Debian Codegraph @

src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-6x8c8-minmax-avx2-madd.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_6x8c8__avx2_madd(
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 <= 6);
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
76 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
77 XNN_FORCE_REALIZATION(vsign_mask);
78 const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point + 128);
79 const __m256i vinput_zero_point1 = _mm256_set1_epi32((int) quantization_params[1].zero_point + 128);
80 const __m256i vinput_zero_point2 = _mm256_set1_epi32((int) quantization_params[2].zero_point + 128);
81 const __m256i vinput_zero_point3 = _mm256_set1_epi32((int) quantization_params[3].zero_point + 128);
82 const __m256i vinput_zero_point4 = _mm256_set1_epi32((int) quantization_params[4].zero_point + 128);
83 const __m256i vinput_zero_point5 = _mm256_set1_epi32((int) quantization_params[5].zero_point + 128);
84 const __m256 voutput_min = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.min));
85 const __m256 voutput_max = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.max));
86 // XNN_FORCE_REALIZATION(voutput_min);
87 // XNN_FORCE_REALIZATION(voutput_max);
88 const __m256i vmask = _mm256_set1_epi8(0x0F);
89 XNN_FORCE_REALIZATION(vmask);
90 do {
91 const __m256i vksum01234567 = _mm256_load_si256(w);
92 __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point0);
93 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
94 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
95 __m256i vsum1x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point1);
96 __m256i vacc1x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 0));
97 __m256i vacc1x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 1));
98 __m256i vsum2x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point2);
99 __m256i vacc2x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 0));
100 __m256i vacc2x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 1));
101 __m256i vsum3x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point3);
102 __m256i vacc3x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 0));
103 __m256i vacc3x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 1));
104 __m256i vsum4x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point4);
105 __m256i vacc4x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum4x01234567, 0));
106 __m256i vacc4x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum4x01234567, 1));
107 __m256i vsum5x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point5);
108 __m256i vacc5x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum5x01234567, 0));
109 __m256i vacc5x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum5x01234567, 1));
110 w = (const int32_t*) w + 8;
111
112 size_t k = kc;
113 while (k >= 16 * sizeof(int8_t)) {
114 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
115 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
116 a0 += 16;
117 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
118 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
119 a1 += 16;
120 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
121 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
122 a2 += 16;
123 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
124 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
125 a3 += 16;
126 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
127 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
128 a4 += 16;
129 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
130 const __m256i va5x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
131 a5 += 16;
132
133 const __m256i vbb01234567x01234567 = _mm256_load_si256(w);
134 const __m256i vbb89ABCDEFx01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
135 const __m256i vbs01234567x0123 = _mm256_srli_epi32(vbb01234567x01234567, 4);
136 const __m256i vbs89ABCDEFx0123 = _mm256_srli_epi32(vbb89ABCDEFx01234567, 4);
137 const __m256i vb01234567x0123 = _mm256_and_si256(vbb01234567x01234567, vmask);
138 const __m256i vb89ABCDEFx0123 = _mm256_and_si256(vbb89ABCDEFx01234567, vmask);
139 const __m256i vb01234567x4567 = _mm256_and_si256(vbs01234567x0123, vmask);
140 const __m256i vb89ABCDEFx4567 = _mm256_and_si256(vbs89ABCDEFx0123, vmask);
141
142 vacc0x0123 = _mm256_dpbusd_epi32_madd(vacc0x0123, va0x01234567, vb01234567x0123);
143 vacc0x4567 = _mm256_dpbusd_epi32_madd(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
144 vacc1x0123 = _mm256_dpbusd_epi32_madd(vacc1x0123, va1x01234567, vb01234567x0123);
145 vacc1x4567 = _mm256_dpbusd_epi32_madd(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
146 vacc2x0123 = _mm256_dpbusd_epi32_madd(vacc2x0123, va2x01234567, vb01234567x0123);
147 vacc2x4567 = _mm256_dpbusd_epi32_madd(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
148 vacc3x0123 = _mm256_dpbusd_epi32_madd(vacc3x0123, va3x01234567, vb01234567x0123);
149 vacc3x4567 = _mm256_dpbusd_epi32_madd(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
150 vacc4x0123 = _mm256_dpbusd_epi32_madd(vacc4x0123, va4x01234567, vb01234567x0123);
151 vacc4x4567 = _mm256_dpbusd_epi32_madd(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
152 vacc5x0123 = _mm256_dpbusd_epi32_madd(vacc5x0123, va5x01234567, vb01234567x0123);
153 vacc5x4567 = _mm256_dpbusd_epi32_madd(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
154 vacc0x0123 = _mm256_dpbusd_epi32_madd(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
155 vacc0x4567 = _mm256_dpbusd_epi32_madd(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
156 vacc1x0123 = _mm256_dpbusd_epi32_madd(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
157 vacc1x4567 = _mm256_dpbusd_epi32_madd(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
158 vacc2x0123 = _mm256_dpbusd_epi32_madd(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
159 vacc2x4567 = _mm256_dpbusd_epi32_madd(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
160 vacc3x0123 = _mm256_dpbusd_epi32_madd(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
161 vacc3x4567 = _mm256_dpbusd_epi32_madd(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
162 vacc4x0123 = _mm256_dpbusd_epi32_madd(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
163 vacc4x4567 = _mm256_dpbusd_epi32_madd(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
164 vacc5x0123 = _mm256_dpbusd_epi32_madd(vacc5x0123, va5x89ABCDEF, vb01234567x4567);
165 vacc5x4567 = _mm256_dpbusd_epi32_madd(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);
166
167 w = (const int8_t*) w + 64;
168 k -= 16 * sizeof(int8_t);
169 }
170
171 if (k != 0) {
172 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
173 a0 += 8;
174 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
175 a1 += 8;
176 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
177 a2 += 8;
178 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
179 a3 += 8;
180 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
181 a4 += 8;
182 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
183 a5 += 8;
184
185 const __m256i vbb01234567x01234567 = _mm256_load_si256(w);
186 const __m256i vbb89ABCDEFx01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
187 const __m256i vb01234567x0123 = _mm256_and_si256(vbb01234567x01234567, vmask);
188 const __m256i vb89ABCDEFx0123 = _mm256_and_si256(vbb89ABCDEFx01234567, vmask);
189
190 vacc0x0123 = _mm256_dpbusd_epi32_madd(vacc0x0123, va0x01234567, vb01234567x0123);
191 vacc0x4567 = _mm256_dpbusd_epi32_madd(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
192 vacc1x0123 = _mm256_dpbusd_epi32_madd(vacc1x0123, va1x01234567, vb01234567x0123);
193 vacc1x4567 = _mm256_dpbusd_epi32_madd(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
194 vacc2x0123 = _mm256_dpbusd_epi32_madd(vacc2x0123, va2x01234567, vb01234567x0123);
195 vacc2x4567 = _mm256_dpbusd_epi32_madd(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
196 vacc3x0123 = _mm256_dpbusd_epi32_madd(vacc3x0123, va3x01234567, vb01234567x0123);
197 vacc3x4567 = _mm256_dpbusd_epi32_madd(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
198 vacc4x0123 = _mm256_dpbusd_epi32_madd(vacc4x0123, va4x01234567, vb01234567x0123);
199 vacc4x4567 = _mm256_dpbusd_epi32_madd(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
200 vacc5x0123 = _mm256_dpbusd_epi32_madd(vacc5x0123, va5x01234567, vb01234567x0123);
201 vacc5x4567 = _mm256_dpbusd_epi32_madd(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
202
203 w = (const int8_t*) w + 64;
204 k -= 8 * sizeof(int8_t);
205 }
206
207 // Add adjacent pairs
208 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
209 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
210 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
211 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
212 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
213 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
214 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
215 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
216 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
217 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
218 const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);
219 __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));
220
221 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
222 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
223 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
224 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
225 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
226 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
227
228 vout0x01234567 = _mm256_mul_ps(vout0x01234567, _mm256_set1_ps(quantization_params[0].inv_scale));
229 vout1x01234567 = _mm256_mul_ps(vout1x01234567, _mm256_set1_ps(quantization_params[1].inv_scale));
230 vout2x01234567 = _mm256_mul_ps(vout2x01234567, _mm256_set1_ps(quantization_params[2].inv_scale));
231 vout3x01234567 = _mm256_mul_ps(vout3x01234567, _mm256_set1_ps(quantization_params[3].inv_scale));
232 vout4x01234567 = _mm256_mul_ps(vout4x01234567, _mm256_set1_ps(quantization_params[4].inv_scale));
233 vout5x01234567 = _mm256_mul_ps(vout5x01234567, _mm256_set1_ps(quantization_params[5].inv_scale));
234
235 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
236 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
237 w = (const float*) w + 16;
238
239 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
240 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
241 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
242 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
243 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
244 vout5x01234567 = _mm256_fmadd_ps(vout5x01234567, vfilter_output_scale01234567, vbias01234567);
245
246 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
247 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
248 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
249 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
250 vout4x01234567 = _mm256_max_ps(vout4x01234567, voutput_min);
251 vout5x01234567 = _mm256_max_ps(vout5x01234567, voutput_min);
252
253 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
254 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
255 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
256 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
257 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max);
258 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max);
259
260 __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT);
261 __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT);
262 __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT);
263 __m128i vfp16out3x01234567 = _mm256_cvtps_ph(vout3x01234567, _MM_FROUND_TO_NEAREST_INT);
264 __m128i vfp16out4x01234567 = _mm256_cvtps_ph(vout4x01234567, _MM_FROUND_TO_NEAREST_INT);
265 __m128i vfp16out5x01234567 = _mm256_cvtps_ph(vout5x01234567, _MM_FROUND_TO_NEAREST_INT);
266 if XNN_LIKELY(nc >= 8) {
267 _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567);
268 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
269 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
270 _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567);
271 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
272 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
273 _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567);
274 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
275 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
276 _mm_storeu_si128((__m128i*) c3, vfp16out3x01234567);
277 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
278 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
279 _mm_storeu_si128((__m128i*) c4, vfp16out4x01234567);
280 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
281 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
282 _mm_storeu_si128((__m128i*) c5, vfp16out5x01234567);
283 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
284 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
285 nc -= 8;
286 } else {
287 if (nc & 4) {
288 _mm_storel_epi64((__m128i*) c0, vfp16out0x01234567);
289 c0 += 4;
290 vfp16out0x01234567 = _mm_unpackhi_epi64(vfp16out0x01234567, vfp16out0x01234567);
291 _mm_storel_epi64((__m128i*) c1, vfp16out1x01234567);
292 c1 += 4;
293 vfp16out1x01234567 = _mm_unpackhi_epi64(vfp16out1x01234567, vfp16out1x01234567);
294 _mm_storel_epi64((__m128i*) c2, vfp16out2x01234567);
295 c2 += 4;
296 vfp16out2x01234567 = _mm_unpackhi_epi64(vfp16out2x01234567, vfp16out2x01234567);
297 _mm_storel_epi64((__m128i*) c3, vfp16out3x01234567);
298 c3 += 4;
299 vfp16out3x01234567 = _mm_unpackhi_epi64(vfp16out3x01234567, vfp16out3x01234567);
300 _mm_storel_epi64((__m128i*) c4, vfp16out4x01234567);
301 c4 += 4;
302 vfp16out4x01234567 = _mm_unpackhi_epi64(vfp16out4x01234567, vfp16out4x01234567);
303 _mm_storel_epi64((__m128i*) c5, vfp16out5x01234567);
304 c5 += 4;
305 vfp16out5x01234567 = _mm_unpackhi_epi64(vfp16out5x01234567, vfp16out5x01234567);
306 }
307 if (nc & 2) {
308 _mm_storeu_si32(c0, vfp16out0x01234567);
309 c0 += 2;
310 vfp16out0x01234567 = _mm_srli_epi64(vfp16out0x01234567, 32);
311 _mm_storeu_si32(c1, vfp16out1x01234567);
312 c1 += 2;
313 vfp16out1x01234567 = _mm_srli_epi64(vfp16out1x01234567, 32);
314 _mm_storeu_si32(c2, vfp16out2x01234567);
315 c2 += 2;
316 vfp16out2x01234567 = _mm_srli_epi64(vfp16out2x01234567, 32);
317 _mm_storeu_si32(c3, vfp16out3x01234567);
318 c3 += 2;
319 vfp16out3x01234567 = _mm_srli_epi64(vfp16out3x01234567, 32);
320 _mm_storeu_si32(c4, vfp16out4x01234567);
321 c4 += 2;
322 vfp16out4x01234567 = _mm_srli_epi64(vfp16out4x01234567, 32);
323 _mm_storeu_si32(c5, vfp16out5x01234567);
324 c5 += 2;
325 vfp16out5x01234567 = _mm_srli_epi64(vfp16out5x01234567, 32);
326 }
327 if (nc & 1) {
328 *c0 = (uint16_t) _mm_extract_epi16(vfp16out0x01234567, 0);
329 *c1 = (uint16_t) _mm_extract_epi16(vfp16out1x01234567, 0);
330 *c2 = (uint16_t) _mm_extract_epi16(vfp16out2x01234567, 0);
331 *c3 = (uint16_t) _mm_extract_epi16(vfp16out3x01234567, 0);
332 *c4 = (uint16_t) _mm_extract_epi16(vfp16out4x01234567, 0);
333 *c5 = (uint16_t) _mm_extract_epi16(vfp16out5x01234567, 0);
334 }
335 nc = 0;
336 }
337 } while (nc != 0);
338 }