Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-8x8c4-minmax-avxvnni-u2-acc2.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx8c4-avxvnni.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2023 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 #include <stdint.h>
12
13 #include <immintrin.h>
14
15 #include "xnnpack/common.h"
16 #include "xnnpack/gemm.h"
17 #include "xnnpack/intrinsics-polyfill.h"
18 #include "xnnpack/math.h"
19 #include "xnnpack/unaligned.h"
20 #include "xnnpack/prefetch.h"
21
22
23 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_8x8c4__avxvnni_u2_acc2(
24 size_t mr,
25 size_t nc,
26 size_t kc,
27 const int8_t* restrict a,
28 size_t a_stride,
29 const void* restrict w,
30 float* restrict c,
31 size_t cm_stride,
32 size_t cn_stride,
33 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
34 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
35 {
36 assert(mr != 0);
37 assert(mr <= 8);
38 assert(nc != 0);
39 assert(kc != 0);
40 assert(kc % sizeof(int8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 kc = round_up_po2(kc, 4 * sizeof(int8_t));
46 const int8_t* a0 = a;
47 float* c0 = c;
48 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
49 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
50 if XNN_UNPREDICTABLE(mr < 2) {
51 a1 = a0;
52 c1 = c0;
53 }
54 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
55 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
56 if XNN_UNPREDICTABLE(mr <= 2) {
57 a2 = a1;
58 c2 = c1;
59 }
60 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
61 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
62 if XNN_UNPREDICTABLE(mr < 4) {
63 a3 = a2;
64 c3 = c2;
65 }
66 const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
67 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
68 if XNN_UNPREDICTABLE(mr <= 4) {
69 a4 = a3;
70 c4 = c3;
71 }
72 const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride);
73 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
74 if XNN_UNPREDICTABLE(mr < 6) {
75 a5 = a4;
76 c5 = c4;
77 }
78 const int8_t* a6 = (const int8_t*) ((uintptr_t) a5 + a_stride);
79 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
80 if XNN_UNPREDICTABLE(mr <= 6) {
81 a6 = a5;
82 c6 = c5;
83 }
84 const int8_t* a7 = (const int8_t*) ((uintptr_t) a6 + a_stride);
85 float* c7 = (float*) ((uintptr_t) c6 + cm_stride);
86 if XNN_UNPREDICTABLE(mr != 8) {
87 a7 = a6;
88 c7 = c6;
89 }
90
91 const __m256i vsign_mask =_mm256_set1_epi8(0x80);
92 XNN_FORCE_REALIZATION(vsign_mask);
93 const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point + 128);
94 const __m256i vinput_zero_point1 = _mm256_set1_epi32((int) quantization_params[1].zero_point + 128);
95 const __m256i vinput_zero_point2 = _mm256_set1_epi32((int) quantization_params[2].zero_point + 128);
96 const __m256i vinput_zero_point3 = _mm256_set1_epi32((int) quantization_params[3].zero_point + 128);
97 const __m256i vinput_zero_point4 = _mm256_set1_epi32((int) quantization_params[4].zero_point + 128);
98 const __m256i vinput_zero_point5 = _mm256_set1_epi32((int) quantization_params[5].zero_point + 128);
99 const __m256i vinput_zero_point6 = _mm256_set1_epi32((int) quantization_params[6].zero_point + 128);
100 const __m256i vinput_zero_point7 = _mm256_set1_epi32((int) quantization_params[7].zero_point + 128);
101 const __m256 voutput_min = _mm256_set1_ps(params->scalar.min);
102 const __m256 voutput_max = _mm256_set1_ps(params->scalar.max);
103 // XNN_FORCE_REALIZATION(voutput_min);
104 // XNN_FORCE_REALIZATION(voutput_max);
105 do {
106 const __m256i vksum01234567 = _mm256_load_si256(w);
107 __m256i vacc0x0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point0);
108 __m256i vacc0x1x01234567 = _mm256_setzero_si256();
109 __m256i vacc1x0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point1);
110 __m256i vacc1x1x01234567 = _mm256_setzero_si256();
111 __m256i vacc2x0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point2);
112 __m256i vacc2x1x01234567 = _mm256_setzero_si256();
113 __m256i vacc3x0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point3);
114 __m256i vacc3x1x01234567 = _mm256_setzero_si256();
115 __m256i vacc4x0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point4);
116 __m256i vacc4x1x01234567 = _mm256_setzero_si256();
117 __m256i vacc5x0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point5);
118 __m256i vacc5x1x01234567 = _mm256_setzero_si256();
119 __m256i vacc6x0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point6);
120 __m256i vacc6x1x01234567 = _mm256_setzero_si256();
121 __m256i vacc7x0x01234567 = _mm256_mullo_epi32(vksum01234567, vinput_zero_point7);
122 __m256i vacc7x1x01234567 = _mm256_setzero_si256();
123 w = (const int32_t*) w + 8;
124
125 size_t k = kc;
126 while (k >= 8 * sizeof(int8_t)) {
127 __m256i va0x0x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a0 + 0));
128 __m256i va0x1x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a0 + 4));
129 a0 += 8;
130 __m256i va1x0x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a1 + 0));
131 __m256i va1x1x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a1 + 4));
132 a1 += 8;
133 __m256i va2x0x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a2 + 0));
134 __m256i va2x1x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a2 + 4));
135 a2 += 8;
136 __m256i va3x0x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a3 + 0));
137 __m256i va3x1x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a3 + 4));
138 a3 += 8;
139 __m256i va4x0x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a4 + 0));
140 __m256i va4x1x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a4 + 4));
141 a4 += 8;
142 __m256i va5x0x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a5 + 0));
143 __m256i va5x1x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a5 + 4));
144 a5 += 8;
145 __m256i va6x0x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a6 + 0));
146 __m256i va6x1x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a6 + 4));
147 a6 += 8;
148 __m256i va7x0x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a7 + 0));
149 __m256i va7x1x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a7 + 4));
150 a7 += 8;
151
152 va0x0x0123 = _mm256_xor_si256(va0x0x0123, vsign_mask);
153 va0x1x0123 = _mm256_xor_si256(va0x1x0123, vsign_mask);
154 va1x0x0123 = _mm256_xor_si256(va1x0x0123, vsign_mask);
155 va1x1x0123 = _mm256_xor_si256(va1x1x0123, vsign_mask);
156 va2x0x0123 = _mm256_xor_si256(va2x0x0123, vsign_mask);
157 va2x1x0123 = _mm256_xor_si256(va2x1x0123, vsign_mask);
158 va3x0x0123 = _mm256_xor_si256(va3x0x0123, vsign_mask);
159 va3x1x0123 = _mm256_xor_si256(va3x1x0123, vsign_mask);
160 va4x0x0123 = _mm256_xor_si256(va4x0x0123, vsign_mask);
161 va4x1x0123 = _mm256_xor_si256(va4x1x0123, vsign_mask);
162 va5x0x0123 = _mm256_xor_si256(va5x0x0123, vsign_mask);
163 va5x1x0123 = _mm256_xor_si256(va5x1x0123, vsign_mask);
164 va6x0x0123 = _mm256_xor_si256(va6x0x0123, vsign_mask);
165 va6x1x0123 = _mm256_xor_si256(va6x1x0123, vsign_mask);
166 va7x0x0123 = _mm256_xor_si256(va7x0x0123, vsign_mask);
167 va7x1x0123 = _mm256_xor_si256(va7x1x0123, vsign_mask);
168
169 const __m256i vb0x01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 0));
170 const __m256i vb1x01234567 = _mm256_load_si256((const __m256i*) ((const int8_t*) w + 32));
171 w = (const int8_t*) w + 64;
172
173 vacc0x0x01234567 = _mm256_dpbusd_avx_epi32(vacc0x0x01234567, va0x0x0123, vb0x01234567);
174 vacc0x1x01234567 = _mm256_dpbusd_avx_epi32(vacc0x1x01234567, va0x1x0123, vb1x01234567);
175 vacc1x0x01234567 = _mm256_dpbusd_avx_epi32(vacc1x0x01234567, va1x0x0123, vb0x01234567);
176 vacc1x1x01234567 = _mm256_dpbusd_avx_epi32(vacc1x1x01234567, va1x1x0123, vb1x01234567);
177 vacc2x0x01234567 = _mm256_dpbusd_avx_epi32(vacc2x0x01234567, va2x0x0123, vb0x01234567);
178 vacc2x1x01234567 = _mm256_dpbusd_avx_epi32(vacc2x1x01234567, va2x1x0123, vb1x01234567);
179 vacc3x0x01234567 = _mm256_dpbusd_avx_epi32(vacc3x0x01234567, va3x0x0123, vb0x01234567);
180 vacc3x1x01234567 = _mm256_dpbusd_avx_epi32(vacc3x1x01234567, va3x1x0123, vb1x01234567);
181 vacc4x0x01234567 = _mm256_dpbusd_avx_epi32(vacc4x0x01234567, va4x0x0123, vb0x01234567);
182 vacc4x1x01234567 = _mm256_dpbusd_avx_epi32(vacc4x1x01234567, va4x1x0123, vb1x01234567);
183 vacc5x0x01234567 = _mm256_dpbusd_avx_epi32(vacc5x0x01234567, va5x0x0123, vb0x01234567);
184 vacc5x1x01234567 = _mm256_dpbusd_avx_epi32(vacc5x1x01234567, va5x1x0123, vb1x01234567);
185 vacc6x0x01234567 = _mm256_dpbusd_avx_epi32(vacc6x0x01234567, va6x0x0123, vb0x01234567);
186 vacc6x1x01234567 = _mm256_dpbusd_avx_epi32(vacc6x1x01234567, va6x1x0123, vb1x01234567);
187 vacc7x0x01234567 = _mm256_dpbusd_avx_epi32(vacc7x0x01234567, va7x0x0123, vb0x01234567);
188 vacc7x1x01234567 = _mm256_dpbusd_avx_epi32(vacc7x1x01234567, va7x1x0123, vb1x01234567);
189
190 k -= 8 * sizeof(int8_t);
191 }
192 vacc0x0x01234567 = _mm256_add_epi32(vacc0x0x01234567, vacc0x1x01234567);
193 vacc1x0x01234567 = _mm256_add_epi32(vacc1x0x01234567, vacc1x1x01234567);
194 vacc2x0x01234567 = _mm256_add_epi32(vacc2x0x01234567, vacc2x1x01234567);
195 vacc3x0x01234567 = _mm256_add_epi32(vacc3x0x01234567, vacc3x1x01234567);
196 vacc4x0x01234567 = _mm256_add_epi32(vacc4x0x01234567, vacc4x1x01234567);
197 vacc5x0x01234567 = _mm256_add_epi32(vacc5x0x01234567, vacc5x1x01234567);
198 vacc6x0x01234567 = _mm256_add_epi32(vacc6x0x01234567, vacc6x1x01234567);
199 vacc7x0x01234567 = _mm256_add_epi32(vacc7x0x01234567, vacc7x1x01234567);
200 while (k != 0) {
201 __m256i va0x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a0));
202 a0 += 4;
203 __m256i va1x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a1));
204 a1 += 4;
205 __m256i va2x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a2));
206 a2 += 4;
207 __m256i va3x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a3));
208 a3 += 4;
209 __m256i va4x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a4));
210 a4 += 4;
211 __m256i va5x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a5));
212 a5 += 4;
213 __m256i va6x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a6));
214 a6 += 4;
215 __m256i va7x0123 = _mm256_set1_epi32((int) unaligned_load_u32(a7));
216 a7 += 4;
217
218 va0x0123 = _mm256_xor_si256(va0x0123, vsign_mask);
219 va1x0123 = _mm256_xor_si256(va1x0123, vsign_mask);
220 va2x0123 = _mm256_xor_si256(va2x0123, vsign_mask);
221 va3x0123 = _mm256_xor_si256(va3x0123, vsign_mask);
222 va4x0123 = _mm256_xor_si256(va4x0123, vsign_mask);
223 va5x0123 = _mm256_xor_si256(va5x0123, vsign_mask);
224 va6x0123 = _mm256_xor_si256(va6x0123, vsign_mask);
225 va7x0123 = _mm256_xor_si256(va7x0123, vsign_mask);
226
227 const __m256i vb01234567 = _mm256_load_si256(w);
228
229 vacc0x0x01234567 = _mm256_dpbusd_avx_epi32(vacc0x0x01234567, va0x0123, vb01234567);
230 vacc1x0x01234567 = _mm256_dpbusd_avx_epi32(vacc1x0x01234567, va1x0123, vb01234567);
231 vacc2x0x01234567 = _mm256_dpbusd_avx_epi32(vacc2x0x01234567, va2x0123, vb01234567);
232 vacc3x0x01234567 = _mm256_dpbusd_avx_epi32(vacc3x0x01234567, va3x0123, vb01234567);
233 vacc4x0x01234567 = _mm256_dpbusd_avx_epi32(vacc4x0x01234567, va4x0123, vb01234567);
234 vacc5x0x01234567 = _mm256_dpbusd_avx_epi32(vacc5x0x01234567, va5x0123, vb01234567);
235 vacc6x0x01234567 = _mm256_dpbusd_avx_epi32(vacc6x0x01234567, va6x0123, vb01234567);
236 vacc7x0x01234567 = _mm256_dpbusd_avx_epi32(vacc7x0x01234567, va7x0123, vb01234567);
237
238 w = (const int8_t*) w + 32;
239 k -= 4 * sizeof(int8_t);
240 }
241
242 __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x0x01234567);
243 __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x0x01234567);
244 __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x0x01234567);
245 __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x0x01234567);
246 __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x0x01234567);
247 __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x0x01234567);
248 __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x0x01234567);
249 __m256 vout7x01234567 = _mm256_cvtepi32_ps(vacc7x0x01234567);
250
251 vout0x01234567 = _mm256_mul_ps(vout0x01234567, _mm256_set1_ps(quantization_params[0].inv_scale));
252 vout1x01234567 = _mm256_mul_ps(vout1x01234567, _mm256_set1_ps(quantization_params[1].inv_scale));
253 vout2x01234567 = _mm256_mul_ps(vout2x01234567, _mm256_set1_ps(quantization_params[2].inv_scale));
254 vout3x01234567 = _mm256_mul_ps(vout3x01234567, _mm256_set1_ps(quantization_params[3].inv_scale));
255 vout4x01234567 = _mm256_mul_ps(vout4x01234567, _mm256_set1_ps(quantization_params[4].inv_scale));
256 vout5x01234567 = _mm256_mul_ps(vout5x01234567, _mm256_set1_ps(quantization_params[5].inv_scale));
257 vout6x01234567 = _mm256_mul_ps(vout6x01234567, _mm256_set1_ps(quantization_params[6].inv_scale));
258 vout7x01234567 = _mm256_mul_ps(vout7x01234567, _mm256_set1_ps(quantization_params[7].inv_scale));
259
260 const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w);
261 const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8);
262 w = (const float*) w + 16;
263
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 vout6x01234567 = _mm256_fmadd_ps(vout6x01234567, vfilter_output_scale01234567, vbias01234567);
271 vout7x01234567 = _mm256_fmadd_ps(vout7x01234567, vfilter_output_scale01234567, vbias01234567);
272
273 vout0x01234567 = _mm256_max_ps(vout0x01234567, voutput_min);
274 vout1x01234567 = _mm256_max_ps(vout1x01234567, voutput_min);
275 vout2x01234567 = _mm256_max_ps(vout2x01234567, voutput_min);
276 vout3x01234567 = _mm256_max_ps(vout3x01234567, voutput_min);
277 vout4x01234567 = _mm256_max_ps(vout4x01234567, voutput_min);
278 vout5x01234567 = _mm256_max_ps(vout5x01234567, voutput_min);
279 vout6x01234567 = _mm256_max_ps(vout6x01234567, voutput_min);
280 vout7x01234567 = _mm256_max_ps(vout7x01234567, voutput_min);
281
282 vout0x01234567 = _mm256_min_ps(vout0x01234567, voutput_max);
283 vout1x01234567 = _mm256_min_ps(vout1x01234567, voutput_max);
284 vout2x01234567 = _mm256_min_ps(vout2x01234567, voutput_max);
285 vout3x01234567 = _mm256_min_ps(vout3x01234567, voutput_max);
286 vout4x01234567 = _mm256_min_ps(vout4x01234567, voutput_max);
287 vout5x01234567 = _mm256_min_ps(vout5x01234567, voutput_max);
288 vout6x01234567 = _mm256_min_ps(vout6x01234567, voutput_max);
289 vout7x01234567 = _mm256_min_ps(vout7x01234567, voutput_max);
290
291 if XNN_LIKELY(nc >= 8) {
292 _mm256_storeu_ps(c0, vout0x01234567);
293 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
294 c0 = (float*) ((uintptr_t) c0 + cn_stride);
295 _mm256_storeu_ps(c1, vout1x01234567);
296 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
297 c1 = (float*) ((uintptr_t) c1 + cn_stride);
298 _mm256_storeu_ps(c2, vout2x01234567);
299 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
300 c2 = (float*) ((uintptr_t) c2 + cn_stride);
301 _mm256_storeu_ps(c3, vout3x01234567);
302 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
303 c3 = (float*) ((uintptr_t) c3 + cn_stride);
304 _mm256_storeu_ps(c4, vout4x01234567);
305 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
306 c4 = (float*) ((uintptr_t) c4 + cn_stride);
307 _mm256_storeu_ps(c5, vout5x01234567);
308 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
309 c5 = (float*) ((uintptr_t) c5 + cn_stride);
310 _mm256_storeu_ps(c6, vout6x01234567);
311 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
312 c6 = (float*) ((uintptr_t) c6 + cn_stride);
313 _mm256_storeu_ps(c7, vout7x01234567);
314 a7 = (const int8_t*) ((uintptr_t) a7 - kc);
315 c7 = (float*) ((uintptr_t) c7 + cn_stride);
316 nc -= 8;
317 } else {
318 __m128 vout0x0123 = _mm256_castps256_ps128(vout0x01234567);
319 __m128 vout1x0123 = _mm256_castps256_ps128(vout1x01234567);
320 __m128 vout2x0123 = _mm256_castps256_ps128(vout2x01234567);
321 __m128 vout3x0123 = _mm256_castps256_ps128(vout3x01234567);
322 __m128 vout4x0123 = _mm256_castps256_ps128(vout4x01234567);
323 __m128 vout5x0123 = _mm256_castps256_ps128(vout5x01234567);
324 __m128 vout6x0123 = _mm256_castps256_ps128(vout6x01234567);
325 __m128 vout7x0123 = _mm256_castps256_ps128(vout7x01234567);
326 if (nc & 4) {
327 _mm_storeu_ps(c0, vout0x0123);
328 c0 += 4;
329 _mm_storeu_ps(c1, vout1x0123);
330 c1 += 4;
331 _mm_storeu_ps(c2, vout2x0123);
332 c2 += 4;
333 _mm_storeu_ps(c3, vout3x0123);
334 c3 += 4;
335 _mm_storeu_ps(c4, vout4x0123);
336 c4 += 4;
337 _mm_storeu_ps(c5, vout5x0123);
338 c5 += 4;
339 _mm_storeu_ps(c6, vout6x0123);
340 c6 += 4;
341 _mm_storeu_ps(c7, vout7x0123);
342 c7 += 4;
343 vout0x0123 = _mm256_extractf128_ps(vout0x01234567, 1);
344 vout1x0123 = _mm256_extractf128_ps(vout1x01234567, 1);
345 vout2x0123 = _mm256_extractf128_ps(vout2x01234567, 1);
346 vout3x0123 = _mm256_extractf128_ps(vout3x01234567, 1);
347 vout4x0123 = _mm256_extractf128_ps(vout4x01234567, 1);
348 vout5x0123 = _mm256_extractf128_ps(vout5x01234567, 1);
349 vout6x0123 = _mm256_extractf128_ps(vout6x01234567, 1);
350 vout7x0123 = _mm256_extractf128_ps(vout7x01234567, 1);
351 }
352 if (nc & 2) {
353 _mm_storel_pi((__m64*) c0, vout0x0123);
354 c0 += 2;
355 _mm_storel_pi((__m64*) c1, vout1x0123);
356 c1 += 2;
357 _mm_storel_pi((__m64*) c2, vout2x0123);
358 c2 += 2;
359 _mm_storel_pi((__m64*) c3, vout3x0123);
360 c3 += 2;
361 _mm_storel_pi((__m64*) c4, vout4x0123);
362 c4 += 2;
363 _mm_storel_pi((__m64*) c5, vout5x0123);
364 c5 += 2;
365 _mm_storel_pi((__m64*) c6, vout6x0123);
366 c6 += 2;
367 _mm_storel_pi((__m64*) c7, vout7x0123);
368 c7 += 2;
369 vout0x0123 = _mm_movehl_ps(vout0x0123, vout0x0123);
370 vout1x0123 = _mm_movehl_ps(vout1x0123, vout1x0123);
371 vout2x0123 = _mm_movehl_ps(vout2x0123, vout2x0123);
372 vout3x0123 = _mm_movehl_ps(vout3x0123, vout3x0123);
373 vout4x0123 = _mm_movehl_ps(vout4x0123, vout4x0123);
374 vout5x0123 = _mm_movehl_ps(vout5x0123, vout5x0123);
375 vout6x0123 = _mm_movehl_ps(vout6x0123, vout6x0123);
376 vout7x0123 = _mm_movehl_ps(vout7x0123, vout7x0123);
377 }
378 if (nc & 1) {
379 _mm_store_ss(c0, vout0x0123);
380 _mm_store_ss(c1, vout1x0123);
381 _mm_store_ss(c2, vout2x0123);
382 _mm_store_ss(c3, vout3x0123);
383 _mm_store_ss(c4, vout4x0123);
384 _mm_store_ss(c5, vout5x0123);
385 _mm_store_ss(c6, vout6x0123);
386 _mm_store_ss(c7, vout7x0123);
387 }
388 nc = 0;
389 }
390 } while (nc != 0);
391 }