1 | namespace dnnl { |
2 | namespace impl { |
3 | namespace gpu { |
4 | namespace ocl { |
5 | const char *gen9_gemm_nocopy_x8x8s32_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/gemm/ocl_gemm_attrs.h" )==" "\n" |
21 | R"==(#include "gpu/ocl/ocl_post_ops.h" )==" "\n" |
22 | R"==(#include "gpu/ocl/ocl_types.h" )==" "\n" |
23 | R"==(#undef GRX )==" "\n" |
24 | R"==(#define GRX 8 )==" "\n" |
25 | R"==(#if defined(S8S8) )==" "\n" |
26 | R"==(#define FLOATA char )==" "\n" |
27 | R"==(#define FLOATA2 char2 )==" "\n" |
28 | R"==(#define FLOATA4 char4 )==" "\n" |
29 | R"==(#define FLOATB char )==" "\n" |
30 | R"==(#define FLOATB4 char4 )==" "\n" |
31 | R"==(#define SHUFFLE(X, Y) as_char4(intel_sub_group_shuffle(as_int(X), Y)) )==" "\n" |
32 | R"==(#endif )==" "\n" |
33 | R"==(#if defined(U8S8) )==" "\n" |
34 | R"==(#define FLOATA uchar )==" "\n" |
35 | R"==(#define FLOATA2 uchar2 )==" "\n" |
36 | R"==(#define FLOATA4 uchar4 )==" "\n" |
37 | R"==(#define FLOATB char )==" "\n" |
38 | R"==(#define FLOATB4 char4 )==" "\n" |
39 | R"==(#define SHUFFLE(X, Y) as_char4(intel_sub_group_shuffle(as_int(X), Y)) )==" "\n" |
40 | R"==(#endif )==" "\n" |
41 | R"==(#if defined(S8U8) )==" "\n" |
42 | R"==(#define FLOATA char )==" "\n" |
43 | R"==(#define FLOATA2 char2 )==" "\n" |
44 | R"==(#define FLOATA4 char4 )==" "\n" |
45 | R"==(#define FLOATB uchar )==" "\n" |
46 | R"==(#define FLOATB4 uchar4 )==" "\n" |
47 | R"==(#define SHUFFLE(X, Y) as_uchar4(intel_sub_group_shuffle(as_int(X), Y)) )==" "\n" |
48 | R"==(#endif )==" "\n" |
49 | R"==(#if defined(U8U8) )==" "\n" |
50 | R"==(#define FLOATA uchar )==" "\n" |
51 | R"==(#define FLOATA2 uchar2 )==" "\n" |
52 | R"==(#define FLOATA4 uchar4 )==" "\n" |
53 | R"==(#define FLOATB uchar )==" "\n" |
54 | R"==(#define FLOATB4 uchar4 )==" "\n" |
55 | R"==(#define SHUFFLE(X, Y) as_uchar4(intel_sub_group_shuffle(as_int(X), Y)) )==" "\n" |
56 | R"==(#endif )==" "\n" |
57 | R"==(#define FLOATC int )==" "\n" |
58 | R"==(#define FLOATC4 int4 )==" "\n" |
59 | R"==(#if WITH_ELTWISE == 1 )==" "\n" |
60 | R"==(#define POST_OP(val) \ )==" "\n" |
61 | R"==(do { \ )==" "\n" |
62 | R"==(if (apply_eltwise) \ )==" "\n" |
63 | R"==(val = fwd_eltwise( \ )==" "\n" |
64 | R"==(val, eltwise_alpha, eltwise_beta, eltwise_scale); \ )==" "\n" |
65 | R"==(} while (0) )==" "\n" |
66 | R"==(#else )==" "\n" |
67 | R"==(#define POST_OP(val) )==" "\n" |
68 | R"==(#endif )==" "\n" |
69 | R"==(#define COMPUTE_C(X, Y, CO_IDX) \ )==" "\n" |
70 | R"==(do { \ )==" "\n" |
71 | R"==(float val = (!beta ? 0 : (X)) + (Y); \ )==" "\n" |
72 | R"==(POST_OP(val); \ )==" "\n" |
73 | R"==((X) = convert_int_sat_rte(val + (apply_co ? co[CO_IDX] : 0)); \ )==" "\n" |
74 | R"==(} while (0) )==" "\n" |
75 | R"==(#ifdef FF )==" "\n" |
76 | R"==(#define ADD_EACH(X, OFF) \ )==" "\n" |
77 | R"==(do { \ )==" "\n" |
78 | R"==(if (n > X + OFF) { \ )==" "\n" |
79 | R"==(if (m > 0) \ )==" "\n" |
80 | R"==(COMPUTE_C(c[0], sc[X / 4 + 0].s##OFF + xa[0] + xb[0], 0); \ )==" "\n" |
81 | R"==(if (m > 1) \ )==" "\n" |
82 | R"==(COMPUTE_C(c[1], sc[X / 4 + 4].s##OFF + xa[1] + xb[0], 0); \ )==" "\n" |
83 | R"==(if (m > 2) \ )==" "\n" |
84 | R"==(COMPUTE_C(c[2], sc[X / 4 + 8].s##OFF + xa[2] + xb[0], 0); \ )==" "\n" |
85 | R"==(if (m > 3) \ )==" "\n" |
86 | R"==(COMPUTE_C(c[3], sc[X / 4 + 12].s##OFF + xa[3] + xb[0], 0); \ )==" "\n" |
87 | R"==(xb++; \ )==" "\n" |
88 | R"==(c += ldc; \ )==" "\n" |
89 | R"==(} \ )==" "\n" |
90 | R"==(} while (0) )==" "\n" |
91 | R"==(#elif defined CC )==" "\n" |
92 | R"==(#define ADD_EACH(X, OFF) \ )==" "\n" |
93 | R"==(do { \ )==" "\n" |
94 | R"==(if (n > X + OFF) { \ )==" "\n" |
95 | R"==(if (m > 0) \ )==" "\n" |
96 | R"==(COMPUTE_C(c[0], sc[X / 4 + 0].s##OFF + xa[0] + xb[0], 0); \ )==" "\n" |
97 | R"==(if (m > 1) \ )==" "\n" |
98 | R"==(COMPUTE_C(c[1], sc[X / 4 + 4].s##OFF + xa[1] + xb[0], 0); \ )==" "\n" |
99 | R"==(if (m > 2) \ )==" "\n" |
100 | R"==(COMPUTE_C(c[2], sc[X / 4 + 8].s##OFF + xa[2] + xb[0], 0); \ )==" "\n" |
101 | R"==(if (m > 3) \ )==" "\n" |
102 | R"==(COMPUTE_C(c[3], sc[X / 4 + 12].s##OFF + xa[3] + xb[0], 0); \ )==" "\n" |
103 | R"==(xb++; \ )==" "\n" |
104 | R"==(c += ldc; \ )==" "\n" |
105 | R"==(co++; \ )==" "\n" |
106 | R"==(} \ )==" "\n" |
107 | R"==(} while (0) )==" "\n" |
108 | R"==(#else )==" "\n" |
109 | R"==(#define ADD_EACH(X, OFF) \ )==" "\n" |
110 | R"==(do { \ )==" "\n" |
111 | R"==(if (n > X + OFF) { \ )==" "\n" |
112 | R"==(if (m > 0) \ )==" "\n" |
113 | R"==(COMPUTE_C(c[0], sc[X / 4 + 0].s##OFF + xa[0] + xb[0], 0); \ )==" "\n" |
114 | R"==(if (m > 1) \ )==" "\n" |
115 | R"==(COMPUTE_C(c[1], sc[X / 4 + 4].s##OFF + xa[1] + xb[0], 1); \ )==" "\n" |
116 | R"==(if (m > 2) \ )==" "\n" |
117 | R"==(COMPUTE_C(c[2], sc[X / 4 + 8].s##OFF + xa[2] + xb[0], 2); \ )==" "\n" |
118 | R"==(if (m > 3) \ )==" "\n" |
119 | R"==(COMPUTE_C(c[3], sc[X / 4 + 12].s##OFF + xa[3] + xb[0], 3); \ )==" "\n" |
120 | R"==(xb++; \ )==" "\n" |
121 | R"==(c += ldc; \ )==" "\n" |
122 | R"==(} \ )==" "\n" |
123 | R"==(} while (0) )==" "\n" |
124 | R"==(#endif )==" "\n" |
125 | R"==(#define ADD_SCALE(X) \ )==" "\n" |
126 | R"==(do { \ )==" "\n" |
127 | R"==(ADD_EACH(X, 0); \ )==" "\n" |
128 | R"==(ADD_EACH(X, 1); \ )==" "\n" |
129 | R"==(ADD_EACH(X, 2); \ )==" "\n" |
130 | R"==(ADD_EACH(X, 3); \ )==" "\n" |
131 | R"==(} while (0) )==" "\n" |
132 | R"==(#define ACCUMULATE_1(a, b) \ )==" "\n" |
133 | R"==(((FLOATC)a.s0 * (FLOATC)b.s0) + ((FLOATC)a.s1 * (FLOATC)b.s1) \ )==" "\n" |
134 | R"==(+ ((FLOATC)a.s2 * (FLOATC)b.s2) + ((FLOATC)a.s3 * (FLOATC)b.s3) )==" "\n" |
135 | R"==(#define ACCUMULATE(a, b0, b1, b2, b3) \ )==" "\n" |
136 | R"==((FLOATC4)(ACCUMULATE_1(a, b0), ACCUMULATE_1(a, b1), ACCUMULATE_1(a, b2), \ )==" "\n" |
137 | R"==(ACCUMULATE_1(a, b3)) )==" "\n" |
138 | R"==(#define GROUPSIZE_M (6 * UNROLL_M) )==" "\n" |
139 | R"==(#define GROUPSIZE_N (4 * UNROLL_N) )==" "\n" |
140 | R"==(__attribute__((intel_reqd_sub_group_size(GRX))) kernel void )==" "\n" |
141 | R"==(gen9_gemm_compute_x8x8s32(global FLOATA *a, global FLOATB *b, global FLOATC *c, )==" "\n" |
142 | R"==(long offsetA, long offsetB, long offsetC, long lda, long ldb, long ldc, )==" "\n" |
143 | R"==(long m, long n, long k, int beta, global int *ao, global int *bo, )==" "\n" |
144 | R"==(global int *co, long offsetCO, int apply_co, local FLOATA *sa, )==" "\n" |
145 | R"==(local FLOATB *sb, int apply_eltwise, float eltwise_alpha, )==" "\n" |
146 | R"==(float eltwise_beta, float eltwise_scale) { )==" "\n" |
147 | R"==(long kk = (k + UNROLL_K - 1) & ~(UNROLL_K - 1); )==" "\n" |
148 | R"==(long i, j, l, ll; )==" "\n" |
149 | R"==(global FLOATC *c_ori; )==" "\n" |
150 | R"==(int lid = get_local_id(0); )==" "\n" |
151 | R"==(int idx = get_local_id(1); )==" "\n" |
152 | R"==(int idy = get_local_id(2); )==" "\n" |
153 | R"==(long gdx = get_group_id(1); )==" "\n" |
154 | R"==(long gdy = get_group_id(2); )==" "\n" |
155 | R"==(long szx = get_local_size(1); )==" "\n" |
156 | R"==(long szy = get_local_size(2); )==" "\n" |
157 | R"==(a += offsetA; )==" "\n" |
158 | R"==(b += offsetB; )==" "\n" |
159 | R"==(c += offsetC + UNROLL_M * idx + GROUPSIZE_M * gdx + UNROLL_M * lid / GRX )==" "\n" |
160 | R"==(+ (UNROLL_N * idy + GROUPSIZE_N * gdy) * ldc; )==" "\n" |
161 | R"==(c_ori = c; )==" "\n" |
162 | R"==(if (apply_co) { )==" "\n" |
163 | R"==(co += offsetCO; )==" "\n" |
164 | R"==(#ifdef RR )==" "\n" |
165 | R"==(co += GROUPSIZE_M * gdx + UNROLL_M * idx + UNROLL_M * lid / GRX; )==" "\n" |
166 | R"==(#endif )==" "\n" |
167 | R"==(#ifdef CC )==" "\n" |
168 | R"==(co += GROUPSIZE_N * gdy + UNROLL_N * idy; )==" "\n" |
169 | R"==(#endif )==" "\n" |
170 | R"==(} )==" "\n" |
171 | R"==(__local FLOATC *xa = (__local FLOATC *)sa; )==" "\n" |
172 | R"==(sa += UNROLL_M * szx * sizeof(FLOATC); )==" "\n" |
173 | R"==(__local FLOATC *xb = (__local FLOATC *)sb; )==" "\n" |
174 | R"==(sb += UNROLL_N * szy * sizeof(FLOATC); )==" "\n" |
175 | R"==(int cid0 = (idy * szx + idx) * get_local_size(0) + lid; )==" "\n" |
176 | R"==(int ctotal = get_local_size(0) * szx * szy; )==" "\n" |
177 | R"==(for (int cid = cid0; cid < szx * UNROLL_M; cid += ctotal) { )==" "\n" |
178 | R"==(long sa_moffset = (cid & ~(UNROLL_M - 1)) * kk )==" "\n" |
179 | R"==(+ (cid & (UNROLL_M - 1)) * UNROLL_K; )==" "\n" |
180 | R"==(long i = cid + GROUPSIZE_M * gdx; )==" "\n" |
181 | R"==(FLOATC sumA = 0; )==" "\n" |
182 | R"==(#if defined(NN) || defined(NT) )==" "\n" |
183 | R"==(long a_offset = i; )==" "\n" |
184 | R"==(#else )==" "\n" |
185 | R"==(long a_offset = i * lda; )==" "\n" |
186 | R"==(#endif )==" "\n" |
187 | R"==(for (l = 0; l < kk; l += UNROLL_K) { )==" "\n" |
188 | R"==(for (ll = 0; ll < UNROLL_K; ll++) { )==" "\n" |
189 | R"==(FLOATA a_val = (((i < m) && (l + ll < k)) ? a[a_offset] : 0); )==" "\n" |
190 | R"==(sa[sa_moffset + l * UNROLL_M + ll] = a_val; )==" "\n" |
191 | R"==(sumA -= a_val; )==" "\n" |
192 | R"==(#if defined(NN) || defined(NT) )==" "\n" |
193 | R"==(a_offset += lda; )==" "\n" |
194 | R"==(#else )==" "\n" |
195 | R"==(a_offset++; )==" "\n" |
196 | R"==(#endif )==" "\n" |
197 | R"==(} )==" "\n" |
198 | R"==(} )==" "\n" |
199 | R"==(xa[cid] = (FLOATC)ATTR_B0 * sumA; )==" "\n" |
200 | R"==(} )==" "\n" |
201 | R"==(for (int cid = cid0; cid < szy * UNROLL_N; cid += ctotal) { )==" "\n" |
202 | R"==(long sb_noffset = (cid & ~(UNROLL_N - 1)) * kk )==" "\n" |
203 | R"==(+ (cid & (UNROLL_N - 1)) * UNROLL_K; )==" "\n" |
204 | R"==(long j = cid + GROUPSIZE_N * gdy; )==" "\n" |
205 | R"==(FLOATC sumB = (FLOATC)ATTR_B0 * k; )==" "\n" |
206 | R"==(#if defined(NN) || defined(TN) )==" "\n" |
207 | R"==(long b_offset = j * ldb; )==" "\n" |
208 | R"==(#else )==" "\n" |
209 | R"==(long b_offset = j; )==" "\n" |
210 | R"==(#endif )==" "\n" |
211 | R"==(for (l = 0; l < kk; l += UNROLL_K) { )==" "\n" |
212 | R"==(for (ll = 0; ll < UNROLL_K; ll++) { )==" "\n" |
213 | R"==(FLOATB b_val = (((j < n) && (l + ll < k)) ? b[b_offset] : 0); )==" "\n" |
214 | R"==(sb[sb_noffset + l * UNROLL_N + ll] = b_val; )==" "\n" |
215 | R"==(sumB -= b_val; )==" "\n" |
216 | R"==(#if defined(NN) || defined(TN) )==" "\n" |
217 | R"==(b_offset++; )==" "\n" |
218 | R"==(#else )==" "\n" |
219 | R"==(b_offset += ldb; )==" "\n" |
220 | R"==(#endif )==" "\n" |
221 | R"==(} )==" "\n" |
222 | R"==(} )==" "\n" |
223 | R"==(xb[cid] = (FLOATC)ATTR_A0 * sumB; )==" "\n" |
224 | R"==(} )==" "\n" |
225 | R"==(barrier(CLK_LOCAL_MEM_FENCE); )==" "\n" |
226 | R"==(m -= GROUPSIZE_M * gdx + UNROLL_M * idx; )==" "\n" |
227 | R"==(if (m > UNROLL_M) m = UNROLL_M; )==" "\n" |
228 | R"==(n -= GROUPSIZE_N * gdy + UNROLL_N * idy; )==" "\n" |
229 | R"==(if (n > UNROLL_N) n = UNROLL_N; )==" "\n" |
230 | R"==(if ((m <= 0) || (n <= 0)) return; )==" "\n" |
231 | R"==(m -= UNROLL_M * lid / GRX; )==" "\n" |
232 | R"==(sa += UNROLL_M * kk * idx + UNROLL_M * UNROLL_K * lid / GRX; )==" "\n" |
233 | R"==(sb += UNROLL_N * kk * idy + UNROLL_K * lid; )==" "\n" |
234 | R"==(xa += UNROLL_M * idx + UNROLL_M * lid / GRX; )==" "\n" |
235 | R"==(xb += UNROLL_N * idy; )==" "\n" |
236 | R"==(FLOATC4 sc[UNROLL_M * UNROLL_N / GRX / 4] = {0}; )==" "\n" |
237 | R"==(for (l = 0; l < kk; l += UNROLL_K) { )==" "\n" |
238 | R"==(FLOATA4 a0, a1, a2, a3; )==" "\n" |
239 | R"==(FLOATB4 bb, b0, b1, b2, b3; )==" "\n" |
240 | R"==(a0 = ((__local FLOATA4 *)sa)[0]; )==" "\n" |
241 | R"==(a1 = ((__local FLOATA4 *)sa)[1]; )==" "\n" |
242 | R"==(a2 = ((__local FLOATA4 *)sa)[2]; )==" "\n" |
243 | R"==(a3 = ((__local FLOATA4 *)sa)[3]; )==" "\n" |
244 | R"==(for (ll = 0; ll < GRX / 4; ll++) { )==" "\n" |
245 | R"==(bb = ((__local FLOATB4 *)sb)[0]; )==" "\n" |
246 | R"==(b0 = SHUFFLE(bb, 0); )==" "\n" |
247 | R"==(b1 = SHUFFLE(bb, 1); )==" "\n" |
248 | R"==(b2 = SHUFFLE(bb, 2); )==" "\n" |
249 | R"==(b3 = SHUFFLE(bb, 3); )==" "\n" |
250 | R"==(sc[ll * 2 + 0] += ACCUMULATE(a0, b0, b1, b2, b3); )==" "\n" |
251 | R"==(sc[ll * 2 + 4] += ACCUMULATE(a1, b0, b1, b2, b3); )==" "\n" |
252 | R"==(sc[ll * 2 + 8] += ACCUMULATE(a2, b0, b1, b2, b3); )==" "\n" |
253 | R"==(sc[ll * 2 + 12] += ACCUMULATE(a3, b0, b1, b2, b3); )==" "\n" |
254 | R"==(b0 = SHUFFLE(bb, 4); )==" "\n" |
255 | R"==(b1 = SHUFFLE(bb, 5); )==" "\n" |
256 | R"==(b2 = SHUFFLE(bb, 6); )==" "\n" |
257 | R"==(b3 = SHUFFLE(bb, 7); )==" "\n" |
258 | R"==(sc[ll * 2 + 1] += ACCUMULATE(a0, b0, b1, b2, b3); )==" "\n" |
259 | R"==(sc[ll * 2 + 5] += ACCUMULATE(a1, b0, b1, b2, b3); )==" "\n" |
260 | R"==(sc[ll * 2 + 9] += ACCUMULATE(a2, b0, b1, b2, b3); )==" "\n" |
261 | R"==(sc[ll * 2 + 13] += ACCUMULATE(a3, b0, b1, b2, b3); )==" "\n" |
262 | R"==(sb += UNROLL_N * GRX / 4; )==" "\n" |
263 | R"==(} )==" "\n" |
264 | R"==(sa += UNROLL_M * UNROLL_K; )==" "\n" |
265 | R"==(} )==" "\n" |
266 | R"==(ADD_SCALE(0); )==" "\n" |
267 | R"==(ADD_SCALE(4); )==" "\n" |
268 | R"==(ADD_SCALE(8); )==" "\n" |
269 | R"==(ADD_SCALE(12); )==" "\n" |
270 | R"==(} )==" "\n" |
271 | R"==()==" ; |
272 | } |
273 | } |
274 | } |
275 | } |