Debian Codegraph @

src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-14x8c8-minmax-avx256skx-madd-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_f16_qc4w_gemm_minmax_ukernel_14x8c8__avx256skx_madd_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 xnn_float16* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 const struct xnn_f16_qc4w_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 uint16_t* c0 = (uint16_t*) c;
46 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
47 uint16_t* c1 = (uint16_t*) ((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 uint16_t* c2 = (uint16_t*) ((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 uint16_t* c3 = (uint16_t*) ((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 uint16_t* c4 = (uint16_t*) ((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 uint16_t* c5 = (uint16_t*) ((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 uint16_t* c6 = (uint16_t*) ((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 uint16_t* c7 = (uint16_t*) ((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 uint16_t* c8 = (uint16_t*) ((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 uint16_t* c9 = (uint16_t*) ((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 uint16_t* c10 = (uint16_t*) ((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 uint16_t* c11 = (uint16_t*) ((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 uint16_t* c12 = (uint16_t*) ((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 uint16_t* c13 = (uint16_t*) ((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_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
142 const __m256 voutput_max = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
143 // XNN_FORCE_REALIZATION(voutput_min);
144 // XNN_FORCE_REALIZATION(voutput_max);
145 const __m256i vmask = _mm256_set1_epi8(0x0F);
146 XNN_FORCE_REALIZATION(vmask);
147 do {
148 const __m256i vksum01234567 = _mm256_load_si256(w);
149 __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point0);
150 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
151 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
152 __m256i vsum1x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point1);
153 __m256i vacc1x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 0));
154 __m256i vacc1x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 1));
155 __m256i vsum2x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point2);
156 __m256i vacc2x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 0));
157 __m256i vacc2x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 1));
158 __m256i vsum3x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point3);
159 __m256i vacc3x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 0));
160 __m256i vacc3x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 1));
161 __m256i vsum4x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point4);
162 __m256i vacc4x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum4x01234567, 0));
163 __m256i vacc4x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum4x01234567, 1));
164 __m256i vsum5x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point5);
165 __m256i vacc5x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum5x01234567, 0));
166 __m256i vacc5x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum5x01234567, 1));
167 __m256i vsum6x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point6);
168 __m256i vacc6x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum6x01234567, 0));
169 __m256i vacc6x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum6x01234567, 1));
170 __m256i vsum7x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point7);
171 __m256i vacc7x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum7x01234567, 0));
172 __m256i vacc7x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum7x01234567, 1));
173 __m256i vsum8x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point8);
174 __m256i vacc8x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum8x01234567, 0));
175 __m256i vacc8x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum8x01234567, 1));
176 __m256i vsum9x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point9);
177 __m256i vacc9x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum9x01234567, 0));
178 __m256i vacc9x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum9x01234567, 1));
179 __m256i vsum10x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point10);
180 __m256i vacc10x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum10x01234567, 0));
181 __m256i vacc10x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum10x01234567, 1));
182 __m256i vsum11x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point11);
183 __m256i vacc11x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum11x01234567, 0));
184 __m256i vacc11x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum11x01234567, 1));
185 __m256i vsum12x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point12);
186 __m256i vacc12x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum12x01234567, 0));
187 __m256i vacc12x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum12x01234567, 1));
188 __m256i vsum13x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point13);
189 __m256i vacc13x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum13x01234567, 0));
190 __m256i vacc13x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum13x01234567, 1));
191 w = (const int32_t*) w + 8;
192
193 size_t k = kc;
194 while (k >= 16 * sizeof(int8_t)) {
195 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
196 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
197 a0 += 16;
198 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
199 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
200 a1 += 16;
201 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
202 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
203 a2 += 16;
204 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
205 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
206 a3 += 16;
207 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
208 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
209 a4 += 16;
210 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
211 const __m256i va5x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
212 a5 += 16;
213 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
214 const __m256i va6x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
215 a6 += 16;
216 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
217 const __m256i va7x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7 + 8)), vsign_mask);
218 a7 += 16;
219 const __m256i va8x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8)), vsign_mask);
220 const __m256i va8x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8 + 8)), vsign_mask);
221 a8 += 16;
222 const __m256i va9x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9)), vsign_mask);
223 const __m256i va9x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9 + 8)), vsign_mask);
224 a9 += 16;
225 const __m256i va10x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a10)), vsign_mask);
226 const __m256i va10x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a10 + 8)), vsign_mask);
227 a10 += 16;
228 const __m256i va11x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a11)), vsign_mask);
229 const __m256i va11x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a11 + 8)), vsign_mask);
230 a11 += 16;
231 const __m256i va12x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a12)), vsign_mask);
232 const __m256i va12x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a12 + 8)), vsign_mask);
233 a12 += 16;
234 const __m256i va13x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a13)), vsign_mask);
235 const __m256i va13x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a13 + 8)), vsign_mask);
236 a13 += 16;
237
238 const __m256i vbb01234567x01234567 = _mm256_load_si256(w);
239 const __m256i vbb89ABCDEFx01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
240 const __m256i vbs01234567x0123 = _mm256_srli_epi32(vbb01234567x01234567, 4);
241 const __m256i vbs89ABCDEFx0123 = _mm256_srli_epi32(vbb89ABCDEFx01234567, 4);
242 const __m256i vb01234567x0123 = _mm256_and_si256(vbb01234567x01234567, vmask);
243 const __m256i vb89ABCDEFx0123 = _mm256_and_si256(vbb89ABCDEFx01234567, vmask);
244 const __m256i vb01234567x4567 = _mm256_and_si256(vbs01234567x0123, vmask);
245 const __m256i vb89ABCDEFx4567 = _mm256_and_si256(vbs89ABCDEFx0123, vmask);
246
247 vacc0x0123 = _mm256_dpbusd_epi32_madd(vacc0x0123, va0x01234567, vb01234567x0123);
248 vacc0x4567 = _mm256_dpbusd_epi32_madd(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
249 vacc1x0123 = _mm256_dpbusd_epi32_madd(vacc1x0123, va1x01234567, vb01234567x0123);
250 vacc1x4567 = _mm256_dpbusd_epi32_madd(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
251 vacc2x0123 = _mm256_dpbusd_epi32_madd(vacc2x0123, va2x01234567, vb01234567x0123);
252 vacc2x4567 = _mm256_dpbusd_epi32_madd(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
253 vacc3x0123 = _mm256_dpbusd_epi32_madd(vacc3x0123, va3x01234567, vb01234567x0123);
254 vacc3x4567 = _mm256_dpbusd_epi32_madd(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
255 vacc4x0123 = _mm256_dpbusd_epi32_madd(vacc4x0123, va4x01234567, vb01234567x0123);
256 vacc4x4567 = _mm256_dpbusd_epi32_madd(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
257 vacc5x0123 = _mm256_dpbusd_epi32_madd(vacc5x0123, va5x01234567, vb01234567x0123);
258 vacc5x4567 = _mm256_dpbusd_epi32_madd(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
259 vacc6x0123 = _mm256_dpbusd_epi32_madd(vacc6x0123, va6x01234567, vb01234567x0123);
260 vacc6x4567 = _mm256_dpbusd_epi32_madd(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
261 vacc7x0123 = _mm256_dpbusd_epi32_madd(vacc7x0123, va7x01234567, vb01234567x0123);
262 vacc7x4567 = _mm256_dpbusd_epi32_madd(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
263 vacc8x0123 = _mm256_dpbusd_epi32_madd(vacc8x0123, va8x01234567, vb01234567x0123);
264 vacc8x4567 = _mm256_dpbusd_epi32_madd(vacc8x4567, va8x01234567, vb89ABCDEFx0123);
265 vacc9x0123 = _mm256_dpbusd_epi32_madd(vacc9x0123, va9x01234567, vb01234567x0123);
266 vacc9x4567 = _mm256_dpbusd_epi32_madd(vacc9x4567, va9x01234567, vb89ABCDEFx0123);
267 vacc10x0123 = _mm256_dpbusd_epi32_madd(vacc10x0123, va10x01234567, vb01234567x0123);
268 vacc10x4567 = _mm256_dpbusd_epi32_madd(vacc10x4567, va10x01234567, vb89ABCDEFx0123);
269 vacc11x0123 = _mm256_dpbusd_epi32_madd(vacc11x0123, va11x01234567, vb01234567x0123);
270 vacc11x4567 = _mm256_dpbusd_epi32_madd(vacc11x4567, va11x01234567, vb89ABCDEFx0123);
271 vacc12x0123 = _mm256_dpbusd_epi32_madd(vacc12x0123, va12x01234567, vb01234567x0123);
272 vacc12x4567 = _mm256_dpbusd_epi32_madd(vacc12x4567, va12x01234567, vb89ABCDEFx0123);
273 vacc13x0123 = _mm256_dpbusd_epi32_madd(vacc13x0123, va13x01234567, vb01234567x0123);
274 vacc13x4567 = _mm256_dpbusd_epi32_madd(vacc13x4567, va13x01234567, vb89ABCDEFx0123);
275 xnn_prefetch_to_l1((const int8_t*) w + 960);
276 vacc0x0123 = _mm256_dpbusd_epi32_madd(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
277 vacc0x4567 = _mm256_dpbusd_epi32_madd(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
278 vacc1x0123 = _mm256_dpbusd_epi32_madd(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
279 vacc1x4567 = _mm256_dpbusd_epi32_madd(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
280 vacc2x0123 = _mm256_dpbusd_epi32_madd(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
281 vacc2x4567 = _mm256_dpbusd_epi32_madd(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
282 vacc3x0123 = _mm256_dpbusd_epi32_madd(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
283 vacc3x4567 = _mm256_dpbusd_epi32_madd(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
284 vacc4x0123 = _mm256_dpbusd_epi32_madd(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
285 vacc4x4567 = _mm256_dpbusd_epi32_madd(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
286 vacc5x0123 = _mm256_dpbusd_epi32_madd(vacc5x0123, va5x89ABCDEF, vb01234567x4567);
287 vacc5x4567 = _mm256_dpbusd_epi32_madd(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);
288 vacc6x0123 = _mm256_dpbusd_epi32_madd(vacc6x0123, va6x89ABCDEF, vb01234567x4567);
289 vacc6x4567 = _mm256_dpbusd_epi32_madd(vacc6x4567, va6x89ABCDEF, vb89ABCDEFx4567);
290 vacc7x0123 = _mm256_dpbusd_epi32_madd(vacc7x0123, va7x89ABCDEF, vb01234567x4567);
291 vacc7x4567 = _mm256_dpbusd_epi32_madd(vacc7x4567, va7x89ABCDEF, vb89ABCDEFx4567);
292 vacc8x0123 = _mm256_dpbusd_epi32_madd(vacc8x0123, va8x89ABCDEF, vb01234567x4567);
293 vacc8x4567 = _mm256_dpbusd_epi32_madd(vacc8x4567, va8x89ABCDEF, vb89ABCDEFx4567);
294 vacc9x0123 = _mm256_dpbusd_epi32_madd(vacc9x0123, va9x89ABCDEF, vb01234567x4567);
295 vacc9x4567 = _mm256_dpbusd_epi32_madd(vacc9x4567, va9x89ABCDEF, vb89ABCDEFx4567);
296 vacc10x0123 = _mm256_dpbusd_epi32_madd(vacc10x0123, va10x89ABCDEF, vb01234567x4567);
297 vacc10x4567 = _mm256_dpbusd_epi32_madd(vacc10x4567, va10x89ABCDEF, vb89ABCDEFx4567);
298 vacc11x0123 = _mm256_dpbusd_epi32_madd(vacc11x0123, va11x89ABCDEF, vb01234567x4567);
299 vacc11x4567 = _mm256_dpbusd_epi32_madd(vacc11x4567, va11x89ABCDEF, vb89ABCDEFx4567);
300 vacc12x0123 = _mm256_dpbusd_epi32_madd(vacc12x0123, va12x89ABCDEF, vb01234567x4567);
301 vacc12x4567 = _mm256_dpbusd_epi32_madd(vacc12x4567, va12x89ABCDEF, vb89ABCDEFx4567);
302 vacc13x0123 = _mm256_dpbusd_epi32_madd(vacc13x0123, va13x89ABCDEF, vb01234567x4567);
303 vacc13x4567 = _mm256_dpbusd_epi32_madd(vacc13x4567, va13x89ABCDEF, vb89ABCDEFx4567);
304
305 w = (const int8_t*) w + 64;
306 k -= 16 * sizeof(int8_t);
307 }
308
309 if (k != 0) {
310 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
311 a0 += 8;
312 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
313 a1 += 8;
314 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
315 a2 += 8;
316 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
317 a3 += 8;
318 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
319 a4 += 8;
320 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
321 a5 += 8;
322 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
323 a6 += 8;
324 const __m256i va7x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a7)), vsign_mask);
325 a7 += 8;
326 const __m256i va8x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a8)), vsign_mask);
327 a8 += 8;
328 const __m256i va9x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a9)), vsign_mask);
329 a9 += 8;
330 const __m256i va10x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a10)), vsign_mask);
331 a10 += 8;
332 const __m256i va11x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a11)), vsign_mask);
333 a11 += 8;
334 const __m256i va12x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a12)), vsign_mask);
335 a12 += 8;
336 const __m256i va13x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a13)), vsign_mask);
337 a13 += 8;
338
339 const __m256i vbb01234567x01234567 = _mm256_load_si256(w);
340 const __m256i vbb89ABCDEFx01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
341 const __m256i vb01234567x0123 = _mm256_and_si256(vbb01234567x01234567, vmask);
342 const __m256i vb89ABCDEFx0123 = _mm256_and_si256(vbb89ABCDEFx01234567, vmask);
343
344 vacc0x0123 = _mm256_dpbusd_epi32_madd(vacc0x0123, va0x01234567, vb01234567x0123);
345 vacc0x4567 = _mm256_dpbusd_epi32_madd(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
346 vacc1x0123 = _mm256_dpbusd_epi32_madd(vacc1x0123, va1x01234567, vb01234567x0123);
347 vacc1x4567 = _mm256_dpbusd_epi32_madd(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
348 vacc2x0123 = _mm256_dpbusd_epi32_madd(vacc2x0123, va2x01234567, vb01234567x0123);
349 vacc2x4567 = _mm256_dpbusd_epi32_madd(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
350 vacc3x0123 = _mm256_dpbusd_epi32_madd(vacc3x0123, va3x01234567, vb01234567x0123);
351 vacc3x4567 = _mm256_dpbusd_epi32_madd(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
352 vacc4x0123 = _mm256_dpbusd_epi32_madd(vacc4x0123, va4x01234567, vb01234567x0123);
353 vacc4x4567 = _mm256_dpbusd_epi32_madd(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
354 vacc5x0123 = _mm256_dpbusd_epi32_madd(vacc5x0123, va5x01234567, vb01234567x0123);
355 vacc5x4567 = _mm256_dpbusd_epi32_madd(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
356 vacc6x0123 = _mm256_dpbusd_epi32_madd(vacc6x0123, va6x01234567, vb01234567x0123);
357 vacc6x4567 = _mm256_dpbusd_epi32_madd(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
358 vacc7x0123 = _mm256_dpbusd_epi32_madd(vacc7x0123, va7x01234567, vb01234567x0123);
359 vacc7x4567 = _mm256_dpbusd_epi32_madd(vacc7x4567, va7x01234567, vb89ABCDEFx0123);
360 vacc8x0123 = _mm256_dpbusd_epi32_madd(vacc8x0123, va8x01234567, vb01234567x0123);
361 vacc8x4567 = _mm256_dpbusd_epi32_madd(vacc8x4567, va8x01234567, vb89ABCDEFx0123);
362 vacc9x0123 = _mm256_dpbusd_epi32_madd(vacc9x0123, va9x01234567, vb01234567x0123);
363 vacc9x4567 = _mm256_dpbusd_epi32_madd(vacc9x4567, va9x01234567, vb89ABCDEFx0123);
364 vacc10x0123 = _mm256_dpbusd_epi32_madd(vacc10x0123, va10x01234567, vb01234567x0123);
365 vacc10x4567 = _mm256_dpbusd_epi32_madd(vacc10x4567, va10x01234567, vb89ABCDEFx0123);
366 vacc11x0123 = _mm256_dpbusd_epi32_madd(vacc11x0123, va11x01234567, vb01234567x0123);
367 vacc11x4567 = _mm256_dpbusd_epi32_madd(vacc11x4567, va11x01234567, vb89ABCDEFx0123);
368 vacc12x0123 = _mm256_dpbusd_epi32_madd(vacc12x0123, va12x01234567, vb01234567x0123);
369 vacc12x4567 = _mm256_dpbusd_epi32_madd(vacc12x4567, va12x01234567, vb89ABCDEFx0123);
370 vacc13x0123 = _mm256_dpbusd_epi32_madd(vacc13x0123, va13x01234567, vb01234567x0123);
371 vacc13x4567 = _mm256_dpbusd_epi32_madd(vacc13x4567, va13x01234567, vb89ABCDEFx0123);
372 xnn_prefetch_to_l1((const int8_t*) w + 960);
373
374 w = (const int8_t*) w + 64;
375 k -= 8 * sizeof(int8_t);
376 }
377
378 // Add adjacent pairs
379 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
380 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
381 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
382 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
383 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
384 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
385 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
386 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
387 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
388 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
389 const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);
390 __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));
391 const __m256i vsum6x02134657 = _mm256_hadd_epi32(vacc6x0123, vacc6x4567);
392 __m256i vacc6x01234567 = _mm256_permute4x64_epi64(vsum6x02134657, _MM_SHUFFLE(3, 1, 2, 0));
393 const __m256i vsum7x02134657 = _mm256_hadd_epi32(vacc7x0123, vacc7x4567);
394 __m256i vacc7x01234567 = _mm256_permute4x64_epi64(vsum7x02134657, _MM_SHUFFLE(3, 1, 2, 0));
395 const __m256i vsum8x02134657 = _mm256_hadd_epi32(vacc8x0123, vacc8x4567);
396 __m256i vacc8x01234567 = _mm256_permute4x64_epi64(vsum8x02134657, _MM_SHUFFLE(3, 1, 2, 0));
397 const __m256i vsum9x02134657 = _mm256_hadd_epi32(vacc9x0123, vacc9x4567);
398 __m256i vacc9x01234567 = _mm256_permute4x64_epi64(vsum9x02134657, _MM_SHUFFLE(3, 1, 2, 0));
399 const __m256i vsum10x02134657 = _mm256_hadd_epi32(vacc10x0123, vacc10x4567);
400 __m256i vacc10x01234567 = _mm256_permute4x64_epi64(vsum10x02134657, _MM_SHUFFLE(3, 1, 2, 0));
401 const __m256i vsum11x02134657 = _mm256_hadd_epi32(vacc11x0123, vacc11x4567);
402 __m256i vacc11x01234567 = _mm256_permute4x64_epi64(vsum11x02134657, _MM_SHUFFLE(3, 1, 2, 0));
403 const __m256i vsum12x02134657 = _mm256_hadd_epi32(vacc12x0123, vacc12x4567);
404 __m256i vacc12x01234567 = _mm256_permute4x64_epi64(vsum12x02134657, _MM_SHUFFLE(3, 1, 2, 0));
405 const __m256i vsum13x02134657 = _mm256_hadd_epi32(vacc13x0123, vacc13x4567);
406 __m256i vacc13x01234567 = _mm256_permute4x64_epi64(vsum13x02134657, _MM_SHUFFLE(3, 1, 2, 0));
407
408 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
409 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
410 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
411 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
412 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
413 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
414 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567);
415 __m256 vout7x01234567 = _mm256_cvtepi32_ps(vacc7x01234567);
416 __m256 vout8x01234567 = _mm256_cvtepi32_ps(vacc8x01234567);
417 __m256 vout9x01234567 = _mm256_cvtepi32_ps(vacc9x01234567);
418 __m256 vout10x01234567 = _mm256_cvtepi32_ps(vacc10x01234567);
419 __m256 vout11x01234567 = _mm256_cvtepi32_ps(vacc11x01234567);
420 __m256 vout12x01234567 = _mm256_cvtepi32_ps(vacc12x01234567);
421 __m256 vout13x01234567 = _mm256_cvtepi32_ps(vacc13x01234567);
422
423 vout0x01234567 = _mm256_mul_ps(vout0x01234567, _mm256_set1_ps(quantization_params[0].inv_scale));
424 vout1x01234567 = _mm256_mul_ps(vout1x01234567, _mm256_set1_ps(quantization_params[1].inv_scale));
425 vout2x01234567 = _mm256_mul_ps(vout2x01234567, _mm256_set1_ps(quantization_params[2].inv_scale));
426 vout3x01234567 = _mm256_mul_ps(vout3x01234567, _mm256_set1_ps(quantization_params[3].inv_scale));
427 vout4x01234567 = _mm256_mul_ps(vout4x01234567, _mm256_set1_ps(quantization_params[4].inv_scale));
428 vout5x01234567 = _mm256_mul_ps(vout5x01234567, _mm256_set1_ps(quantization_params[5].inv_scale));
429 vout6x01234567 = _mm256_mul_ps(vout6x01234567, _mm256_set1_ps(quantization_params[6].inv_scale));
430 vout7x01234567 = _mm256_mul_ps(vout7x01234567, _mm256_set1_ps(quantization_params[7].inv_scale));
431 vout8x01234567 = _mm256_mul_ps(vout8x01234567, _mm256_set1_ps(quantization_params[8].inv_scale));
432 vout9x01234567 = _mm256_mul_ps(vout9x01234567, _mm256_set1_ps(quantization_params[9].inv_scale));
433 vout10x01234567 = _mm256_mul_ps(vout10x01234567, _mm256_set1_ps(quantization_params[10].inv_scale));
434 vout11x01234567 = _mm256_mul_ps(vout11x01234567, _mm256_set1_ps(quantization_params[11].inv_scale));
435 vout12x01234567 = _mm256_mul_ps(vout12x01234567, _mm256_set1_ps(quantization_params[12].inv_scale));
436 vout13x01234567 = _mm256_mul_ps(vout13x01234567, _mm256_set1_ps(quantization_params[13].inv_scale));
437
438 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
439 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
440 w = (const float*) w + 16;
441
442 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
443 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
444 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
445 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
446 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
447 vout5x01234567 = _mm256_fmadd_ps(vout5x01234567, vfilter_output_scale01234567, vbias01234567);
448 vout6x01234567 = _mm256_fmadd_ps(vout6x01234567, vfilter_output_scale01234567, vbias01234567);
449 vout7x01234567 = _mm256_fmadd_ps(vout7x01234567, vfilter_output_scale01234567, vbias01234567);
450 vout8x01234567 = _mm256_fmadd_ps(vout8x01234567, vfilter_output_scale01234567, vbias01234567);
451 vout9x01234567 = _mm256_fmadd_ps(vout9x01234567, vfilter_output_scale01234567, vbias01234567);
452 vout10x01234567 = _mm256_fmadd_ps(vout10x01234567, vfilter_output_scale01234567, vbias01234567);
453 vout11x01234567 = _mm256_fmadd_ps(vout11x01234567, vfilter_output_scale01234567, vbias01234567);
454 vout12x01234567 = _mm256_fmadd_ps(vout12x01234567, vfilter_output_scale01234567, vbias01234567);
455 vout13x01234567 = _mm256_fmadd_ps(vout13x01234567, vfilter_output_scale01234567, vbias01234567);
456
457 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
458 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
459 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
460 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
461 vout4x01234567 = _mm256_max_ps(vout4x01234567, voutput_min);
462 vout5x01234567 = _mm256_max_ps(vout5x01234567, voutput_min);
463 vout6x01234567 = _mm256_max_ps(vout6x01234567, voutput_min);
464 vout7x01234567 = _mm256_max_ps(vout7x01234567, voutput_min);
465 vout8x01234567 = _mm256_max_ps(vout8x01234567, voutput_min);
466 vout9x01234567 = _mm256_max_ps(vout9x01234567, voutput_min);
467 vout10x01234567 = _mm256_max_ps(vout10x01234567, voutput_min);
468 vout11x01234567 = _mm256_max_ps(vout11x01234567, voutput_min);
469 vout12x01234567 = _mm256_max_ps(vout12x01234567, voutput_min);
470 vout13x01234567 = _mm256_max_ps(vout13x01234567, voutput_min);
471
472 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
473 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
474 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
475 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
476 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max);
477 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max);
478 vout6x01234567 = _mm256_min_ps(vout6x01234567, voutput_max);
479 vout7x01234567 = _mm256_min_ps(vout7x01234567, voutput_max);
480 vout8x01234567 = _mm256_min_ps(vout8x01234567, voutput_max);
481 vout9x01234567 = _mm256_min_ps(vout9x01234567, voutput_max);
482 vout10x01234567 = _mm256_min_ps(vout10x01234567, voutput_max);
483 vout11x01234567 = _mm256_min_ps(vout11x01234567, voutput_max);
484 vout12x01234567 = _mm256_min_ps(vout12x01234567, voutput_max);
485 vout13x01234567 = _mm256_min_ps(vout13x01234567, voutput_max);
486
487 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
488 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
489 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
490 __m128i vfp16out3x01234567 = _mm256_cvtps_ph(vout3x01234567, _MM_FROUND_TO_NEAREST_INT);
491 __m128i vfp16out4x01234567 = _mm256_cvtps_ph(vout4x01234567, _MM_FROUND_TO_NEAREST_INT);
492 __m128i vfp16out5x01234567 = _mm256_cvtps_ph(vout5x01234567, _MM_FROUND_TO_NEAREST_INT);
493 __m128i vfp16out6x01234567 = _mm256_cvtps_ph(vout6x01234567, _MM_FROUND_TO_NEAREST_INT);
494 __m128i vfp16out7x01234567 = _mm256_cvtps_ph(vout7x01234567, _MM_FROUND_TO_NEAREST_INT);
495 __m128i vfp16out8x01234567 = _mm256_cvtps_ph(vout8x01234567, _MM_FROUND_TO_NEAREST_INT);
496 __m128i vfp16out9x01234567 = _mm256_cvtps_ph(vout9x01234567, _MM_FROUND_TO_NEAREST_INT);
497 __m128i vfp16out10x01234567 = _mm256_cvtps_ph(vout10x01234567, _MM_FROUND_TO_NEAREST_INT);
498 __m128i vfp16out11x01234567 = _mm256_cvtps_ph(vout11x01234567, _MM_FROUND_TO_NEAREST_INT);
499 __m128i vfp16out12x01234567 = _mm256_cvtps_ph(vout12x01234567, _MM_FROUND_TO_NEAREST_INT);
500 __m128i vfp16out13x01234567 = _mm256_cvtps_ph(vout13x01234567, _MM_FROUND_TO_NEAREST_INT);
501 if XNN_LIKELY(nc >= 8) {
502 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
503 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
504 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
505 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
506 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
507 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
508 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
509 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
510 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
511 _mm_storeu_si128((__m128i*) c3, vfp16out3x01234567);
512 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
513 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
514 _mm_storeu_si128((__m128i*) c4, vfp16out4x01234567);
515 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
516 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
517 _mm_storeu_si128((__m128i*) c5, vfp16out5x01234567);
518 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
519 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
520 _mm_storeu_si128((__m128i*) c6, vfp16out6x01234567);
521 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
522 c6 = (uint16_t*) ((uintptr_t) c6 + cn_stride);
523 _mm_storeu_si128((__m128i*) c7, vfp16out7x01234567);
524 a7 = (const int8_t*) ((uintptr_t) a7 - kc);
525 c7 = (uint16_t*) ((uintptr_t) c7 + cn_stride);
526 _mm_storeu_si128((__m128i*) c8, vfp16out8x01234567);
527 a8 = (const int8_t*) ((uintptr_t) a8 - kc);
528 c8 = (uint16_t*) ((uintptr_t) c8 + cn_stride);
529 _mm_storeu_si128((__m128i*) c9, vfp16out9x01234567);
530 a9 = (const int8_t*) ((uintptr_t) a9 - kc);
531 c9 = (uint16_t*) ((uintptr_t) c9 + cn_stride);
532 _mm_storeu_si128((__m128i*) c10, vfp16out10x01234567);
533 a10 = (const int8_t*) ((uintptr_t) a10 - kc);
534 c10 = (uint16_t*) ((uintptr_t) c10 + cn_stride);
535 _mm_storeu_si128((__m128i*) c11, vfp16out11x01234567);
536 a11 = (const int8_t*) ((uintptr_t) a11 - kc);
537 c11 = (uint16_t*) ((uintptr_t) c11 + cn_stride);
538 _mm_storeu_si128((__m128i*) c12, vfp16out12x01234567);
539 a12 = (const int8_t*) ((uintptr_t) a12 - kc);
540 c12 = (uint16_t*) ((uintptr_t) c12 + cn_stride);
541 _mm_storeu_si128((__m128i*) c13, vfp16out13x01234567);
542 a13 = (const int8_t*) ((uintptr_t) a13 - kc);
543 c13 = (uint16_t*) ((uintptr_t) c13 + cn_stride);
544 nc -= 8;
545 } else {
546 // Prepare mask for valid 16-bit elements (depends on nc).
547 const __mmask8 vmask = _cvtu32_mask8((UINT32_C(1) << nc) - 1);
548 _mm_mask_storeu_epi16(c0, vmask, vfp16out0x01234567);
549 _mm_mask_storeu_epi16(c1, vmask, vfp16out1x01234567);
550 _mm_mask_storeu_epi16(c2, vmask, vfp16out2x01234567);
551 _mm_mask_storeu_epi16(c3, vmask, vfp16out3x01234567);
552 _mm_mask_storeu_epi16(c4, vmask, vfp16out4x01234567);
553 _mm_mask_storeu_epi16(c5, vmask, vfp16out5x01234567);
554 _mm_mask_storeu_epi16(c6, vmask, vfp16out6x01234567);
555 _mm_mask_storeu_epi16(c7, vmask, vfp16out7x01234567);
556 _mm_mask_storeu_epi16(c8, vmask, vfp16out8x01234567);
557 _mm_mask_storeu_epi16(c9, vmask, vfp16out9x01234567);
558 _mm_mask_storeu_epi16(c10, vmask, vfp16out10x01234567);
559 _mm_mask_storeu_epi16(c11, vmask, vfp16out11x01234567);
560 _mm_mask_storeu_epi16(c12, vmask, vfp16out12x01234567);
561 _mm_mask_storeu_epi16(c13, vmask, vfp16out13x01234567);
562 nc = 0;
563 }
564 } while (nc != 0);
565 }