Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-5x8c8-minmax-fp32-avxvnniint8-prfm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx8c8-avxvnni.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2024 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <immintrin.h>
13
14 #include "xnnpack/common.h"
15 #include "xnnpack/gemm.h"
16 #include "xnnpack/intrinsics-polyfill.h"
17 #include "xnnpack/math.h"
18 #include "xnnpack/unaligned.h"
19 #include "xnnpack/prefetch.h"
20
21
22 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_5x8c8__avxvnniint8_prfm(
23 size_t mr,
24 size_t nc,
25 size_t kc,
26 size_t ks,
27 const int8_t** restrict a,
28 const void* restrict w,
29 int8_t* restrict c,
30 size_t cm_stride,
31 size_t cn_stride,
32 size_t a_offset,
33 const int8_t* zero,
34 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
35 {
36 assert(mr != 0);
37 assert(mr <= 5);
38 assert(nc != 0);
39 assert(kc != 0);
40 assert(kc % sizeof(int8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 kc = round_up_po2(kc, 8 * sizeof(int8_t));
46 int8_t* c0 = c;
47 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 c1 = c0;
50 }
51 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 c2 = c1;
54 }
55 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr < 4) {
57 c3 = c2;
58 }
59 int8_t* c4 = (int8_t*) ((uintptr_t) c3 + cm_stride);
60 if XNN_UNPREDICTABLE(mr <= 4) {
61 c4 = c3;
62 }
63
64 const __m256 voutput_max_less_zero_point = _mm256_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point);
65 const __m256i voutput_zero_point = _mm256_set1_epi32(params->fp32_scalar.output_zero_point);
66 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
67 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
68 // XNN_FORCE_REALIZATION(voutput_zero_point);
69 // XNN_FORCE_REALIZATION(voutput_min);
70 do {
71 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm_load_si128((const __m128i*) w));
72 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm_load_si128((const __m128i*) ((const int32_t*) w + 4)));
73 __m256i vacc1x0123 = vacc0x0123;
74 __m256i vacc1x4567 = vacc0x4567;
75 __m256i vacc2x0123 = vacc0x0123;
76 __m256i vacc2x4567 = vacc0x4567;
77 __m256i vacc3x0123 = vacc0x0123;
78 __m256i vacc3x4567 = vacc0x4567;
79 __m256i vacc4x0123 = vacc0x0123;
80 __m256i vacc4x4567 = vacc0x4567;
81 w = (const int32_t*) w + 8;
82
83 size_t p = ks;
84 do {
85 const int8_t* restrict a0 = a[0];
86 if XNN_UNPREDICTABLE(a0 != zero) {
87 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
88 }
89 const int8_t* restrict a1 = a[1];
90 if XNN_UNPREDICTABLE(a1 != zero) {
91 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
92 }
93 const int8_t* restrict a2 = a[2];
94 if XNN_UNPREDICTABLE(a2 != zero) {
95 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
96 }
97 const int8_t* restrict a3 = a[3];
98 if XNN_UNPREDICTABLE(a3 != zero) {
99 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
100 }
101 const int8_t* restrict a4 = a[4];
102 if XNN_UNPREDICTABLE(a4 != zero) {
103 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
104 }
105 a += 5;
106
107 size_t k = kc;
108 while (k >= 16 * sizeof(int8_t)) {
109 const __m256i va0x01234567 = _mm256_set1_epi64x((int64_t) unaligned_load_u64(a0));
110 const __m256i va0x89ABCDEF = _mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8));
111 a0 += 16;
112 const __m256i va1x01234567 = _mm256_set1_epi64x((int64_t) unaligned_load_u64(a1));
113 const __m256i va1x89ABCDEF = _mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8));
114 a1 += 16;
115 const __m256i va2x01234567 = _mm256_set1_epi64x((int64_t) unaligned_load_u64(a2));
116 const __m256i va2x89ABCDEF = _mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8));
117 a2 += 16;
118 const __m256i va3x01234567 = _mm256_set1_epi64x((int64_t) unaligned_load_u64(a3));
119 const __m256i va3x89ABCDEF = _mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8));
120 a3 += 16;
121 const __m256i va4x01234567 = _mm256_set1_epi64x((int64_t) unaligned_load_u64(a4));
122 const __m256i va4x89ABCDEF = _mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8));
123 a4 += 16;
124
125 const __m256i vb01234567x0123 = _mm256_load_si256(w);
126 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
127 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
128 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
129 xnn_prefetch_to_l1((const int8_t*) w + 896);
130
131 vacc0x0123 = _mm256_dpbssd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
132 vacc0x4567 = _mm256_dpbssd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
133 vacc1x0123 = _mm256_dpbssd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
134 vacc1x4567 = _mm256_dpbssd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
135 vacc2x0123 = _mm256_dpbssd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
136 vacc2x4567 = _mm256_dpbssd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
137 vacc3x0123 = _mm256_dpbssd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
138 vacc3x4567 = _mm256_dpbssd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
139 vacc4x0123 = _mm256_dpbssd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
140 vacc4x4567 = _mm256_dpbssd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
141 xnn_prefetch_to_l1((const int8_t*) w + 960);
142 vacc0x0123 = _mm256_dpbssd_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
143 vacc0x4567 = _mm256_dpbssd_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
144 vacc1x0123 = _mm256_dpbssd_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
145 vacc1x4567 = _mm256_dpbssd_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
146 vacc2x0123 = _mm256_dpbssd_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
147 vacc2x4567 = _mm256_dpbssd_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
148 vacc3x0123 = _mm256_dpbssd_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
149 vacc3x4567 = _mm256_dpbssd_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
150 vacc4x0123 = _mm256_dpbssd_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
151 vacc4x4567 = _mm256_dpbssd_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
152
153 w = (const int8_t*) w + 128;
154 k -= 16 * sizeof(int8_t);
155 }
156
157 if (k != 0) {
158 const __m256i va0x01234567 = _mm256_set1_epi64x((int64_t) unaligned_load_u64(a0));
159 a0 += 8;
160 const __m256i va1x01234567 = _mm256_set1_epi64x((int64_t) unaligned_load_u64(a1));
161 a1 += 8;
162 const __m256i va2x01234567 = _mm256_set1_epi64x((int64_t) unaligned_load_u64(a2));
163 a2 += 8;
164 const __m256i va3x01234567 = _mm256_set1_epi64x((int64_t) unaligned_load_u64(a3));
165 a3 += 8;
166 const __m256i va4x01234567 = _mm256_set1_epi64x((int64_t) unaligned_load_u64(a4));
167 a4 += 8;
168
169 const __m256i vb01234567x0123 = _mm256_load_si256(w);
170 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
171
172 vacc0x0123 = _mm256_dpbssd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
173 vacc0x4567 = _mm256_dpbssd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
174 vacc1x0123 = _mm256_dpbssd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
175 vacc1x4567 = _mm256_dpbssd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
176 vacc2x0123 = _mm256_dpbssd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
177 vacc2x4567 = _mm256_dpbssd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
178 vacc3x0123 = _mm256_dpbssd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
179 vacc3x4567 = _mm256_dpbssd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
180 vacc4x0123 = _mm256_dpbssd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
181 vacc4x4567 = _mm256_dpbssd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
182 xnn_prefetch_to_l1((const int8_t*) w + 960);
183
184 w = (const int8_t*) w + 64;
185 k -= 8 * sizeof(int8_t);
186 }
187
188 p -= 5 * sizeof(void*);
189 } while (p != 0);
190
191
192 // Add adjacent pairs
193 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
194 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
195 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
196 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
197 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
198 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
199 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
200 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
201 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
202 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
203
204 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
205 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
206 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
207 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
208 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
209
210 const __m256 vscale01234567 = _mm256_load_ps(w);
211 w = (const float*) w + 8;
212 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vscale01234567);
213 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vscale01234567);
214 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vscale01234567);
215 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vscale01234567);
216 vout4x01234567 = _mm256_mul_ps(vout4x01234567, vscale01234567);
217
218 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max_less_zero_point);
219 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max_less_zero_point);
220 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max_less_zero_point);
221 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max_less_zero_point);
222 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max_less_zero_point);
223
224 vacc0x01234567 = _mm256_cvtps_epi32(vout0x01234567);
225 vacc1x01234567 = _mm256_cvtps_epi32(vout1x01234567);
226 vacc2x01234567 = _mm256_cvtps_epi32(vout2x01234567);
227 vacc3x01234567 = _mm256_cvtps_epi32(vout3x01234567);
228 vacc4x01234567 = _mm256_cvtps_epi32(vout4x01234567);
229
230 vacc0x01234567 = _mm256_add_epi32(vacc0x01234567, voutput_zero_point);
231 vacc1x01234567 = _mm256_add_epi32(vacc1x01234567, voutput_zero_point);
232 vacc2x01234567 = _mm256_add_epi32(vacc2x01234567, voutput_zero_point);
233 vacc3x01234567 = _mm256_add_epi32(vacc3x01234567, voutput_zero_point);
234 vacc4x01234567 = _mm256_add_epi32(vacc4x01234567, voutput_zero_point);
235
236 vacc0x01234567 = _mm256_packs_epi32(vacc0x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc0x01234567, 1)));
237 __m128i voutb0x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc0x01234567, vacc0x01234567));
238 vacc1x01234567 = _mm256_packs_epi32(vacc1x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc1x01234567, 1)));
239 __m128i voutb1x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc1x01234567, vacc1x01234567));
240 vacc2x01234567 = _mm256_packs_epi32(vacc2x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc2x01234567, 1)));
241 __m128i voutb2x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc2x01234567, vacc2x01234567));
242 vacc3x01234567 = _mm256_packs_epi32(vacc3x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc3x01234567, 1)));
243 __m128i voutb3x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc3x01234567, vacc3x01234567));
244 vacc4x01234567 = _mm256_packs_epi32(vacc4x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc4x01234567, 1)));
245 __m128i voutb4x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc4x01234567, vacc4x01234567));
246
247 voutb0x01234567 = _mm_max_epi8(voutb0x01234567, voutput_min);
248 voutb1x01234567 = _mm_max_epi8(voutb1x01234567, voutput_min);
249 voutb2x01234567 = _mm_max_epi8(voutb2x01234567, voutput_min);
250 voutb3x01234567 = _mm_max_epi8(voutb3x01234567, voutput_min);
251 voutb4x01234567 = _mm_max_epi8(voutb4x01234567, voutput_min);
252
253 if (nc >= 8) {
254 _mm_storel_epi64((__m128i*) c4, voutb4x01234567);
255 c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
256 _mm_storel_epi64((__m128i*) c3, voutb3x01234567);
257 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
258 _mm_storel_epi64((__m128i*) c2, voutb2x01234567);
259 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
260 _mm_storel_epi64((__m128i*) c1, voutb1x01234567);
261 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
262 _mm_storel_epi64((__m128i*) c0, voutb0x01234567);
263 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
264 a = (const int8_t**restrict) ((uintptr_t) a - ks);
265 nc -= 8;
266 } else {
267 if (nc & 4) {
268 _mm_storeu_si32(c4, voutb4x01234567);
269 c4 += 4;
270 _mm_storeu_si32(c3, voutb3x01234567);
271 c3 += 4;
272 _mm_storeu_si32(c2, voutb2x01234567);
273 c2 += 4;
274 _mm_storeu_si32(c1, voutb1x01234567);
275 c1 += 4;
276 _mm_storeu_si32(c0, voutb0x01234567);
277 c0 += 4;
278 voutb4x01234567 = _mm_srli_epi64(voutb4x01234567, 32);
279 voutb3x01234567 = _mm_srli_epi64(voutb3x01234567, 32);
280 voutb2x01234567 = _mm_srli_epi64(voutb2x01234567, 32);
281 voutb1x01234567 = _mm_srli_epi64(voutb1x01234567, 32);
282 voutb0x01234567 = _mm_srli_epi64(voutb0x01234567, 32);
283 }
284 if (nc & 2) {
285 unaligned_store_u16(c4, (uint16_t) _mm_extract_epi16(voutb4x01234567, 0));
286 c4 += 2;
287 unaligned_store_u16(c3, (uint16_t) _mm_extract_epi16(voutb3x01234567, 0));
288 c3 += 2;
289 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(voutb2x01234567, 0));
290 c2 += 2;
291 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(voutb1x01234567, 0));
292 c1 += 2;
293 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(voutb0x01234567, 0));
294 c0 += 2;
295 voutb4x01234567 = _mm_srli_epi32(voutb4x01234567, 16);
296 voutb3x01234567 = _mm_srli_epi32(voutb3x01234567, 16);
297 voutb2x01234567 = _mm_srli_epi32(voutb2x01234567, 16);
298 voutb1x01234567 = _mm_srli_epi32(voutb1x01234567, 16);
299 voutb0x01234567 = _mm_srli_epi32(voutb0x01234567, 16);
300 }
301 if (nc & 1) {
302 *c4 = (int8_t) _mm_extract_epi8(voutb4x01234567, 0);
303 *c3 = (int8_t) _mm_extract_epi8(voutb3x01234567, 0);
304 *c2 = (int8_t) _mm_extract_epi8(voutb2x01234567, 0);
305 *c1 = (int8_t) _mm_extract_epi8(voutb1x01234567, 0);
306 *c0 = (int8_t) _mm_extract_epi8(voutb0x01234567, 0);
307 }
308 nc = 0;
309 }
310 } while (nc != 0);
311 }