Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-9x8c8-minmax-avx256vnni-prfm.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 #include "xnnpack/prefetch.h"
19
20
21 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_9x8c8__avx256vnni_prfm(
22 size_t mr,
23 size_t nc,
24 size_t kc,
25 const int8_t* restrict a,
26 size_t a_stride,
27 const void* restrict w,
28 float* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
32 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(mr != 0);
35 assert(mr <= 9);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(kc % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 8 * sizeof(int8_t));
44 const int8_t* a0 = a;
45 float* c0 = c;
46 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
47 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 a1 = a0;
50 c1 = c0;
51 }
52 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
53 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
54 if XNN_UNPREDICTABLE(mr <= 2) {
55 a2 = a1;
56 c2 = c1;
57 }
58 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
59 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
60 if XNN_UNPREDICTABLE(mr < 4) {
61 a3 = a2;
62 c3 = c2;
63 }
64 const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
65 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
66 if XNN_UNPREDICTABLE(mr <= 4) {
67 a4 = a3;
68 c4 = c3;
69 }
70 const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride);
71 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
72 if XNN_UNPREDICTABLE(mr < 6) {
73 a5 = a4;
74 c5 = c4;
75 }
76 const int8_t* a6 = (const int8_t*) ((uintptr_t) a5 + a_stride);
77 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
78 if XNN_UNPREDICTABLE(mr <= 6) {
79 a6 = a5;
80 c6 = c5;
81 }
82 const int8_t* a7 = (const int8_t*) ((uintptr_t) a6 + a_stride);
83 float* c7 = (float*) ((uintptr_t) c6 + cm_stride);
84 if XNN_UNPREDICTABLE(mr < 8) {
85 a7 = a6;
86 c7 = c6;
87 }
88 const int8_t* a8 = (const int8_t*) ((uintptr_t) a7 + a_stride);
89 float* c8 = (float*) ((uintptr_t) c7 + cm_stride);
90 if XNN_UNPREDICTABLE(mr <= 8) {
91 a8 = a7;
92 c8 = c7;
93 }
94
95 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
96 XNN_FORCE_REALIZATION(vsign_mask);
97 const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point + 128);
98 const __m256i vinput_zero_point1 = _mm256_set1_epi32((int) quantization_params[1].zero_point + 128);
99 const __m256i vinput_zero_point2 = _mm256_set1_epi32((int) quantization_params[2].zero_point + 128);
100 const __m256i vinput_zero_point3 = _mm256_set1_epi32((int) quantization_params[3].zero_point + 128);
101 const __m256i vinput_zero_point4 = _mm256_set1_epi32((int) quantization_params[4].zero_point + 128);
102 const __m256i vinput_zero_point5 = _mm256_set1_epi32((int) quantization_params[5].zero_point + 128);
103 const __m256i vinput_zero_point6 = _mm256_set1_epi32((int) quantization_params[6].zero_point + 128);
104 const __m256i vinput_zero_point7 = _mm256_set1_epi32((int) quantization_params[7].zero_point + 128);
105 const __m256i vinput_zero_point8 = _mm256_set1_epi32((int) quantization_params[8].zero_point + 128);
106 const __m256 voutput_min = _mm256_set1_ps(params->scalar.min);
107 const __m256 voutput_max = _mm256_set1_ps(params->scalar.max);
108 // XNN_FORCE_REALIZATION(voutput_min);
109 // XNN_FORCE_REALIZATION(voutput_max);
110 do {
111 const __m256i vksum01234567 = _mm256_load_si256(w);
112 __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point0);
113 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
114 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
115 __m256i vsum1x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point1);
116 __m256i vacc1x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 0));
117 __m256i vacc1x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 1));
118 __m256i vsum2x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point2);
119 __m256i vacc2x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 0));
120 __m256i vacc2x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 1));
121 __m256i vsum3x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point3);
122 __m256i vacc3x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 0));
123 __m256i vacc3x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 1));
124 __m256i vsum4x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point4);
125 __m256i vacc4x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum4x01234567, 0));
126 __m256i vacc4x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum4x01234567, 1));
127 __m256i vsum5x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point5);
128 __m256i vacc5x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum5x01234567, 0));
129 __m256i vacc5x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum5x01234567, 1));
130 __m256i vsum6x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point6);
131 __m256i vacc6x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum6x01234567, 0));
132 __m256i vacc6x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum6x01234567, 1));
133 __m256i vsum7x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point7);
134 __m256i vacc7x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum7x01234567, 0));
135 __m256i vacc7x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum7x01234567, 1));
136 __m256i vsum8x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point8);
137 __m256i vacc8x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum8x01234567, 0));
138 __m256i vacc8x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum8x01234567, 1));
139 w = (const int32_t*) w + 8;
140
141 size_t k = kc;
142 while (k >= 16 * sizeof(int8_t)) {
143 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
144 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
145 a0 += 16;
146 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
147 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
148 a1 += 16;
149 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
150 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
151 a2 += 16;
152 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
153 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
154 a3 += 16;
155 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
156 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
157 a4 += 16;
158 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
159 const __m256i va5x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
160 a5 += 16;
161 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
162 const __m256i va6x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
163 a6 += 16;
164 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
165 const __m256i va7x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7 + 8)), vsign_mask);
166 a7 += 16;
167 const __m256i va8x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8)), vsign_mask);
168 const __m256i va8x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8 + 8)), vsign_mask);
169 a8 += 16;
170
171 const __m256i vb01234567x0123 = _mm256_load_si256(w);
172 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
173 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
174 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
175 xnn_prefetch_to_l1((const int8_t*) w + 896);
176
177 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
178 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
179 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
180 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
181 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
182 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
183 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
184 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
185 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
186 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
187 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
188 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
189 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
190 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
191 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
192 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
193 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x01234567, vb01234567x0123);
194 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x01234567, vb89ABCDEFx0123);
195 xnn_prefetch_to_l1((const int8_t*) w + 960);
196 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
197 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
198 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
199 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
200 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
201 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
202 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
203 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
204 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
205 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
206 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x89ABCDEF, vb01234567x4567);
207 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);
208 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x89ABCDEF, vb01234567x4567);
209 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x89ABCDEF, vb89ABCDEFx4567);
210 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x89ABCDEF, vb01234567x4567);
211 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x89ABCDEF, vb89ABCDEFx4567);
212 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x89ABCDEF, vb01234567x4567);
213 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x89ABCDEF, vb89ABCDEFx4567);
214
215 w = (const int8_t*) w + 128;
216 k -= 16 * sizeof(int8_t);
217 }
218
219 if (k != 0) {
220 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
221 a0 += 8;
222 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
223 a1 += 8;
224 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
225 a2 += 8;
226 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
227 a3 += 8;
228 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
229 a4 += 8;
230 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
231 a5 += 8;
232 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
233 a6 += 8;
234 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
235 a7 += 8;
236 const __m256i va8x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8)), vsign_mask);
237 a8 += 8;
238
239 const __m256i vb01234567x0123 = _mm256_load_si256(w);
240 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
241
242 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
243 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
244 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
245 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
246 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
247 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
248 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
249 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
250 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
251 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
252 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
253 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
254 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
255 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
256 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
257 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
258 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x01234567, vb01234567x0123);
259 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x01234567, vb89ABCDEFx0123);
260 xnn_prefetch_to_l1((const int8_t*) w + 960);
261
262 w = (const int8_t*) w + 64;
263 k -= 8 * sizeof(int8_t);
264 }
265
266 // Add adjacent pairs
267 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
268 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
269 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
270 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
271 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
272 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
273 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
274 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
275 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
276 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
277 const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);
278 __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));
279 const __m256i vsum6x02134657 = _mm256_hadd_epi32(vacc6x0123, vacc6x4567);
280 __m256i vacc6x01234567 = _mm256_permute4x64_epi64(vsum6x02134657, _MM_SHUFFLE(3, 1, 2, 0));
281 const __m256i vsum7x02134657 = _mm256_hadd_epi32(vacc7x0123, vacc7x4567);
282 __m256i vacc7x01234567 = _mm256_permute4x64_epi64(vsum7x02134657, _MM_SHUFFLE(3, 1, 2, 0));
283 const __m256i vsum8x02134657 = _mm256_hadd_epi32(vacc8x0123, vacc8x4567);
284 __m256i vacc8x01234567 = _mm256_permute4x64_epi64(vsum8x02134657, _MM_SHUFFLE(3, 1, 2, 0));
285
286 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
287 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
288 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
289 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
290 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
291 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
292 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567);
293 __m256 vout7x01234567 = _mm256_cvtepi32_ps(vacc7x01234567);
294 __m256 vout8x01234567 = _mm256_cvtepi32_ps(vacc8x01234567);
295
296 vout0x01234567 = _mm256_mul_ps(vout0x01234567, _mm256_set1_ps(quantization_params[0].inv_scale));
297 vout1x01234567 = _mm256_mul_ps(vout1x01234567, _mm256_set1_ps(quantization_params[1].inv_scale));
298 vout2x01234567 = _mm256_mul_ps(vout2x01234567, _mm256_set1_ps(quantization_params[2].inv_scale));
299 vout3x01234567 = _mm256_mul_ps(vout3x01234567, _mm256_set1_ps(quantization_params[3].inv_scale));
300 vout4x01234567 = _mm256_mul_ps(vout4x01234567, _mm256_set1_ps(quantization_params[4].inv_scale));
301 vout5x01234567 = _mm256_mul_ps(vout5x01234567, _mm256_set1_ps(quantization_params[5].inv_scale));
302 vout6x01234567 = _mm256_mul_ps(vout6x01234567, _mm256_set1_ps(quantization_params[6].inv_scale));
303 vout7x01234567 = _mm256_mul_ps(vout7x01234567, _mm256_set1_ps(quantization_params[7].inv_scale));
304 vout8x01234567 = _mm256_mul_ps(vout8x01234567, _mm256_set1_ps(quantization_params[8].inv_scale));
305
306 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
307 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
308 w = (const float*) w + 16;
309
310 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
311 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
312 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
313 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
314 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
315 vout5x01234567 = _mm256_fmadd_ps(vout5x01234567, vfilter_output_scale01234567, vbias01234567);
316 vout6x01234567 = _mm256_fmadd_ps(vout6x01234567, vfilter_output_scale01234567, vbias01234567);
317 vout7x01234567 = _mm256_fmadd_ps(vout7x01234567, vfilter_output_scale01234567, vbias01234567);
318 vout8x01234567 = _mm256_fmadd_ps(vout8x01234567, vfilter_output_scale01234567, vbias01234567);
319
320 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
321 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
322 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
323 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
324 vout4x01234567 = _mm256_max_ps(vout4x01234567, voutput_min);
325 vout5x01234567 = _mm256_max_ps(vout5x01234567, voutput_min);
326 vout6x01234567 = _mm256_max_ps(vout6x01234567, voutput_min);
327 vout7x01234567 = _mm256_max_ps(vout7x01234567, voutput_min);
328 vout8x01234567 = _mm256_max_ps(vout8x01234567, voutput_min);
329
330 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
331 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
332 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
333 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
334 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max);
335 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max);
336 vout6x01234567 = _mm256_min_ps(vout6x01234567, voutput_max);
337 vout7x01234567 = _mm256_min_ps(vout7x01234567, voutput_max);
338 vout8x01234567 = _mm256_min_ps(vout8x01234567, voutput_max);
339
340 if XNN_LIKELY(nc >= 8) {
341 _mm256_storeu_ps(c0, vout0x01234567);
342 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
343 c0 = (float*) ((uintptr_t) c0 + cn_stride);
344 _mm256_storeu_ps(c1, vout1x01234567);
345 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
346 c1 = (float*) ((uintptr_t) c1 + cn_stride);
347 _mm256_storeu_ps(c2, vout2x01234567);
348 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
349 c2 = (float*) ((uintptr_t) c2 + cn_stride);
350 _mm256_storeu_ps(c3, vout3x01234567);
351 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
352 c3 = (float*) ((uintptr_t) c3 + cn_stride);
353 _mm256_storeu_ps(c4, vout4x01234567);
354 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
355 c4 = (float*) ((uintptr_t) c4 + cn_stride);
356 _mm256_storeu_ps(c5, vout5x01234567);
357 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
358 c5 = (float*) ((uintptr_t) c5 + cn_stride);
359 _mm256_storeu_ps(c6, vout6x01234567);
360 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
361 c6 = (float*) ((uintptr_t) c6 + cn_stride);
362 _mm256_storeu_ps(c7, vout7x01234567);
363 a7 = (const int8_t*) ((uintptr_t) a7 - kc);
364 c7 = (float*) ((uintptr_t) c7 + cn_stride);
365 _mm256_storeu_ps(c8, vout8x01234567);
366 a8 = (const int8_t*) ((uintptr_t) a8 - kc);
367 c8 = (float*) ((uintptr_t) c8 + cn_stride);
368 nc -= 8;
369 } else {
370 // Prepare mask for valid 32-bit elements (depends on nc).
371 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
372 _mm256_mask_storeu_ps(c0, vmask, vout0x01234567);
373 _mm256_mask_storeu_ps(c1, vmask, vout1x01234567);
374 _mm256_mask_storeu_ps(c2, vmask, vout2x01234567);
375 _mm256_mask_storeu_ps(c3, vmask, vout3x01234567);
376 _mm256_mask_storeu_ps(c4, vmask, vout4x01234567);
377 _mm256_mask_storeu_ps(c5, vmask, vout5x01234567);
378 _mm256_mask_storeu_ps(c6, vmask, vout6x01234567);
379 _mm256_mask_storeu_ps(c7, vmask, vout7x01234567);
380 _mm256_mask_storeu_ps(c8, vmask, vout8x01234567);
381 nc = 0;
382 }
383 } while (nc != 0);
384 }