Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-14x8c8-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_14x8c8__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 <= 14);
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 const int8_t* a9 = (const int8_t*) ((uintptr_t) a8 + a_stride);
95 float* c9 = (float*) ((uintptr_t) c8 + cm_stride);
96 if XNN_UNPREDICTABLE(mr < 10) {
97 a9 = a8;
98 c9 = c8;
99 }
100 const int8_t* a10 = (const int8_t*) ((uintptr_t) a9 + a_stride);
101 float* c10 = (float*) ((uintptr_t) c9 + cm_stride);
102 if XNN_UNPREDICTABLE(mr <= 10) {
103 a10 = a9;
104 c10 = c9;
105 }
106 const int8_t* a11 = (const int8_t*) ((uintptr_t) a10 + a_stride);
107 float* c11 = (float*) ((uintptr_t) c10 + cm_stride);
108 if XNN_UNPREDICTABLE(mr < 12) {
109 a11 = a10;
110 c11 = c10;
111 }
112 const int8_t* a12 = (const int8_t*) ((uintptr_t) a11 + a_stride);
113 float* c12 = (float*) ((uintptr_t) c11 + cm_stride);
114 if XNN_UNPREDICTABLE(mr <= 12) {
115 a12 = a11;
116 c12 = c11;
117 }
118 const int8_t* a13 = (const int8_t*) ((uintptr_t) a12 + a_stride);
119 float* c13 = (float*) ((uintptr_t) c12 + cm_stride);
120 if XNN_UNPREDICTABLE(mr != 14) {
121 a13 = a12;
122 c13 = c12;
123 }
124
125 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
126 XNN_FORCE_REALIZATION(vsign_mask);
127 const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point + 128);
128 const __m256i vinput_zero_point1 = _mm256_set1_epi32((int) quantization_params[1].zero_point + 128);
129 const __m256i vinput_zero_point2 = _mm256_set1_epi32((int) quantization_params[2].zero_point + 128);
130 const __m256i vinput_zero_point3 = _mm256_set1_epi32((int) quantization_params[3].zero_point + 128);
131 const __m256i vinput_zero_point4 = _mm256_set1_epi32((int) quantization_params[4].zero_point + 128);
132 const __m256i vinput_zero_point5 = _mm256_set1_epi32((int) quantization_params[5].zero_point + 128);
133 const __m256i vinput_zero_point6 = _mm256_set1_epi32((int) quantization_params[6].zero_point + 128);
134 const __m256i vinput_zero_point7 = _mm256_set1_epi32((int) quantization_params[7].zero_point + 128);
135 const __m256i vinput_zero_point8 = _mm256_set1_epi32((int) quantization_params[8].zero_point + 128);
136 const __m256i vinput_zero_point9 = _mm256_set1_epi32((int) quantization_params[9].zero_point + 128);
137 const __m256i vinput_zero_point10 = _mm256_set1_epi32((int) quantization_params[10].zero_point + 128);
138 const __m256i vinput_zero_point11 = _mm256_set1_epi32((int) quantization_params[11].zero_point + 128);
139 const __m256i vinput_zero_point12 = _mm256_set1_epi32((int) quantization_params[12].zero_point + 128);
140 const __m256i vinput_zero_point13 = _mm256_set1_epi32((int) quantization_params[13].zero_point + 128);
141 const __m256 voutput_min = _mm256_set1_ps(params->scalar.min);
142 const __m256 voutput_max = _mm256_set1_ps(params->scalar.max);
143 // XNN_FORCE_REALIZATION(voutput_min);
144 // XNN_FORCE_REALIZATION(voutput_max);
145 do {
146 const __m256i vksum01234567 = _mm256_load_si256(w);
147 __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point0);
148 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
149 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
150 __m256i vsum1x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point1);
151 __m256i vacc1x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 0));
152 __m256i vacc1x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 1));
153 __m256i vsum2x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point2);
154 __m256i vacc2x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 0));
155 __m256i vacc2x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 1));
156 __m256i vsum3x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point3);
157 __m256i vacc3x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 0));
158 __m256i vacc3x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 1));
159 __m256i vsum4x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point4);
160 __m256i vacc4x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum4x01234567, 0));
161 __m256i vacc4x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum4x01234567, 1));
162 __m256i vsum5x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point5);
163 __m256i vacc5x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum5x01234567, 0));
164 __m256i vacc5x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum5x01234567, 1));
165 __m256i vsum6x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point6);
166 __m256i vacc6x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum6x01234567, 0));
167 __m256i vacc6x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum6x01234567, 1));
168 __m256i vsum7x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point7);
169 __m256i vacc7x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum7x01234567, 0));
170 __m256i vacc7x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum7x01234567, 1));
171 __m256i vsum8x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point8);
172 __m256i vacc8x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum8x01234567, 0));
173 __m256i vacc8x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum8x01234567, 1));
174 __m256i vsum9x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point9);
175 __m256i vacc9x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum9x01234567, 0));
176 __m256i vacc9x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum9x01234567, 1));
177 __m256i vsum10x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point10);
178 __m256i vacc10x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum10x01234567, 0));
179 __m256i vacc10x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum10x01234567, 1));
180 __m256i vsum11x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point11);
181 __m256i vacc11x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum11x01234567, 0));
182 __m256i vacc11x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum11x01234567, 1));
183 __m256i vsum12x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point12);
184 __m256i vacc12x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum12x01234567, 0));
185 __m256i vacc12x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum12x01234567, 1));
186 __m256i vsum13x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point13);
187 __m256i vacc13x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum13x01234567, 0));
188 __m256i vacc13x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum13x01234567, 1));
189 w = (const int32_t*) w + 8;
190
191 size_t k = kc;
192 while (k >= 16 * sizeof(int8_t)) {
193 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
194 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
195 a0 += 16;
196 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
197 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
198 a1 += 16;
199 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
200 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
201 a2 += 16;
202 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
203 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
204 a3 += 16;
205 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
206 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
207 a4 += 16;
208 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
209 const __m256i va5x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
210 a5 += 16;
211 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
212 const __m256i va6x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
213 a6 += 16;
214 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
215 const __m256i va7x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7 + 8)), vsign_mask);
216 a7 += 16;
217 const __m256i va8x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8)), vsign_mask);
218 const __m256i va8x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8 + 8)), vsign_mask);
219 a8 += 16;
220 const __m256i va9x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9)), vsign_mask);
221 const __m256i va9x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9 + 8)), vsign_mask);
222 a9 += 16;
223 const __m256i va10x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a10)), vsign_mask);
224 const __m256i va10x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a10 + 8)), vsign_mask);
225 a10 += 16;
226 const __m256i va11x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a11)), vsign_mask);
227 const __m256i va11x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a11 + 8)), vsign_mask);
228 a11 += 16;
229 const __m256i va12x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a12)), vsign_mask);
230 const __m256i va12x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a12 + 8)), vsign_mask);
231 a12 += 16;
232 const __m256i va13x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a13)), vsign_mask);
233 const __m256i va13x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a13 + 8)), vsign_mask);
234 a13 += 16;
235
236 const __m256i vb01234567x0123 = _mm256_load_si256(w);
237 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
238 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
239 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
240 xnn_prefetch_to_l1((const int8_t*) w + 896);
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 vacc9x0123 = _mm256_dpbusd_epi32(vacc9x0123, va9x01234567, vb01234567x0123);
261 vacc9x4567 = _mm256_dpbusd_epi32(vacc9x4567, va9x01234567, vb89ABCDEFx0123);
262 vacc10x0123 = _mm256_dpbusd_epi32(vacc10x0123, va10x01234567, vb01234567x0123);
263 vacc10x4567 = _mm256_dpbusd_epi32(vacc10x4567, va10x01234567, vb89ABCDEFx0123);
264 vacc11x0123 = _mm256_dpbusd_epi32(vacc11x0123, va11x01234567, vb01234567x0123);
265 vacc11x4567 = _mm256_dpbusd_epi32(vacc11x4567, va11x01234567, vb89ABCDEFx0123);
266 vacc12x0123 = _mm256_dpbusd_epi32(vacc12x0123, va12x01234567, vb01234567x0123);
267 vacc12x4567 = _mm256_dpbusd_epi32(vacc12x4567, va12x01234567, vb89ABCDEFx0123);
268 vacc13x0123 = _mm256_dpbusd_epi32(vacc13x0123, va13x01234567, vb01234567x0123);
269 vacc13x4567 = _mm256_dpbusd_epi32(vacc13x4567, va13x01234567, vb89ABCDEFx0123);
270 xnn_prefetch_to_l1((const int8_t*) w + 960);
271 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
272 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
273 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
274 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
275 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
276 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
277 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
278 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
279 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
280 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
281 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x89ABCDEF, vb01234567x4567);
282 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);
283 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x89ABCDEF, vb01234567x4567);
284 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x89ABCDEF, vb89ABCDEFx4567);
285 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x89ABCDEF, vb01234567x4567);
286 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x89ABCDEF, vb89ABCDEFx4567);
287 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x89ABCDEF, vb01234567x4567);
288 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x89ABCDEF, vb89ABCDEFx4567);
289 vacc9x0123 = _mm256_dpbusd_epi32(vacc9x0123, va9x89ABCDEF, vb01234567x4567);
290 vacc9x4567 = _mm256_dpbusd_epi32(vacc9x4567, va9x89ABCDEF, vb89ABCDEFx4567);
291 vacc10x0123 = _mm256_dpbusd_epi32(vacc10x0123, va10x89ABCDEF, vb01234567x4567);
292 vacc10x4567 = _mm256_dpbusd_epi32(vacc10x4567, va10x89ABCDEF, vb89ABCDEFx4567);
293 vacc11x0123 = _mm256_dpbusd_epi32(vacc11x0123, va11x89ABCDEF, vb01234567x4567);
294 vacc11x4567 = _mm256_dpbusd_epi32(vacc11x4567, va11x89ABCDEF, vb89ABCDEFx4567);
295 vacc12x0123 = _mm256_dpbusd_epi32(vacc12x0123, va12x89ABCDEF, vb01234567x4567);
296 vacc12x4567 = _mm256_dpbusd_epi32(vacc12x4567, va12x89ABCDEF, vb89ABCDEFx4567);
297 vacc13x0123 = _mm256_dpbusd_epi32(vacc13x0123, va13x89ABCDEF, vb01234567x4567);
298 vacc13x4567 = _mm256_dpbusd_epi32(vacc13x4567, va13x89ABCDEF, vb89ABCDEFx4567);
299
300 w = (const int8_t*) w + 128;
301 k -= 16 * sizeof(int8_t);
302 }
303
304 if (k != 0) {
305 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
306 a0 += 8;
307 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
308 a1 += 8;
309 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
310 a2 += 8;
311 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
312 a3 += 8;
313 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
314 a4 += 8;
315 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
316 a5 += 8;
317 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
318 a6 += 8;
319 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
320 a7 += 8;
321 const __m256i va8x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8)), vsign_mask);
322 a8 += 8;
323 const __m256i va9x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9)), vsign_mask);
324 a9 += 8;
325 const __m256i va10x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a10)), vsign_mask);
326 a10 += 8;
327 const __m256i va11x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a11)), vsign_mask);
328 a11 += 8;
329 const __m256i va12x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a12)), vsign_mask);
330 a12 += 8;
331 const __m256i va13x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a13)), vsign_mask);
332 a13 += 8;
333
334 const __m256i vb01234567x0123 = _mm256_load_si256(w);
335 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
336
337 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
338 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
339 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
340 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
341 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
342 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
343 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
344 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
345 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
346 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
347 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
348 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
349 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
350 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
351 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
352 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
353 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x01234567, vb01234567x0123);
354 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x01234567, vb89ABCDEFx0123);
355 vacc9x0123 = _mm256_dpbusd_epi32(vacc9x0123, va9x01234567, vb01234567x0123);
356 vacc9x4567 = _mm256_dpbusd_epi32(vacc9x4567, va9x01234567, vb89ABCDEFx0123);
357 vacc10x0123 = _mm256_dpbusd_epi32(vacc10x0123, va10x01234567, vb01234567x0123);
358 vacc10x4567 = _mm256_dpbusd_epi32(vacc10x4567, va10x01234567, vb89ABCDEFx0123);
359 vacc11x0123 = _mm256_dpbusd_epi32(vacc11x0123, va11x01234567, vb01234567x0123);
360 vacc11x4567 = _mm256_dpbusd_epi32(vacc11x4567, va11x01234567, vb89ABCDEFx0123);
361 vacc12x0123 = _mm256_dpbusd_epi32(vacc12x0123, va12x01234567, vb01234567x0123);
362 vacc12x4567 = _mm256_dpbusd_epi32(vacc12x4567, va12x01234567, vb89ABCDEFx0123);
363 vacc13x0123 = _mm256_dpbusd_epi32(vacc13x0123, va13x01234567, vb01234567x0123);
364 vacc13x4567 = _mm256_dpbusd_epi32(vacc13x4567, va13x01234567, vb89ABCDEFx0123);
365 xnn_prefetch_to_l1((const int8_t*) w + 960);
366
367 w = (const int8_t*) w + 64;
368 k -= 8 * sizeof(int8_t);
369 }
370
371 // Add adjacent pairs
372 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
373 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
374 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
375 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
376 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
377 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
378 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
379 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
380 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
381 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
382 const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);
383 __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));
384 const __m256i vsum6x02134657 = _mm256_hadd_epi32(vacc6x0123, vacc6x4567);
385 __m256i vacc6x01234567 = _mm256_permute4x64_epi64(vsum6x02134657, _MM_SHUFFLE(3, 1, 2, 0));
386 const __m256i vsum7x02134657 = _mm256_hadd_epi32(vacc7x0123, vacc7x4567);
387 __m256i vacc7x01234567 = _mm256_permute4x64_epi64(vsum7x02134657, _MM_SHUFFLE(3, 1, 2, 0));
388 const __m256i vsum8x02134657 = _mm256_hadd_epi32(vacc8x0123, vacc8x4567);
389 __m256i vacc8x01234567 = _mm256_permute4x64_epi64(vsum8x02134657, _MM_SHUFFLE(3, 1, 2, 0));
390 const __m256i vsum9x02134657 = _mm256_hadd_epi32(vacc9x0123, vacc9x4567);
391 __m256i vacc9x01234567 = _mm256_permute4x64_epi64(vsum9x02134657, _MM_SHUFFLE(3, 1, 2, 0));
392 const __m256i vsum10x02134657 = _mm256_hadd_epi32(vacc10x0123, vacc10x4567);
393 __m256i vacc10x01234567 = _mm256_permute4x64_epi64(vsum10x02134657, _MM_SHUFFLE(3, 1, 2, 0));
394 const __m256i vsum11x02134657 = _mm256_hadd_epi32(vacc11x0123, vacc11x4567);
395 __m256i vacc11x01234567 = _mm256_permute4x64_epi64(vsum11x02134657, _MM_SHUFFLE(3, 1, 2, 0));
396 const __m256i vsum12x02134657 = _mm256_hadd_epi32(vacc12x0123, vacc12x4567);
397 __m256i vacc12x01234567 = _mm256_permute4x64_epi64(vsum12x02134657, _MM_SHUFFLE(3, 1, 2, 0));
398 const __m256i vsum13x02134657 = _mm256_hadd_epi32(vacc13x0123, vacc13x4567);
399 __m256i vacc13x01234567 = _mm256_permute4x64_epi64(vsum13x02134657, _MM_SHUFFLE(3, 1, 2, 0));
400
401 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
402 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
403 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
404 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
405 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
406 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
407 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567);
408 __m256 vout7x01234567 = _mm256_cvtepi32_ps(vacc7x01234567);
409 __m256 vout8x01234567 = _mm256_cvtepi32_ps(vacc8x01234567);
410 __m256 vout9x01234567 = _mm256_cvtepi32_ps(vacc9x01234567);
411 __m256 vout10x01234567 = _mm256_cvtepi32_ps(vacc10x01234567);
412 __m256 vout11x01234567 = _mm256_cvtepi32_ps(vacc11x01234567);
413 __m256 vout12x01234567 = _mm256_cvtepi32_ps(vacc12x01234567);
414 __m256 vout13x01234567 = _mm256_cvtepi32_ps(vacc13x01234567);
415
416 vout0x01234567 = _mm256_mul_ps(vout0x01234567, _mm256_set1_ps(quantization_params[0].inv_scale));
417 vout1x01234567 = _mm256_mul_ps(vout1x01234567, _mm256_set1_ps(quantization_params[1].inv_scale));
418 vout2x01234567 = _mm256_mul_ps(vout2x01234567, _mm256_set1_ps(quantization_params[2].inv_scale));
419 vout3x01234567 = _mm256_mul_ps(vout3x01234567, _mm256_set1_ps(quantization_params[3].inv_scale));
420 vout4x01234567 = _mm256_mul_ps(vout4x01234567, _mm256_set1_ps(quantization_params[4].inv_scale));
421 vout5x01234567 = _mm256_mul_ps(vout5x01234567, _mm256_set1_ps(quantization_params[5].inv_scale));
422 vout6x01234567 = _mm256_mul_ps(vout6x01234567, _mm256_set1_ps(quantization_params[6].inv_scale));
423 vout7x01234567 = _mm256_mul_ps(vout7x01234567, _mm256_set1_ps(quantization_params[7].inv_scale));
424 vout8x01234567 = _mm256_mul_ps(vout8x01234567, _mm256_set1_ps(quantization_params[8].inv_scale));
425 vout9x01234567 = _mm256_mul_ps(vout9x01234567, _mm256_set1_ps(quantization_params[9].inv_scale));
426 vout10x01234567 = _mm256_mul_ps(vout10x01234567, _mm256_set1_ps(quantization_params[10].inv_scale));
427 vout11x01234567 = _mm256_mul_ps(vout11x01234567, _mm256_set1_ps(quantization_params[11].inv_scale));
428 vout12x01234567 = _mm256_mul_ps(vout12x01234567, _mm256_set1_ps(quantization_params[12].inv_scale));
429 vout13x01234567 = _mm256_mul_ps(vout13x01234567, _mm256_set1_ps(quantization_params[13].inv_scale));
430
431 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
432 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
433 w = (const float*) w + 16;
434
435 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
436 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
437 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
438 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
439 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
440 vout5x01234567 = _mm256_fmadd_ps(vout5x01234567, vfilter_output_scale01234567, vbias01234567);
441 vout6x01234567 = _mm256_fmadd_ps(vout6x01234567, vfilter_output_scale01234567, vbias01234567);
442 vout7x01234567 = _mm256_fmadd_ps(vout7x01234567, vfilter_output_scale01234567, vbias01234567);
443 vout8x01234567 = _mm256_fmadd_ps(vout8x01234567, vfilter_output_scale01234567, vbias01234567);
444 vout9x01234567 = _mm256_fmadd_ps(vout9x01234567, vfilter_output_scale01234567, vbias01234567);
445 vout10x01234567 = _mm256_fmadd_ps(vout10x01234567, vfilter_output_scale01234567, vbias01234567);
446 vout11x01234567 = _mm256_fmadd_ps(vout11x01234567, vfilter_output_scale01234567, vbias01234567);
447 vout12x01234567 = _mm256_fmadd_ps(vout12x01234567, vfilter_output_scale01234567, vbias01234567);
448 vout13x01234567 = _mm256_fmadd_ps(vout13x01234567, vfilter_output_scale01234567, vbias01234567);
449
450 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
451 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
452 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
453 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
454 vout4x01234567 = _mm256_max_ps(vout4x01234567, voutput_min);
455 vout5x01234567 = _mm256_max_ps(vout5x01234567, voutput_min);
456 vout6x01234567 = _mm256_max_ps(vout6x01234567, voutput_min);
457 vout7x01234567 = _mm256_max_ps(vout7x01234567, voutput_min);
458 vout8x01234567 = _mm256_max_ps(vout8x01234567, voutput_min);
459 vout9x01234567 = _mm256_max_ps(vout9x01234567, voutput_min);
460 vout10x01234567 = _mm256_max_ps(vout10x01234567, voutput_min);
461 vout11x01234567 = _mm256_max_ps(vout11x01234567, voutput_min);
462 vout12x01234567 = _mm256_max_ps(vout12x01234567, voutput_min);
463 vout13x01234567 = _mm256_max_ps(vout13x01234567, voutput_min);
464
465 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
466 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
467 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
468 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
469 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max);
470 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max);
471 vout6x01234567 = _mm256_min_ps(vout6x01234567, voutput_max);
472 vout7x01234567 = _mm256_min_ps(vout7x01234567, voutput_max);
473 vout8x01234567 = _mm256_min_ps(vout8x01234567, voutput_max);
474 vout9x01234567 = _mm256_min_ps(vout9x01234567, voutput_max);
475 vout10x01234567 = _mm256_min_ps(vout10x01234567, voutput_max);
476 vout11x01234567 = _mm256_min_ps(vout11x01234567, voutput_max);
477 vout12x01234567 = _mm256_min_ps(vout12x01234567, voutput_max);
478 vout13x01234567 = _mm256_min_ps(vout13x01234567, voutput_max);
479
480 if XNN_LIKELY(nc >= 8) {
481 _mm256_storeu_ps(c0, vout0x01234567);
482 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
483 c0 = (float*) ((uintptr_t) c0 + cn_stride);
484 _mm256_storeu_ps(c1, vout1x01234567);
485 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
486 c1 = (float*) ((uintptr_t) c1 + cn_stride);
487 _mm256_storeu_ps(c2, vout2x01234567);
488 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
489 c2 = (float*) ((uintptr_t) c2 + cn_stride);
490 _mm256_storeu_ps(c3, vout3x01234567);
491 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
492 c3 = (float*) ((uintptr_t) c3 + cn_stride);
493 _mm256_storeu_ps(c4, vout4x01234567);
494 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
495 c4 = (float*) ((uintptr_t) c4 + cn_stride);
496 _mm256_storeu_ps(c5, vout5x01234567);
497 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
498 c5 = (float*) ((uintptr_t) c5 + cn_stride);
499 _mm256_storeu_ps(c6, vout6x01234567);
500 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
501 c6 = (float*) ((uintptr_t) c6 + cn_stride);
502 _mm256_storeu_ps(c7, vout7x01234567);
503 a7 = (const int8_t*) ((uintptr_t) a7 - kc);
504 c7 = (float*) ((uintptr_t) c7 + cn_stride);
505 _mm256_storeu_ps(c8, vout8x01234567);
506 a8 = (const int8_t*) ((uintptr_t) a8 - kc);
507 c8 = (float*) ((uintptr_t) c8 + cn_stride);
508 _mm256_storeu_ps(c9, vout9x01234567);
509 a9 = (const int8_t*) ((uintptr_t) a9 - kc);
510 c9 = (float*) ((uintptr_t) c9 + cn_stride);
511 _mm256_storeu_ps(c10, vout10x01234567);
512 a10 = (const int8_t*) ((uintptr_t) a10 - kc);
513 c10 = (float*) ((uintptr_t) c10 + cn_stride);
514 _mm256_storeu_ps(c11, vout11x01234567);
515 a11 = (const int8_t*) ((uintptr_t) a11 - kc);
516 c11 = (float*) ((uintptr_t) c11 + cn_stride);
517 _mm256_storeu_ps(c12, vout12x01234567);
518 a12 = (const int8_t*) ((uintptr_t) a12 - kc);
519 c12 = (float*) ((uintptr_t) c12 + cn_stride);
520 _mm256_storeu_ps(c13, vout13x01234567);
521 a13 = (const int8_t*) ((uintptr_t) a13 - kc);
522 c13 = (float*) ((uintptr_t) c13 + cn_stride);
523 nc -= 8;
524 } else {
525 // Prepare mask for valid 32-bit elements (depends on nc).
526 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
527 _mm256_mask_storeu_ps(c0, vmask, vout0x01234567);
528 _mm256_mask_storeu_ps(c1, vmask, vout1x01234567);
529 _mm256_mask_storeu_ps(c2, vmask, vout2x01234567);
530 _mm256_mask_storeu_ps(c3, vmask, vout3x01234567);
531 _mm256_mask_storeu_ps(c4, vmask, vout4x01234567);
532 _mm256_mask_storeu_ps(c5, vmask, vout5x01234567);
533 _mm256_mask_storeu_ps(c6, vmask, vout6x01234567);
534 _mm256_mask_storeu_ps(c7, vmask, vout7x01234567);
535 _mm256_mask_storeu_ps(c8, vmask, vout8x01234567);
536 _mm256_mask_storeu_ps(c9, vmask, vout9x01234567);
537 _mm256_mask_storeu_ps(c10, vmask, vout10x01234567);
538 _mm256_mask_storeu_ps(c11, vmask, vout11x01234567);
539 _mm256_mask_storeu_ps(c12, vmask, vout12x01234567);
540 _mm256_mask_storeu_ps(c13, vmask, vout13x01234567);
541 nc = 0;
542 }
543 } while (nc != 0);
544 }