1 | namespace dnnl { |
2 | namespace impl { |
3 | namespace gpu { |
4 | namespace ocl { |
5 | const char *gen9_pooling_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_post_ops.h" )==" "\n" |
21 | R"==(#include "gpu/ocl/ocl_types.h" )==" "\n" |
22 | R"==(inline VECT_DATA_T read_vect_c_block(int idx, const __global DATA_T *ptr, int c, )==" "\n" |
23 | R"==(int blocks_stride, int chunks_per_block); )==" "\n" |
24 | R"==(inline VECT_INT_T read_vect_c_block_int(int idx, const __global int *ptr, int c, )==" "\n" |
25 | R"==(int blocks_stride, int chunks_per_block); )==" "\n" |
26 | R"==(inline void write_vect_c_block(int idx, __global DATA_T *ptr, int c, )==" "\n" |
27 | R"==(int blocks_stride, int chunks_per_block, VECT_DATA_T block); )==" "\n" |
28 | R"==(inline void write_vect_c_block_int(int idx, __global int *ptr, int c, )==" "\n" |
29 | R"==(int blocks_stride, int chunks_per_block, VECT_INT_T block); )==" "\n" |
30 | R"==(#if DT_BF16 || DT_F16 )==" "\n" |
31 | R"==(#define USE_FLOATS true )==" "\n" |
32 | R"==(#else )==" "\n" |
33 | R"==(#define USE_FLOATS (ALG_AVG_NP || ALG_AVG_P) )==" "\n" |
34 | R"==(#endif )==" "\n" |
35 | R"==(#if IS_FWD )==" "\n" |
36 | R"==(KERNEL_ATTR )==" "\n" |
37 | R"==(__kernel void gen9_pooling_fwd(__global DATA_T *src, __global int *ws, )==" "\n" |
38 | R"==(__global DATA_T *dst, const int batch_id POST_OP_ARGS) { )==" "\n" |
39 | R"==(const int mb0 = MB_BLOCK_SIZE * batch_id + GWS_GET_MB(); )==" "\n" |
40 | R"==(#if UNROLL_MB )==" "\n" |
41 | R"==(const int mb1 = mb0 + MB / 2; )==" "\n" |
42 | R"==(#endif )==" "\n" |
43 | R"==(const int c = GWS_GET_C(); )==" "\n" |
44 | R"==(const int od = GWS_GET_OD(); )==" "\n" |
45 | R"==(const int oh = GWS_GET_OH(); )==" "\n" |
46 | R"==(const int ow = GWS_GET_OW(); )==" "\n" |
47 | R"==(#if USE_MB_C_BLOCK )==" "\n" |
48 | R"==(const int src_stride = (SRC_SB0 > 1) ? SRC_SB0 : SRC_S0; )==" "\n" |
49 | R"==(const int dst_stride = (DST_SB0 > 1) ? DST_SB0 : DST_S0; )==" "\n" |
50 | R"==(const int src_chunks_per_c_block = CHUNKS_PER_C_BLOCK; )==" "\n" |
51 | R"==(const int dst_chunks_per_c_block = CHUNKS_PER_C_BLOCK; )==" "\n" |
52 | R"==(#elif USE_ONLY_C_BLOCK )==" "\n" |
53 | R"==(const int src_stride = (SRC_B1 > 1) ? SRC_S1 : SUB_GROUP_SIZE; )==" "\n" |
54 | R"==(const int dst_stride = (DST_B1 > 1) ? DST_S1 : SUB_GROUP_SIZE; )==" "\n" |
55 | R"==(const int src_chunks_per_c_block )==" "\n" |
56 | R"==(= (SRC_B1 > 1) ? (SRC_B1 / SUB_GROUP_SIZE) : 1; )==" "\n" |
57 | R"==(const int dst_chunks_per_c_block )==" "\n" |
58 | R"==(= (DST_B1 > 1) ? (DST_B1 / SUB_GROUP_SIZE) : 1; )==" "\n" |
59 | R"==(#endif )==" "\n" |
60 | R"==(const int ws_stride = dst_stride; )==" "\n" |
61 | R"==(const int ws_chunks_per_c_block = dst_chunks_per_c_block; )==" "\n" |
62 | R"==(if (mb0 >= SRC_D0) { )==" "\n" |
63 | R"==(VECT_DATA_T dst_zero = DATA_ZERO; )==" "\n" |
64 | R"==(VECT_INT_T ws_zero = 0; )==" "\n" |
65 | R"==(int off = DST_OFF(mb0, c, od, oh, ow); )==" "\n" |
66 | R"==(write_vect_c_block( )==" "\n" |
67 | R"==(0, &dst[off], c, dst_stride, dst_chunks_per_c_block, dst_zero); )==" "\n" |
68 | R"==(write_vect_c_block( )==" "\n" |
69 | R"==(1, &dst[off], c, dst_stride, dst_chunks_per_c_block, dst_zero); )==" "\n" |
70 | R"==(#if ALG_MAX && IS_TRAINING )==" "\n" |
71 | R"==(write_vect_c_block_int( )==" "\n" |
72 | R"==(0, &ws[off], c, ws_stride, ws_chunks_per_c_block, ws_zero); )==" "\n" |
73 | R"==(write_vect_c_block_int( )==" "\n" |
74 | R"==(1, &ws[off], c, ws_stride, ws_chunks_per_c_block, ws_zero); )==" "\n" |
75 | R"==(#endif )==" "\n" |
76 | R"==(return; )==" "\n" |
77 | R"==(} )==" "\n" |
78 | R"==(const int id = od * SD - PD; )==" "\n" |
79 | R"==(const int ih = oh * SH - PH; )==" "\n" |
80 | R"==(const int iw = ow * SW - PW; )==" "\n" |
81 | R"==(#if USE_FLOATS )==" "\n" |
82 | R"==(VECT_FLOAT_T D0 = ALG_MAX ? CONVERT_FLOAT_T(DATA_MIN) : 0.0f; )==" "\n" |
83 | R"==(VECT_FLOAT_T D1 = ALG_MAX ? CONVERT_FLOAT_T(DATA_MIN) : 0.0f; )==" "\n" |
84 | R"==(#else )==" "\n" |
85 | R"==(VECT_DATA_T D0 = ALG_MAX ? DATA_MIN : DATA_ZERO; )==" "\n" |
86 | R"==(VECT_DATA_T D1 = ALG_MAX ? DATA_MIN : DATA_ZERO; )==" "\n" |
87 | R"==(#endif )==" "\n" |
88 | R"==(VECT_INT_T WS0 = 0, WS1 = 0; )==" "\n" |
89 | R"==(for (int kd = 0; kd < KD; ++kd) { )==" "\n" |
90 | R"==(if (id + kd < 0 || id + kd >= ID) continue; )==" "\n" |
91 | R"==(for (int kh = 0; kh < KH; ++kh) { )==" "\n" |
92 | R"==(if (ih + kh < 0 || ih + kh >= IH) continue; )==" "\n" |
93 | R"==(for (int kw = 0; kw < KW; ++kw) { )==" "\n" |
94 | R"==(if (iw + kw < 0 || iw + kw >= IW) continue; )==" "\n" |
95 | R"==(int src_off0 = SRC_OFF(mb0, c, id + kd, ih + kh, iw + kw); )==" "\n" |
96 | R"==(#if UNROLL_MB )==" "\n" |
97 | R"==(int src_off1 = SRC_OFF(mb1, c, id + kd, ih + kh, iw + kw); )==" "\n" |
98 | R"==(#endif )==" "\n" |
99 | R"==(#if USE_FLOATS )==" "\n" |
100 | R"==(VECT_FLOAT_T S0 = CONVERT_VECT_FLOAT_T(read_vect_c_block(0, )==" "\n" |
101 | R"==(&src[src_off0], c, src_stride, src_chunks_per_c_block)); )==" "\n" |
102 | R"==(#if UNROLL_MB )==" "\n" |
103 | R"==(VECT_FLOAT_T S1 = CONVERT_VECT_FLOAT_T(read_vect_c_block(0, )==" "\n" |
104 | R"==(&src[src_off1], c, src_stride, src_chunks_per_c_block)); )==" "\n" |
105 | R"==(#else )==" "\n" |
106 | R"==(VECT_FLOAT_T S1 = CONVERT_VECT_FLOAT_T(read_vect_c_block(1, )==" "\n" |
107 | R"==(&src[src_off0], c, src_stride, src_chunks_per_c_block)); )==" "\n" |
108 | R"==(#endif )==" "\n" |
109 | R"==(#else )==" "\n" |
110 | R"==(VECT_DATA_T S0 = read_vect_c_block(0, &src[src_off0], c, )==" "\n" |
111 | R"==(src_stride, src_chunks_per_c_block); )==" "\n" |
112 | R"==(#if UNROLL_MB )==" "\n" |
113 | R"==(VECT_DATA_T S1 = read_vect_c_block(0, &src[src_off1], c, )==" "\n" |
114 | R"==(src_stride, src_chunks_per_c_block); )==" "\n" |
115 | R"==(#else )==" "\n" |
116 | R"==(VECT_DATA_T S1 = read_vect_c_block(1, &src[src_off0], c, )==" "\n" |
117 | R"==(src_stride, src_chunks_per_c_block); )==" "\n" |
118 | R"==(#endif )==" "\n" |
119 | R"==(#endif )==" "\n" |
120 | R"==(#if ALG_MAX )==" "\n" |
121 | R"==(#if IS_TRAINING )==" "\n" |
122 | R"==(VECT_INT_T CMP0 = isless(D0, S0); )==" "\n" |
123 | R"==(WS0 = select(WS0, kd * KH * KW + kh * KW + kw, CMP0); )==" "\n" |
124 | R"==(D0 = select(D0, S0, CMP0); )==" "\n" |
125 | R"==(VECT_INT_T CMP1 = isless(D1, S1); )==" "\n" |
126 | R"==(WS1 = select(WS1, kd * KH * KW + kh * KW + kw, CMP1); )==" "\n" |
127 | R"==(D1 = select(D1, S1, CMP1); )==" "\n" |
128 | R"==(#else )==" "\n" |
129 | R"==(D0 = max(D0, S0); )==" "\n" |
130 | R"==(D1 = max(D1, S1); )==" "\n" |
131 | R"==(#endif )==" "\n" |
132 | R"==(#else )==" "\n" |
133 | R"==(D0 += S0; )==" "\n" |
134 | R"==(D1 += S1; )==" "\n" |
135 | R"==(#endif )==" "\n" |
136 | R"==(} )==" "\n" |
137 | R"==(} )==" "\n" |
138 | R"==(} )==" "\n" |
139 | R"==(#if ALG_AVG_P )==" "\n" |
140 | R"==(D0 = D0 / (KD * KH * KW); )==" "\n" |
141 | R"==(D1 = D1 / (KD * KH * KW); )==" "\n" |
142 | R"==(#endif )==" "\n" |
143 | R"==(#if ALG_AVG_NP )==" "\n" |
144 | R"==(const int id_start = max(od * SD - PD, 0); )==" "\n" |
145 | R"==(const int ih_start = max(oh * SH - PH, 0); )==" "\n" |
146 | R"==(const int iw_start = max(ow * SW - PW, 0); )==" "\n" |
147 | R"==(const int id_end = min(od * SD - PD + KD, ID); )==" "\n" |
148 | R"==(const int ih_end = min(oh * SH - PH + KH, IH); )==" "\n" |
149 | R"==(const int iw_end = min(ow * SW - PW + KW, IW); )==" "\n" |
150 | R"==(const int num_summands )==" "\n" |
151 | R"==(= (ih_end - ih_start) * (iw_end - iw_start) * (id_end - id_start); )==" "\n" |
152 | R"==(D0 = D0 / num_summands; )==" "\n" |
153 | R"==(D1 = D1 / num_summands; )==" "\n" |
154 | R"==(#endif )==" "\n" |
155 | R"==(int dst_off0 = DST_OFF(mb0, c, od, oh, ow); )==" "\n" |
156 | R"==(#if UNROLL_MB )==" "\n" |
157 | R"==(int dst_off1 = DST_OFF(mb1, c, od, oh, ow); )==" "\n" |
158 | R"==(#endif )==" "\n" |
159 | R"==(VECT_DATA_T sum0; )==" "\n" |
160 | R"==(VECT_DATA_T sum1; )==" "\n" |
161 | R"==(#if WITH_SUM )==" "\n" |
162 | R"==(sum0 = read_vect_c_block( )==" "\n" |
163 | R"==(0, &dst[dst_off0], c, dst_stride, dst_chunks_per_c_block); )==" "\n" |
164 | R"==(#if UNROLL_MB )==" "\n" |
165 | R"==(sum1 = read_vect_c_block( )==" "\n" |
166 | R"==(0, &dst[dst_off1], c, dst_stride, dst_chunks_per_c_block); )==" "\n" |
167 | R"==(#else )==" "\n" |
168 | R"==(sum1 = read_vect_c_block( )==" "\n" |
169 | R"==(1, &dst[dst_off0], c, dst_stride, dst_chunks_per_c_block); )==" "\n" |
170 | R"==(#endif )==" "\n" |
171 | R"==(#endif )==" "\n" |
172 | R"==(const int local_id = get_sub_group_local_id(); )==" "\n" |
173 | R"==(#if VECT_DT_N == 1 )==" "\n" |
174 | R"==(const int po_mb = mb0; )==" "\n" |
175 | R"==(const int po_oc = c + local_id; )==" "\n" |
176 | R"==(if (po_oc < C_WO_PADDING) { )==" "\n" |
177 | R"==(POST_OP_DATA_T po_sum0 = DATA_TO_REF(sum0); )==" "\n" |
178 | R"==(float po_D0 = USE_FLOATS ? D0 : CONVERT_FLOAT_T(D0); )==" "\n" |
179 | R"==(APPLY_POST_OPS_SERIAL_BINARY_2D( )==" "\n" |
180 | R"==(po_D0, float, po_sum0, POST_OP_DATA_T, po_mb, 1, po_oc, 1); )==" "\n" |
181 | R"==(D0 = USE_FLOATS ? po_D0 : CONVERT_DATA_T(po_D0); )==" "\n" |
182 | R"==(POST_OP_DATA_T po_sum1 = DATA_TO_REF(sum1); )==" "\n" |
183 | R"==(float po_D1 = USE_FLOATS ? D1 : CONVERT_FLOAT_T(D1); )==" "\n" |
184 | R"==(APPLY_POST_OPS_SERIAL_BINARY_2D( )==" "\n" |
185 | R"==(po_D1, float, po_sum1, POST_OP_DATA_T, po_mb, 1, po_oc, 1); )==" "\n" |
186 | R"==(D1 = USE_FLOATS ? po_D1 : CONVERT_DATA_T(po_D1); )==" "\n" |
187 | R"==(} )==" "\n" |
188 | R"==(#else )==" "\n" |
189 | R"==(for (int idx = 0; idx < VECT_DT_N; ++idx) { )==" "\n" |
190 | R"==(#if USE_MB_C_BLOCK )==" "\n" |
191 | R"==(int c_sub_block_id = idx % CHUNKS_PER_C_BLOCK; )==" "\n" |
192 | R"==(int mb_sub_block_id = idx / CHUNKS_PER_C_BLOCK; )==" "\n" |
193 | R"==(const int po_oc = c + c_sub_block_id * SUB_GROUP_SIZE + local_id; )==" "\n" |
194 | R"==(int po_mb = (mb0 + mb_sub_block_id) % MB; )==" "\n" |
195 | R"==(#else )==" "\n" |
196 | R"==(const int po_oc = c + idx * SUB_GROUP_SIZE + local_id; )==" "\n" |
197 | R"==(int po_mb = mb0; )==" "\n" |
198 | R"==(#endif )==" "\n" |
199 | R"==(if (po_mb >= MB || po_oc >= C_WO_PADDING) continue; )==" "\n" |
200 | R"==(float d0_i = USE_FLOATS ? D0[idx] : CONVERT_FLOAT_T(D0[idx]); )==" "\n" |
201 | R"==(POST_OP_DATA_T sum0_i = DATA_TO_REF(sum0[idx]); )==" "\n" |
202 | R"==(APPLY_POST_OPS_SERIAL_BINARY_2D( )==" "\n" |
203 | R"==(d0_i, float, sum0_i, POST_OP_DATA_T, po_mb, 1, po_oc, 1); )==" "\n" |
204 | R"==(D0[idx] = USE_FLOATS ? d0_i : CONVERT_DATA_T(d0_i); )==" "\n" |
205 | R"==(float d1_i = USE_FLOATS ? D1[idx] : CONVERT_FLOAT_T(D1[idx]); )==" "\n" |
206 | R"==(POST_OP_DATA_T sum1_i = DATA_TO_REF(sum1[idx]); )==" "\n" |
207 | R"==(po_mb += VECT_DT_N; )==" "\n" |
208 | R"==(APPLY_POST_OPS_SERIAL_BINARY_2D( )==" "\n" |
209 | R"==(d1_i, float, sum1_i, POST_OP_DATA_T, po_mb, 1, po_oc, 1); )==" "\n" |
210 | R"==(D1[idx] = USE_FLOATS ? d1_i : CONVERT_DATA_T(d1_i); )==" "\n" |
211 | R"==(} )==" "\n" |
212 | R"==(#endif )==" "\n" |
213 | R"==(#if USE_FLOATS )==" "\n" |
214 | R"==(VECT_DATA_T res0 = CONVERT_VECTOR_DATA_T(D0); )==" "\n" |
215 | R"==(VECT_DATA_T res1 = CONVERT_VECTOR_DATA_T(D1); )==" "\n" |
216 | R"==(#else )==" "\n" |
217 | R"==(VECT_DATA_T res0 = D0; )==" "\n" |
218 | R"==(VECT_DATA_T res1 = D1; )==" "\n" |
219 | R"==(#endif )==" "\n" |
220 | R"==(write_vect_c_block( )==" "\n" |
221 | R"==(0, &dst[dst_off0], c, dst_stride, dst_chunks_per_c_block, res0); )==" "\n" |
222 | R"==(#if UNROLL_MB )==" "\n" |
223 | R"==(write_vect_c_block( )==" "\n" |
224 | R"==(0, &dst[dst_off1], c, dst_stride, dst_chunks_per_c_block, res1); )==" "\n" |
225 | R"==(#else )==" "\n" |
226 | R"==(write_vect_c_block( )==" "\n" |
227 | R"==(1, &dst[dst_off0], c, dst_stride, dst_chunks_per_c_block, res1); )==" "\n" |
228 | R"==(#endif )==" "\n" |
229 | R"==(#if ALG_MAX && IS_TRAINING )==" "\n" |
230 | R"==(int ws_off0 = dst_off0; )==" "\n" |
231 | R"==(#if UNROLL_MB )==" "\n" |
232 | R"==(int ws_off1 = dst_off1; )==" "\n" |
233 | R"==(#endif )==" "\n" |
234 | R"==(write_vect_c_block_int( )==" "\n" |
235 | R"==(0, &ws[ws_off0], c, ws_stride, ws_chunks_per_c_block, WS0); )==" "\n" |
236 | R"==(#if UNROLL_MB )==" "\n" |
237 | R"==(write_vect_c_block_int( )==" "\n" |
238 | R"==(0, &ws[ws_off1], c, ws_stride, ws_chunks_per_c_block, WS1); )==" "\n" |
239 | R"==(#else )==" "\n" |
240 | R"==(write_vect_c_block_int( )==" "\n" |
241 | R"==(1, &ws[ws_off0], c, ws_stride, ws_chunks_per_c_block, WS1); )==" "\n" |
242 | R"==(#endif )==" "\n" |
243 | R"==(#endif )==" "\n" |
244 | R"==(} )==" "\n" |
245 | R"==(#endif )==" "\n" |
246 | R"==(#if IS_BWD )==" "\n" |
247 | R"==(KERNEL_ATTR )==" "\n" |
248 | R"==(__kernel void gen9_pooling_bwd(__global DATA_T *diff_src, __global int *ws, )==" "\n" |
249 | R"==(__global DATA_T *diff_dst) { )==" "\n" |
250 | R"==(const int mb0 = GWS_GET_MB(); )==" "\n" |
251 | R"==(#if UNROLL_MB )==" "\n" |
252 | R"==(const int mb1 = mb0 + MB / 4; )==" "\n" |
253 | R"==(const int mb2 = mb1 + MB / 4; )==" "\n" |
254 | R"==(const int mb3 = mb2 + MB / 4; )==" "\n" |
255 | R"==(#endif )==" "\n" |
256 | R"==(const int c = GWS_GET_C(); )==" "\n" |
257 | R"==(const int id = GWS_GET_ID(); )==" "\n" |
258 | R"==(const int ih = GWS_GET_IH(); )==" "\n" |
259 | R"==(const int iw = GWS_GET_IW(); )==" "\n" |
260 | R"==(#if USE_MB_C_BLOCK )==" "\n" |
261 | R"==(const int src_stride = (SRC_SB0 > 1) ? SRC_SB0 : SRC_S0; )==" "\n" |
262 | R"==(const int dst_stride = (DST_SB0 > 1) ? DST_SB0 : DST_S0; )==" "\n" |
263 | R"==(const int src_chunks_per_c_block = CHUNKS_PER_C_BLOCK; )==" "\n" |
264 | R"==(const int dst_chunks_per_c_block = CHUNKS_PER_C_BLOCK; )==" "\n" |
265 | R"==(#elif USE_ONLY_C_BLOCK )==" "\n" |
266 | R"==(const int src_stride = (SRC_B1 > 1) ? SRC_S1 : SUB_GROUP_SIZE; )==" "\n" |
267 | R"==(const int dst_stride = (DST_B1 > 1) ? DST_S1 : SUB_GROUP_SIZE; )==" "\n" |
268 | R"==(const int src_chunks_per_c_block )==" "\n" |
269 | R"==(= (SRC_B1 > 1) ? (SRC_B1 / SUB_GROUP_SIZE) : 1; )==" "\n" |
270 | R"==(const int dst_chunks_per_c_block )==" "\n" |
271 | R"==(= (DST_B1 > 1) ? (DST_B1 / SUB_GROUP_SIZE) : 1; )==" "\n" |
272 | R"==(#endif )==" "\n" |
273 | R"==(const int ws_stride = dst_stride; )==" "\n" |
274 | R"==(const int ws_chunks_per_c_block = dst_chunks_per_c_block; )==" "\n" |
275 | R"==(VECT_FLOAT_T S0 = 0, S1 = 0; )==" "\n" |
276 | R"==(#if UNROLL_MB )==" "\n" |
277 | R"==(VECT_FLOAT_T S2 = 0, S3 = 0; )==" "\n" |
278 | R"==(#endif )==" "\n" |
279 | R"==(for (int kd = 0; kd < KD; kd++) { )==" "\n" |
280 | R"==(int od = (id + PD - kd); )==" "\n" |
281 | R"==(if (od % SD != 0) continue; )==" "\n" |
282 | R"==(od /= SD; )==" "\n" |
283 | R"==(if (od < 0 || od >= OD) continue; )==" "\n" |
284 | R"==(for (int kh = 0; kh < KH; kh++) { )==" "\n" |
285 | R"==(int oh = (ih + PH - kh); )==" "\n" |
286 | R"==(if (oh % SH != 0) continue; )==" "\n" |
287 | R"==(oh /= SH; )==" "\n" |
288 | R"==(if (oh < 0 || oh >= OH) continue; )==" "\n" |
289 | R"==(for (int kw = 0; kw < KW; kw++) { )==" "\n" |
290 | R"==(int ow = (iw + PW - kw); )==" "\n" |
291 | R"==(if (ow % SW != 0) continue; )==" "\n" |
292 | R"==(ow /= SW; )==" "\n" |
293 | R"==(if (ow < 0 || ow >= OW) continue; )==" "\n" |
294 | R"==(const int dst_off0 = DST_OFF(mb0, c, od, oh, ow); )==" "\n" |
295 | R"==(#if UNROLL_MB )==" "\n" |
296 | R"==(const int dst_off1 = DST_OFF(mb1, c, od, oh, ow); )==" "\n" |
297 | R"==(const int dst_off2 = DST_OFF(mb2, c, od, oh, ow); )==" "\n" |
298 | R"==(const int dst_off3 = DST_OFF(mb3, c, od, oh, ow); )==" "\n" |
299 | R"==(#endif )==" "\n" |
300 | R"==(VECT_FLOAT_T D0 = CONVERT_VECT_FLOAT_T( )==" "\n" |
301 | R"==(read_vect_c_block(0, &diff_dst[dst_off0], c, dst_stride, )==" "\n" |
302 | R"==(dst_chunks_per_c_block)); )==" "\n" |
303 | R"==(#if UNROLL_MB )==" "\n" |
304 | R"==(VECT_FLOAT_T D1 = CONVERT_VECT_FLOAT_T( )==" "\n" |
305 | R"==(read_vect_c_block(0, &diff_dst[dst_off1], c, dst_stride, )==" "\n" |
306 | R"==(dst_chunks_per_c_block)); )==" "\n" |
307 | R"==(VECT_FLOAT_T D2 = CONVERT_VECT_FLOAT_T( )==" "\n" |
308 | R"==(read_vect_c_block(0, &diff_dst[dst_off2], c, dst_stride, )==" "\n" |
309 | R"==(dst_chunks_per_c_block)); )==" "\n" |
310 | R"==(VECT_FLOAT_T D3 = CONVERT_VECT_FLOAT_T( )==" "\n" |
311 | R"==(read_vect_c_block(0, &diff_dst[dst_off3], c, dst_stride, )==" "\n" |
312 | R"==(dst_chunks_per_c_block)); )==" "\n" |
313 | R"==(#else )==" "\n" |
314 | R"==(VECT_FLOAT_T D1 = CONVERT_VECT_FLOAT_T( )==" "\n" |
315 | R"==(read_vect_c_block(1, &diff_dst[dst_off0], c, dst_stride, )==" "\n" |
316 | R"==(dst_chunks_per_c_block)); )==" "\n" |
317 | R"==(#endif )==" "\n" |
318 | R"==(#if ALG_MAX )==" "\n" |
319 | R"==(VECT_INT_T WS0 = read_vect_c_block_int( )==" "\n" |
320 | R"==(0, &ws[dst_off0], c, ws_stride, ws_chunks_per_c_block); )==" "\n" |
321 | R"==(#if UNROLL_MB )==" "\n" |
322 | R"==(VECT_INT_T WS1 = read_vect_c_block_int( )==" "\n" |
323 | R"==(0, &ws[dst_off1], c, ws_stride, ws_chunks_per_c_block); )==" "\n" |
324 | R"==(VECT_INT_T WS2 = read_vect_c_block_int( )==" "\n" |
325 | R"==(0, &ws[dst_off2], c, ws_stride, ws_chunks_per_c_block); )==" "\n" |
326 | R"==(VECT_INT_T WS3 = read_vect_c_block_int( )==" "\n" |
327 | R"==(0, &ws[dst_off3], c, ws_stride, ws_chunks_per_c_block); )==" "\n" |
328 | R"==(#else )==" "\n" |
329 | R"==(VECT_INT_T WS1 = read_vect_c_block_int( )==" "\n" |
330 | R"==(1, &ws[dst_off0], c, ws_stride, ws_chunks_per_c_block); )==" "\n" |
331 | R"==(#endif )==" "\n" |
332 | R"==(VECT_INT_T CMP0 = isnotequal( )==" "\n" |
333 | R"==(AS_VECT_FLOAT_T(WS0 - kd * KH * KW - kh * KW - kw), )==" "\n" |
334 | R"==((VECT_FLOAT_T)0); )==" "\n" |
335 | R"==(D0 = select(D0, (VECT_FLOAT_T)0, CMP0); )==" "\n" |
336 | R"==(VECT_INT_T CMP1 = isnotequal( )==" "\n" |
337 | R"==(AS_VECT_FLOAT_T(WS1 - kd * KH * KW - kh * KW - kw), )==" "\n" |
338 | R"==((VECT_FLOAT_T)0); )==" "\n" |
339 | R"==(D1 = select(D1, (VECT_FLOAT_T)0, CMP1); )==" "\n" |
340 | R"==(#if UNROLL_MB )==" "\n" |
341 | R"==(VECT_INT_T CMP2 = isnotequal( )==" "\n" |
342 | R"==(AS_VECT_FLOAT_T(WS2 - kd * KH * KW - kh * KW - kw), )==" "\n" |
343 | R"==((VECT_FLOAT_T)0); )==" "\n" |
344 | R"==(D2 = select(D2, (VECT_FLOAT_T)0, CMP2); )==" "\n" |
345 | R"==(VECT_INT_T CMP3 = isnotequal( )==" "\n" |
346 | R"==(AS_VECT_FLOAT_T(WS3 - kd * KH * KW - kh * KW - kw), )==" "\n" |
347 | R"==((VECT_FLOAT_T)0); )==" "\n" |
348 | R"==(D3 = select(D3, (VECT_FLOAT_T)0, CMP3); )==" "\n" |
349 | R"==(#endif )==" "\n" |
350 | R"==(#endif )==" "\n" |
351 | R"==(#if ALG_AVG_NP )==" "\n" |
352 | R"==(const int id_start = max(id - kd, 0); )==" "\n" |
353 | R"==(const int ih_start = max(ih - kh, 0); )==" "\n" |
354 | R"==(const int iw_start = max(iw - kw, 0); )==" "\n" |
355 | R"==(const int id_end = min(id - kd + KD, ID); )==" "\n" |
356 | R"==(const int ih_end = min(ih - kh + KH, IH); )==" "\n" |
357 | R"==(const int iw_end = min(iw - kw + KW, IW); )==" "\n" |
358 | R"==(const int num_summands = (ih_end - ih_start) )==" "\n" |
359 | R"==(* (iw_end - iw_start) * (id_end - id_start); )==" "\n" |
360 | R"==(D0 /= num_summands; )==" "\n" |
361 | R"==(D1 /= num_summands; )==" "\n" |
362 | R"==(#endif )==" "\n" |
363 | R"==(S0 += D0; )==" "\n" |
364 | R"==(S1 += D1; )==" "\n" |
365 | R"==(#if UNROLL_MB )==" "\n" |
366 | R"==(S2 += D2; )==" "\n" |
367 | R"==(S3 += D3; )==" "\n" |
368 | R"==(#endif )==" "\n" |
369 | R"==(} )==" "\n" |
370 | R"==(} )==" "\n" |
371 | R"==(} )==" "\n" |
372 | R"==(#if ALG_AVG_P )==" "\n" |
373 | R"==(S0 /= KD * KH * KW; )==" "\n" |
374 | R"==(S1 /= KD * KH * KW; )==" "\n" |
375 | R"==(#if UNROLL_MB )==" "\n" |
376 | R"==(S2 /= KD * KH * KW; )==" "\n" |
377 | R"==(S3 /= KD * KH * KW; )==" "\n" |
378 | R"==(#endif )==" "\n" |
379 | R"==(#endif )==" "\n" |
380 | R"==(int src_off0 = SRC_OFF(mb0, c, id, ih, iw); )==" "\n" |
381 | R"==(#if UNROLL_MB )==" "\n" |
382 | R"==(int src_off1 = SRC_OFF(mb1, c, id, ih, iw); )==" "\n" |
383 | R"==(int src_off2 = SRC_OFF(mb2, c, id, ih, iw); )==" "\n" |
384 | R"==(int src_off3 = SRC_OFF(mb3, c, id, ih, iw); )==" "\n" |
385 | R"==(#endif )==" "\n" |
386 | R"==(write_vect_c_block(0, &diff_src[src_off0], c, src_stride, )==" "\n" |
387 | R"==(src_chunks_per_c_block, CONVERT_VECTOR_DATA_T(S0)); )==" "\n" |
388 | R"==(#if UNROLL_MB )==" "\n" |
389 | R"==(write_vect_c_block(0, &diff_src[src_off1], c, src_stride, )==" "\n" |
390 | R"==(src_chunks_per_c_block, CONVERT_VECTOR_DATA_T(S1)); )==" "\n" |
391 | R"==(write_vect_c_block(0, &diff_src[src_off2], c, src_stride, )==" "\n" |
392 | R"==(src_chunks_per_c_block, CONVERT_VECTOR_DATA_T(S2)); )==" "\n" |
393 | R"==(write_vect_c_block(0, &diff_src[src_off3], c, src_stride, )==" "\n" |
394 | R"==(src_chunks_per_c_block, CONVERT_VECTOR_DATA_T(S3)); )==" "\n" |
395 | R"==(#else )==" "\n" |
396 | R"==(write_vect_c_block(1, &diff_src[src_off0], c, src_stride, )==" "\n" |
397 | R"==(src_chunks_per_c_block, CONVERT_VECTOR_DATA_T(S1)); )==" "\n" |
398 | R"==(#endif )==" "\n" |
399 | R"==(} )==" "\n" |
400 | R"==(#endif )==" "\n" |
401 | R"==(inline DATA_T read_c_block(const __global DATA_T *ptr, int c) { )==" "\n" |
402 | R"==(#if C_W_PADDING % SUB_GROUP_SIZE != 0 )==" "\n" |
403 | R"==(int local_id = get_sub_group_local_id(); )==" "\n" |
404 | R"==(int tail = C_WO_PADDING - c; )==" "\n" |
405 | R"==(return (local_id < tail) ? ptr[local_id] : 0; )==" "\n" |
406 | R"==(#else )==" "\n" |
407 | R"==(return AS_DATA_T(BLOCK_READ((const __global BLOCK_DATA_T *)ptr)); )==" "\n" |
408 | R"==(#endif )==" "\n" |
409 | R"==(} )==" "\n" |
410 | R"==(#define CALC_VECT_LEN() \ )==" "\n" |
411 | R"==(({ \ )==" "\n" |
412 | R"==(int size; \ )==" "\n" |
413 | R"==(if (USE_ONLY_C_BLOCK == 1 \ )==" "\n" |
414 | R"==(&& VECT_DT_N > C_WO_PADDING / SUB_GROUP_SIZE + 1) \ )==" "\n" |
415 | R"==(size = C_WO_PADDING / SUB_GROUP_SIZE + 1; \ )==" "\n" |
416 | R"==(else \ )==" "\n" |
417 | R"==(size = VECT_DT_N; \ )==" "\n" |
418 | R"==(size; \ )==" "\n" |
419 | R"==(}) )==" "\n" |
420 | R"==(inline VECT_DATA_T read_vect_c_block(int idx, const __global DATA_T *ptr, int c, )==" "\n" |
421 | R"==(int blocks_stride, int chunks_per_block) { )==" "\n" |
422 | R"==(if (idx >= NVECT) return 0; )==" "\n" |
423 | R"==(if ((blocks_stride == chunks_per_block * SUB_GROUP_SIZE) )==" "\n" |
424 | R"==(&& (C_WO_PADDING % (chunks_per_block * SUB_GROUP_SIZE) == 0)) { )==" "\n" |
425 | R"==(return AS_VECT_DATA_T(VECT_BLOCK_READ((const __global BLOCK_DATA_T *)ptr )==" "\n" |
426 | R"==(+ idx * VECT_DT_N * SUB_GROUP_SIZE)); )==" "\n" |
427 | R"==(} else { )==" "\n" |
428 | R"==(VECT_DATA_T ret; )==" "\n" |
429 | R"==(for (int i = 0; i < CALC_VECT_LEN(); i++) { )==" "\n" |
430 | R"==(const int offset_index = (idx * VECT_DT_N + i); )==" "\n" |
431 | R"==(const int local_c_block_index = offset_index % chunks_per_block; )==" "\n" |
432 | R"==(const int global_c_block_index = offset_index / chunks_per_block; )==" "\n" |
433 | R"==(const int ptr_offset = local_c_block_index * SUB_GROUP_SIZE )==" "\n" |
434 | R"==(+ global_c_block_index * blocks_stride; )==" "\n" |
435 | R"==(const int c_off )==" "\n" |
436 | R"==(= (USE_ONLY_C_BLOCK ? offset_index * SUB_GROUP_SIZE )==" "\n" |
437 | R"==(: local_c_block_index * SUB_GROUP_SIZE); )==" "\n" |
438 | R"==(#if VECT_DT_N == 1 )==" "\n" |
439 | R"==(ret = read_c_block(ptr + ptr_offset, c + c_off); )==" "\n" |
440 | R"==(#else )==" "\n" |
441 | R"==(ret[i] = read_c_block(ptr + ptr_offset, c + c_off); )==" "\n" |
442 | R"==(#endif )==" "\n" |
443 | R"==(} )==" "\n" |
444 | R"==(#if VECT_DT_N > 1 )==" "\n" |
445 | R"==(for (int i = CALC_VECT_LEN(); i < VECT_DT_N; ++i) { )==" "\n" |
446 | R"==(ret[i] = 0; )==" "\n" |
447 | R"==(} )==" "\n" |
448 | R"==(#endif )==" "\n" |
449 | R"==(return ret; )==" "\n" |
450 | R"==(} )==" "\n" |
451 | R"==(} )==" "\n" |
452 | R"==(inline int read_c_block_int(const __global int *ptr, int c) { )==" "\n" |
453 | R"==(#if C_W_PADDING % SUB_GROUP_SIZE != 0 )==" "\n" |
454 | R"==(int local_id = get_sub_group_local_id(); )==" "\n" |
455 | R"==(int tail = C_WO_PADDING - c; )==" "\n" |
456 | R"==(return (local_id < tail) ? ptr[local_id] : 0; )==" "\n" |
457 | R"==(#else )==" "\n" |
458 | R"==(return as_int(intel_sub_group_block_read((const __global uint *)ptr)); )==" "\n" |
459 | R"==(#endif )==" "\n" |
460 | R"==(} )==" "\n" |
461 | R"==(inline VECT_INT_T read_vect_c_block_int(int idx, const __global int *ptr, int c, )==" "\n" |
462 | R"==(int blocks_stride, int chunks_per_block) { )==" "\n" |
463 | R"==(if (idx >= NVECT) return 0; )==" "\n" |
464 | R"==(if ((blocks_stride == chunks_per_block * SUB_GROUP_SIZE) )==" "\n" |
465 | R"==(&& (C_WO_PADDING % (chunks_per_block * SUB_GROUP_SIZE) == 0)) { )==" "\n" |
466 | R"==(return AS_VECT_INT_T(VECT_UINT_READ( )==" "\n" |
467 | R"==((const __global uint *)ptr + idx * VECT_DT_N * SUB_GROUP_SIZE)); )==" "\n" |
468 | R"==(} else { )==" "\n" |
469 | R"==(VECT_INT_T ret; )==" "\n" |
470 | R"==(for (int i = 0; i < VECT_DT_N; i++) { )==" "\n" |
471 | R"==(const int offset_index = (idx * VECT_DT_N + i); )==" "\n" |
472 | R"==(const int local_c_block_index = offset_index % chunks_per_block; )==" "\n" |
473 | R"==(const int global_c_block_index = offset_index / chunks_per_block; )==" "\n" |
474 | R"==(const int ptr_offset = local_c_block_index * SUB_GROUP_SIZE )==" "\n" |
475 | R"==(+ global_c_block_index * blocks_stride; )==" "\n" |
476 | R"==(const int c_off )==" "\n" |
477 | R"==(= (USE_ONLY_C_BLOCK ? offset_index * SUB_GROUP_SIZE )==" "\n" |
478 | R"==(: local_c_block_index * SUB_GROUP_SIZE); )==" "\n" |
479 | R"==(#if VECT_DT_N == 1 )==" "\n" |
480 | R"==(ret = read_c_block_int(ptr + ptr_offset, c + c_off); )==" "\n" |
481 | R"==(#else )==" "\n" |
482 | R"==(ret[i] = read_c_block_int(ptr + ptr_offset, c + c_off); )==" "\n" |
483 | R"==(#endif )==" "\n" |
484 | R"==(} )==" "\n" |
485 | R"==(return ret; )==" "\n" |
486 | R"==(} )==" "\n" |
487 | R"==(} )==" "\n" |
488 | R"==(inline void write_c_block(__global DATA_T *ptr, int c, DATA_T value) { )==" "\n" |
489 | R"==(#if C_W_PADDING % SUB_GROUP_SIZE != 0 )==" "\n" |
490 | R"==(int local_id = get_sub_group_local_id(); )==" "\n" |
491 | R"==(int tail = C_WO_PADDING - c; )==" "\n" |
492 | R"==(if (local_id < tail) ptr[local_id] = value; )==" "\n" |
493 | R"==(#else )==" "\n" |
494 | R"==(#if C_WO_PADDING % SUB_GROUP_SIZE != 0 )==" "\n" |
495 | R"==(int local_id = get_sub_group_local_id(); )==" "\n" |
496 | R"==(if (local_id >= C_WO_PADDING - c && local_id < C_W_PADDING - c) value = 0; )==" "\n" |
497 | R"==(#endif )==" "\n" |
498 | R"==(if (c >= C_WO_PADDING) { )==" "\n" |
499 | R"==(BLOCK_WRITE((__global BLOCK_DATA_T *)ptr, )==" "\n" |
500 | R"==(AS_BLOCK_DATA_T(CONVERT_DATA_T(DATA_ZERO))); )==" "\n" |
501 | R"==(return; )==" "\n" |
502 | R"==(} )==" "\n" |
503 | R"==(BLOCK_WRITE((__global BLOCK_DATA_T *)ptr, AS_BLOCK_DATA_T(value)); )==" "\n" |
504 | R"==(#endif )==" "\n" |
505 | R"==(} )==" "\n" |
506 | R"==(inline void write_vect_c_block(int idx, __global DATA_T *ptr, int c, )==" "\n" |
507 | R"==(int blocks_stride, int chunks_per_block, VECT_DATA_T block) { )==" "\n" |
508 | R"==(if (idx >= NVECT) return; )==" "\n" |
509 | R"==(if ((blocks_stride == chunks_per_block * SUB_GROUP_SIZE) )==" "\n" |
510 | R"==(&& (C_WO_PADDING % (chunks_per_block * SUB_GROUP_SIZE) == 0)) { )==" "\n" |
511 | R"==(VECT_BLOCK_WRITE( )==" "\n" |
512 | R"==((__global BLOCK_DATA_T *)ptr + idx * VECT_DT_N * SUB_GROUP_SIZE, )==" "\n" |
513 | R"==(AS_VECT_BLOCK_DATA_T(block)); )==" "\n" |
514 | R"==(} else { )==" "\n" |
515 | R"==(for (int i = 0; i < VECT_DT_N; i++) { )==" "\n" |
516 | R"==(const int offset_index = (idx * VECT_DT_N + i); )==" "\n" |
517 | R"==(const int local_c_block_index = offset_index % chunks_per_block; )==" "\n" |
518 | R"==(const int global_c_block_index = offset_index / chunks_per_block; )==" "\n" |
519 | R"==(const int ptr_offset = local_c_block_index * SUB_GROUP_SIZE )==" "\n" |
520 | R"==(+ global_c_block_index * blocks_stride; )==" "\n" |
521 | R"==(const int c_off )==" "\n" |
522 | R"==(= (USE_ONLY_C_BLOCK ? offset_index * SUB_GROUP_SIZE )==" "\n" |
523 | R"==(: local_c_block_index * SUB_GROUP_SIZE); )==" "\n" |
524 | R"==(#if VECT_DT_N == 1 )==" "\n" |
525 | R"==(write_c_block(ptr + ptr_offset, c + c_off, block); )==" "\n" |
526 | R"==(#else )==" "\n" |
527 | R"==(write_c_block(ptr + ptr_offset, c + c_off, block[i]); )==" "\n" |
528 | R"==(#endif )==" "\n" |
529 | R"==(} )==" "\n" |
530 | R"==(} )==" "\n" |
531 | R"==(} )==" "\n" |
532 | R"==(inline void write_c_block_int(__global int *ptr, int c, int value) { )==" "\n" |
533 | R"==(#if C_WO_PADDING % SUB_GROUP_SIZE != 0 )==" "\n" |
534 | R"==(int local_id = get_sub_group_local_id(); )==" "\n" |
535 | R"==(int tail = C_WO_PADDING - c; )==" "\n" |
536 | R"==(if (local_id < tail) )==" "\n" |
537 | R"==(ptr[local_id] = value; )==" "\n" |
538 | R"==(else if (local_id < C_W_PADDING - c) { )==" "\n" |
539 | R"==(ptr[local_id] = 0; )==" "\n" |
540 | R"==(} else )==" "\n" |
541 | R"==(return; )==" "\n" |
542 | R"==(#else )==" "\n" |
543 | R"==(if (c >= C_WO_PADDING) { )==" "\n" |
544 | R"==(intel_sub_group_block_write((__global uint *)ptr, 0); )==" "\n" |
545 | R"==(return; )==" "\n" |
546 | R"==(} )==" "\n" |
547 | R"==(intel_sub_group_block_write((__global uint *)ptr, as_uint(value)); )==" "\n" |
548 | R"==(#endif )==" "\n" |
549 | R"==(} )==" "\n" |
550 | R"==(inline void write_vect_c_block_int(int idx, __global int *ptr, int c, )==" "\n" |
551 | R"==(int blocks_stride, int chunks_per_block, VECT_INT_T block) { )==" "\n" |
552 | R"==(if (idx >= NVECT) return; )==" "\n" |
553 | R"==(if ((blocks_stride == chunks_per_block * SUB_GROUP_SIZE) )==" "\n" |
554 | R"==(&& (C_WO_PADDING % (chunks_per_block * SUB_GROUP_SIZE) == 0)) { )==" "\n" |
555 | R"==(VECT_UINT_WRITE((__global uint *)ptr + idx * VECT_DT_N * SUB_GROUP_SIZE, )==" "\n" |
556 | R"==(AS_VECT_UINT_T(block)); )==" "\n" |
557 | R"==(} else { )==" "\n" |
558 | R"==(for (int i = 0; i < VECT_DT_N; i++) { )==" "\n" |
559 | R"==(const int offset_index = (idx * VECT_DT_N + i); )==" "\n" |
560 | R"==(const int local_c_block_index = offset_index % chunks_per_block; )==" "\n" |
561 | R"==(const int global_c_block_index = offset_index / chunks_per_block; )==" "\n" |
562 | R"==(const int ptr_offset = local_c_block_index * SUB_GROUP_SIZE )==" "\n" |
563 | R"==(+ global_c_block_index * blocks_stride; )==" "\n" |
564 | R"==(const int c_off )==" "\n" |
565 | R"==(= (USE_ONLY_C_BLOCK ? offset_index * SUB_GROUP_SIZE )==" "\n" |
566 | R"==(: local_c_block_index * SUB_GROUP_SIZE); )==" "\n" |
567 | R"==(#if VECT_DT_N == 1 )==" "\n" |
568 | R"==(write_c_block_int(ptr + ptr_offset, c + c_off, block); )==" "\n" |
569 | R"==(#else )==" "\n" |
570 | R"==(write_c_block_int(ptr + ptr_offset, c + c_off, block[i]); )==" "\n" |
571 | R"==(#endif )==" "\n" |
572 | R"==(} )==" "\n" |
573 | R"==(} )==" "\n" |
574 | R"==(} )==" "\n" |
575 | R"==()==" ; |
576 | } |
577 | } |
578 | } |
579 | } |