1 // Auto-generated file. Do not edit!
2 // Template: src/f32-gemm/avx-shuffle4.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2019 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
16
xnn_f32_gemminc_minmax_ukernel_4x16s4__fma3_broadcast(size_t mr,size_t nc,size_t kc,const float * restrict a,size_t a_stride,const float * restrict w,float * restrict c,size_t cm_stride,size_t cn_stride,const float * restrict acc,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_gemminc_minmax_ukernel_4x16s4__fma3_broadcast(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 const float*restrict a,
22 size_t a_stride,
23 const float*restrict w,
24 float*restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 const float*restrict acc,
28 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(mr != 0);
31 assert(mr <= 4);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(float) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38 assert(acc != NULL);
39
40 const float* a0 = a;
41 float* c0 = c;
42 const float* a1 = (const float*) ((uintptr_t) a0 + a_stride);
43 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
44 if XNN_UNPREDICTABLE(mr < 2) {
45 a1 = a0;
46 c1 = c0;
47 }
48 const float* a2 = (const float*) ((uintptr_t) a1 + a_stride);
49 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 a2 = a1;
52 c2 = c1;
53 }
54 const float* a3 = (const float*) ((uintptr_t) a2 + a_stride);
55 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr != 4) {
57 a3 = a2;
58 c3 = c2;
59 }
60
61 do {
62 __m256 vacc0x01234567 = _mm256_load_ps(acc + 0);
63 __m256 vacc0x89ABCDEF = _mm256_load_ps(acc + 8);
64 __m256 vacc1x01234567 = _mm256_load_ps(acc + 16);
65 __m256 vacc1x89ABCDEF = _mm256_load_ps(acc + 24);
66 __m256 vacc2x01234567 = _mm256_load_ps(acc + 32);
67 __m256 vacc2x89ABCDEF = _mm256_load_ps(acc + 40);
68 __m256 vacc3x01234567 = _mm256_load_ps(acc + 48);
69 __m256 vacc3x89ABCDEF = _mm256_load_ps(acc + 56);
70 acc += 64;
71
72 size_t k = kc;
73 while (k >= 4 * sizeof(float)) {
74 __m256 va0 = _mm256_broadcast_ps((const __m128*) a0);
75 a0 += 4;
76 __m256 va1 = _mm256_broadcast_ps((const __m128*) a1);
77 a1 += 4;
78 __m256 va2 = _mm256_broadcast_ps((const __m128*) a2);
79 a2 += 4;
80 __m256 va3 = _mm256_broadcast_ps((const __m128*) a3);
81 a3 += 4;
82
83
84 const __m256 vb01234567c0 = _mm256_load_ps(w + 0);
85 const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8);
86
87 vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c0, vacc0x01234567);
88 vacc1x01234567 = _mm256_fmadd_ps(va1, vb01234567c0, vacc1x01234567);
89 vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567c0, vacc2x01234567);
90 vacc3x01234567 = _mm256_fmadd_ps(va3, vb01234567c0, vacc3x01234567);
91 vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc0, vacc0x89ABCDEF);
92 vacc1x89ABCDEF = _mm256_fmadd_ps(va1, vb89ABCDEFc0, vacc1x89ABCDEF);
93 vacc2x89ABCDEF = _mm256_fmadd_ps(va2, vb89ABCDEFc0, vacc2x89ABCDEF);
94 vacc3x89ABCDEF = _mm256_fmadd_ps(va3, vb89ABCDEFc0, vacc3x89ABCDEF);
95
96 va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
97 va1 = _mm256_permute_ps(va1, _MM_SHUFFLE(0, 3, 2, 1));
98 va2 = _mm256_permute_ps(va2, _MM_SHUFFLE(0, 3, 2, 1));
99 va3 = _mm256_permute_ps(va3, _MM_SHUFFLE(0, 3, 2, 1));
100
101 const __m256 vb01234567c1 = _mm256_load_ps(w + 16);
102 const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24);
103
104 vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c1, vacc0x01234567);
105 vacc1x01234567 = _mm256_fmadd_ps(va1, vb01234567c1, vacc1x01234567);
106 vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567c1, vacc2x01234567);
107 vacc3x01234567 = _mm256_fmadd_ps(va3, vb01234567c1, vacc3x01234567);
108 vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc1, vacc0x89ABCDEF);
109 vacc1x89ABCDEF = _mm256_fmadd_ps(va1, vb89ABCDEFc1, vacc1x89ABCDEF);
110 vacc2x89ABCDEF = _mm256_fmadd_ps(va2, vb89ABCDEFc1, vacc2x89ABCDEF);
111 vacc3x89ABCDEF = _mm256_fmadd_ps(va3, vb89ABCDEFc1, vacc3x89ABCDEF);
112
113 va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
114 va1 = _mm256_permute_ps(va1, _MM_SHUFFLE(0, 3, 2, 1));
115 va2 = _mm256_permute_ps(va2, _MM_SHUFFLE(0, 3, 2, 1));
116 va3 = _mm256_permute_ps(va3, _MM_SHUFFLE(0, 3, 2, 1));
117
118 const __m256 vb01234567c2 = _mm256_load_ps(w + 32);
119 const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40);
120
121 vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c2, vacc0x01234567);
122 vacc1x01234567 = _mm256_fmadd_ps(va1, vb01234567c2, vacc1x01234567);
123 vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567c2, vacc2x01234567);
124 vacc3x01234567 = _mm256_fmadd_ps(va3, vb01234567c2, vacc3x01234567);
125 vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc2, vacc0x89ABCDEF);
126 vacc1x89ABCDEF = _mm256_fmadd_ps(va1, vb89ABCDEFc2, vacc1x89ABCDEF);
127 vacc2x89ABCDEF = _mm256_fmadd_ps(va2, vb89ABCDEFc2, vacc2x89ABCDEF);
128 vacc3x89ABCDEF = _mm256_fmadd_ps(va3, vb89ABCDEFc2, vacc3x89ABCDEF);
129
130 va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
131 va1 = _mm256_permute_ps(va1, _MM_SHUFFLE(0, 3, 2, 1));
132 va2 = _mm256_permute_ps(va2, _MM_SHUFFLE(0, 3, 2, 1));
133 va3 = _mm256_permute_ps(va3, _MM_SHUFFLE(0, 3, 2, 1));
134
135 const __m256 vb01234567c3 = _mm256_load_ps(w + 48);
136 const __m256 vb89ABCDEFc3 = _mm256_load_ps(w + 56);
137
138 vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c3, vacc0x01234567);
139 vacc1x01234567 = _mm256_fmadd_ps(va1, vb01234567c3, vacc1x01234567);
140 vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567c3, vacc2x01234567);
141 vacc3x01234567 = _mm256_fmadd_ps(va3, vb01234567c3, vacc3x01234567);
142 vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc3, vacc0x89ABCDEF);
143 vacc1x89ABCDEF = _mm256_fmadd_ps(va1, vb89ABCDEFc3, vacc1x89ABCDEF);
144 vacc2x89ABCDEF = _mm256_fmadd_ps(va2, vb89ABCDEFc3, vacc2x89ABCDEF);
145 vacc3x89ABCDEF = _mm256_fmadd_ps(va3, vb89ABCDEFc3, vacc3x89ABCDEF);
146
147
148 w += 64;
149 k -= 4 * sizeof(float);
150 }
151 if XNN_UNLIKELY(k != 0) {
152 __m256 va0 = _mm256_broadcast_ps((const __m128*) a0);
153 a0 = (const float*) ((uintptr_t) a0 + k);
154 __m256 va1 = _mm256_broadcast_ps((const __m128*) a1);
155 a1 = (const float*) ((uintptr_t) a1 + k);
156 __m256 va2 = _mm256_broadcast_ps((const __m128*) a2);
157 a2 = (const float*) ((uintptr_t) a2 + k);
158 __m256 va3 = _mm256_broadcast_ps((const __m128*) a3);
159 a3 = (const float*) ((uintptr_t) a3 + k);
160
161 const __m256 vzero = _mm256_setzero_ps();
162
163 const __m256 vb01234567c0 = _mm256_load_ps(w + 0);
164 const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8);
165
166 vacc0x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb01234567c0, vzero, _CMP_NEQ_OQ)), vb01234567c0, vacc0x01234567);
167 vacc1x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb01234567c0, vzero, _CMP_NEQ_OQ)), vb01234567c0, vacc1x01234567);
168 vacc2x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb01234567c0, vzero, _CMP_NEQ_OQ)), vb01234567c0, vacc2x01234567);
169 vacc3x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va3, _mm256_cmp_ps(vb01234567c0, vzero, _CMP_NEQ_OQ)), vb01234567c0, vacc3x01234567);
170 vacc0x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb89ABCDEFc0, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc0, vacc0x89ABCDEF);
171 vacc1x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb89ABCDEFc0, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc0, vacc1x89ABCDEF);
172 vacc2x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb89ABCDEFc0, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc0, vacc2x89ABCDEF);
173 vacc3x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va3, _mm256_cmp_ps(vb89ABCDEFc0, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc0, vacc3x89ABCDEF);
174
175 va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
176 va1 = _mm256_permute_ps(va1, _MM_SHUFFLE(0, 3, 2, 1));
177 va2 = _mm256_permute_ps(va2, _MM_SHUFFLE(0, 3, 2, 1));
178 va3 = _mm256_permute_ps(va3, _MM_SHUFFLE(0, 3, 2, 1));
179
180 const __m256 vb01234567c1 = _mm256_load_ps(w + 16);
181 const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24);
182
183 vacc0x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb01234567c1, vzero, _CMP_NEQ_OQ)), vb01234567c1, vacc0x01234567);
184 vacc1x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb01234567c1, vzero, _CMP_NEQ_OQ)), vb01234567c1, vacc1x01234567);
185 vacc2x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb01234567c1, vzero, _CMP_NEQ_OQ)), vb01234567c1, vacc2x01234567);
186 vacc3x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va3, _mm256_cmp_ps(vb01234567c1, vzero, _CMP_NEQ_OQ)), vb01234567c1, vacc3x01234567);
187 vacc0x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb89ABCDEFc1, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc1, vacc0x89ABCDEF);
188 vacc1x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb89ABCDEFc1, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc1, vacc1x89ABCDEF);
189 vacc2x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb89ABCDEFc1, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc1, vacc2x89ABCDEF);
190 vacc3x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va3, _mm256_cmp_ps(vb89ABCDEFc1, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc1, vacc3x89ABCDEF);
191
192 va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
193 va1 = _mm256_permute_ps(va1, _MM_SHUFFLE(0, 3, 2, 1));
194 va2 = _mm256_permute_ps(va2, _MM_SHUFFLE(0, 3, 2, 1));
195 va3 = _mm256_permute_ps(va3, _MM_SHUFFLE(0, 3, 2, 1));
196
197 const __m256 vb01234567c2 = _mm256_load_ps(w + 32);
198 const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40);
199
200 vacc0x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb01234567c2, vzero, _CMP_NEQ_OQ)), vb01234567c2, vacc0x01234567);
201 vacc1x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb01234567c2, vzero, _CMP_NEQ_OQ)), vb01234567c2, vacc1x01234567);
202 vacc2x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb01234567c2, vzero, _CMP_NEQ_OQ)), vb01234567c2, vacc2x01234567);
203 vacc3x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va3, _mm256_cmp_ps(vb01234567c2, vzero, _CMP_NEQ_OQ)), vb01234567c2, vacc3x01234567);
204 vacc0x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb89ABCDEFc2, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc2, vacc0x89ABCDEF);
205 vacc1x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb89ABCDEFc2, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc2, vacc1x89ABCDEF);
206 vacc2x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb89ABCDEFc2, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc2, vacc2x89ABCDEF);
207 vacc3x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va3, _mm256_cmp_ps(vb89ABCDEFc2, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc2, vacc3x89ABCDEF);
208
209 va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
210 va1 = _mm256_permute_ps(va1, _MM_SHUFFLE(0, 3, 2, 1));
211 va2 = _mm256_permute_ps(va2, _MM_SHUFFLE(0, 3, 2, 1));
212 va3 = _mm256_permute_ps(va3, _MM_SHUFFLE(0, 3, 2, 1));
213
214 const __m256 vb01234567c3 = _mm256_load_ps(w + 48);
215 const __m256 vb89ABCDEFc3 = _mm256_load_ps(w + 56);
216
217 vacc0x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb01234567c3, vzero, _CMP_NEQ_OQ)), vb01234567c3, vacc0x01234567);
218 vacc1x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb01234567c3, vzero, _CMP_NEQ_OQ)), vb01234567c3, vacc1x01234567);
219 vacc2x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb01234567c3, vzero, _CMP_NEQ_OQ)), vb01234567c3, vacc2x01234567);
220 vacc3x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va3, _mm256_cmp_ps(vb01234567c3, vzero, _CMP_NEQ_OQ)), vb01234567c3, vacc3x01234567);
221 vacc0x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb89ABCDEFc3, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc3, vacc0x89ABCDEF);
222 vacc1x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb89ABCDEFc3, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc3, vacc1x89ABCDEF);
223 vacc2x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb89ABCDEFc3, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc3, vacc2x89ABCDEF);
224 vacc3x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va3, _mm256_cmp_ps(vb89ABCDEFc3, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc3, vacc3x89ABCDEF);
225
226
227 w += 64;
228 }
229
230 const __m256 vmin = _mm256_load_ps(params->avx.min);
231 vacc0x01234567 = _mm256_max_ps(vacc0x01234567, vmin);
232 vacc1x01234567 = _mm256_max_ps(vacc1x01234567, vmin);
233 vacc2x01234567 = _mm256_max_ps(vacc2x01234567, vmin);
234 vacc3x01234567 = _mm256_max_ps(vacc3x01234567, vmin);
235 vacc0x89ABCDEF = _mm256_max_ps(vacc0x89ABCDEF, vmin);
236 vacc1x89ABCDEF = _mm256_max_ps(vacc1x89ABCDEF, vmin);
237 vacc2x89ABCDEF = _mm256_max_ps(vacc2x89ABCDEF, vmin);
238 vacc3x89ABCDEF = _mm256_max_ps(vacc3x89ABCDEF, vmin);
239
240 const __m256 vmax = _mm256_load_ps(params->avx.max);
241 vacc0x01234567 = _mm256_min_ps(vacc0x01234567, vmax);
242 vacc1x01234567 = _mm256_min_ps(vacc1x01234567, vmax);
243 vacc2x01234567 = _mm256_min_ps(vacc2x01234567, vmax);
244 vacc3x01234567 = _mm256_min_ps(vacc3x01234567, vmax);
245 vacc0x89ABCDEF = _mm256_min_ps(vacc0x89ABCDEF, vmax);
246 vacc1x89ABCDEF = _mm256_min_ps(vacc1x89ABCDEF, vmax);
247 vacc2x89ABCDEF = _mm256_min_ps(vacc2x89ABCDEF, vmax);
248 vacc3x89ABCDEF = _mm256_min_ps(vacc3x89ABCDEF, vmax);
249
250 if XNN_LIKELY(nc >= 16) {
251 _mm256_storeu_ps(c3, vacc3x01234567);
252 _mm256_storeu_ps(c3 + 8, vacc3x89ABCDEF);
253 c3 = (float*) ((uintptr_t) c3 + cn_stride);
254 _mm256_storeu_ps(c2, vacc2x01234567);
255 _mm256_storeu_ps(c2 + 8, vacc2x89ABCDEF);
256 c2 = (float*) ((uintptr_t) c2 + cn_stride);
257 _mm256_storeu_ps(c1, vacc1x01234567);
258 _mm256_storeu_ps(c1 + 8, vacc1x89ABCDEF);
259 c1 = (float*) ((uintptr_t) c1 + cn_stride);
260 _mm256_storeu_ps(c0, vacc0x01234567);
261 _mm256_storeu_ps(c0 + 8, vacc0x89ABCDEF);
262 c0 = (float*) ((uintptr_t) c0 + cn_stride);
263
264 a3 = (const float*) ((uintptr_t) a3 - kc);
265 a2 = (const float*) ((uintptr_t) a2 - kc);
266 a1 = (const float*) ((uintptr_t) a1 - kc);
267 a0 = (const float*) ((uintptr_t) a0 - kc);
268
269 nc -= 16;
270 } else {
271 if (nc & 8) {
272 _mm256_storeu_ps(c3, vacc3x01234567);
273 _mm256_storeu_ps(c2, vacc2x01234567);
274 _mm256_storeu_ps(c1, vacc1x01234567);
275 _mm256_storeu_ps(c0, vacc0x01234567);
276
277 vacc3x01234567 = vacc3x89ABCDEF;
278 vacc2x01234567 = vacc2x89ABCDEF;
279 vacc1x01234567 = vacc1x89ABCDEF;
280 vacc0x01234567 = vacc0x89ABCDEF;
281
282 c3 += 8;
283 c2 += 8;
284 c1 += 8;
285 c0 += 8;
286 }
287 __m128 vacc3x0123 = _mm256_castps256_ps128(vacc3x01234567);
288 __m128 vacc2x0123 = _mm256_castps256_ps128(vacc2x01234567);
289 __m128 vacc1x0123 = _mm256_castps256_ps128(vacc1x01234567);
290 __m128 vacc0x0123 = _mm256_castps256_ps128(vacc0x01234567);
291 if (nc & 4) {
292 _mm_storeu_ps(c3, vacc3x0123);
293 _mm_storeu_ps(c2, vacc2x0123);
294 _mm_storeu_ps(c1, vacc1x0123);
295 _mm_storeu_ps(c0, vacc0x0123);
296
297 vacc3x0123 = _mm256_extractf128_ps(vacc3x01234567, 1);
298 vacc2x0123 = _mm256_extractf128_ps(vacc2x01234567, 1);
299 vacc1x0123 = _mm256_extractf128_ps(vacc1x01234567, 1);
300 vacc0x0123 = _mm256_extractf128_ps(vacc0x01234567, 1);
301
302 c3 += 4;
303 c2 += 4;
304 c1 += 4;
305 c0 += 4;
306 }
307 if (nc & 2) {
308 _mm_storel_pi((__m64*) c3, vacc3x0123);
309 _mm_storel_pi((__m64*) c2, vacc2x0123);
310 _mm_storel_pi((__m64*) c1, vacc1x0123);
311 _mm_storel_pi((__m64*) c0, vacc0x0123);
312
313 vacc3x0123 = _mm_movehl_ps(vacc3x0123, vacc3x0123);
314 vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123);
315 vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
316 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
317
318 c3 += 2;
319 c2 += 2;
320 c1 += 2;
321 c0 += 2;
322 }
323 if (nc & 1) {
324 _mm_store_ss(c3, vacc3x0123);
325 _mm_store_ss(c2, vacc2x0123);
326 _mm_store_ss(c1, vacc1x0123);
327 _mm_store_ss(c0, vacc0x0123);
328 }
329
330 nc = 0;
331 }
332 } while (nc != 0);
333 }
334