Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-14x8c8-minmax-fp32-avx256vnni.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_14x8c8__avx256vnni(
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 <= 14);
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 int8_t* c5 = (int8_t*) ((uintptr_t) c4 + cm_stride);
63 if XNN_UNPREDICTABLE(mr < 6) {
64 c5 = c4;
65 }
66 int8_t* c6 = (int8_t*) ((uintptr_t) c5 + cm_stride);
67 if XNN_UNPREDICTABLE(mr <= 6) {
68 c6 = c5;
69 }
70 int8_t* c7 = (int8_t*) ((uintptr_t) c6 + cm_stride);
71 if XNN_UNPREDICTABLE(mr < 8) {
72 c7 = c6;
73 }
74 int8_t* c8 = (int8_t*) ((uintptr_t) c7 + cm_stride);
75 if XNN_UNPREDICTABLE(mr <= 8) {
76 c8 = c7;
77 }
78 int8_t* c9 = (int8_t*) ((uintptr_t) c8 + cm_stride);
79 if XNN_UNPREDICTABLE(mr < 10) {
80 c9 = c8;
81 }
82 int8_t* c10 = (int8_t*) ((uintptr_t) c9 + cm_stride);
83 if XNN_UNPREDICTABLE(mr <= 10) {
84 c10 = c9;
85 }
86 int8_t* c11 = (int8_t*) ((uintptr_t) c10 + cm_stride);
87 if XNN_UNPREDICTABLE(mr < 12) {
88 c11 = c10;
89 }
90 int8_t* c12 = (int8_t*) ((uintptr_t) c11 + cm_stride);
91 if XNN_UNPREDICTABLE(mr <= 12) {
92 c12 = c11;
93 }
94 int8_t* c13 = (int8_t*) ((uintptr_t) c12 + cm_stride);
95 if XNN_UNPREDICTABLE(mr != 14) {
96 c13 = c12;
97 }
98
99 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
100 XNN_FORCE_REALIZATION(vsign_mask);
101 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);
102 const __m256i voutput_zero_point = _mm256_set1_epi32(params->fp32_scalar.output_zero_point);
103 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
104 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
105 // XNN_FORCE_REALIZATION(voutput_zero_point);
106 // XNN_FORCE_REALIZATION(voutput_min);
107 do {
108 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm_load_si128((const __m128i*) w));
109 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm_load_si128((const __m128i*) ((const int32_t*) w + 4)));
110 __m256i vacc1x0123 = vacc0x0123;
111 __m256i vacc1x4567 = vacc0x4567;
112 __m256i vacc2x0123 = vacc0x0123;
113 __m256i vacc2x4567 = vacc0x4567;
114 __m256i vacc3x0123 = vacc0x0123;
115 __m256i vacc3x4567 = vacc0x4567;
116 __m256i vacc4x0123 = vacc0x0123;
117 __m256i vacc4x4567 = vacc0x4567;
118 __m256i vacc5x0123 = vacc0x0123;
119 __m256i vacc5x4567 = vacc0x4567;
120 __m256i vacc6x0123 = vacc0x0123;
121 __m256i vacc6x4567 = vacc0x4567;
122 __m256i vacc7x0123 = vacc0x0123;
123 __m256i vacc7x4567 = vacc0x4567;
124 __m256i vacc8x0123 = vacc0x0123;
125 __m256i vacc8x4567 = vacc0x4567;
126 __m256i vacc9x0123 = vacc0x0123;
127 __m256i vacc9x4567 = vacc0x4567;
128 __m256i vacc10x0123 = vacc0x0123;
129 __m256i vacc10x4567 = vacc0x4567;
130 __m256i vacc11x0123 = vacc0x0123;
131 __m256i vacc11x4567 = vacc0x4567;
132 __m256i vacc12x0123 = vacc0x0123;
133 __m256i vacc12x4567 = vacc0x4567;
134 __m256i vacc13x0123 = vacc0x0123;
135 __m256i vacc13x4567 = vacc0x4567;
136 w = (const int32_t*) w + 8;
137
138 size_t p = ks;
139 do {
140 const int8_t* restrict a0 = a[0];
141 if XNN_UNPREDICTABLE(a0 != zero) {
142 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
143 }
144 const int8_t* restrict a1 = a[1];
145 if XNN_UNPREDICTABLE(a1 != zero) {
146 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
147 }
148 const int8_t* restrict a2 = a[2];
149 if XNN_UNPREDICTABLE(a2 != zero) {
150 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
151 }
152 const int8_t* restrict a3 = a[3];
153 if XNN_UNPREDICTABLE(a3 != zero) {
154 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
155 }
156 const int8_t* restrict a4 = a[4];
157 if XNN_UNPREDICTABLE(a4 != zero) {
158 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
159 }
160 const int8_t* restrict a5 = a[5];
161 if XNN_UNPREDICTABLE(a5 != zero) {
162 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
163 }
164 const int8_t* restrict a6 = a[6];
165 if XNN_UNPREDICTABLE(a6 != zero) {
166 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
167 }
168 const int8_t* restrict a7 = a[7];
169 if XNN_UNPREDICTABLE(a7 != zero) {
170 a7 = (const int8_t*) ((uintptr_t) a7 + a_offset);
171 }
172 const int8_t* restrict a8 = a[8];
173 if XNN_UNPREDICTABLE(a8 != zero) {
174 a8 = (const int8_t*) ((uintptr_t) a8 + a_offset);
175 }
176 const int8_t* restrict a9 = a[9];
177 if XNN_UNPREDICTABLE(a9 != zero) {
178 a9 = (const int8_t*) ((uintptr_t) a9 + a_offset);
179 }
180 const int8_t* restrict a10 = a[10];
181 if XNN_UNPREDICTABLE(a10 != zero) {
182 a10 = (const int8_t*) ((uintptr_t) a10 + a_offset);
183 }
184 const int8_t* restrict a11 = a[11];
185 if XNN_UNPREDICTABLE(a11 != zero) {
186 a11 = (const int8_t*) ((uintptr_t) a11 + a_offset);
187 }
188 const int8_t* restrict a12 = a[12];
189 if XNN_UNPREDICTABLE(a12 != zero) {
190 a12 = (const int8_t*) ((uintptr_t) a12 + a_offset);
191 }
192 const int8_t* restrict a13 = a[13];
193 if XNN_UNPREDICTABLE(a13 != zero) {
194 a13 = (const int8_t*) ((uintptr_t) a13 + a_offset);
195 }
196 a += 14;
197
198 size_t k = kc;
199 while (k >= 16 * sizeof(int8_t)) {
200 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
201 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
202 a0 += 16;
203 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
204 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
205 a1 += 16;
206 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
207 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
208 a2 += 16;
209 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
210 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
211 a3 += 16;
212 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
213 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
214 a4 += 16;
215 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
216 const __m256i va5x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
217 a5 += 16;
218 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
219 const __m256i va6x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
220 a6 += 16;
221 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
222 const __m256i va7x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7 + 8)), vsign_mask);
223 a7 += 16;
224 const __m256i va8x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8)), vsign_mask);
225 const __m256i va8x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8 + 8)), vsign_mask);
226 a8 += 16;
227 const __m256i va9x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9)), vsign_mask);
228 const __m256i va9x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9 + 8)), vsign_mask);
229 a9 += 16;
230 const __m256i va10x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a10)), vsign_mask);
231 const __m256i va10x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a10 + 8)), vsign_mask);
232 a10 += 16;
233 const __m256i va11x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a11)), vsign_mask);
234 const __m256i va11x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a11 + 8)), vsign_mask);
235 a11 += 16;
236 const __m256i va12x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a12)), vsign_mask);
237 const __m256i va12x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a12 + 8)), vsign_mask);
238 a12 += 16;
239 const __m256i va13x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a13)), vsign_mask);
240 const __m256i va13x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a13 + 8)), vsign_mask);
241 a13 += 16;
242
243 const __m256i vb01234567x0123 = _mm256_load_si256(w);
244 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
245 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
246 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
247
248 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
249 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
250 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
251 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
252 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
253 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
254 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
255 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
256 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
257 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
258 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
259 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
260 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
261 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
262 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
263 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
264 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x01234567, vb01234567x0123);
265 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x01234567, vb89ABCDEFx0123);
266 vacc9x0123 = _mm256_dpbusd_epi32(vacc9x0123, va9x01234567, vb01234567x0123);
267 vacc9x4567 = _mm256_dpbusd_epi32(vacc9x4567, va9x01234567, vb89ABCDEFx0123);
268 vacc10x0123 = _mm256_dpbusd_epi32(vacc10x0123, va10x01234567, vb01234567x0123);
269 vacc10x4567 = _mm256_dpbusd_epi32(vacc10x4567, va10x01234567, vb89ABCDEFx0123);
270 vacc11x0123 = _mm256_dpbusd_epi32(vacc11x0123, va11x01234567, vb01234567x0123);
271 vacc11x4567 = _mm256_dpbusd_epi32(vacc11x4567, va11x01234567, vb89ABCDEFx0123);
272 vacc12x0123 = _mm256_dpbusd_epi32(vacc12x0123, va12x01234567, vb01234567x0123);
273 vacc12x4567 = _mm256_dpbusd_epi32(vacc12x4567, va12x01234567, vb89ABCDEFx0123);
274 vacc13x0123 = _mm256_dpbusd_epi32(vacc13x0123, va13x01234567, vb01234567x0123);
275 vacc13x4567 = _mm256_dpbusd_epi32(vacc13x4567, va13x01234567, vb89ABCDEFx0123);
276 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
277 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
278 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
279 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
280 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
281 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
282 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
283 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
284 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
285 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
286 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x89ABCDEF, vb01234567x4567);
287 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);
288 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x89ABCDEF, vb01234567x4567);
289 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x89ABCDEF, vb89ABCDEFx4567);
290 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x89ABCDEF, vb01234567x4567);
291 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x89ABCDEF, vb89ABCDEFx4567);
292 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x89ABCDEF, vb01234567x4567);
293 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x89ABCDEF, vb89ABCDEFx4567);
294 vacc9x0123 = _mm256_dpbusd_epi32(vacc9x0123, va9x89ABCDEF, vb01234567x4567);
295 vacc9x4567 = _mm256_dpbusd_epi32(vacc9x4567, va9x89ABCDEF, vb89ABCDEFx4567);
296 vacc10x0123 = _mm256_dpbusd_epi32(vacc10x0123, va10x89ABCDEF, vb01234567x4567);
297 vacc10x4567 = _mm256_dpbusd_epi32(vacc10x4567, va10x89ABCDEF, vb89ABCDEFx4567);
298 vacc11x0123 = _mm256_dpbusd_epi32(vacc11x0123, va11x89ABCDEF, vb01234567x4567);
299 vacc11x4567 = _mm256_dpbusd_epi32(vacc11x4567, va11x89ABCDEF, vb89ABCDEFx4567);
300 vacc12x0123 = _mm256_dpbusd_epi32(vacc12x0123, va12x89ABCDEF, vb01234567x4567);
301 vacc12x4567 = _mm256_dpbusd_epi32(vacc12x4567, va12x89ABCDEF, vb89ABCDEFx4567);
302 vacc13x0123 = _mm256_dpbusd_epi32(vacc13x0123, va13x89ABCDEF, vb01234567x4567);
303 vacc13x4567 = _mm256_dpbusd_epi32(vacc13x4567, va13x89ABCDEF, vb89ABCDEFx4567);
304
305 w = (const int8_t*) w + 128;
306 k -= 16 * sizeof(int8_t);
307 }
308
309 if (k != 0) {
310 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
311 a0 += 8;
312 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
313 a1 += 8;
314 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
315 a2 += 8;
316 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
317 a3 += 8;
318 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
319 a4 += 8;
320 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
321 a5 += 8;
322 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
323 a6 += 8;
324 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
325 a7 += 8;
326 const __m256i va8x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8)), vsign_mask);
327 a8 += 8;
328 const __m256i va9x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9)), vsign_mask);
329 a9 += 8;
330 const __m256i va10x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a10)), vsign_mask);
331 a10 += 8;
332 const __m256i va11x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a11)), vsign_mask);
333 a11 += 8;
334 const __m256i va12x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a12)), vsign_mask);
335 a12 += 8;
336 const __m256i va13x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a13)), vsign_mask);
337 a13 += 8;
338
339 const __m256i vb01234567x0123 = _mm256_load_si256(w);
340 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
341
342 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
343 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
344 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
345 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
346 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
347 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
348 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
349 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
350 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
351 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
352 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
353 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
354 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
355 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
356 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
357 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
358 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x01234567, vb01234567x0123);
359 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x01234567, vb89ABCDEFx0123);
360 vacc9x0123 = _mm256_dpbusd_epi32(vacc9x0123, va9x01234567, vb01234567x0123);
361 vacc9x4567 = _mm256_dpbusd_epi32(vacc9x4567, va9x01234567, vb89ABCDEFx0123);
362 vacc10x0123 = _mm256_dpbusd_epi32(vacc10x0123, va10x01234567, vb01234567x0123);
363 vacc10x4567 = _mm256_dpbusd_epi32(vacc10x4567, va10x01234567, vb89ABCDEFx0123);
364 vacc11x0123 = _mm256_dpbusd_epi32(vacc11x0123, va11x01234567, vb01234567x0123);
365 vacc11x4567 = _mm256_dpbusd_epi32(vacc11x4567, va11x01234567, vb89ABCDEFx0123);
366 vacc12x0123 = _mm256_dpbusd_epi32(vacc12x0123, va12x01234567, vb01234567x0123);
367 vacc12x4567 = _mm256_dpbusd_epi32(vacc12x4567, va12x01234567, vb89ABCDEFx0123);
368 vacc13x0123 = _mm256_dpbusd_epi32(vacc13x0123, va13x01234567, vb01234567x0123);
369 vacc13x4567 = _mm256_dpbusd_epi32(vacc13x4567, va13x01234567, vb89ABCDEFx0123);
370
371 w = (const int8_t*) w + 64;
372 k -= 8 * sizeof(int8_t);
373 }
374
375 p -= 14 * sizeof(void*);
376 } while (p != 0);
377
378
379 // Add adjacent pairs
380 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
381 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
382 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
383 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
384 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
385 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
386 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
387 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
388 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
389 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
390 const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);
391 __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));
392 const __m256i vsum6x02134657 = _mm256_hadd_epi32(vacc6x0123, vacc6x4567);
393 __m256i vacc6x01234567 = _mm256_permute4x64_epi64(vsum6x02134657, _MM_SHUFFLE(3, 1, 2, 0));
394 const __m256i vsum7x02134657 = _mm256_hadd_epi32(vacc7x0123, vacc7x4567);
395 __m256i vacc7x01234567 = _mm256_permute4x64_epi64(vsum7x02134657, _MM_SHUFFLE(3, 1, 2, 0));
396 const __m256i vsum8x02134657 = _mm256_hadd_epi32(vacc8x0123, vacc8x4567);
397 __m256i vacc8x01234567 = _mm256_permute4x64_epi64(vsum8x02134657, _MM_SHUFFLE(3, 1, 2, 0));
398 const __m256i vsum9x02134657 = _mm256_hadd_epi32(vacc9x0123, vacc9x4567);
399 __m256i vacc9x01234567 = _mm256_permute4x64_epi64(vsum9x02134657, _MM_SHUFFLE(3, 1, 2, 0));
400 const __m256i vsum10x02134657 = _mm256_hadd_epi32(vacc10x0123, vacc10x4567);
401 __m256i vacc10x01234567 = _mm256_permute4x64_epi64(vsum10x02134657, _MM_SHUFFLE(3, 1, 2, 0));
402 const __m256i vsum11x02134657 = _mm256_hadd_epi32(vacc11x0123, vacc11x4567);
403 __m256i vacc11x01234567 = _mm256_permute4x64_epi64(vsum11x02134657, _MM_SHUFFLE(3, 1, 2, 0));
404 const __m256i vsum12x02134657 = _mm256_hadd_epi32(vacc12x0123, vacc12x4567);
405 __m256i vacc12x01234567 = _mm256_permute4x64_epi64(vsum12x02134657, _MM_SHUFFLE(3, 1, 2, 0));
406 const __m256i vsum13x02134657 = _mm256_hadd_epi32(vacc13x0123, vacc13x4567);
407 __m256i vacc13x01234567 = _mm256_permute4x64_epi64(vsum13x02134657, _MM_SHUFFLE(3, 1, 2, 0));
408
409 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
410 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
411 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
412 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
413 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
414 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
415 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567);
416 __m256 vout7x01234567 = _mm256_cvtepi32_ps(vacc7x01234567);
417 __m256 vout8x01234567 = _mm256_cvtepi32_ps(vacc8x01234567);
418 __m256 vout9x01234567 = _mm256_cvtepi32_ps(vacc9x01234567);
419 __m256 vout10x01234567 = _mm256_cvtepi32_ps(vacc10x01234567);
420 __m256 vout11x01234567 = _mm256_cvtepi32_ps(vacc11x01234567);
421 __m256 vout12x01234567 = _mm256_cvtepi32_ps(vacc12x01234567);
422 __m256 vout13x01234567 = _mm256_cvtepi32_ps(vacc13x01234567);
423
424 const __m256 vscale01234567 = _mm256_load_ps(w);
425 w = (const float*) w + 8;
426 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vscale01234567);
427 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vscale01234567);
428 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vscale01234567);
429 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vscale01234567);
430 vout4x01234567 = _mm256_mul_ps(vout4x01234567, vscale01234567);
431 vout5x01234567 = _mm256_mul_ps(vout5x01234567, vscale01234567);
432 vout6x01234567 = _mm256_mul_ps(vout6x01234567, vscale01234567);
433 vout7x01234567 = _mm256_mul_ps(vout7x01234567, vscale01234567);
434 vout8x01234567 = _mm256_mul_ps(vout8x01234567, vscale01234567);
435 vout9x01234567 = _mm256_mul_ps(vout9x01234567, vscale01234567);
436 vout10x01234567 = _mm256_mul_ps(vout10x01234567, vscale01234567);
437 vout11x01234567 = _mm256_mul_ps(vout11x01234567, vscale01234567);
438 vout12x01234567 = _mm256_mul_ps(vout12x01234567, vscale01234567);
439 vout13x01234567 = _mm256_mul_ps(vout13x01234567, vscale01234567);
440
441 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max_less_zero_point);
442 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max_less_zero_point);
443 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max_less_zero_point);
444 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max_less_zero_point);
445 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max_less_zero_point);
446 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max_less_zero_point);
447 vout6x01234567 = _mm256_min_ps(vout6x01234567, voutput_max_less_zero_point);
448 vout7x01234567 = _mm256_min_ps(vout7x01234567, voutput_max_less_zero_point);
449 vout8x01234567 = _mm256_min_ps(vout8x01234567, voutput_max_less_zero_point);
450 vout9x01234567 = _mm256_min_ps(vout9x01234567, voutput_max_less_zero_point);
451 vout10x01234567 = _mm256_min_ps(vout10x01234567, voutput_max_less_zero_point);
452 vout11x01234567 = _mm256_min_ps(vout11x01234567, voutput_max_less_zero_point);
453 vout12x01234567 = _mm256_min_ps(vout12x01234567, voutput_max_less_zero_point);
454 vout13x01234567 = _mm256_min_ps(vout13x01234567, voutput_max_less_zero_point);
455
456 vacc0x01234567 = _mm256_cvtps_epi32(vout0x01234567);
457 vacc1x01234567 = _mm256_cvtps_epi32(vout1x01234567);
458 vacc2x01234567 = _mm256_cvtps_epi32(vout2x01234567);
459 vacc3x01234567 = _mm256_cvtps_epi32(vout3x01234567);
460 vacc4x01234567 = _mm256_cvtps_epi32(vout4x01234567);
461 vacc5x01234567 = _mm256_cvtps_epi32(vout5x01234567);
462 vacc6x01234567 = _mm256_cvtps_epi32(vout6x01234567);
463 vacc7x01234567 = _mm256_cvtps_epi32(vout7x01234567);
464 vacc8x01234567 = _mm256_cvtps_epi32(vout8x01234567);
465 vacc9x01234567 = _mm256_cvtps_epi32(vout9x01234567);
466 vacc10x01234567 = _mm256_cvtps_epi32(vout10x01234567);
467 vacc11x01234567 = _mm256_cvtps_epi32(vout11x01234567);
468 vacc12x01234567 = _mm256_cvtps_epi32(vout12x01234567);
469 vacc13x01234567 = _mm256_cvtps_epi32(vout13x01234567);
470
471 vacc0x01234567 = _mm256_add_epi32(vacc0x01234567, voutput_zero_point);
472 vacc1x01234567 = _mm256_add_epi32(vacc1x01234567, voutput_zero_point);
473 vacc2x01234567 = _mm256_add_epi32(vacc2x01234567, voutput_zero_point);
474 vacc3x01234567 = _mm256_add_epi32(vacc3x01234567, voutput_zero_point);
475 vacc4x01234567 = _mm256_add_epi32(vacc4x01234567, voutput_zero_point);
476 vacc5x01234567 = _mm256_add_epi32(vacc5x01234567, voutput_zero_point);
477 vacc6x01234567 = _mm256_add_epi32(vacc6x01234567, voutput_zero_point);
478 vacc7x01234567 = _mm256_add_epi32(vacc7x01234567, voutput_zero_point);
479 vacc8x01234567 = _mm256_add_epi32(vacc8x01234567, voutput_zero_point);
480 vacc9x01234567 = _mm256_add_epi32(vacc9x01234567, voutput_zero_point);
481 vacc10x01234567 = _mm256_add_epi32(vacc10x01234567, voutput_zero_point);
482 vacc11x01234567 = _mm256_add_epi32(vacc11x01234567, voutput_zero_point);
483 vacc12x01234567 = _mm256_add_epi32(vacc12x01234567, voutput_zero_point);
484 vacc13x01234567 = _mm256_add_epi32(vacc13x01234567, voutput_zero_point);
485
486 __m128i voutb0x01234567 = _mm256_cvtsepi32_epi8(vacc0x01234567);
487 __m128i voutb1x01234567 = _mm256_cvtsepi32_epi8(vacc1x01234567);
488 __m128i voutb2x01234567 = _mm256_cvtsepi32_epi8(vacc2x01234567);
489 __m128i voutb3x01234567 = _mm256_cvtsepi32_epi8(vacc3x01234567);
490 __m128i voutb4x01234567 = _mm256_cvtsepi32_epi8(vacc4x01234567);
491 __m128i voutb5x01234567 = _mm256_cvtsepi32_epi8(vacc5x01234567);
492 __m128i voutb6x01234567 = _mm256_cvtsepi32_epi8(vacc6x01234567);
493 __m128i voutb7x01234567 = _mm256_cvtsepi32_epi8(vacc7x01234567);
494 __m128i voutb8x01234567 = _mm256_cvtsepi32_epi8(vacc8x01234567);
495 __m128i voutb9x01234567 = _mm256_cvtsepi32_epi8(vacc9x01234567);
496 __m128i voutb10x01234567 = _mm256_cvtsepi32_epi8(vacc10x01234567);
497 __m128i voutb11x01234567 = _mm256_cvtsepi32_epi8(vacc11x01234567);
498 __m128i voutb12x01234567 = _mm256_cvtsepi32_epi8(vacc12x01234567);
499 __m128i voutb13x01234567 = _mm256_cvtsepi32_epi8(vacc13x01234567);
500
501 voutb0x01234567 = _mm_max_epi8(voutb0x01234567, voutput_min);
502 voutb1x01234567 = _mm_max_epi8(voutb1x01234567, voutput_min);
503 voutb2x01234567 = _mm_max_epi8(voutb2x01234567, voutput_min);
504 voutb3x01234567 = _mm_max_epi8(voutb3x01234567, voutput_min);
505 voutb4x01234567 = _mm_max_epi8(voutb4x01234567, voutput_min);
506 voutb5x01234567 = _mm_max_epi8(voutb5x01234567, voutput_min);
507 voutb6x01234567 = _mm_max_epi8(voutb6x01234567, voutput_min);
508 voutb7x01234567 = _mm_max_epi8(voutb7x01234567, voutput_min);
509 voutb8x01234567 = _mm_max_epi8(voutb8x01234567, voutput_min);
510 voutb9x01234567 = _mm_max_epi8(voutb9x01234567, voutput_min);
511 voutb10x01234567 = _mm_max_epi8(voutb10x01234567, voutput_min);
512 voutb11x01234567 = _mm_max_epi8(voutb11x01234567, voutput_min);
513 voutb12x01234567 = _mm_max_epi8(voutb12x01234567, voutput_min);
514 voutb13x01234567 = _mm_max_epi8(voutb13x01234567, voutput_min);
515
516 if (nc >= 8) {
517 _mm_storel_epi64((__m128i*) c13, voutb13x01234567);
518 c13 = (int8_t*) ((uintptr_t) c13 + cn_stride);
519 _mm_storel_epi64((__m128i*) c12, voutb12x01234567);
520 c12 = (int8_t*) ((uintptr_t) c12 + cn_stride);
521 _mm_storel_epi64((__m128i*) c11, voutb11x01234567);
522 c11 = (int8_t*) ((uintptr_t) c11 + cn_stride);
523 _mm_storel_epi64((__m128i*) c10, voutb10x01234567);
524 c10 = (int8_t*) ((uintptr_t) c10 + cn_stride);
525 _mm_storel_epi64((__m128i*) c9, voutb9x01234567);
526 c9 = (int8_t*) ((uintptr_t) c9 + cn_stride);
527 _mm_storel_epi64((__m128i*) c8, voutb8x01234567);
528 c8 = (int8_t*) ((uintptr_t) c8 + cn_stride);
529 _mm_storel_epi64((__m128i*) c7, voutb7x01234567);
530 c7 = (int8_t*) ((uintptr_t) c7 + cn_stride);
531 _mm_storel_epi64((__m128i*) c6, voutb6x01234567);
532 c6 = (int8_t*) ((uintptr_t) c6 + cn_stride);
533 _mm_storel_epi64((__m128i*) c5, voutb5x01234567);
534 c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);
535 _mm_storel_epi64((__m128i*) c4, voutb4x01234567);
536 c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
537 _mm_storel_epi64((__m128i*) c3, voutb3x01234567);
538 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
539 _mm_storel_epi64((__m128i*) c2, voutb2x01234567);
540 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
541 _mm_storel_epi64((__m128i*) c1, voutb1x01234567);
542 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
543 _mm_storel_epi64((__m128i*) c0, voutb0x01234567);
544 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
545 a = (const int8_t**restrict) ((uintptr_t) a - ks);
546 nc -= 8;
547 } else {
548 // Prepare mask for valid 8-bit elements (depends on nc).
549 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - UINT32_C(1));
550 _mm_mask_storeu_epi8(c13, vmask, voutb13x01234567);
551 _mm_mask_storeu_epi8(c12, vmask, voutb12x01234567);
552 _mm_mask_storeu_epi8(c11, vmask, voutb11x01234567);
553 _mm_mask_storeu_epi8(c10, vmask, voutb10x01234567);
554 _mm_mask_storeu_epi8(c9, vmask, voutb9x01234567);
555 _mm_mask_storeu_epi8(c8, vmask, voutb8x01234567);
556 _mm_mask_storeu_epi8(c7, vmask, voutb7x01234567);
557 _mm_mask_storeu_epi8(c6, vmask, voutb6x01234567);
558 _mm_mask_storeu_epi8(c5, vmask, voutb5x01234567);
559 _mm_mask_storeu_epi8(c4, vmask, voutb4x01234567);
560 _mm_mask_storeu_epi8(c3, vmask, voutb3x01234567);
561 _mm_mask_storeu_epi8(c2, vmask, voutb2x01234567);
562 _mm_mask_storeu_epi8(c1, vmask, voutb1x01234567);
563 _mm_mask_storeu_epi8(c0, vmask, voutb0x01234567);
564 nc = 0;
565 }
566 } while (nc != 0);
567 }