Debian Codegraph @

src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-6x8c8-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_6x8c8__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 <= 6);
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
69 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
70 XNN_FORCE_REALIZATION(vsign_mask);
71 const __m256i vinput_zero_point = _mm256_set1_epi32((int) quantization_params->zero_point + 128);
72 const __m256 vinput_inv_scale = _mm256_set1_ps(quantization_params->inv_scale);
73 const __m256 voutput_min = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
74 const __m256 voutput_max = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
75 // XNN_FORCE_REALIZATION(vinput_zero_point);
76 // XNN_FORCE_REALIZATION(vinput_inv_scale);
77 // XNN_FORCE_REALIZATION(voutput_min);
78 // XNN_FORCE_REALIZATION(voutput_max);
79 do {
80 const __m256i vksum01234567 = _mm256_load_si256(w);
81 const __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point);
82 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
83 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
84 __m256i vacc1x0123 = vacc0x0123;
85 __m256i vacc1x4567 = vacc0x4567;
86 __m256i vacc2x0123 = vacc0x0123;
87 __m256i vacc2x4567 = vacc0x4567;
88 __m256i vacc3x0123 = vacc0x0123;
89 __m256i vacc3x4567 = vacc0x4567;
90 __m256i vacc4x0123 = vacc0x0123;
91 __m256i vacc4x4567 = vacc0x4567;
92 __m256i vacc5x0123 = vacc0x0123;
93 __m256i vacc5x4567 = vacc0x4567;
94 w = (const int32_t*) w + 8;
95
96 size_t p = ks;
97 do {
98 const int8_t* restrict a0 = a[0];
99 if XNN_UNPREDICTABLE(a0 != zero) {
100 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
101 } else {
102 a0 = zero_data;
103 }
104 const int8_t* restrict a1 = a[1];
105 if XNN_UNPREDICTABLE(a1 != zero) {
106 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
107 } else {
108 a1 = zero_data;
109 }
110 const int8_t* restrict a2 = a[2];
111 if XNN_UNPREDICTABLE(a2 != zero) {
112 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
113 } else {
114 a2 = zero_data;
115 }
116 const int8_t* restrict a3 = a[3];
117 if XNN_UNPREDICTABLE(a3 != zero) {
118 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
119 } else {
120 a3 = zero_data;
121 }
122 const int8_t* restrict a4 = a[4];
123 if XNN_UNPREDICTABLE(a4 != zero) {
124 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
125 } else {
126 a4 = zero_data;
127 }
128 const int8_t* restrict a5 = a[5];
129 if XNN_UNPREDICTABLE(a5 != zero) {
130 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
131 } else {
132 a5 = zero_data;
133 }
134 a += 6;
135
136 size_t k = kc;
137 while (k >= 16 * sizeof(int8_t)) {
138 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
139 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
140 a0 += 16;
141 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
142 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
143 a1 += 16;
144 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
145 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
146 a2 += 16;
147 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
148 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
149 a3 += 16;
150 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
151 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
152 a4 += 16;
153 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
154 const __m256i va5x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
155 a5 += 16;
156
157 const __m256i vb01234567x0123 = _mm256_load_si256(w);
158 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
159 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
160 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
161
162 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
163 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
164 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
165 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
166 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
167 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
168 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
169 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
170 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
171 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
172 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
173 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
174 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
175 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
176 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
177 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
178 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
179 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
180 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
181 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
182 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
183 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
184 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x89ABCDEF, vb01234567x4567);
185 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);
186
187 w = (const int8_t*) w + 128;
188 k -= 16 * sizeof(int8_t);
189 }
190
191 if (k != 0) {
192 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
193 a0 += 8;
194 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
195 a1 += 8;
196 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
197 a2 += 8;
198 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
199 a3 += 8;
200 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
201 a4 += 8;
202 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
203 a5 += 8;
204
205 const __m256i vb01234567x0123 = _mm256_load_si256(w);
206 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
207
208 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
209 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
210 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
211 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
212 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
213 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
214 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
215 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
216 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
217 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
218 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
219 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
220
221 w = (const int8_t*) w + 64;
222 k -= 8 * sizeof(int8_t);
223 }
224
225 p -= 6 * sizeof(void*);
226 } while (p != 0);
227
228
229 // Add adjacent pairs
230 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
231 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
232 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
233 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
234 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
235 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
236 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
237 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
238 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
239 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
240 const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);
241 __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));
242
243 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
244 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
245 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
246 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
247 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
248 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
249
250 vout0x01234567 = _mm256_mul_ps(vout0x01234567, vinput_inv_scale);
251 vout1x01234567 = _mm256_mul_ps(vout1x01234567, vinput_inv_scale);
252 vout2x01234567 = _mm256_mul_ps(vout2x01234567, vinput_inv_scale);
253 vout3x01234567 = _mm256_mul_ps(vout3x01234567, vinput_inv_scale);
254 vout4x01234567 = _mm256_mul_ps(vout4x01234567, vinput_inv_scale);
255 vout5x01234567 = _mm256_mul_ps(vout5x01234567, vinput_inv_scale);
256
257 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
258 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
259 w = (const float*) w + 16;
260 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
261 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
262 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
263 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
264 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
265 vout5x01234567 = _mm256_fmadd_ps(vout5x01234567, vfilter_output_scale01234567, vbias01234567);
266
267 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
268 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
269 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
270 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
271 vout4x01234567 = _mm256_max_ps(vout4x01234567, voutput_min);
272 vout5x01234567 = _mm256_max_ps(vout5x01234567, voutput_min);
273
274 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
275 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
276 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
277 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
278 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max);
279 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max);
280
281 __m128i vfp16out5x01234567 = _mm256_cvtps_ph(vout5x01234567, _MM_FROUND_TO_NEAREST_INT);
282 __m128i vfp16out4x01234567 = _mm256_cvtps_ph(vout4x01234567, _MM_FROUND_TO_NEAREST_INT);
283 __m128i vfp16out3x01234567 = _mm256_cvtps_ph(vout3x01234567, _MM_FROUND_TO_NEAREST_INT);
284 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
285 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
286 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
287 if XNN_LIKELY(nc >= 8) {
288 _mm_storeu_si128((__m128i*) c5, vfp16out5x01234567);
289 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
290 _mm_storeu_si128((__m128i*) c4, vfp16out4x01234567);
291 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
292 _mm_storeu_si128((__m128i*) c3, vfp16out3x01234567);
293 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
294 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
295 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
296 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
297 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
298 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
299 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
300
301 a = (const int8_t**restrict) ((uintptr_t) a - ks);
302 nc -= 8;
303 } else {
304 if (nc & 4) {
305 _mm_storel_epi64((__m128i*) c5, vfp16out5x01234567);
306 c5 += 4;
307 vfp16out5x01234567 = _mm_unpackhi_epi64(vfp16out5x01234567, vfp16out5x01234567);
308 _mm_storel_epi64((__m128i*) c4, vfp16out4x01234567);
309 c4 += 4;
310 vfp16out4x01234567 = _mm_unpackhi_epi64(vfp16out4x01234567, vfp16out4x01234567);
311 _mm_storel_epi64((__m128i*) c3, vfp16out3x01234567);
312 c3 += 4;
313 vfp16out3x01234567 = _mm_unpackhi_epi64(vfp16out3x01234567, vfp16out3x01234567);
314 _mm_storel_epi64((__m128i*) c2, vfp16out2x01234567);
315 c2 += 4;
316 vfp16out2x01234567 = _mm_unpackhi_epi64(vfp16out2x01234567, vfp16out2x01234567);
317 _mm_storel_epi64((__m128i*) c1, vfp16out1x01234567);
318 c1 += 4;
319 vfp16out1x01234567 = _mm_unpackhi_epi64(vfp16out1x01234567, vfp16out1x01234567);
320 _mm_storel_epi64((__m128i*) c0, vfp16out0x01234567);
321 c0 += 4;
322 vfp16out0x01234567 = _mm_unpackhi_epi64(vfp16out0x01234567, vfp16out0x01234567);
323 }
324 if (nc & 2) {
325 _mm_storeu_si32(c5, vfp16out5x01234567);
326 vfp16out5x01234567 = _mm_srli_epi64(vfp16out5x01234567, 32);
327 c5 += 2;
328 _mm_storeu_si32(c4, vfp16out4x01234567);
329 vfp16out4x01234567 = _mm_srli_epi64(vfp16out4x01234567, 32);
330 c4 += 2;
331 _mm_storeu_si32(c3, vfp16out3x01234567);
332 vfp16out3x01234567 = _mm_srli_epi64(vfp16out3x01234567, 32);
333 c3 += 2;
334 _mm_storeu_si32(c2, vfp16out2x01234567);
335 vfp16out2x01234567 = _mm_srli_epi64(vfp16out2x01234567, 32);
336 c2 += 2;
337 _mm_storeu_si32(c1, vfp16out1x01234567);
338 vfp16out1x01234567 = _mm_srli_epi64(vfp16out1x01234567, 32);
339 c1 += 2;
340 _mm_storeu_si32(c0, vfp16out0x01234567);
341 vfp16out0x01234567 = _mm_srli_epi64(vfp16out0x01234567, 32);
342 c0 += 2;
343 }
344 if (nc & 1) {
345 *c5 = (uint16_t) _mm_extract_epi16(vfp16out5x01234567, 0);
346 *c4 = (uint16_t) _mm_extract_epi16(vfp16out4x01234567, 0);
347 *c3 = (uint16_t) _mm_extract_epi16(vfp16out3x01234567, 0);
348 *c2 = (uint16_t) _mm_extract_epi16(vfp16out2x01234567, 0);
349 *c1 = (uint16_t) _mm_extract_epi16(vfp16out1x01234567, 0);
350 *c0 = (uint16_t) _mm_extract_epi16(vfp16out0x01234567, 0);
351 }
352 nc = 0;
353 }
354 } while (nc != 0);
355 }