Debian Codegraph @

src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-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_f16_qc4w_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 xnn_float16* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 const struct xnn_f16_qc4w_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 uint16_t* c0 = (uint16_t*) c;
45 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
46 uint16_t* c1 = (uint16_t*) ((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 uint16_t* c2 = (uint16_t*) ((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 uint16_t* c3 = (uint16_t*) ((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 uint16_t* c4 = (uint16_t*) ((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 uint16_t* c5 = (uint16_t*) ((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 uint16_t* c6 = (uint16_t*) ((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 uint16_t* c7 = (uint16_t*) ((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 uint16_t* c8 = (uint16_t*) ((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 uint16_t* c9 = (uint16_t*) ((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 uint16_t* c10 = (uint16_t*) ((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 uint16_t* c11 = (uint16_t*) ((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 uint16_t* c12 = (uint16_t*) ((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 uint16_t* c13 = (uint16_t*) ((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_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
141 const __m256 voutput_max = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
142 // XNN_FORCE_REALIZATION(voutput_min);
143 // XNN_FORCE_REALIZATION(voutput_max);
144 const __m256i vmask = _mm256_set1_epi8(0xF0);
145 XNN_FORCE_REALIZATION(vmask);
146 do {
147 const __m256i vksum01234567 = _mm256_load_si256(w);
148 __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point0);
149 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
150 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
151 __m256i vsum1x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point1);
152 __m256i vacc1x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 0));
153 __m256i vacc1x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 1));
154 __m256i vsum2x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point2);
155 __m256i vacc2x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 0));
156 __m256i vacc2x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 1));
157 __m256i vsum3x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point3);
158 __m256i vacc3x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 0));
159 __m256i vacc3x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 1));
160 __m256i vsum4x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point4);
161 __m256i vacc4x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum4x01234567, 0));
162 __m256i vacc4x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum4x01234567, 1));
163 __m256i vsum5x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point5);
164 __m256i vacc5x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum5x01234567, 0));
165 __m256i vacc5x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum5x01234567, 1));
166 __m256i vsum6x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point6);
167 __m256i vacc6x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum6x01234567, 0));
168 __m256i vacc6x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum6x01234567, 1));
169 __m256i vsum7x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point7);
170 __m256i vacc7x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum7x01234567, 0));
171 __m256i vacc7x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum7x01234567, 1));
172 __m256i vsum8x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point8);
173 __m256i vacc8x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum8x01234567, 0));
174 __m256i vacc8x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum8x01234567, 1));
175 __m256i vsum9x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point9);
176 __m256i vacc9x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum9x01234567, 0));
177 __m256i vacc9x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum9x01234567, 1));
178 __m256i vsum10x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point10);
179 __m256i vacc10x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum10x01234567, 0));
180 __m256i vacc10x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum10x01234567, 1));
181 __m256i vsum11x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point11);
182 __m256i vacc11x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum11x01234567, 0));
183 __m256i vacc11x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum11x01234567, 1));
184 __m256i vsum12x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point12);
185 __m256i vacc12x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum12x01234567, 0));
186 __m256i vacc12x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum12x01234567, 1));
187 __m256i vsum13x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point13);
188 __m256i vacc13x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum13x01234567, 0));
189 __m256i vacc13x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum13x01234567, 1));
190 w = (const int32_t*) w + 8;
191
192 size_t k = kc;
193 while (k >= 16 * sizeof(int8_t)) {
194 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
195 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
196 a0 += 16;
197 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
198 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
199 a1 += 16;
200 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
201 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
202 a2 += 16;
203 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
204 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
205 a3 += 16;
206 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
207 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
208 a4 += 16;
209 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
210 const __m256i va5x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
211 a5 += 16;
212 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
213 const __m256i va6x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
214 a6 += 16;
215 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
216 const __m256i va7x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7 + 8)), vsign_mask);
217 a7 += 16;
218 const __m256i va8x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8)), vsign_mask);
219 const __m256i va8x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8 + 8)), vsign_mask);
220 a8 += 16;
221 const __m256i va9x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9)), vsign_mask);
222 const __m256i va9x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9 + 8)), vsign_mask);
223 a9 += 16;
224 const __m256i va10x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a10)), vsign_mask);
225 const __m256i va10x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a10 + 8)), vsign_mask);
226 a10 += 16;
227 const __m256i va11x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a11)), vsign_mask);
228 const __m256i va11x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a11 + 8)), vsign_mask);
229 a11 += 16;
230 const __m256i va12x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a12)), vsign_mask);
231 const __m256i va12x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a12 + 8)), vsign_mask);
232 a12 += 16;
233 const __m256i va13x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a13)), vsign_mask);
234 const __m256i va13x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a13 + 8)), vsign_mask);
235 a13 += 16;
236
237 const __m256i vbb01234567x01234567 = _mm256_load_si256(w);
238 const __m256i vbb89ABCDEFx01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
239 const __m256i vbs01234567x0123 = _mm256_slli_epi32(vbb01234567x01234567, 4);
240 const __m256i vbs89ABCDEFx0123 = _mm256_slli_epi32(vbb89ABCDEFx01234567, 4);
241 const __m256i vb01234567x4567 = _mm256_and_si256(vbb01234567x01234567, vmask);
242 const __m256i vb89ABCDEFx4567 = _mm256_and_si256(vbb89ABCDEFx01234567, vmask);
243 const __m256i vb01234567x0123 = _mm256_and_si256(vbs01234567x0123, vmask);
244 const __m256i vb89ABCDEFx0123 = _mm256_and_si256(vbs89ABCDEFx0123, vmask);
245
246 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
247 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
248 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
249 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
250 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
251 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
252 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
253 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
254 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
255 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
256 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
257 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
258 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
259 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
260 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
261 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
262 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x01234567, vb01234567x0123);
263 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x01234567, vb89ABCDEFx0123);
264 vacc9x0123 = _mm256_dpbusd_epi32(vacc9x0123, va9x01234567, vb01234567x0123);
265 vacc9x4567 = _mm256_dpbusd_epi32(vacc9x4567, va9x01234567, vb89ABCDEFx0123);
266 vacc10x0123 = _mm256_dpbusd_epi32(vacc10x0123, va10x01234567, vb01234567x0123);
267 vacc10x4567 = _mm256_dpbusd_epi32(vacc10x4567, va10x01234567, vb89ABCDEFx0123);
268 vacc11x0123 = _mm256_dpbusd_epi32(vacc11x0123, va11x01234567, vb01234567x0123);
269 vacc11x4567 = _mm256_dpbusd_epi32(vacc11x4567, va11x01234567, vb89ABCDEFx0123);
270 vacc12x0123 = _mm256_dpbusd_epi32(vacc12x0123, va12x01234567, vb01234567x0123);
271 vacc12x4567 = _mm256_dpbusd_epi32(vacc12x4567, va12x01234567, vb89ABCDEFx0123);
272 vacc13x0123 = _mm256_dpbusd_epi32(vacc13x0123, va13x01234567, vb01234567x0123);
273 vacc13x4567 = _mm256_dpbusd_epi32(vacc13x4567, va13x01234567, vb89ABCDEFx0123);
274 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
275 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
276 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
277 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
278 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
279 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
280 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
281 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
282 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
283 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
284 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x89ABCDEF, vb01234567x4567);
285 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);
286 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x89ABCDEF, vb01234567x4567);
287 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x89ABCDEF, vb89ABCDEFx4567);
288 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x89ABCDEF, vb01234567x4567);
289 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x89ABCDEF, vb89ABCDEFx4567);
290 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x89ABCDEF, vb01234567x4567);
291 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x89ABCDEF, vb89ABCDEFx4567);
292 vacc9x0123 = _mm256_dpbusd_epi32(vacc9x0123, va9x89ABCDEF, vb01234567x4567);
293 vacc9x4567 = _mm256_dpbusd_epi32(vacc9x4567, va9x89ABCDEF, vb89ABCDEFx4567);
294 vacc10x0123 = _mm256_dpbusd_epi32(vacc10x0123, va10x89ABCDEF, vb01234567x4567);
295 vacc10x4567 = _mm256_dpbusd_epi32(vacc10x4567, va10x89ABCDEF, vb89ABCDEFx4567);
296 vacc11x0123 = _mm256_dpbusd_epi32(vacc11x0123, va11x89ABCDEF, vb01234567x4567);
297 vacc11x4567 = _mm256_dpbusd_epi32(vacc11x4567, va11x89ABCDEF, vb89ABCDEFx4567);
298 vacc12x0123 = _mm256_dpbusd_epi32(vacc12x0123, va12x89ABCDEF, vb01234567x4567);
299 vacc12x4567 = _mm256_dpbusd_epi32(vacc12x4567, va12x89ABCDEF, vb89ABCDEFx4567);
300 vacc13x0123 = _mm256_dpbusd_epi32(vacc13x0123, va13x89ABCDEF, vb01234567x4567);
301 vacc13x4567 = _mm256_dpbusd_epi32(vacc13x4567, va13x89ABCDEF, vb89ABCDEFx4567);
302
303 w = (const int8_t*) w + 64;
304 k -= 16 * sizeof(int8_t);
305 }
306
307 if (k != 0) {
308 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
309 a0 += 8;
310 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
311 a1 += 8;
312 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
313 a2 += 8;
314 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
315 a3 += 8;
316 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
317 a4 += 8;
318 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
319 a5 += 8;
320 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
321 a6 += 8;
322 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
323 a7 += 8;
324 const __m256i va8x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8)), vsign_mask);
325 a8 += 8;
326 const __m256i va9x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9)), vsign_mask);
327 a9 += 8;
328 const __m256i va10x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a10)), vsign_mask);
329 a10 += 8;
330 const __m256i va11x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a11)), vsign_mask);
331 a11 += 8;
332 const __m256i va12x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a12)), vsign_mask);
333 a12 += 8;
334 const __m256i va13x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a13)), vsign_mask);
335 a13 += 8;
336
337 const __m256i vbb01234567x01234567 = _mm256_load_si256(w);
338 const __m256i vbb89ABCDEFx01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
339 const __m256i vb01234567x0123 = _mm256_slli_epi32(vbb01234567x01234567, 4);
340 const __m256i vb89ABCDEFx0123 = _mm256_slli_epi32(vbb89ABCDEFx01234567, 4);
341
342 vacc0x0123 = _mm256_dpbusd_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
343 vacc0x4567 = _mm256_dpbusd_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
344 vacc1x0123 = _mm256_dpbusd_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
345 vacc1x4567 = _mm256_dpbusd_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
346 vacc2x0123 = _mm256_dpbusd_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
347 vacc2x4567 = _mm256_dpbusd_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
348 vacc3x0123 = _mm256_dpbusd_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
349 vacc3x4567 = _mm256_dpbusd_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
350 vacc4x0123 = _mm256_dpbusd_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
351 vacc4x4567 = _mm256_dpbusd_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
352 vacc5x0123 = _mm256_dpbusd_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
353 vacc5x4567 = _mm256_dpbusd_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
354 vacc6x0123 = _mm256_dpbusd_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
355 vacc6x4567 = _mm256_dpbusd_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
356 vacc7x0123 = _mm256_dpbusd_epi32(vacc7x0123, va7x01234567, vb01234567x0123);
357 vacc7x4567 = _mm256_dpbusd_epi32(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
358 vacc8x0123 = _mm256_dpbusd_epi32(vacc8x0123, va8x01234567, vb01234567x0123);
359 vacc8x4567 = _mm256_dpbusd_epi32(vacc8x4567, va8x01234567, vb89ABCDEFx0123);
360 vacc9x0123 = _mm256_dpbusd_epi32(vacc9x0123, va9x01234567, vb01234567x0123);
361 vacc9x4567 = _mm256_dpbusd_epi32(vacc9x4567, va9x01234567, vb89ABCDEFx0123);
362 vacc10x0123 = _mm256_dpbusd_epi32(vacc10x0123, va10x01234567, vb01234567x0123);
363 vacc10x4567 = _mm256_dpbusd_epi32(vacc10x4567, va10x01234567, vb89ABCDEFx0123);
364 vacc11x0123 = _mm256_dpbusd_epi32(vacc11x0123, va11x01234567, vb01234567x0123);
365 vacc11x4567 = _mm256_dpbusd_epi32(vacc11x4567, va11x01234567, vb89ABCDEFx0123);
366 vacc12x0123 = _mm256_dpbusd_epi32(vacc12x0123, va12x01234567, vb01234567x0123);
367 vacc12x4567 = _mm256_dpbusd_epi32(vacc12x4567, va12x01234567, vb89ABCDEFx0123);
368 vacc13x0123 = _mm256_dpbusd_epi32(vacc13x0123, va13x01234567, vb01234567x0123);
369 vacc13x4567 = _mm256_dpbusd_epi32(vacc13x4567, va13x01234567, vb89ABCDEFx0123);
370
371 w = (const int8_t*) w + 64;
372 k -= 8 * sizeof(int8_t);
373 }
374
375 // Add adjacent pairs
376 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
377 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
378 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
379 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
380 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
381 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
382 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
383 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
384 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
385 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
386 const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);
387 __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));
388 const __m256i vsum6x02134657 = _mm256_hadd_epi32(vacc6x0123, vacc6x4567);
389 __m256i vacc6x01234567 = _mm256_permute4x64_epi64(vsum6x02134657, _MM_SHUFFLE(3, 1, 2, 0));
390 const __m256i vsum7x02134657 = _mm256_hadd_epi32(vacc7x0123, vacc7x4567);
391 __m256i vacc7x01234567 = _mm256_permute4x64_epi64(vsum7x02134657, _MM_SHUFFLE(3, 1, 2, 0));
392 const __m256i vsum8x02134657 = _mm256_hadd_epi32(vacc8x0123, vacc8x4567);
393 __m256i vacc8x01234567 = _mm256_permute4x64_epi64(vsum8x02134657, _MM_SHUFFLE(3, 1, 2, 0));
394 const __m256i vsum9x02134657 = _mm256_hadd_epi32(vacc9x0123, vacc9x4567);
395 __m256i vacc9x01234567 = _mm256_permute4x64_epi64(vsum9x02134657, _MM_SHUFFLE(3, 1, 2, 0));
396 const __m256i vsum10x02134657 = _mm256_hadd_epi32(vacc10x0123, vacc10x4567);
397 __m256i vacc10x01234567 = _mm256_permute4x64_epi64(vsum10x02134657, _MM_SHUFFLE(3, 1, 2, 0));
398 const __m256i vsum11x02134657 = _mm256_hadd_epi32(vacc11x0123, vacc11x4567);
399 __m256i vacc11x01234567 = _mm256_permute4x64_epi64(vsum11x02134657, _MM_SHUFFLE(3, 1, 2, 0));
400 const __m256i vsum12x02134657 = _mm256_hadd_epi32(vacc12x0123, vacc12x4567);
401 __m256i vacc12x01234567 = _mm256_permute4x64_epi64(vsum12x02134657, _MM_SHUFFLE(3, 1, 2, 0));
402 const __m256i vsum13x02134657 = _mm256_hadd_epi32(vacc13x0123, vacc13x4567);
403 __m256i vacc13x01234567 = _mm256_permute4x64_epi64(vsum13x02134657, _MM_SHUFFLE(3, 1, 2, 0));
404
405 vacc0x01234567 = _mm256_srai_epi32(vacc0x01234567, 4);
406 vacc1x01234567 = _mm256_srai_epi32(vacc1x01234567, 4);
407 vacc2x01234567 = _mm256_srai_epi32(vacc2x01234567, 4);
408 vacc3x01234567 = _mm256_srai_epi32(vacc3x01234567, 4);
409 vacc4x01234567 = _mm256_srai_epi32(vacc4x01234567, 4);
410 vacc5x01234567 = _mm256_srai_epi32(vacc5x01234567, 4);
411 vacc6x01234567 = _mm256_srai_epi32(vacc6x01234567, 4);
412 vacc7x01234567 = _mm256_srai_epi32(vacc7x01234567, 4);
413 vacc8x01234567 = _mm256_srai_epi32(vacc8x01234567, 4);
414 vacc9x01234567 = _mm256_srai_epi32(vacc9x01234567, 4);
415 vacc10x01234567 = _mm256_srai_epi32(vacc10x01234567, 4);
416 vacc11x01234567 = _mm256_srai_epi32(vacc11x01234567, 4);
417 vacc12x01234567 = _mm256_srai_epi32(vacc12x01234567, 4);
418 vacc13x01234567 = _mm256_srai_epi32(vacc13x01234567, 4);
419 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
420 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
421 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
422 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
423 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
424 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
425 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567);
426 __m256 vout7x01234567 = _mm256_cvtepi32_ps(vacc7x01234567);
427 __m256 vout8x01234567 = _mm256_cvtepi32_ps(vacc8x01234567);
428 __m256 vout9x01234567 = _mm256_cvtepi32_ps(vacc9x01234567);
429 __m256 vout10x01234567 = _mm256_cvtepi32_ps(vacc10x01234567);
430 __m256 vout11x01234567 = _mm256_cvtepi32_ps(vacc11x01234567);
431 __m256 vout12x01234567 = _mm256_cvtepi32_ps(vacc12x01234567);
432 __m256 vout13x01234567 = _mm256_cvtepi32_ps(vacc13x01234567);
433
434 vout0x01234567 = _mm256_mul_ps(vout0x01234567, _mm256_set1_ps(quantization_params[0].inv_scale));
435 vout1x01234567 = _mm256_mul_ps(vout1x01234567, _mm256_set1_ps(quantization_params[1].inv_scale));
436 vout2x01234567 = _mm256_mul_ps(vout2x01234567, _mm256_set1_ps(quantization_params[2].inv_scale));
437 vout3x01234567 = _mm256_mul_ps(vout3x01234567, _mm256_set1_ps(quantization_params[3].inv_scale));
438 vout4x01234567 = _mm256_mul_ps(vout4x01234567, _mm256_set1_ps(quantization_params[4].inv_scale));
439 vout5x01234567 = _mm256_mul_ps(vout5x01234567, _mm256_set1_ps(quantization_params[5].inv_scale));
440 vout6x01234567 = _mm256_mul_ps(vout6x01234567, _mm256_set1_ps(quantization_params[6].inv_scale));
441 vout7x01234567 = _mm256_mul_ps(vout7x01234567, _mm256_set1_ps(quantization_params[7].inv_scale));
442 vout8x01234567 = _mm256_mul_ps(vout8x01234567, _mm256_set1_ps(quantization_params[8].inv_scale));
443 vout9x01234567 = _mm256_mul_ps(vout9x01234567, _mm256_set1_ps(quantization_params[9].inv_scale));
444 vout10x01234567 = _mm256_mul_ps(vout10x01234567, _mm256_set1_ps(quantization_params[10].inv_scale));
445 vout11x01234567 = _mm256_mul_ps(vout11x01234567, _mm256_set1_ps(quantization_params[11].inv_scale));
446 vout12x01234567 = _mm256_mul_ps(vout12x01234567, _mm256_set1_ps(quantization_params[12].inv_scale));
447 vout13x01234567 = _mm256_mul_ps(vout13x01234567, _mm256_set1_ps(quantization_params[13].inv_scale));
448
449 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
450 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
451 w = (const float*) w + 16;
452
453 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
454 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
455 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
456 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
457 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
458 vout5x01234567 = _mm256_fmadd_ps(vout5x01234567, vfilter_output_scale01234567, vbias01234567);
459 vout6x01234567 = _mm256_fmadd_ps(vout6x01234567, vfilter_output_scale01234567, vbias01234567);
460 vout7x01234567 = _mm256_fmadd_ps(vout7x01234567, vfilter_output_scale01234567, vbias01234567);
461 vout8x01234567 = _mm256_fmadd_ps(vout8x01234567, vfilter_output_scale01234567, vbias01234567);
462 vout9x01234567 = _mm256_fmadd_ps(vout9x01234567, vfilter_output_scale01234567, vbias01234567);
463 vout10x01234567 = _mm256_fmadd_ps(vout10x01234567, vfilter_output_scale01234567, vbias01234567);
464 vout11x01234567 = _mm256_fmadd_ps(vout11x01234567, vfilter_output_scale01234567, vbias01234567);
465 vout12x01234567 = _mm256_fmadd_ps(vout12x01234567, vfilter_output_scale01234567, vbias01234567);
466 vout13x01234567 = _mm256_fmadd_ps(vout13x01234567, vfilter_output_scale01234567, vbias01234567);
467
468 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
469 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
470 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
471 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
472 vout4x01234567 = _mm256_max_ps(vout4x01234567, voutput_min);
473 vout5x01234567 = _mm256_max_ps(vout5x01234567, voutput_min);
474 vout6x01234567 = _mm256_max_ps(vout6x01234567, voutput_min);
475 vout7x01234567 = _mm256_max_ps(vout7x01234567, voutput_min);
476 vout8x01234567 = _mm256_max_ps(vout8x01234567, voutput_min);
477 vout9x01234567 = _mm256_max_ps(vout9x01234567, voutput_min);
478 vout10x01234567 = _mm256_max_ps(vout10x01234567, voutput_min);
479 vout11x01234567 = _mm256_max_ps(vout11x01234567, voutput_min);
480 vout12x01234567 = _mm256_max_ps(vout12x01234567, voutput_min);
481 vout13x01234567 = _mm256_max_ps(vout13x01234567, voutput_min);
482
483 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
484 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
485 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
486 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
487 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max);
488 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max);
489 vout6x01234567 = _mm256_min_ps(vout6x01234567, voutput_max);
490 vout7x01234567 = _mm256_min_ps(vout7x01234567, voutput_max);
491 vout8x01234567 = _mm256_min_ps(vout8x01234567, voutput_max);
492 vout9x01234567 = _mm256_min_ps(vout9x01234567, voutput_max);
493 vout10x01234567 = _mm256_min_ps(vout10x01234567, voutput_max);
494 vout11x01234567 = _mm256_min_ps(vout11x01234567, voutput_max);
495 vout12x01234567 = _mm256_min_ps(vout12x01234567, voutput_max);
496 vout13x01234567 = _mm256_min_ps(vout13x01234567, voutput_max);
497
498 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
499 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
500 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
501 __m128i vfp16out3x01234567 = _mm256_cvtps_ph(vout3x01234567, _MM_FROUND_TO_NEAREST_INT);
502 __m128i vfp16out4x01234567 = _mm256_cvtps_ph(vout4x01234567, _MM_FROUND_TO_NEAREST_INT);
503 __m128i vfp16out5x01234567 = _mm256_cvtps_ph(vout5x01234567, _MM_FROUND_TO_NEAREST_INT);
504 __m128i vfp16out6x01234567 = _mm256_cvtps_ph(vout6x01234567, _MM_FROUND_TO_NEAREST_INT);
505 __m128i vfp16out7x01234567 = _mm256_cvtps_ph(vout7x01234567, _MM_FROUND_TO_NEAREST_INT);
506 __m128i vfp16out8x01234567 = _mm256_cvtps_ph(vout8x01234567, _MM_FROUND_TO_NEAREST_INT);
507 __m128i vfp16out9x01234567 = _mm256_cvtps_ph(vout9x01234567, _MM_FROUND_TO_NEAREST_INT);
508 __m128i vfp16out10x01234567 = _mm256_cvtps_ph(vout10x01234567, _MM_FROUND_TO_NEAREST_INT);
509 __m128i vfp16out11x01234567 = _mm256_cvtps_ph(vout11x01234567, _MM_FROUND_TO_NEAREST_INT);
510 __m128i vfp16out12x01234567 = _mm256_cvtps_ph(vout12x01234567, _MM_FROUND_TO_NEAREST_INT);
511 __m128i vfp16out13x01234567 = _mm256_cvtps_ph(vout13x01234567, _MM_FROUND_TO_NEAREST_INT);
512 if XNN_LIKELY(nc >= 8) {
513 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
514 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
515 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
516 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
517 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
518 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
519 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
520 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
521 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
522 _mm_storeu_si128((__m128i*) c3, vfp16out3x01234567);
523 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
524 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
525 _mm_storeu_si128((__m128i*) c4, vfp16out4x01234567);
526 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
527 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
528 _mm_storeu_si128((__m128i*) c5, vfp16out5x01234567);
529 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
530 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
531 _mm_storeu_si128((__m128i*) c6, vfp16out6x01234567);
532 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
533 c6 = (uint16_t*) ((uintptr_t) c6 + cn_stride);
534 _mm_storeu_si128((__m128i*) c7, vfp16out7x01234567);
535 a7 = (const int8_t*) ((uintptr_t) a7 - kc);
536 c7 = (uint16_t*) ((uintptr_t) c7 + cn_stride);
537 _mm_storeu_si128((__m128i*) c8, vfp16out8x01234567);
538 a8 = (const int8_t*) ((uintptr_t) a8 - kc);
539 c8 = (uint16_t*) ((uintptr_t) c8 + cn_stride);
540 _mm_storeu_si128((__m128i*) c9, vfp16out9x01234567);
541 a9 = (const int8_t*) ((uintptr_t) a9 - kc);
542 c9 = (uint16_t*) ((uintptr_t) c9 + cn_stride);
543 _mm_storeu_si128((__m128i*) c10, vfp16out10x01234567);
544 a10 = (const int8_t*) ((uintptr_t) a10 - kc);
545 c10 = (uint16_t*) ((uintptr_t) c10 + cn_stride);
546 _mm_storeu_si128((__m128i*) c11, vfp16out11x01234567);
547 a11 = (const int8_t*) ((uintptr_t) a11 - kc);
548 c11 = (uint16_t*) ((uintptr_t) c11 + cn_stride);
549 _mm_storeu_si128((__m128i*) c12, vfp16out12x01234567);
550 a12 = (const int8_t*) ((uintptr_t) a12 - kc);
551 c12 = (uint16_t*) ((uintptr_t) c12 + cn_stride);
552 _mm_storeu_si128((__m128i*) c13, vfp16out13x01234567);
553 a13 = (const int8_t*) ((uintptr_t) a13 - kc);
554 c13 = (uint16_t*) ((uintptr_t) c13 + cn_stride);
555 nc -= 8;
556 } else {
557 // Prepare mask for valid 16-bit elements (depends on nc).
558 const __mmask8 vmask = _cvtu32_mask8((UINT32_C(1) << nc) - 1);
559 _mm_mask_storeu_epi16(c0, vmask, vfp16out0x01234567);
560 _mm_mask_storeu_epi16(c1, vmask, vfp16out1x01234567);
561 _mm_mask_storeu_epi16(c2, vmask, vfp16out2x01234567);
562 _mm_mask_storeu_epi16(c3, vmask, vfp16out3x01234567);
563 _mm_mask_storeu_epi16(c4, vmask, vfp16out4x01234567);
564 _mm_mask_storeu_epi16(c5, vmask, vfp16out5x01234567);
565 _mm_mask_storeu_epi16(c6, vmask, vfp16out6x01234567);
566 _mm_mask_storeu_epi16(c7, vmask, vfp16out7x01234567);
567 _mm_mask_storeu_epi16(c8, vmask, vfp16out8x01234567);
568 _mm_mask_storeu_epi16(c9, vmask, vfp16out9x01234567);
569 _mm_mask_storeu_epi16(c10, vmask, vfp16out10x01234567);
570 _mm_mask_storeu_epi16(c11, vmask, vfp16out11x01234567);
571 _mm_mask_storeu_epi16(c12, vmask, vfp16out12x01234567);
572 _mm_mask_storeu_epi16(c13, vmask, vfp16out13x01234567);
573 nc = 0;
574 }
575 } while (nc != 0);
576 }