Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-6x8c8-minmax-avxvnni.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_f32_qc8w_gemm_minmax_ukernel_6x8c8__avxvnni(
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 float* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 const union xnn_f32_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 float* c0 = c;
45 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
46 float* c1 = (float*) ((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 float* c2 = (float*) ((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 float* c3 = (float*) ((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 float* c4 = (float*) ((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 float* c5 = (float*) ((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_set1_ps(params->scalar.min);
85 const __m256 voutput_max = _mm256_set1_ps(params->scalar.max);
86 // XNN_FORCE_REALIZATION(voutput_min);
87 // XNN_FORCE_REALIZATION(voutput_max);
88 do {
89 const __m256i vksum01234567 = _mm256_load_si256(w);
90 __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point0);
91 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
92 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
93 __m256i vsum1x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point1);
94 __m256i vacc1x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 0));
95 __m256i vacc1x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 1));
96 __m256i vsum2x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point2);
97 __m256i vacc2x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 0));
98 __m256i vacc2x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 1));
99 __m256i vsum3x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point3);
100 __m256i vacc3x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 0));
101 __m256i vacc3x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 1));
102 __m256i vsum4x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point4);
103 __m256i vacc4x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum4x01234567, 0));
104 __m256i vacc4x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum4x01234567, 1));
105 __m256i vsum5x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point5);
106 __m256i vacc5x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum5x01234567, 0));
107 __m256i vacc5x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum5x01234567, 1));
108 w = (const int32_t*) w + 8;
109
110 size_t k = kc;
111 while (k >= 16 * sizeof(int8_t)) {
112 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
113 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
114 a0 += 16;
115 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
116 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
117 a1 += 16;
118 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
119 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
120 a2 += 16;
121 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
122 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
123 a3 += 16;
124 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
125 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
126 a4 += 16;
127 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
128 const __m256i va5x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
129 a5 += 16;
130
131 const __m256i vb01234567x0123 = _mm256_load_si256(w);
132 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
133 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
134 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
135
136 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
137 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
138 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
139 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
140 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
141 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
142 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
143 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
144 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
145 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
146 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
147 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
148 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
149 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
150 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
151 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
152 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
153 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
154 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
155 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
156 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
157 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
158 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x89ABCDEF, vb01234567x4567);
159 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);
160
161 w = (const int8_t*) w + 128;
162 k -= 16 * sizeof(int8_t);
163 }
164
165 if (k != 0) {
166 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
167 a0 += 8;
168 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
169 a1 += 8;
170 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
171 a2 += 8;
172 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
173 a3 += 8;
174 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
175 a4 += 8;
176 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
177 a5 += 8;
178
179 const __m256i vb01234567x0123 = _mm256_load_si256(w);
180 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
181
182 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
183 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
184 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
185 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
186 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
187 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
188 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
189 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
190 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
191 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
192 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
193 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
194
195 w = (const int8_t*) w + 64;
196 k -= 8 * sizeof(int8_t);
197 }
198
199 // Add adjacent pairs
200 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
201 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
202 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
203 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
204 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
205 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
206 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
207 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
208 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
209 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
210 const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);
211 __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));
212
213 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
214 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
215 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
216 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
217 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
218 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
219
220 vout0x01234567 = _mm256_mul_ps(vout0x01234567, _mm256_set1_ps(quantization_params[0].inv_scale));
221 vout1x01234567 = _mm256_mul_ps(vout1x01234567, _mm256_set1_ps(quantization_params[1].inv_scale));
222 vout2x01234567 = _mm256_mul_ps(vout2x01234567, _mm256_set1_ps(quantization_params[2].inv_scale));
223 vout3x01234567 = _mm256_mul_ps(vout3x01234567, _mm256_set1_ps(quantization_params[3].inv_scale));
224 vout4x01234567 = _mm256_mul_ps(vout4x01234567, _mm256_set1_ps(quantization_params[4].inv_scale));
225 vout5x01234567 = _mm256_mul_ps(vout5x01234567, _mm256_set1_ps(quantization_params[5].inv_scale));
226
227 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
228 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
229 w = (const float*) w + 16;
230
231 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
232 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
233 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
234 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
235 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
236 vout5x01234567 = _mm256_fmadd_ps(vout5x01234567, vfilter_output_scale01234567, vbias01234567);
237
238 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
239 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
240 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
241 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
242 vout4x01234567 = _mm256_max_ps(vout4x01234567, voutput_min);
243 vout5x01234567 = _mm256_max_ps(vout5x01234567, voutput_min);
244
245 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
246 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
247 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
248 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
249 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max);
250 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max);
251
252 if XNN_LIKELY(nc >= 8) {
253 _mm256_storeu_ps(c0, vout0x01234567);
254 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
255 c0 = (float*) ((uintptr_t) c0 + cn_stride);
256 _mm256_storeu_ps(c1, vout1x01234567);
257 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
258 c1 = (float*) ((uintptr_t) c1 + cn_stride);
259 _mm256_storeu_ps(c2, vout2x01234567);
260 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
261 c2 = (float*) ((uintptr_t) c2 + cn_stride);
262 _mm256_storeu_ps(c3, vout3x01234567);
263 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
264 c3 = (float*) ((uintptr_t) c3 + cn_stride);
265 _mm256_storeu_ps(c4, vout4x01234567);
266 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
267 c4 = (float*) ((uintptr_t) c4 + cn_stride);
268 _mm256_storeu_ps(c5, vout5x01234567);
269 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
270 c5 = (float*) ((uintptr_t) c5 + cn_stride);
271 nc -= 8;
272 } else {
273 __m128 vout0x0123 = _mm256_castps256_ps128(vout0x01234567);
274 __m128 vout1x0123 = _mm256_castps256_ps128(vout1x01234567);
275 __m128 vout2x0123 = _mm256_castps256_ps128(vout2x01234567);
276 __m128 vout3x0123 = _mm256_castps256_ps128(vout3x01234567);
277 __m128 vout4x0123 = _mm256_castps256_ps128(vout4x01234567);
278 __m128 vout5x0123 = _mm256_castps256_ps128(vout5x01234567);
279 if (nc & 4) {
280 _mm_storeu_ps(c0, vout0x0123);
281 c0 += 4;
282 vout0x0123 = _mm256_extractf128_ps(vout0x01234567, 1);
283 _mm_storeu_ps(c1, vout1x0123);
284 c1 += 4;
285 vout1x0123 = _mm256_extractf128_ps(vout1x01234567, 1);
286 _mm_storeu_ps(c2, vout2x0123);
287 c2 += 4;
288 vout2x0123 = _mm256_extractf128_ps(vout2x01234567, 1);
289 _mm_storeu_ps(c3, vout3x0123);
290 c3 += 4;
291 vout3x0123 = _mm256_extractf128_ps(vout3x01234567, 1);
292 _mm_storeu_ps(c4, vout4x0123);
293 c4 += 4;
294 vout4x0123 = _mm256_extractf128_ps(vout4x01234567, 1);
295 _mm_storeu_ps(c5, vout5x0123);
296 c5 += 4;
297 vout5x0123 = _mm256_extractf128_ps(vout5x01234567, 1);
298 }
299 if (nc & 2) {
300 _mm_storel_pi((__m64*) c0, vout0x0123);
301 c0 += 2;
302 vout0x0123 = _mm_movehl_ps(vout0x0123, vout0x0123);
303 _mm_storel_pi((__m64*) c1, vout1x0123);
304 c1 += 2;
305 vout1x0123 = _mm_movehl_ps(vout1x0123, vout1x0123);
306 _mm_storel_pi((__m64*) c2, vout2x0123);
307 c2 += 2;
308 vout2x0123 = _mm_movehl_ps(vout2x0123, vout2x0123);
309 _mm_storel_pi((__m64*) c3, vout3x0123);
310 c3 += 2;
311 vout3x0123 = _mm_movehl_ps(vout3x0123, vout3x0123);
312 _mm_storel_pi((__m64*) c4, vout4x0123);
313 c4 += 2;
314 vout4x0123 = _mm_movehl_ps(vout4x0123, vout4x0123);
315 _mm_storel_pi((__m64*) c5, vout5x0123);
316 c5 += 2;
317 vout5x0123 = _mm_movehl_ps(vout5x0123, vout5x0123);
318 }
319 if (nc & 1) {
320 _mm_store_ss(c0, vout0x0123);
321 _mm_store_ss(c1, vout1x0123);
322 _mm_store_ss(c2, vout2x0123);
323 _mm_store_ss(c3, vout3x0123);
324 _mm_store_ss(c4, vout4x0123);
325 _mm_store_ss(c5, vout5x0123);
326 }
327 nc = 0;
328 }
329 } while (nc != 0);
330 }