Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-6x8c8-minmax-avxvnni-prfm.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 #include "xnnpack/prefetch.h"
19
20
21 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_6x8c8__avxvnni_prfm(
22 size_t mr,
23 size_t nc,
24 size_t kc,
25 const int8_t* restrict a,
26 size_t a_stride,
27 const void* restrict w,
28 float* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
32 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(mr != 0);
35 assert(mr <= 6);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(kc % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 8 * sizeof(int8_t));
44 const int8_t* a0 = a;
45 float* c0 = c;
46 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
47 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 a1 = a0;
50 c1 = c0;
51 }
52 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
53 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
54 if XNN_UNPREDICTABLE(mr <= 2) {
55 a2 = a1;
56 c2 = c1;
57 }
58 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
59 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
60 if XNN_UNPREDICTABLE(mr < 4) {
61 a3 = a2;
62 c3 = c2;
63 }
64 const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
65 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
66 if XNN_UNPREDICTABLE(mr <= 4) {
67 a4 = a3;
68 c4 = c3;
69 }
70 const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride);
71 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
72 if XNN_UNPREDICTABLE(mr != 6) {
73 a5 = a4;
74 c5 = c4;
75 }
76
77 const __m256i vsign_mask = _mm256_set1_epi8(0x80);
78 XNN_FORCE_REALIZATION(vsign_mask);
79 const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point + 128);
80 const __m256i vinput_zero_point1 = _mm256_set1_epi32((int) quantization_params[1].zero_point + 128);
81 const __m256i vinput_zero_point2 = _mm256_set1_epi32((int) quantization_params[2].zero_point + 128);
82 const __m256i vinput_zero_point3 = _mm256_set1_epi32((int) quantization_params[3].zero_point + 128);
83 const __m256i vinput_zero_point4 = _mm256_set1_epi32((int) quantization_params[4].zero_point + 128);
84 const __m256i vinput_zero_point5 = _mm256_set1_epi32((int) quantization_params[5].zero_point + 128);
85 const __m256 voutput_min = _mm256_set1_ps(params->scalar.min);
86 const __m256 voutput_max = _mm256_set1_ps(params->scalar.max);
87 // XNN_FORCE_REALIZATION(voutput_min);
88 // XNN_FORCE_REALIZATION(voutput_max);
89 do {
90 const __m256i vksum01234567 = _mm256_load_si256(w);
91 __m256i vsum0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point0);
92 __m256i vacc0x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 0));
93 __m256i vacc0x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum0x01234567, 1));
94 __m256i vsum1x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point1);
95 __m256i vacc1x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 0));
96 __m256i vacc1x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum1x01234567, 1));
97 __m256i vsum2x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point2);
98 __m256i vacc2x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 0));
99 __m256i vacc2x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum2x01234567, 1));
100 __m256i vsum3x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point3);
101 __m256i vacc3x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 0));
102 __m256i vacc3x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum3x01234567, 1));
103 __m256i vsum4x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point4);
104 __m256i vacc4x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum4x01234567, 0));
105 __m256i vacc4x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum4x01234567, 1));
106 __m256i vsum5x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point5);
107 __m256i vacc5x0123 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum5x01234567, 0));
108 __m256i vacc5x4567 = _mm256_cvtepu32_epi64(_mm256_extracti128_si256(vsum5x01234567, 1));
109 w = (const int32_t*) w + 8;
110
111 size_t k = kc;
112 while (k >= 16 * sizeof(int8_t)) {
113 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
114 const __m256i va0x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
115 a0 += 16;
116 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
117 const __m256i va1x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
118 a1 += 16;
119 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
120 const __m256i va2x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
121 a2 += 16;
122 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
123 const __m256i va3x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
124 a3 += 16;
125 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
126 const __m256i va4x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask);
127 a4 += 16;
128 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
129 const __m256i va5x89ABCDEF = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask);
130 a5 += 16;
131
132 const __m256i vb01234567x0123 = _mm256_load_si256(w);
133 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
134 const __m256i vb01234567x4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 64));
135 const __m256i vb89ABCDEFx4567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 96));
136 xnn_prefetch_to_l1((const int8_t*) w + 896);
137
138 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
139 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
140 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
141 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
142 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
143 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
144 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
145 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
146 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
147 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
148 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
149 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
150 xnn_prefetch_to_l1((const int8_t*) w + 960);
151 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x89ABCDEF, vb01234567x4567);
152 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x89ABCDEF, vb89ABCDEFx4567);
153 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x89ABCDEF, vb01234567x4567);
154 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x89ABCDEF, vb89ABCDEFx4567);
155 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x89ABCDEF, vb01234567x4567);
156 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x89ABCDEF, vb89ABCDEFx4567);
157 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x89ABCDEF, vb01234567x4567);
158 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x89ABCDEF, vb89ABCDEFx4567);
159 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x89ABCDEF, vb01234567x4567);
160 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x89ABCDEF, vb89ABCDEFx4567);
161 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x89ABCDEF, vb01234567x4567);
162 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x89ABCDEF, vb89ABCDEFx4567);
163
164 w = (const int8_t*) w + 128;
165 k -= 16 * sizeof(int8_t);
166 }
167
168 if (k != 0) {
169 const __m256i va0x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
170 a0 += 8;
171 const __m256i va1x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
172 a1 += 8;
173 const __m256i va2x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
174 a2 += 8;
175 const __m256i va3x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
176 a3 += 8;
177 const __m256i va4x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a4)), vsign_mask);
178 a4 += 8;
179 const __m256i va5x01234567 = _mm256_xor_si256(_mm256_set1_epi64x((int64_t) unaligned_load_u64(a5)), vsign_mask);
180 a5 += 8;
181
182 const __m256i vb01234567x0123 = _mm256_load_si256(w);
183 const __m256i vb89ABCDEFx0123 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
184
185 vacc0x0123 = _mm256_dpbusd_avx_epi32(vacc0x0123, va0x01234567, vb01234567x0123);
186 vacc0x4567 = _mm256_dpbusd_avx_epi32(vacc0x4567, va0x01234567, vb89ABCDEFx0123);
187 vacc1x0123 = _mm256_dpbusd_avx_epi32(vacc1x0123, va1x01234567, vb01234567x0123);
188 vacc1x4567 = _mm256_dpbusd_avx_epi32(vacc1x4567, va1x01234567, vb89ABCDEFx0123);
189 vacc2x0123 = _mm256_dpbusd_avx_epi32(vacc2x0123, va2x01234567, vb01234567x0123);
190 vacc2x4567 = _mm256_dpbusd_avx_epi32(vacc2x4567, va2x01234567, vb89ABCDEFx0123);
191 vacc3x0123 = _mm256_dpbusd_avx_epi32(vacc3x0123, va3x01234567, vb01234567x0123);
192 vacc3x4567 = _mm256_dpbusd_avx_epi32(vacc3x4567, va3x01234567, vb89ABCDEFx0123);
193 vacc4x0123 = _mm256_dpbusd_avx_epi32(vacc4x0123, va4x01234567, vb01234567x0123);
194 vacc4x4567 = _mm256_dpbusd_avx_epi32(vacc4x4567, va4x01234567, vb89ABCDEFx0123);
195 vacc5x0123 = _mm256_dpbusd_avx_epi32(vacc5x0123, va5x01234567, vb01234567x0123);
196 vacc5x4567 = _mm256_dpbusd_avx_epi32(vacc5x4567, va5x01234567, vb89ABCDEFx0123);
197 xnn_prefetch_to_l1((const int8_t*) w + 960);
198
199 w = (const int8_t*) w + 64;
200 k -= 8 * sizeof(int8_t);
201 }
202
203 // Add adjacent pairs
204 const __m256i vsum0x02134657 = _mm256_hadd_epi32(vacc0x0123, vacc0x4567);
205 __m256i vacc0x01234567 = _mm256_permute4x64_epi64(vsum0x02134657, _MM_SHUFFLE(3, 1, 2, 0));
206 const __m256i vsum1x02134657 = _mm256_hadd_epi32(vacc1x0123, vacc1x4567);
207 __m256i vacc1x01234567 = _mm256_permute4x64_epi64(vsum1x02134657, _MM_SHUFFLE(3, 1, 2, 0));
208 const __m256i vsum2x02134657 = _mm256_hadd_epi32(vacc2x0123, vacc2x4567);
209 __m256i vacc2x01234567 = _mm256_permute4x64_epi64(vsum2x02134657, _MM_SHUFFLE(3, 1, 2, 0));
210 const __m256i vsum3x02134657 = _mm256_hadd_epi32(vacc3x0123, vacc3x4567);
211 __m256i vacc3x01234567 = _mm256_permute4x64_epi64(vsum3x02134657, _MM_SHUFFLE(3, 1, 2, 0));
212 const __m256i vsum4x02134657 = _mm256_hadd_epi32(vacc4x0123, vacc4x4567);
213 __m256i vacc4x01234567 = _mm256_permute4x64_epi64(vsum4x02134657, _MM_SHUFFLE(3, 1, 2, 0));
214 const __m256i vsum5x02134657 = _mm256_hadd_epi32(vacc5x0123, vacc5x4567);
215 __m256i vacc5x01234567 = _mm256_permute4x64_epi64(vsum5x02134657, _MM_SHUFFLE(3, 1, 2, 0));
216
217 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
218 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
219 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
220 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567);
221 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567);
222 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567);
223
224 vout0x01234567 = _mm256_mul_ps(vout0x01234567, _mm256_set1_ps(quantization_params[0].inv_scale));
225 vout1x01234567 = _mm256_mul_ps(vout1x01234567, _mm256_set1_ps(quantization_params[1].inv_scale));
226 vout2x01234567 = _mm256_mul_ps(vout2x01234567, _mm256_set1_ps(quantization_params[2].inv_scale));
227 vout3x01234567 = _mm256_mul_ps(vout3x01234567, _mm256_set1_ps(quantization_params[3].inv_scale));
228 vout4x01234567 = _mm256_mul_ps(vout4x01234567, _mm256_set1_ps(quantization_params[4].inv_scale));
229 vout5x01234567 = _mm256_mul_ps(vout5x01234567, _mm256_set1_ps(quantization_params[5].inv_scale));
230
231 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
232 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
233 w = (const float*) w + 16;
234
235 vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567);
236 vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567);
237 vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567);
238 vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567);
239 vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567);
240 vout5x01234567 = _mm256_fmadd_ps(vout5x01234567, vfilter_output_scale01234567, vbias01234567);
241
242 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
243 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
244 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
245 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
246 vout4x01234567 = _mm256_max_ps(vout4x01234567, voutput_min);
247 vout5x01234567 = _mm256_max_ps(vout5x01234567, voutput_min);
248
249 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
250 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
251 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
252 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
253 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max);
254 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max);
255
256 if XNN_LIKELY(nc >= 8) {
257 _mm256_storeu_ps(c0, vout0x01234567);
258 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
259 c0 = (float*) ((uintptr_t) c0 + cn_stride);
260 _mm256_storeu_ps(c1, vout1x01234567);
261 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
262 c1 = (float*) ((uintptr_t) c1 + cn_stride);
263 _mm256_storeu_ps(c2, vout2x01234567);
264 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
265 c2 = (float*) ((uintptr_t) c2 + cn_stride);
266 _mm256_storeu_ps(c3, vout3x01234567);
267 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
268 c3 = (float*) ((uintptr_t) c3 + cn_stride);
269 _mm256_storeu_ps(c4, vout4x01234567);
270 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
271 c4 = (float*) ((uintptr_t) c4 + cn_stride);
272 _mm256_storeu_ps(c5, vout5x01234567);
273 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
274 c5 = (float*) ((uintptr_t) c5 + cn_stride);
275 nc -= 8;
276 } else {
277 __m128 vout0x0123 = _mm256_castps256_ps128(vout0x01234567);
278 __m128 vout1x0123 = _mm256_castps256_ps128(vout1x01234567);
279 __m128 vout2x0123 = _mm256_castps256_ps128(vout2x01234567);
280 __m128 vout3x0123 = _mm256_castps256_ps128(vout3x01234567);
281 __m128 vout4x0123 = _mm256_castps256_ps128(vout4x01234567);
282 __m128 vout5x0123 = _mm256_castps256_ps128(vout5x01234567);
283 if (nc & 4) {
284 _mm_storeu_ps(c0, vout0x0123);
285 c0 += 4;
286 vout0x0123 = _mm256_extractf128_ps(vout0x01234567, 1);
287 _mm_storeu_ps(c1, vout1x0123);
288 c1 += 4;
289 vout1x0123 = _mm256_extractf128_ps(vout1x01234567, 1);
290 _mm_storeu_ps(c2, vout2x0123);
291 c2 += 4;
292 vout2x0123 = _mm256_extractf128_ps(vout2x01234567, 1);
293 _mm_storeu_ps(c3, vout3x0123);
294 c3 += 4;
295 vout3x0123 = _mm256_extractf128_ps(vout3x01234567, 1);
296 _mm_storeu_ps(c4, vout4x0123);
297 c4 += 4;
298 vout4x0123 = _mm256_extractf128_ps(vout4x01234567, 1);
299 _mm_storeu_ps(c5, vout5x0123);
300 c5 += 4;
301 vout5x0123 = _mm256_extractf128_ps(vout5x01234567, 1);
302 }
303 if (nc & 2) {
304 _mm_storel_pi((__m64*) c0, vout0x0123);
305 c0 += 2;
306 vout0x0123 = _mm_movehl_ps(vout0x0123, vout0x0123);
307 _mm_storel_pi((__m64*) c1, vout1x0123);
308 c1 += 2;
309 vout1x0123 = _mm_movehl_ps(vout1x0123, vout1x0123);
310 _mm_storel_pi((__m64*) c2, vout2x0123);
311 c2 += 2;
312 vout2x0123 = _mm_movehl_ps(vout2x0123, vout2x0123);
313 _mm_storel_pi((__m64*) c3, vout3x0123);
314 c3 += 2;
315 vout3x0123 = _mm_movehl_ps(vout3x0123, vout3x0123);
316 _mm_storel_pi((__m64*) c4, vout4x0123);
317 c4 += 2;
318 vout4x0123 = _mm_movehl_ps(vout4x0123, vout4x0123);
319 _mm_storel_pi((__m64*) c5, vout5x0123);
320 c5 += 2;
321 vout5x0123 = _mm_movehl_ps(vout5x0123, vout5x0123);
322 }
323 if (nc & 1) {
324 _mm_store_ss(c0, vout0x0123);
325 _mm_store_ss(c1, vout1x0123);
326 _mm_store_ss(c2, vout2x0123);
327 _mm_store_ss(c3, vout3x0123);
328 _mm_store_ss(c4, vout4x0123);
329 _mm_store_ss(c5, vout5x0123);
330 }
331 nc = 0;
332 }
333 } while (nc != 0);
334 }