Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-12x8c8-minmax-avx256vnni.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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/gemm.h"
15 #include "xnnpack/intrinsics-polyfill.h"
16 #include "xnnpack/math.h"
17 #include "xnnpack/unaligned.h"
18
19
20 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_12x8c8__avx256vnni(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 const int8_t* restrict a,
25 size_t a_stride,
26 const void* restrict w,
27 float* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
31 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(mr != 0);
34 assert(mr <= 12);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(kc % sizeof(int8_t) == 0);
38 assert(a != NULL);
39 assert(w != NULL);
40 assert(c != NULL);
41
42 kc = round_up_po2(kc, 8 * sizeof(int8_t));
43 const int8_t* a0 = a;
44 float* c0 = c;
45 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
46 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr < 2) {
48 a1 = a0;
49 c1 = c0;
50 }
51 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
52 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
53 if XNN_UNPREDICTABLE(mr <= 2) {
54 a2 = a1;
55 c2 = c1;
56 }
57 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
58 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
59 if XNN_UNPREDICTABLE(mr < 4) {
60 a3 = a2;
61 c3 = c2;
62 }
63 const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
64 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
65 if XNN_UNPREDICTABLE(mr <= 4) {
66 a4 = a3;
67 c4 = c3;
68 }
69 const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride);
70 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
71 if XNN_UNPREDICTABLE(mr < 6) {
72 a5 = a4;
73 c5 = c4;
74 }
75 const int8_t* a6 = (const int8_t*) ((uintptr_t) a5 + a_stride);
76 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
77 if XNN_UNPREDICTABLE(mr <= 6) {
78 a6 = a5;
79 c6 = c5;
80 }
81 const int8_t* a7 = (const int8_t*) ((uintptr_t) a6 + a_stride);
82 float* c7 = (float*) ((uintptr_t) c6 + cm_stride);
83 if XNN_UNPREDICTABLE(mr < 8) {
84 a7 = a6;
85 c7 = c6;
86 }
87 const int8_t* a8 = (const int8_t*) ((uintptr_t) a7 + a_stride);
88 float* c8 = (float*) ((uintptr_t) c7 + cm_stride);
89 if XNN_UNPREDICTABLE(mr <= 8) {
90 a8 = a7;
91 c8 = c7;
92 }
93 const int8_t* a9 = (const int8_t*) ((uintptr_t) a8 + a_stride);
94 float* c9 = (float*) ((uintptr_t) c8 + cm_stride);
95 if XNN_UNPREDICTABLE(mr < 10) {
96 a9 = a8;
97 c9 = c8;
98 }
99 const int8_t* a10 = (const int8_t*) ((uintptr_t) a9 + a_stride);
100 float* c10 = (float*) ((uintptr_t) c9 + cm_stride);
101 if XNN_UNPREDICTABLE(mr <= 10) {
102 a10 = a9;
103 c10 = c9;
104 }
105 const int8_t* a11 = (const int8_t*) ((uintptr_t) a10 + a_stride);
106 float* c11 = (float*) ((uintptr_t) c10 + cm_stride);
107 if XNN_UNPREDICTABLE(mr != 12) {
108 a11 = a10;
109 c11 = c10;
110 }
111
112 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
113 XNN_FORCE_REALIZATION(vsign_mask);
114 const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point + 128);
115 const __m256i vinput_zero_point1 = _mm256_set1_epi32((int) quantization_params[1].zero_point + 128);
116 const __m256i vinput_zero_point2 = _mm256_set1_epi32((int) quantization_params[2].zero_point + 128);
117 const __m256i vinput_zero_point3 = _mm256_set1_epi32((int) quantization_params[3].zero_point + 128);
118 const __m256i vinput_zero_point4 = _mm256_set1_epi32((int) quantization_params[4].zero_point + 128);
119 const __m256i vinput_zero_point5 = _mm256_set1_epi32((int) quantization_params[5].zero_point + 128);
120 const __m256i vinput_zero_point6 = _mm256_set1_epi32((int) quantization_params[6].zero_point + 128);
121 const __m256i vinput_zero_point7 = _mm256_set1_epi32((int) quantization_params[7].zero_point + 128);
122 const __m256i vinput_zero_point8 = _mm256_set1_epi32((int) quantization_params[8].zero_point + 128);
123 const __m256i vinput_zero_point9 = _mm256_set1_epi32((int) quantization_params[9].zero_point + 128);
124 const __m256i vinput_zero_point10 = _mm256_set1_epi32((int) quantization_params[10].zero_point + 128);
125 const __m256i vinput_zero_point11 = _mm256_set1_epi32((int) quantization_params[11].zero_point + 128);
126 const __m256 voutput_min = _mm256_set1_ps(params->scalar.min);
127 const __m256 voutput_max = _mm256_set1_ps(params->scalar.max);
128 // XNN_FORCE_REALIZATION(voutput_min);
129 // XNN_FORCE_REALIZATION(voutput_max);
130 do {
131 const __m256i vksum01234567 = _mm256_load_si256(w);
132 __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point0);
133 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
134 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
135 __m256i vsum1x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point1);
136 __m256i vacc1x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 0));
137 __m256i vacc1x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 1));
138 __m256i vsum2x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point2);
139 __m256i vacc2x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 0));
140 __m256i vacc2x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 1));
141 __m256i vsum3x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point3);
142 __m256i vacc3x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 0));
143 __m256i vacc3x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 1));
144 __m256i vsum4x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point4);
145 __m256i vacc4x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum4x01234567, 0));
146 __m256i vacc4x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum4x01234567, 1));
147 __m256i vsum5x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point5);
148 __m256i vacc5x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum5x01234567, 0));
149 __m256i vacc5x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum5x01234567, 1));
150 __m256i vsum6x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point6);
151 __m256i vacc6x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum6x01234567, 0));
152 __m256i vacc6x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum6x01234567, 1));
153 __m256i vsum7x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point7);
154 __m256i vacc7x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum7x01234567, 0));
155 __m256i vacc7x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum7x01234567, 1));
156 __m256i vsum8x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point8);
157 __m256i vacc8x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum8x01234567, 0));
158 __m256i vacc8x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum8x01234567, 1));
159 __m256i vsum9x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point9);
160 __m256i vacc9x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum9x01234567, 0));
161 __m256i vacc9x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum9x01234567, 1));
162 __m256i vsum10x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point10);
163 __m256i vacc10x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum10x01234567, 0));
164 __m256i vacc10x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum10x01234567, 1));
165 __m256i vsum11x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point11);
166 __m256i vacc11x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum11x01234567, 0));
167 __m256i vacc11x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum11x01234567, 1));
168 w = (const int32_t*) w + 8;
169
170 size_t k = kc;
171 while (k >= 16 * sizeof(int8_t)) {
172 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
173 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
174 a0 += 16;
175 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
176 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
177 a1 += 16;
178 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
179 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
180 a2 += 16;
181 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
182 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
183 a3 += 16;
184 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
185 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
186 a4 += 16;
187 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
188 const __m256i va5x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
189 a5 += 16;
190 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
191 const __m256i va6x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
192 a6 += 16;
193 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
194 const __m256i va7x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7 + 8)), vsign_mask);
195 a7 += 16;
196 const __m256i va8x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8)), vsign_mask);
197 const __m256i va8x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8 + 8)), vsign_mask);
198 a8 += 16;
199 const __m256i va9x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9)), vsign_mask);
200 const __m256i va9x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9 + 8)), vsign_mask);
201 a9 += 16;
202 const __m256i va10x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a10)), vsign_mask);
203 const __m256i va10x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a10 + 8)), vsign_mask);
204 a10 += 16;
205 const __m256i va11x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a11)), vsign_mask);
206 const __m256i va11x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a11 + 8)), vsign_mask);
207 a11 += 16;
208
209 const __m256i vb01234567x0123 = _mm256_load_si256(w);
210 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
211 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
212 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
213
214 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
215 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
216 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
217 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
218 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
219 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
220 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
221 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
222 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
223 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
224 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
225 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
226 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
227 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
228 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
229 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
230 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x01234567, vb01234567x0123);
231 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x01234567, vb89ABCDEFx0123);
232 vacc9x0123 = _mm256_dpbusd_epi32(vacc9x0123, va9x01234567, vb01234567x0123);
233 vacc9x4567 = _mm256_dpbusd_epi32(vacc9x4567, va9x01234567, vb89ABCDEFx0123);
234 vacc10x0123 = _mm256_dpbusd_epi32(vacc10x0123, va10x01234567, vb01234567x0123);
235 vacc10x4567 = _mm256_dpbusd_epi32(vacc10x4567, va10x01234567, vb89ABCDEFx0123);
236 vacc11x0123 = _mm256_dpbusd_epi32(vacc11x0123, va11x01234567, vb01234567x0123);
237 vacc11x4567 = _mm256_dpbusd_epi32(vacc11x4567, va11x01234567, vb89ABCDEFx0123);
238 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
239 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
240 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
241 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
242 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
243 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
244 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
245 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
246 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
247 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
248 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x89ABCDEF, vb01234567x4567);
249 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);
250 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x89ABCDEF, vb01234567x4567);
251 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x89ABCDEF, vb89ABCDEFx4567);
252 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x89ABCDEF, vb01234567x4567);
253 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x89ABCDEF, vb89ABCDEFx4567);
254 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x89ABCDEF, vb01234567x4567);
255 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x89ABCDEF, vb89ABCDEFx4567);
256 vacc9x0123 = _mm256_dpbusd_epi32(vacc9x0123, va9x89ABCDEF, vb01234567x4567);
257 vacc9x4567 = _mm256_dpbusd_epi32(vacc9x4567, va9x89ABCDEF, vb89ABCDEFx4567);
258 vacc10x0123 = _mm256_dpbusd_epi32(vacc10x0123, va10x89ABCDEF, vb01234567x4567);
259 vacc10x4567 = _mm256_dpbusd_epi32(vacc10x4567, va10x89ABCDEF, vb89ABCDEFx4567);
260 vacc11x0123 = _mm256_dpbusd_epi32(vacc11x0123, va11x89ABCDEF, vb01234567x4567);
261 vacc11x4567 = _mm256_dpbusd_epi32(vacc11x4567, va11x89ABCDEF, vb89ABCDEFx4567);
262
263 w = (const int8_t*) w + 128;
264 k -= 16 * sizeof(int8_t);
265 }
266
267 if (k != 0) {
268 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
269 a0 += 8;
270 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
271 a1 += 8;
272 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
273 a2 += 8;
274 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
275 a3 += 8;
276 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
277 a4 += 8;
278 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
279 a5 += 8;
280 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
281 a6 += 8;
282 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
283 a7 += 8;
284 const __m256i va8x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8)), vsign_mask);
285 a8 += 8;
286 const __m256i va9x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9)), vsign_mask);
287 a9 += 8;
288 const __m256i va10x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a10)), vsign_mask);
289 a10 += 8;
290 const __m256i va11x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a11)), vsign_mask);
291 a11 += 8;
292
293 const __m256i vb01234567x0123 = _mm256_load_si256(w);
294 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
295
296 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
297 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
298 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
299 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
300 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
301 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
302 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
303 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
304 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
305 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
306 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
307 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
308 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
309 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
310 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
311 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
312 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x01234567, vb01234567x0123);
313 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x01234567, vb89ABCDEFx0123);
314 vacc9x0123 = _mm256_dpbusd_epi32(vacc9x0123, va9x01234567, vb01234567x0123);
315 vacc9x4567 = _mm256_dpbusd_epi32(vacc9x4567, va9x01234567, vb89ABCDEFx0123);
316 vacc10x0123 = _mm256_dpbusd_epi32(vacc10x0123, va10x01234567, vb01234567x0123);
317 vacc10x4567 = _mm256_dpbusd_epi32(vacc10x4567, va10x01234567, vb89ABCDEFx0123);
318 vacc11x0123 = _mm256_dpbusd_epi32(vacc11x0123, va11x01234567, vb01234567x0123);
319 vacc11x4567 = _mm256_dpbusd_epi32(vacc11x4567, va11x01234567, vb89ABCDEFx0123);
320
321 w = (const int8_t*) w + 64;
322 k -= 8 * sizeof(int8_t);
323 }
324
325 // Add adjacent pairs
326 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
327 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
328 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
329 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
330 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
331 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
332 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
333 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
334 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
335 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
336 const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);
337 __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));
338 const __m256i vsum6x02134657 = _mm256_hadd_epi32(vacc6x0123, vacc6x4567);
339 __m256i vacc6x01234567 = _mm256_permute4x64_epi64(vsum6x02134657, _MM_SHUFFLE(3, 1, 2, 0));
340 const __m256i vsum7x02134657 = _mm256_hadd_epi32(vacc7x0123, vacc7x4567);
341 __m256i vacc7x01234567 = _mm256_permute4x64_epi64(vsum7x02134657, _MM_SHUFFLE(3, 1, 2, 0));
342 const __m256i vsum8x02134657 = _mm256_hadd_epi32(vacc8x0123, vacc8x4567);
343 __m256i vacc8x01234567 = _mm256_permute4x64_epi64(vsum8x02134657, _MM_SHUFFLE(3, 1, 2, 0));
344 const __m256i vsum9x02134657 = _mm256_hadd_epi32(vacc9x0123, vacc9x4567);
345 __m256i vacc9x01234567 = _mm256_permute4x64_epi64(vsum9x02134657, _MM_SHUFFLE(3, 1, 2, 0));
346 const __m256i vsum10x02134657 = _mm256_hadd_epi32(vacc10x0123, vacc10x4567);
347 __m256i vacc10x01234567 = _mm256_permute4x64_epi64(vsum10x02134657, _MM_SHUFFLE(3, 1, 2, 0));
348 const __m256i vsum11x02134657 = _mm256_hadd_epi32(vacc11x0123, vacc11x4567);
349 __m256i vacc11x01234567 = _mm256_permute4x64_epi64(vsum11x02134657, _MM_SHUFFLE(3, 1, 2, 0));
350
351 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
352 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
353 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
354 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
355 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
356 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
357 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567);
358 __m256 vout7x01234567 = _mm256_cvtepi32_ps(vacc7x01234567);
359 __m256 vout8x01234567 = _mm256_cvtepi32_ps(vacc8x01234567);
360 __m256 vout9x01234567 = _mm256_cvtepi32_ps(vacc9x01234567);
361 __m256 vout10x01234567 = _mm256_cvtepi32_ps(vacc10x01234567);
362 __m256 vout11x01234567 = _mm256_cvtepi32_ps(vacc11x01234567);
363
364 vout0x01234567 = _mm256_mul_ps(vout0x01234567, _mm256_set1_ps(quantization_params[0].inv_scale));
365 vout1x01234567 = _mm256_mul_ps(vout1x01234567, _mm256_set1_ps(quantization_params[1].inv_scale));
366 vout2x01234567 = _mm256_mul_ps(vout2x01234567, _mm256_set1_ps(quantization_params[2].inv_scale));
367 vout3x01234567 = _mm256_mul_ps(vout3x01234567, _mm256_set1_ps(quantization_params[3].inv_scale));
368 vout4x01234567 = _mm256_mul_ps(vout4x01234567, _mm256_set1_ps(quantization_params[4].inv_scale));
369 vout5x01234567 = _mm256_mul_ps(vout5x01234567, _mm256_set1_ps(quantization_params[5].inv_scale));
370 vout6x01234567 = _mm256_mul_ps(vout6x01234567, _mm256_set1_ps(quantization_params[6].inv_scale));
371 vout7x01234567 = _mm256_mul_ps(vout7x01234567, _mm256_set1_ps(quantization_params[7].inv_scale));
372 vout8x01234567 = _mm256_mul_ps(vout8x01234567, _mm256_set1_ps(quantization_params[8].inv_scale));
373 vout9x01234567 = _mm256_mul_ps(vout9x01234567, _mm256_set1_ps(quantization_params[9].inv_scale));
374 vout10x01234567 = _mm256_mul_ps(vout10x01234567, _mm256_set1_ps(quantization_params[10].inv_scale));
375 vout11x01234567 = _mm256_mul_ps(vout11x01234567, _mm256_set1_ps(quantization_params[11].inv_scale));
376
377 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
378 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
379 w = (const float*) w + 16;
380
381 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
382 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
383 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
384 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
385 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
386 vout5x01234567 = _mm256_fmadd_ps(vout5x01234567, vfilter_output_scale01234567, vbias01234567);
387 vout6x01234567 = _mm256_fmadd_ps(vout6x01234567, vfilter_output_scale01234567, vbias01234567);
388 vout7x01234567 = _mm256_fmadd_ps(vout7x01234567, vfilter_output_scale01234567, vbias01234567);
389 vout8x01234567 = _mm256_fmadd_ps(vout8x01234567, vfilter_output_scale01234567, vbias01234567);
390 vout9x01234567 = _mm256_fmadd_ps(vout9x01234567, vfilter_output_scale01234567, vbias01234567);
391 vout10x01234567 = _mm256_fmadd_ps(vout10x01234567, vfilter_output_scale01234567, vbias01234567);
392 vout11x01234567 = _mm256_fmadd_ps(vout11x01234567, vfilter_output_scale01234567, vbias01234567);
393
394 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
395 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
396 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
397 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
398 vout4x01234567 = _mm256_max_ps(vout4x01234567, voutput_min);
399 vout5x01234567 = _mm256_max_ps(vout5x01234567, voutput_min);
400 vout6x01234567 = _mm256_max_ps(vout6x01234567, voutput_min);
401 vout7x01234567 = _mm256_max_ps(vout7x01234567, voutput_min);
402 vout8x01234567 = _mm256_max_ps(vout8x01234567, voutput_min);
403 vout9x01234567 = _mm256_max_ps(vout9x01234567, voutput_min);
404 vout10x01234567 = _mm256_max_ps(vout10x01234567, voutput_min);
405 vout11x01234567 = _mm256_max_ps(vout11x01234567, voutput_min);
406
407 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
408 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
409 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
410 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
411 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max);
412 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max);
413 vout6x01234567 = _mm256_min_ps(vout6x01234567, voutput_max);
414 vout7x01234567 = _mm256_min_ps(vout7x01234567, voutput_max);
415 vout8x01234567 = _mm256_min_ps(vout8x01234567, voutput_max);
416 vout9x01234567 = _mm256_min_ps(vout9x01234567, voutput_max);
417 vout10x01234567 = _mm256_min_ps(vout10x01234567, voutput_max);
418 vout11x01234567 = _mm256_min_ps(vout11x01234567, voutput_max);
419
420 if XNN_LIKELY(nc >= 8) {
421 _mm256_storeu_ps(c0, vout0x01234567);
422 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
423 c0 = (float*) ((uintptr_t) c0 + cn_stride);
424 _mm256_storeu_ps(c1, vout1x01234567);
425 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
426 c1 = (float*) ((uintptr_t) c1 + cn_stride);
427 _mm256_storeu_ps(c2, vout2x01234567);
428 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
429 c2 = (float*) ((uintptr_t) c2 + cn_stride);
430 _mm256_storeu_ps(c3, vout3x01234567);
431 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
432 c3 = (float*) ((uintptr_t) c3 + cn_stride);
433 _mm256_storeu_ps(c4, vout4x01234567);
434 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
435 c4 = (float*) ((uintptr_t) c4 + cn_stride);
436 _mm256_storeu_ps(c5, vout5x01234567);
437 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
438 c5 = (float*) ((uintptr_t) c5 + cn_stride);
439 _mm256_storeu_ps(c6, vout6x01234567);
440 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
441 c6 = (float*) ((uintptr_t) c6 + cn_stride);
442 _mm256_storeu_ps(c7, vout7x01234567);
443 a7 = (const int8_t*) ((uintptr_t) a7 - kc);
444 c7 = (float*) ((uintptr_t) c7 + cn_stride);
445 _mm256_storeu_ps(c8, vout8x01234567);
446 a8 = (const int8_t*) ((uintptr_t) a8 - kc);
447 c8 = (float*) ((uintptr_t) c8 + cn_stride);
448 _mm256_storeu_ps(c9, vout9x01234567);
449 a9 = (const int8_t*) ((uintptr_t) a9 - kc);
450 c9 = (float*) ((uintptr_t) c9 + cn_stride);
451 _mm256_storeu_ps(c10, vout10x01234567);
452 a10 = (const int8_t*) ((uintptr_t) a10 - kc);
453 c10 = (float*) ((uintptr_t) c10 + cn_stride);
454 _mm256_storeu_ps(c11, vout11x01234567);
455 a11 = (const int8_t*) ((uintptr_t) a11 - kc);
456 c11 = (float*) ((uintptr_t) c11 + cn_stride);
457 nc -= 8;
458 } else {
459 // Prepare mask for valid 32-bit elements (depends on nc).
460 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
461 _mm256_mask_storeu_ps(c0, vmask, vout0x01234567);
462 _mm256_mask_storeu_ps(c1, vmask, vout1x01234567);
463 _mm256_mask_storeu_ps(c2, vmask, vout2x01234567);
464 _mm256_mask_storeu_ps(c3, vmask, vout3x01234567);
465 _mm256_mask_storeu_ps(c4, vmask, vout4x01234567);
466 _mm256_mask_storeu_ps(c5, vmask, vout5x01234567);
467 _mm256_mask_storeu_ps(c6, vmask, vout6x01234567);
468 _mm256_mask_storeu_ps(c7, vmask, vout7x01234567);
469 _mm256_mask_storeu_ps(c8, vmask, vout8x01234567);
470 _mm256_mask_storeu_ps(c9, vmask, vout9x01234567);
471 _mm256_mask_storeu_ps(c10, vmask, vout10x01234567);
472 _mm256_mask_storeu_ps(c11, vmask, vout11x01234567);
473 nc = 0;
474 }
475 } while (nc != 0);
476 }