Debian Codegraph @

src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-7x8c8-minmax-avxvnni.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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/common.h"
15 #include "xnnpack/gemm.h"
16 #include "xnnpack/intrinsics-polyfill.h"
17 #include "xnnpack/math.h"
18 #include "xnnpack/unaligned.h"
19
20
21 void xnn_qd8_f16_qc8w_igemm_minmax_ukernel_7x8c8__avxvnni(
22 size_t mr,
23 size_t nc,
24 size_t kc,
25 size_t ks,
26 const int8_t** restrict a,
27 const void* restrict w,
28 xnn_float16* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 size_t a_offset,
32 const int8_t* zero,
33 const int8_t* zero_data,
34 const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
35 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
36 {
37 assert(mr != 0);
38 assert(mr <= 7);
39 assert(nc != 0);
40 assert(kc != 0);
41 assert(kc % sizeof(int8_t) == 0);
42 assert(a != NULL);
43 assert(w != NULL);
44 assert(c != NULL);
45
46 kc = round_up_po2(kc, 8 * sizeof(int8_t));
47 uint16_t* c0 = (uint16_t*) c;
48 uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
49 if XNN_UNPREDICTABLE(mr < 2) {
50 c1 = c0;
51 }
52 uint16_t* c2 = (uint16_t*) ((uintptr_t) c1 + cm_stride);
53 if XNN_UNPREDICTABLE(mr <= 2) {
54 c2 = c1;
55 }
56 uint16_t* c3 = (uint16_t*) ((uintptr_t) c2 + cm_stride);
57 if XNN_UNPREDICTABLE(mr < 4) {
58 c3 = c2;
59 }
60 uint16_t* c4 = (uint16_t*) ((uintptr_t) c3 + cm_stride);
61 if XNN_UNPREDICTABLE(mr <= 4) {
62 c4 = c3;
63 }
64 uint16_t* c5 = (uint16_t*) ((uintptr_t) c4 + cm_stride);
65 if XNN_UNPREDICTABLE(mr < 6) {
66 c5 = c4;
67 }
68 uint16_t* c6 = (uint16_t*) ((uintptr_t) c5 + cm_stride);
69 if XNN_UNPREDICTABLE(mr <= 6) {
70 c6 = c5;
71 }
72
73 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
74 XNN_FORCE_REALIZATION(vsign_mask);
75 const __m256i vinput_zero_point = _mm256_set1_epi32((int) quantization_params->zero_point + 128);
76 const __m256 vinput_inv_scale = _mm256_set1_ps(quantization_params->inv_scale);
77 const __m256 voutput_min = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
78 const __m256 voutput_max = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
79 // XNN_FORCE_REALIZATION(vinput_zero_point);
80 // XNN_FORCE_REALIZATION(vinput_inv_scale);
81 // XNN_FORCE_REALIZATION(voutput_min);
82 // XNN_FORCE_REALIZATION(voutput_max);
83 do {
84 const __m256i vksum01234567 = _mm256_load_si256(w);
85 const __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point);
86 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
87 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
88 __m256i vacc1x0123 = vacc0x0123;
89 __m256i vacc1x4567 = vacc0x4567;
90 __m256i vacc2x0123 = vacc0x0123;
91 __m256i vacc2x4567 = vacc0x4567;
92 __m256i vacc3x0123 = vacc0x0123;
93 __m256i vacc3x4567 = vacc0x4567;
94 __m256i vacc4x0123 = vacc0x0123;
95 __m256i vacc4x4567 = vacc0x4567;
96 __m256i vacc5x0123 = vacc0x0123;
97 __m256i vacc5x4567 = vacc0x4567;
98 __m256i vacc6x0123 = vacc0x0123;
99 __m256i vacc6x4567 = vacc0x4567;
100 w = (const int32_t*) w + 8;
101
102 size_t p = ks;
103 do {
104 const int8_t* restrict a0 = a[0];
105 if XNN_UNPREDICTABLE(a0 != zero) {
106 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
107 } else {
108 a0 = zero_data;
109 }
110 const int8_t* restrict a1 = a[1];
111 if XNN_UNPREDICTABLE(a1 != zero) {
112 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
113 } else {
114 a1 = zero_data;
115 }
116 const int8_t* restrict a2 = a[2];
117 if XNN_UNPREDICTABLE(a2 != zero) {
118 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
119 } else {
120 a2 = zero_data;
121 }
122 const int8_t* restrict a3 = a[3];
123 if XNN_UNPREDICTABLE(a3 != zero) {
124 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
125 } else {
126 a3 = zero_data;
127 }
128 const int8_t* restrict a4 = a[4];
129 if XNN_UNPREDICTABLE(a4 != zero) {
130 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
131 } else {
132 a4 = zero_data;
133 }
134 const int8_t* restrict a5 = a[5];
135 if XNN_UNPREDICTABLE(a5 != zero) {
136 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
137 } else {
138 a5 = zero_data;
139 }
140 const int8_t* restrict a6 = a[6];
141 if XNN_UNPREDICTABLE(a6 != zero) {
142 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
143 } else {
144 a6 = zero_data;
145 }
146 a += 7;
147
148 size_t k = kc;
149 while (k >= 16 * sizeof(int8_t)) {
150 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
151 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
152 a0 += 16;
153 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
154 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
155 a1 += 16;
156 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
157 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
158 a2 += 16;
159 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
160 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
161 a3 += 16;
162 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
163 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
164 a4 += 16;
165 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
166 const __m256i va5x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
167 a5 += 16;
168 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
169 const __m256i va6x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask);
170 a6 += 16;
171
172 const __m256i vb01234567x0123 = _mm256_load_si256(w);
173 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
174 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
175 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
176
177 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
178 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
179 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
180 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
181 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
182 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
183 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
184 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
185 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
186 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
187 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
188 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
189 vacc6x0123 = _mm256_dpbusd_avx_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
190 vacc6x4567 = _mm256_dpbusd_avx_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
191 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
192 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
193 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
194 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
195 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
196 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
197 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
198 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
199 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
200 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
201 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x89ABCDEF, vb01234567x4567);
202 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);
203 vacc6x0123 = _mm256_dpbusd_avx_epi32(vacc6x0123, va6x89ABCDEF, vb01234567x4567);
204 vacc6x4567 = _mm256_dpbusd_avx_epi32(vacc6x4567, va6x89ABCDEF, vb89ABCDEFx4567);
205
206 w = (const int8_t*) w + 128;
207 k -= 16 * sizeof(int8_t);
208 }
209
210 if (k != 0) {
211 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
212 a0 += 8;
213 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
214 a1 += 8;
215 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
216 a2 += 8;
217 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
218 a3 += 8;
219 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
220 a4 += 8;
221 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
222 a5 += 8;
223 const __m256i va6x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a6)), vsign_mask);
224 a6 += 8;
225
226 const __m256i vb01234567x0123 = _mm256_load_si256(w);
227 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
228
229 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
230 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
231 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
232 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
233 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
234 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
235 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
236 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
237 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
238 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
239 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
240 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
241 vacc6x0123 = _mm256_dpbusd_avx_epi32(vacc6x0123, va6x01234567, vb01234567x0123);
242 vacc6x4567 = _mm256_dpbusd_avx_epi32(vacc6x4567, va6x01234567, vb89ABCDEFx0123);
243
244 w = (const int8_t*) w + 64;
245 k -= 8 * sizeof(int8_t);
246 }
247
248 p -= 7 * sizeof(void*);
249 } while (p != 0);
250
251
252 // Add adjacent pairs
253 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
254 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
255 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
256 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
257 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
258 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
259 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
260 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
261 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
262 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
263 const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);
264 __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));
265 const __m256i vsum6x02134657 = _mm256_hadd_epi32(vacc6x0123, vacc6x4567);
266 __m256i vacc6x01234567 = _mm256_permute4x64_epi64(vsum6x02134657, _MM_SHUFFLE(3, 1, 2, 0));
267
268 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
269 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
270 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
271 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
272 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
273 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
274 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567);
275
276 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vinput_inv_scale);
277 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vinput_inv_scale);
278 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vinput_inv_scale);
279 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vinput_inv_scale);
280 vout4x01234567 = _mm256_mul_ps(vout4x01234567, vinput_inv_scale);
281 vout5x01234567 = _mm256_mul_ps(vout5x01234567, vinput_inv_scale);
282 vout6x01234567 = _mm256_mul_ps(vout6x01234567, vinput_inv_scale);
283
284 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
285 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
286 w = (const float*) w + 16;
287 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
288 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
289 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
290 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
291 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
292 vout5x01234567 = _mm256_fmadd_ps(vout5x01234567, vfilter_output_scale01234567, vbias01234567);
293 vout6x01234567 = _mm256_fmadd_ps(vout6x01234567, vfilter_output_scale01234567, vbias01234567);
294
295 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
296 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
297 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
298 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
299 vout4x01234567 = _mm256_max_ps(vout4x01234567, voutput_min);
300 vout5x01234567 = _mm256_max_ps(vout5x01234567, voutput_min);
301 vout6x01234567 = _mm256_max_ps(vout6x01234567, voutput_min);
302
303 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
304 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
305 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
306 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
307 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max);
308 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max);
309 vout6x01234567 = _mm256_min_ps(vout6x01234567, voutput_max);
310
311 __m128i vfp16out6x01234567 = _mm256_cvtps_ph(vout6x01234567, _MM_FROUND_TO_NEAREST_INT);
312 __m128i vfp16out5x01234567 = _mm256_cvtps_ph(vout5x01234567, _MM_FROUND_TO_NEAREST_INT);
313 __m128i vfp16out4x01234567 = _mm256_cvtps_ph(vout4x01234567, _MM_FROUND_TO_NEAREST_INT);
314 __m128i vfp16out3x01234567 = _mm256_cvtps_ph(vout3x01234567, _MM_FROUND_TO_NEAREST_INT);
315 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
316 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
317 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
318 if XNN_LIKELY(nc >= 8) {
319 _mm_storeu_si128((__m128i*) c6, vfp16out6x01234567);
320 c6 = (uint16_t*) ((uintptr_t) c6 + cn_stride);
321 _mm_storeu_si128((__m128i*) c5, vfp16out5x01234567);
322 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
323 _mm_storeu_si128((__m128i*) c4, vfp16out4x01234567);
324 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
325 _mm_storeu_si128((__m128i*) c3, vfp16out3x01234567);
326 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
327 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
328 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
329 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
330 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
331 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
332 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
333
334 a = (const int8_t**restrict) ((uintptr_t) a - ks);
335 nc -= 8;
336 } else {
337 if (nc & 4) {
338 _mm_storel_epi64((__m128i*) c6, vfp16out6x01234567);
339 c6 += 4;
340 vfp16out6x01234567 = _mm_unpackhi_epi64(vfp16out6x01234567, vfp16out6x01234567);
341 _mm_storel_epi64((__m128i*) c5, vfp16out5x01234567);
342 c5 += 4;
343 vfp16out5x01234567 = _mm_unpackhi_epi64(vfp16out5x01234567, vfp16out5x01234567);
344 _mm_storel_epi64((__m128i*) c4, vfp16out4x01234567);
345 c4 += 4;
346 vfp16out4x01234567 = _mm_unpackhi_epi64(vfp16out4x01234567, vfp16out4x01234567);
347 _mm_storel_epi64((__m128i*) c3, vfp16out3x01234567);
348 c3 += 4;
349 vfp16out3x01234567 = _mm_unpackhi_epi64(vfp16out3x01234567, vfp16out3x01234567);
350 _mm_storel_epi64((__m128i*) c2, vfp16out2x01234567);
351 c2 += 4;
352 vfp16out2x01234567 = _mm_unpackhi_epi64(vfp16out2x01234567, vfp16out2x01234567);
353 _mm_storel_epi64((__m128i*) c1, vfp16out1x01234567);
354 c1 += 4;
355 vfp16out1x01234567 = _mm_unpackhi_epi64(vfp16out1x01234567, vfp16out1x01234567);
356 _mm_storel_epi64((__m128i*) c0, vfp16out0x01234567);
357 c0 += 4;
358 vfp16out0x01234567 = _mm_unpackhi_epi64(vfp16out0x01234567, vfp16out0x01234567);
359 }
360 if (nc & 2) {
361 _mm_storeu_si32(c6, vfp16out6x01234567);
362 vfp16out6x01234567 = _mm_srli_epi64(vfp16out6x01234567, 32);
363 c6 += 2;
364 _mm_storeu_si32(c5, vfp16out5x01234567);
365 vfp16out5x01234567 = _mm_srli_epi64(vfp16out5x01234567, 32);
366 c5 += 2;
367 _mm_storeu_si32(c4, vfp16out4x01234567);
368 vfp16out4x01234567 = _mm_srli_epi64(vfp16out4x01234567, 32);
369 c4 += 2;
370 _mm_storeu_si32(c3, vfp16out3x01234567);
371 vfp16out3x01234567 = _mm_srli_epi64(vfp16out3x01234567, 32);
372 c3 += 2;
373 _mm_storeu_si32(c2, vfp16out2x01234567);
374 vfp16out2x01234567 = _mm_srli_epi64(vfp16out2x01234567, 32);
375 c2 += 2;
376 _mm_storeu_si32(c1, vfp16out1x01234567);
377 vfp16out1x01234567 = _mm_srli_epi64(vfp16out1x01234567, 32);
378 c1 += 2;
379 _mm_storeu_si32(c0, vfp16out0x01234567);
380 vfp16out0x01234567 = _mm_srli_epi64(vfp16out0x01234567, 32);
381 c0 += 2;
382 }
383 if (nc & 1) {
384 *c6 = (uint16_t) _mm_extract_epi16(vfp16out6x01234567, 0);
385 *c5 = (uint16_t) _mm_extract_epi16(vfp16out5x01234567, 0);
386 *c4 = (uint16_t) _mm_extract_epi16(vfp16out4x01234567, 0);
387 *c3 = (uint16_t) _mm_extract_epi16(vfp16out3x01234567, 0);
388 *c2 = (uint16_t) _mm_extract_epi16(vfp16out2x01234567, 0);
389 *c1 = (uint16_t) _mm_extract_epi16(vfp16out1x01234567, 0);
390 *c0 = (uint16_t) _mm_extract_epi16(vfp16out0x01234567, 0);
391 }
392 nc = 0;
393 }
394 } while (nc != 0);
395 }