Debian Codegraph @

src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-6x8c8-minmax-avxvnni-prfm.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 #include "xnnpack/prefetch.h"
20
21
22 void xnn_qd8_f16_qc8w_igemm_minmax_ukernel_6x8c8__avxvnni_prfm(
23 size_t mr,
24 size_t nc,
25 size_t kc,
26 size_t ks,
27 const int8_t** restrict a,
28 const void* restrict w,
29 xnn_float16* restrict c,
30 size_t cm_stride,
31 size_t cn_stride,
32 size_t a_offset,
33 const int8_t* zero,
34 const int8_t* zero_data,
35 const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
36 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
37 {
38 assert(mr != 0);
39 assert(mr <= 6);
40 assert(nc != 0);
41 assert(kc != 0);
42 assert(kc % sizeof(int8_t) == 0);
43 assert(a != NULL);
44 assert(w != NULL);
45 assert(c != NULL);
46
47 kc = round_up_po2(kc, 8 * sizeof(int8_t));
48 uint16_t* c0 = (uint16_t*) c;
49 uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
50 if XNN_UNPREDICTABLE(mr < 2) {
51 c1 = c0;
52 }
53 uint16_t* c2 = (uint16_t*) ((uintptr_t) c1 + cm_stride);
54 if XNN_UNPREDICTABLE(mr <= 2) {
55 c2 = c1;
56 }
57 uint16_t* c3 = (uint16_t*) ((uintptr_t) c2 + cm_stride);
58 if XNN_UNPREDICTABLE(mr < 4) {
59 c3 = c2;
60 }
61 uint16_t* c4 = (uint16_t*) ((uintptr_t) c3 + cm_stride);
62 if XNN_UNPREDICTABLE(mr <= 4) {
63 c4 = c3;
64 }
65 uint16_t* c5 = (uint16_t*) ((uintptr_t) c4 + cm_stride);
66 if XNN_UNPREDICTABLE(mr != 6) {
67 c5 = c4;
68 }
69
70 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
71 XNN_FORCE_REALIZATION(vsign_mask);
72 const __m256i vinput_zero_point = _mm256_set1_epi32((int) quantization_params->zero_point + 128);
73 const __m256 vinput_inv_scale = _mm256_set1_ps(quantization_params->inv_scale);
74 const __m256 voutput_min = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
75 const __m256 voutput_max = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
76 // XNN_FORCE_REALIZATION(vinput_zero_point);
77 // XNN_FORCE_REALIZATION(vinput_inv_scale);
78 // XNN_FORCE_REALIZATION(voutput_min);
79 // XNN_FORCE_REALIZATION(voutput_max);
80 do {
81 const __m256i vksum01234567 = _mm256_load_si256(w);
82 const __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point);
83 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
84 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
85 __m256i vacc1x0123 = vacc0x0123;
86 __m256i vacc1x4567 = vacc0x4567;
87 __m256i vacc2x0123 = vacc0x0123;
88 __m256i vacc2x4567 = vacc0x4567;
89 __m256i vacc3x0123 = vacc0x0123;
90 __m256i vacc3x4567 = vacc0x4567;
91 __m256i vacc4x0123 = vacc0x0123;
92 __m256i vacc4x4567 = vacc0x4567;
93 __m256i vacc5x0123 = vacc0x0123;
94 __m256i vacc5x4567 = vacc0x4567;
95 w = (const int32_t*) w + 8;
96
97 size_t p = ks;
98 do {
99 const int8_t* restrict a0 = a[0];
100 if XNN_UNPREDICTABLE(a0 != zero) {
101 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
102 } else {
103 a0 = zero_data;
104 }
105 const int8_t* restrict a1 = a[1];
106 if XNN_UNPREDICTABLE(a1 != zero) {
107 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
108 } else {
109 a1 = zero_data;
110 }
111 const int8_t* restrict a2 = a[2];
112 if XNN_UNPREDICTABLE(a2 != zero) {
113 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
114 } else {
115 a2 = zero_data;
116 }
117 const int8_t* restrict a3 = a[3];
118 if XNN_UNPREDICTABLE(a3 != zero) {
119 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
120 } else {
121 a3 = zero_data;
122 }
123 const int8_t* restrict a4 = a[4];
124 if XNN_UNPREDICTABLE(a4 != zero) {
125 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
126 } else {
127 a4 = zero_data;
128 }
129 const int8_t* restrict a5 = a[5];
130 if XNN_UNPREDICTABLE(a5 != zero) {
131 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
132 } else {
133 a5 = zero_data;
134 }
135 a += 6;
136
137 size_t k = kc;
138 while (k >= 16 * sizeof(int8_t)) {
139 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
140 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
141 a0 += 16;
142 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
143 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
144 a1 += 16;
145 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
146 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
147 a2 += 16;
148 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
149 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
150 a3 += 16;
151 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
152 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
153 a4 += 16;
154 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
155 const __m256i va5x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
156 a5 += 16;
157
158 const __m256i vb01234567x0123 = _mm256_load_si256(w);
159 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
160 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
161 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
162 xnn_prefetch_to_l1((const int8_t*) w + 896);
163
164 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
165 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
166 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
167 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
168 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
169 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
170 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
171 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
172 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
173 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
174 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
175 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
176 xnn_prefetch_to_l1((const int8_t*) w + 960);
177 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
178 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
179 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
180 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
181 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
182 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
183 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
184 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
185 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
186 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
187 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x89ABCDEF, vb01234567x4567);
188 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);
189
190 w = (const int8_t*) w + 128;
191 k -= 16 * sizeof(int8_t);
192 }
193
194 if (k != 0) {
195 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
196 a0 += 8;
197 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
198 a1 += 8;
199 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
200 a2 += 8;
201 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
202 a3 += 8;
203 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
204 a4 += 8;
205 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
206 a5 += 8;
207
208 const __m256i vb01234567x0123 = _mm256_load_si256(w);
209 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
210
211 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
212 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
213 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
214 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
215 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
216 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
217 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
218 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
219 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
220 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
221 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
222 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
223 xnn_prefetch_to_l1((const int8_t*) w + 960);
224
225 w = (const int8_t*) w + 64;
226 k -= 8 * sizeof(int8_t);
227 }
228
229 p -= 6 * sizeof(void*);
230 } while (p != 0);
231
232
233 // Add adjacent pairs
234 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
235 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
236 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
237 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
238 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
239 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
240 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
241 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
242 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
243 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
244 const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);
245 __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));
246
247 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
248 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
249 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
250 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
251 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
252 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
253
254 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vinput_inv_scale);
255 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vinput_inv_scale);
256 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vinput_inv_scale);
257 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vinput_inv_scale);
258 vout4x01234567 = _mm256_mul_ps(vout4x01234567, vinput_inv_scale);
259 vout5x01234567 = _mm256_mul_ps(vout5x01234567, vinput_inv_scale);
260
261 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
262 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
263 w = (const float*) w + 16;
264 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
265 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
266 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
267 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
268 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
269 vout5x01234567 = _mm256_fmadd_ps(vout5x01234567, vfilter_output_scale01234567, vbias01234567);
270
271 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
272 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
273 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
274 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
275 vout4x01234567 = _mm256_max_ps(vout4x01234567, voutput_min);
276 vout5x01234567 = _mm256_max_ps(vout5x01234567, voutput_min);
277
278 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
279 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
280 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
281 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
282 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max);
283 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max);
284
285 __m128i vfp16out5x01234567 = _mm256_cvtps_ph(vout5x01234567, _MM_FROUND_TO_NEAREST_INT);
286 __m128i vfp16out4x01234567 = _mm256_cvtps_ph(vout4x01234567, _MM_FROUND_TO_NEAREST_INT);
287 __m128i vfp16out3x01234567 = _mm256_cvtps_ph(vout3x01234567, _MM_FROUND_TO_NEAREST_INT);
288 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
289 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
290 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
291 if XNN_LIKELY(nc >= 8) {
292 _mm_storeu_si128((__m128i*) c5, vfp16out5x01234567);
293 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
294 _mm_storeu_si128((__m128i*) c4, vfp16out4x01234567);
295 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
296 _mm_storeu_si128((__m128i*) c3, vfp16out3x01234567);
297 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
298 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
299 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
300 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
301 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
302 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
303 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
304
305 a = (const int8_t**restrict) ((uintptr_t) a - ks);
306 nc -= 8;
307 } else {
308 if (nc & 4) {
309 _mm_storel_epi64((__m128i*) c5, vfp16out5x01234567);
310 c5 += 4;
311 vfp16out5x01234567 = _mm_unpackhi_epi64(vfp16out5x01234567, vfp16out5x01234567);
312 _mm_storel_epi64((__m128i*) c4, vfp16out4x01234567);
313 c4 += 4;
314 vfp16out4x01234567 = _mm_unpackhi_epi64(vfp16out4x01234567, vfp16out4x01234567);
315 _mm_storel_epi64((__m128i*) c3, vfp16out3x01234567);
316 c3 += 4;
317 vfp16out3x01234567 = _mm_unpackhi_epi64(vfp16out3x01234567, vfp16out3x01234567);
318 _mm_storel_epi64((__m128i*) c2, vfp16out2x01234567);
319 c2 += 4;
320 vfp16out2x01234567 = _mm_unpackhi_epi64(vfp16out2x01234567, vfp16out2x01234567);
321 _mm_storel_epi64((__m128i*) c1, vfp16out1x01234567);
322 c1 += 4;
323 vfp16out1x01234567 = _mm_unpackhi_epi64(vfp16out1x01234567, vfp16out1x01234567);
324 _mm_storel_epi64((__m128i*) c0, vfp16out0x01234567);
325 c0 += 4;
326 vfp16out0x01234567 = _mm_unpackhi_epi64(vfp16out0x01234567, vfp16out0x01234567);
327 }
328 if (nc & 2) {
329 _mm_storeu_si32(c5, vfp16out5x01234567);
330 vfp16out5x01234567 = _mm_srli_epi64(vfp16out5x01234567, 32);
331 c5 += 2;
332 _mm_storeu_si32(c4, vfp16out4x01234567);
333 vfp16out4x01234567 = _mm_srli_epi64(vfp16out4x01234567, 32);
334 c4 += 2;
335 _mm_storeu_si32(c3, vfp16out3x01234567);
336 vfp16out3x01234567 = _mm_srli_epi64(vfp16out3x01234567, 32);
337 c3 += 2;
338 _mm_storeu_si32(c2, vfp16out2x01234567);
339 vfp16out2x01234567 = _mm_srli_epi64(vfp16out2x01234567, 32);
340 c2 += 2;
341 _mm_storeu_si32(c1, vfp16out1x01234567);
342 vfp16out1x01234567 = _mm_srli_epi64(vfp16out1x01234567, 32);
343 c1 += 2;
344 _mm_storeu_si32(c0, vfp16out0x01234567);
345 vfp16out0x01234567 = _mm_srli_epi64(vfp16out0x01234567, 32);
346 c0 += 2;
347 }
348 if (nc & 1) {
349 *c5 = (uint16_t) _mm_extract_epi16(vfp16out5x01234567, 0);
350 *c4 = (uint16_t) _mm_extract_epi16(vfp16out4x01234567, 0);
351 *c3 = (uint16_t) _mm_extract_epi16(vfp16out3x01234567, 0);
352 *c2 = (uint16_t) _mm_extract_epi16(vfp16out2x01234567, 0);
353 *c1 = (uint16_t) _mm_extract_epi16(vfp16out1x01234567, 0);
354 *c0 = (uint16_t) _mm_extract_epi16(vfp16out0x01234567, 0);
355 }
356 nc = 0;
357 }
358 } while (nc != 0);
359 }