Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-5x8c8-minmax-fp32-avxvnni.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
20
21 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_5x8c8__avxvnni(
22 size_t mr,
23 size_t nc,
24 size_t kc,
25 size_t ks,
26 const int8_t** restrict a,
27 const void* restrict w,
28 int8_t* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 size_t a_offset,
32 const int8_t* zero,
33 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
34 {
35 assert(mr != 0);
36 assert(mr <= 5);
37 assert(nc != 0);
38 assert(kc != 0);
39 assert(kc % sizeof(int8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 kc = round_up_po2(kc, 8 * sizeof(int8_t));
45 int8_t* c0 = c;
46 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr < 2) {
48 c1 = c0;
49 }
50 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 c2 = c1;
53 }
54 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
55 if XNN_UNPREDICTABLE(mr < 4) {
56 c3 = c2;
57 }
58 int8_t* c4 = (int8_t*) ((uintptr_t) c3 + cm_stride);
59 if XNN_UNPREDICTABLE(mr <= 4) {
60 c4 = c3;
61 }
62
63 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
64 XNN_FORCE_REALIZATION(vsign_mask);
65 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);
66 const __m256i voutput_zero_point = _mm256_set1_epi32(params->fp32_scalar.output_zero_point);
67 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
68 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
69 // XNN_FORCE_REALIZATION(voutput_zero_point);
70 // XNN_FORCE_REALIZATION(voutput_min);
71 do {
72 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm_load_si128((const __m128i*) w));
73 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm_load_si128((const __m128i*) ((const int32_t*) w + 4)));
74 __m256i vacc1x0123 = vacc0x0123;
75 __m256i vacc1x4567 = vacc0x4567;
76 __m256i vacc2x0123 = vacc0x0123;
77 __m256i vacc2x4567 = vacc0x4567;
78 __m256i vacc3x0123 = vacc0x0123;
79 __m256i vacc3x4567 = vacc0x4567;
80 __m256i vacc4x0123 = vacc0x0123;
81 __m256i vacc4x4567 = vacc0x4567;
82 w = (const int32_t*) w + 8;
83
84 size_t p = ks;
85 do {
86 const int8_t* restrict a0 = a[0];
87 if XNN_UNPREDICTABLE(a0 != zero) {
88 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
89 }
90 const int8_t* restrict a1 = a[1];
91 if XNN_UNPREDICTABLE(a1 != zero) {
92 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
93 }
94 const int8_t* restrict a2 = a[2];
95 if XNN_UNPREDICTABLE(a2 != zero) {
96 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
97 }
98 const int8_t* restrict a3 = a[3];
99 if XNN_UNPREDICTABLE(a3 != zero) {
100 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
101 }
102 const int8_t* restrict a4 = a[4];
103 if XNN_UNPREDICTABLE(a4 != zero) {
104 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
105 }
106 a += 5;
107
108 size_t k = kc;
109 while (k >= 16 * sizeof(int8_t)) {
110 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
111 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
112 a0 += 16;
113 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
114 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
115 a1 += 16;
116 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
117 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
118 a2 += 16;
119 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
120 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
121 a3 += 16;
122 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
123 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
124 a4 += 16;
125
126 const __m256i vb01234567x0123 = _mm256_load_si256(w);
127 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
128 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
129 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
130
131 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
132 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
133 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
134 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
135 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
136 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
137 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
138 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
139 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
140 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
141 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
142 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
143 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
144 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
145 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
146 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
147 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
148 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
149 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
150 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
151
152 w = (const int8_t*) w + 128;
153 k -= 16 * sizeof(int8_t);
154 }
155
156 if (k != 0) {
157 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
158 a0 += 8;
159 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
160 a1 += 8;
161 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
162 a2 += 8;
163 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
164 a3 += 8;
165 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
166 a4 += 8;
167
168 const __m256i vb01234567x0123 = _mm256_load_si256(w);
169 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
170
171 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
172 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
173 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
174 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
175 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
176 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
177 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
178 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
179 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
180 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
181
182 w = (const int8_t*) w + 64;
183 k -= 8 * sizeof(int8_t);
184 }
185
186 p -= 5 * sizeof(void*);
187 } while (p != 0);
188
189
190 // Add adjacent pairs
191 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
192 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
193 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
194 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
195 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
196 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
197 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
198 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
199 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
200 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
201
202 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
203 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
204 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
205 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
206 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
207
208 const __m256 vscale01234567 = _mm256_load_ps(w);
209 w = (const float*) w + 8;
210 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vscale01234567);
211 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vscale01234567);
212 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vscale01234567);
213 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vscale01234567);
214 vout4x01234567 = _mm256_mul_ps(vout4x01234567, vscale01234567);
215
216 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max_less_zero_point);
217 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max_less_zero_point);
218 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max_less_zero_point);
219 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max_less_zero_point);
220 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max_less_zero_point);
221
222 vacc0x01234567 = _mm256_cvtps_epi32(vout0x01234567);
223 vacc1x01234567 = _mm256_cvtps_epi32(vout1x01234567);
224 vacc2x01234567 = _mm256_cvtps_epi32(vout2x01234567);
225 vacc3x01234567 = _mm256_cvtps_epi32(vout3x01234567);
226 vacc4x01234567 = _mm256_cvtps_epi32(vout4x01234567);
227
228 vacc0x01234567 = _mm256_add_epi32(vacc0x01234567, voutput_zero_point);
229 vacc1x01234567 = _mm256_add_epi32(vacc1x01234567, voutput_zero_point);
230 vacc2x01234567 = _mm256_add_epi32(vacc2x01234567, voutput_zero_point);
231 vacc3x01234567 = _mm256_add_epi32(vacc3x01234567, voutput_zero_point);
232 vacc4x01234567 = _mm256_add_epi32(vacc4x01234567, voutput_zero_point);
233
234 vacc0x01234567 = _mm256_packs_epi32(vacc0x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc0x01234567, 1)));
235 __m128i voutb0x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc0x01234567, vacc0x01234567));
236 vacc1x01234567 = _mm256_packs_epi32(vacc1x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc1x01234567, 1)));
237 __m128i voutb1x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc1x01234567, vacc1x01234567));
238 vacc2x01234567 = _mm256_packs_epi32(vacc2x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc2x01234567, 1)));
239 __m128i voutb2x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc2x01234567, vacc2x01234567));
240 vacc3x01234567 = _mm256_packs_epi32(vacc3x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc3x01234567, 1)));
241 __m128i voutb3x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc3x01234567, vacc3x01234567));
242 vacc4x01234567 = _mm256_packs_epi32(vacc4x01234567, _mm256_castsi128_si256(_mm256_extracti128_si256(vacc4x01234567, 1)));
243 __m128i voutb4x01234567 = _mm256_castsi256_si128(_mm256_packs_epi16(vacc4x01234567, vacc4x01234567));
244
245 voutb0x01234567 = _mm_max_epi8(voutb0x01234567, voutput_min);
246 voutb1x01234567 = _mm_max_epi8(voutb1x01234567, voutput_min);
247 voutb2x01234567 = _mm_max_epi8(voutb2x01234567, voutput_min);
248 voutb3x01234567 = _mm_max_epi8(voutb3x01234567, voutput_min);
249 voutb4x01234567 = _mm_max_epi8(voutb4x01234567, voutput_min);
250
251 if (nc >= 8) {
252 _mm_storel_epi64((__m128i*) c4, voutb4x01234567);
253 c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
254 _mm_storel_epi64((__m128i*) c3, voutb3x01234567);
255 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
256 _mm_storel_epi64((__m128i*) c2, voutb2x01234567);
257 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
258 _mm_storel_epi64((__m128i*) c1, voutb1x01234567);
259 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
260 _mm_storel_epi64((__m128i*) c0, voutb0x01234567);
261 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
262 a = (const int8_t**restrict) ((uintptr_t) a - ks);
263 nc -= 8;
264 } else {
265 if (nc & 4) {
266 _mm_storeu_si32(c4, voutb4x01234567);
267 c4 += 4;
268 _mm_storeu_si32(c3, voutb3x01234567);
269 c3 += 4;
270 _mm_storeu_si32(c2, voutb2x01234567);
271 c2 += 4;
272 _mm_storeu_si32(c1, voutb1x01234567);
273 c1 += 4;
274 _mm_storeu_si32(c0, voutb0x01234567);
275 c0 += 4;
276 voutb4x01234567 = _mm_srli_epi64(voutb4x01234567, 32);
277 voutb3x01234567 = _mm_srli_epi64(voutb3x01234567, 32);
278 voutb2x01234567 = _mm_srli_epi64(voutb2x01234567, 32);
279 voutb1x01234567 = _mm_srli_epi64(voutb1x01234567, 32);
280 voutb0x01234567 = _mm_srli_epi64(voutb0x01234567, 32);
281 }
282 if (nc & 2) {
283 unaligned_store_u16(c4, (uint16_t) _mm_extract_epi16(voutb4x01234567, 0));
284 c4 += 2;
285 unaligned_store_u16(c3, (uint16_t) _mm_extract_epi16(voutb3x01234567, 0));
286 c3 += 2;
287 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(voutb2x01234567, 0));
288 c2 += 2;
289 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(voutb1x01234567, 0));
290 c1 += 2;
291 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(voutb0x01234567, 0));
292 c0 += 2;
293 voutb4x01234567 = _mm_srli_epi32(voutb4x01234567, 16);
294 voutb3x01234567 = _mm_srli_epi32(voutb3x01234567, 16);
295 voutb2x01234567 = _mm_srli_epi32(voutb2x01234567, 16);
296 voutb1x01234567 = _mm_srli_epi32(voutb1x01234567, 16);
297 voutb0x01234567 = _mm_srli_epi32(voutb0x01234567, 16);
298 }
299 if (nc & 1) {
300 *c4 = (int8_t) _mm_extract_epi8(voutb4x01234567, 0);
301 *c3 = (int8_t) _mm_extract_epi8(voutb3x01234567, 0);
302 *c2 = (int8_t) _mm_extract_epi8(voutb2x01234567, 0);
303 *c1 = (int8_t) _mm_extract_epi8(voutb1x01234567, 0);
304 *c0 = (int8_t) _mm_extract_epi8(voutb0x01234567, 0);
305 }
306 nc = 0;
307 }
308 } while (nc != 0);
309 }