Debian Codegraph @

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