | R"==(A10 = (intel_sub_group_block_read_uc16( )==" "\n" |
92 | R"==((const __global uchar *)(&src[src_index.s1]))); )==" "\n" |
93 | R"==(} )==" "\n" |
94 | R"==(if (index.s2) { )==" "\n" |
95 | R"==(A20 = 0; )==" "\n" |
96 | R"==(} else { )==" "\n" |
97 | R"==(A20 = (intel_sub_group_block_read_uc16( )==" "\n" |
98 | R"==((const __global uchar *)(&src[src_index.s2]))); )==" "\n" |
99 | R"==(} )==" "\n" |
100 | R"==(if (index.s3) { )==" "\n" |
101 | R"==(A30 = 0; )==" "\n" |
102 | R"==(} else { )==" "\n" |
103 | R"==(A30 = (intel_sub_group_block_read_uc16( )==" "\n" |
104 | R"==((const __global uchar *)(&src[src_index.s3]))); )==" "\n" |
105 | R"==(} )==" "\n" |
106 | R"==(char8 W = as_char8( )==" "\n" |
107 | R"==(intel_sub_group_block_read_uc8((const __global uchar *)(wei))); )==" "\n" |
108 | R"==(char4 W1 = W.s0246; )==" "\n" |
109 | R"==(char4 W2 = W.s1357; )==" "\n" |
110 | R"==(#if WITH_SRC_ZPOINTS )==" "\n" |
111 | R"==(if (index.s0) { )==" "\n" |
112 | R"==(int2 src_comp; )==" "\n" |
113 | R"==(src_comp.s0 = z.s0 * W.s0; )==" "\n" |
114 | R"==(src_comp.s1 = z.s1 * W.s1; )==" "\n" |
115 | R"==(S00 += src_comp.s01010101; )==" "\n" |
116 | R"==(S01 += src_comp.s01010101; )==" "\n" |
117 | R"==(} )==" "\n" |
118 | R"==(if (index.s1) { )==" "\n" |
119 | R"==(int2 src_comp; )==" "\n" |
120 | R"==(src_comp.s0 = z.s0 * W.s2; )==" "\n" |
121 | R"==(src_comp.s1 = z.s1 * W.s3; )==" "\n" |
122 | R"==(S00 += src_comp.s01010101; )==" "\n" |
123 | R"==(S01 += src_comp.s01010101; )==" "\n" |
124 | R"==(} )==" "\n" |
125 | R"==(if (index.s2) { )==" "\n" |
126 | R"==(int2 src_comp; )==" "\n" |
127 | R"==(src_comp.s0 = z.s0 * W.s4; )==" "\n" |
128 | R"==(src_comp.s1 = z.s1 * W.s5; )==" "\n" |
129 | R"==(S00 += src_comp.s01010101; )==" "\n" |
130 | R"==(S01 += src_comp.s01010101; )==" "\n" |
131 | R"==(} )==" "\n" |
132 | R"==(if (index.s3) { )==" "\n" |
133 | R"==(int2 src_comp; )==" "\n" |
134 | R"==(src_comp.s0 = z.s0 * W.s6; )==" "\n" |
135 | R"==(src_comp.s1 = z.s1 * W.s7; )==" "\n" |
136 | R"==(S00 += src_comp.s01010101; )==" "\n" |
137 | R"==(S01 += src_comp.s01010101; )==" "\n" |
138 | R"==(} )==" "\n" |
139 | R"==(#endif )==" "\n" |
140 | R"==(S00.s0 = idot4( )==" "\n" |
141 | R"==((SRC_DATA4_T)(A00.s0, A10.s0, A20.s0, A30.s0), W1, S00.s0); )==" "\n" |
142 | R"==(S00.s1 = idot4( )==" "\n" |
143 | R"==((SRC_DATA4_T)(A00.s1, A10.s1, A20.s1, A30.s1), W2, S00.s1); )==" "\n" |
144 | R"==(S00.s2 = idot4( )==" "\n" |
145 | R"==((SRC_DATA4_T)(A00.s2, A10.s2, A20.s2, A30.s2), W1, S00.s2); )==" "\n" |
146 | R"==(S00.s3 = idot4( )==" "\n" |
147 | R"==((SRC_DATA4_T)(A00.s3, A10.s3, A20.s3, A30.s3), W2, S00.s3); )==" "\n" |
148 | R"==(S00.s4 = idot4( )==" "\n" |
149 | R"==((SRC_DATA4_T)(A00.s4, A10.s4, A20.s4, A30.s4), W1, S00.s4); )==" "\n" |
150 | R"==(S00.s5 = idot4( )==" "\n" |
151 | R"==((SRC_DATA4_T)(A00.s5, A10.s5, A20.s5, A30.s5), W2, S00.s5); )==" "\n" |
152 | R"==(S00.s6 = idot4( )==" "\n" |
153 | R"==((SRC_DATA4_T)(A00.s6, A10.s6, A20.s6, A30.s6), W1, S00.s6); )==" "\n" |
154 | R"==(S00.s7 = idot4( )==" "\n" |
155 | R"==((SRC_DATA4_T)(A00.s7, A10.s7, A20.s7, A30.s7), W2, S00.s7); )==" "\n" |
156 | R"==(S01.s0 = idot4( )==" "\n" |
157 | R"==((SRC_DATA4_T)(A00.s8, A10.s8, A20.s8, A30.s8), W1, S01.s0); )==" "\n" |
158 | R"==(S01.s1 = idot4( )==" "\n" |
159 | R"==((SRC_DATA4_T)(A00.s9, A10.s9, A20.s9, A30.s9), W2, S01.s1); )==" "\n" |
160 | R"==(S01.s2 = idot4( )==" "\n" |
161 | R"==((SRC_DATA4_T)(A00.sa, A10.sa, A20.sa, A30.sa), W1, S01.s2); )==" "\n" |
162 | R"==(S01.s3 = idot4( )==" "\n" |
163 | R"==((SRC_DATA4_T)(A00.sb, A10.sb, A20.sb, A30.sb), W2, S01.s3); )==" "\n" |
164 | R"==(S01.s4 = idot4( )==" "\n" |
165 | R"==((SRC_DATA4_T)(A00.sc, A10.sc, A20.sc, A30.sc), W1, S01.s4); )==" "\n" |
166 | R"==(S01.s5 = idot4( )==" "\n" |
167 | R"==((SRC_DATA4_T)(A00.sd, A10.sd, A20.sd, A30.sd), W2, S01.s5); )==" "\n" |
168 | R"==(S01.s6 = idot4( )==" "\n" |
169 | R"==((SRC_DATA4_T)(A00.se, A10.se, A20.se, A30.se), W1, S01.s6); )==" "\n" |
170 | R"==(S01.s7 = idot4( )==" "\n" |
171 | R"==((SRC_DATA4_T)(A00.sf, A10.sf, A20.sf, A30.sf), W2, S01.s7); )==" "\n" |
172 | R"==(wei += 4 * OC_BLOCK; )==" "\n" |
173 | R"==(} )==" "\n" |
174 | R"==(unroll_for(int sp = KDHW_SIZE - KDHW_SIZE % 4; sp < KDHW_SIZE; sp++) { )==" "\n" |
175 | R"==(const int kd = sp / (KH * KW); )==" "\n" |
176 | R"==(const int kh = (sp % (KH * KW)) / KW; )==" "\n" |
177 | R"==(const int kw = (sp % (KH * KW)) % KW; )==" "\n" |
178 | R"==(const int src_index )==" "\n" |
179 | R"==(= (kd * (1 + DD) * IH * IW + kh * (1 + DH) * IW + kw * (1 + DW)) )==" "\n" |
180 | R"==(* MB_BLOCK * IC_BLOCK; )==" "\n" |
181 | R"==(const int index = id + kd * (1 + DD) < 0 || id + kd * (1 + DD) >= ID )==" "\n" |
182 | R"==(|| ih + kh * (1 + DH) < 0 || ih + kh * (1 + DH) >= IH )==" "\n" |
183 | R"==(|| iw + kw * (1 + DW) < 0 || iw + kw * (1 + DW) >= IW; )==" "\n" |
184 | R"==(if (index) { )==" "\n" |
185 | R"==(A00 = 0; )==" "\n" |
186 | R"==(} else { )==" "\n" |
187 | R"==(A00 = (intel_sub_group_block_read_uc16( )==" "\n" |
188 | R"==((const __global uchar *)(&src[src_index]))); )==" "\n" |
189 | R"==(} )==" "\n" |
190 | R"==(const char2 W = as_char2( )==" "\n" |
191 | R"==(intel_sub_group_block_read_uc2((const __global uchar *)(wei))); )==" "\n" |
192 | R"==(const char W1 = W.s0; )==" "\n" |
193 | R"==(const char W2 = W.s1; )==" "\n" |
194 | R"==(#if WITH_SRC_ZPOINTS )==" "\n" |
195 | R"==(if (index) { )==" "\n" |
196 | R"==(int2 src_comp; )==" "\n" |
197 | R"==(src_comp.s0 = z.s0 * W.s0; )==" "\n" |
198 | R"==(src_comp.s1 = z.s1 * W.s1; )==" "\n" |
199 | R"==(S00 += src_comp.s01010101; )==" "\n" |
200 | R"==(S01 += src_comp.s01010101; )==" "\n" |
201 | R"==(} )==" "\n" |
202 | R"==(#endif )==" "\n" |
203 | R"==(S00.s0 += ((SRC_DATA_T)A00.s0) * W1; )==" "\n" |
204 | R"==(S00.s1 += ((SRC_DATA_T)A00.s1) * W2; )==" "\n" |
205 | R"==(S00.s2 += ((SRC_DATA_T)A00.s2) * W1; )==" "\n" |
206 | R"==(S00.s3 += ((SRC_DATA_T)A00.s3) * W2; )==" "\n" |
207 | R"==(S00.s4 += ((SRC_DATA_T)A00.s4) * W1; )==" "\n" |
208 | R"==(S00.s5 += ((SRC_DATA_T)A00.s5) * W2; )==" "\n" |
209 | R"==(S00.s6 += ((SRC_DATA_T)A00.s6) * W1; )==" "\n" |
210 | R"==(S00.s7 += ((SRC_DATA_T)A00.s7) * W2; )==" "\n" |
211 | R"==(S01.s0 += ((SRC_DATA_T)A00.s8) * W1; )==" "\n" |
212 | R"==(S01.s1 += ((SRC_DATA_T)A00.s9) * W2; )==" "\n" |
213 | R"==(S01.s2 += ((SRC_DATA_T)A00.sa) * W1; )==" "\n" |
214 | R"==(S01.s3 += ((SRC_DATA_T)A00.sb) * W2; )==" "\n" |
215 | R"==(S01.s4 += ((SRC_DATA_T)A00.sc) * W1; )==" "\n" |
216 | R"==(S01.s5 += ((SRC_DATA_T)A00.sd) * W2; )==" "\n" |
217 | R"==(S01.s6 += ((SRC_DATA_T)A00.se) * W1; )==" "\n" |
218 | R"==(S01.s7 += ((SRC_DATA_T)A00.sf) * W2; )==" "\n" |
219 | R"==(wei += OC_BLOCK; )==" "\n" |
220 | R"==(} )==" "\n" |
221 | R"==(#if WITH_SRC_ZPOINTS )==" "\n" |
222 | R"==(const int2 src_comp = as_int2(intel_sub_group_block_read2( )==" "\n" |
223 | R"==((__global uint *)(&src_compensation[g]))); )==" "\n" |
224 | R"==(S00 -= src_comp.s01010101; )==" "\n" |
225 | R"==(S01 -= src_comp.s01010101; )==" "\n" |
226 | R"==(#endif )==" "\n" |
227 | R"==(float8 tmp00 = convert_float8(S00); )==" "\n" |
228 | R"==(float8 tmp01 = convert_float8(S01); )==" "\n" |
229 | R"==(#if SCALES_PER_OC )==" "\n" |
230 | R"==(float2 scales = 1; )==" "\n" |
231 | R"==(if (g < G - 2) { )==" "\n" |
232 | R"==(scales = as_float2(intel_sub_group_block_read2( )==" "\n" |
233 | R"==((const __global uint *)&runtime_scales[g])); )==" "\n" |
234 | R"==(} else { )==" "\n" |
235 | R"==(scales.s1 = as_float(intel_sub_group_block_read( )==" "\n" |
236 | R"==((const __global uint *)&runtime_scales[g])); )==" "\n" |
237 | R"==(} )==" "\n" |
238 | R"==(#endif )==" "\n" |
239 | R"==(#if WITH_BIAS )==" "\n" |
240 | R"==(#if OC % OC_BLOCK == 0 )==" "\n" |
241 | R"==(float2 B = as_float2( )==" "\n" |
242 | R"==(intel_sub_group_block_read2((const __global uint *)&bias[g])); )==" "\n" |
243 | R"==(#else )==" "\n" |
244 | R"==(float2 B = 0; )==" "\n" |
245 | R"==(int i = 0; )==" "\n" |
246 | R"==(for (i = 0; i < (G - g) / SUB_GROUP_SIZE && i < 2; ++i) { )==" "\n" |
247 | R"==(B[i] = as_float(intel_sub_group_block_read( )==" "\n" |
248 | R"==((const __global uint *)&bias[g + (i * SUB_GROUP_SIZE)])); )==" "\n" |
249 | R"==(} )==" "\n" |
250 | R"==(if (i < 2 && get_sub_group_local_id() < (G - g) % SUB_GROUP_SIZE) { )==" "\n" |
251 | R"==(B[i] = bias[g + (i * SUB_GROUP_SIZE) + get_sub_group_local_id()]; )==" "\n" |
252 | R"==(} )==" "\n" |
253 | R"==(#endif )==" "\n" |
254 | R"==(float8 B0123 = B.s01010101; )==" "\n" |
255 | R"==(float8 B4567 = B.s01010101; )==" "\n" |
256 | R"==(if (MB % MB_BLOCK != 0) { )==" "\n" |
257 | R"==(int8 mb_vec = mb + mb_half; )==" "\n" |
258 | R"==(int8 m0123 = (mb_vec + (int8)(0, 0, 1, 1, 2, 2, 3, 3)) < MB; )==" "\n" |
259 | R"==(int8 m4567 = (mb_vec + (int8)(4, 4, 5, 5, 6, 6, 7, 7)) < MB; )==" "\n" |
260 | R"==(B0123 = select((float8)0, B0123, m0123); )==" "\n" |
261 | R"==(B4567 = select((float8)0, B4567, m4567); )==" "\n" |
262 | R"==(} )==" "\n" |
263 | R"==(tmp00 = SCALE_VEC8 * fma(tmp00, (float8)1, B0123); )==" "\n" |
264 | R"==(tmp01 = SCALE_VEC8 * fma(tmp01, (float8)1, B4567); )==" "\n" |
265 | R"==(#else )==" "\n" |
266 | R"==(tmp00 *= SCALE_VEC8; )==" "\n" |
267 | R"==(tmp01 *= SCALE_VEC8; )==" "\n" |
268 | R"==(#endif )==" "\n" |
269 | R"==(SUM_DATA16_T D00; )==" "\n" |
270 | R"==(#if WITH_SUM )==" "\n" |
271 | R"==(D00 = AS_SUM_DATA16_T(BLOCK_READ_DST16(dst)); )==" "\n" |
272 | R"==(#endif )==" "\n" |
273 | R"==(float16 tmp_x16 = (float16)(tmp00, tmp01); )==" "\n" |
274 | R"==(const int po_mb = mb + mb_half; )==" "\n" |
275 | R"==(const int po_oc = g * OC; )==" "\n" |
276 | R"==(APPLY_POST_OPS_TRY_BURST(tmp_x16, float, D00, SUM_DATA_T, po_mb, 8, po_oc, )==" "\n" |
277 | R"==(2 * SUB_GROUP_SIZE, ocl_local_id); )==" "\n" |
278 | R"==(#if WITH_DST_ZPOINTS )==" "\n" |
279 | R"==(float2 dst_zp )==" "\n" |
280 | R"==(= convert_float2(read_dst_zero_points_32g(dst_compensation, g)); )==" "\n" |
281 | R"==(float8 tmp_zp = dst_zp.s01010101; )==" "\n" |
282 | R"==(tmp_x16 += (float16)(tmp_zp, tmp_zp); )==" "\n" |
283 | R"==(#endif )==" "\n" |
284 | R"==(DST_DATA16_T R0 = CONVERT_DST_DATA16_T(tmp_x16); )==" "\n" |
285 | R"==(BLOCK_WRITE_DST16(dst, R0); )==" "\n" |
286 | R"==(} )==" "\n" |
287 | R"==()==" ; |
288 | } |
289 | } |
290 | } |
291 | } |