1 // Auto-generated file. Do not edit!
2 // Template: src/f32-gemm/sse-shuffle.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 <xmmintrin.h>
13
14 #include <xnnpack/gemm.h>
15
16
xnn_f32_gemminc_minmax_ukernel_5x8s4__sse(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_5x8s4__sse(
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 <= 5);
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 const float* a4 = (const float*) ((uintptr_t) a3 + a_stride);
61 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
62 if XNN_UNPREDICTABLE(mr <= 4) {
63 a4 = a3;
64 c4 = c3;
65 }
66
67 do {
68 __m128 vacc0x0123 = _mm_load_ps(acc + 0);
69 __m128 vacc0x4567 = _mm_load_ps(acc + 4);
70 __m128 vacc1x0123 = _mm_load_ps(acc + 8);
71 __m128 vacc1x4567 = _mm_load_ps(acc + 12);
72 __m128 vacc2x0123 = _mm_load_ps(acc + 16);
73 __m128 vacc2x4567 = _mm_load_ps(acc + 20);
74 __m128 vacc3x0123 = _mm_load_ps(acc + 24);
75 __m128 vacc3x4567 = _mm_load_ps(acc + 28);
76 __m128 vacc4x0123 = _mm_load_ps(acc + 32);
77 __m128 vacc4x4567 = _mm_load_ps(acc + 36);
78 acc += 40;
79
80 size_t k = kc;
81 while (k >= 4 * sizeof(float)) {
82 __m128 va0 = _mm_loadu_ps(a0);
83 a0 += 4;
84 __m128 va1 = _mm_loadu_ps(a1);
85 a1 += 4;
86 __m128 va2 = _mm_loadu_ps(a2);
87 a2 += 4;
88 __m128 va3 = _mm_loadu_ps(a3);
89 a3 += 4;
90 __m128 va4 = _mm_loadu_ps(a4);
91 a4 += 4;
92
93
94 const __m128 vb0123c0 = _mm_load_ps(w + 0);
95 const __m128 vb4567c0 = _mm_load_ps(w + 4);
96
97 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c0));
98 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123c0));
99 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123c0));
100 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3, vb0123c0));
101 vacc4x0123 = _mm_add_ps(vacc4x0123, _mm_mul_ps(va4, vb0123c0));
102 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567c0));
103 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1, vb4567c0));
104 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2, vb4567c0));
105 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3, vb4567c0));
106 vacc4x4567 = _mm_add_ps(vacc4x4567, _mm_mul_ps(va4, vb4567c0));
107
108 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
109 va1 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 3, 2, 1));
110 va2 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 3, 2, 1));
111 va3 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(0, 3, 2, 1));
112 va4 = _mm_shuffle_ps(va4, va4, _MM_SHUFFLE(0, 3, 2, 1));
113
114 const __m128 vb0123c1 = _mm_load_ps(w + 8);
115 const __m128 vb4567c1 = _mm_load_ps(w + 12);
116
117 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c1));
118 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123c1));
119 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123c1));
120 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3, vb0123c1));
121 vacc4x0123 = _mm_add_ps(vacc4x0123, _mm_mul_ps(va4, vb0123c1));
122 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567c1));
123 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1, vb4567c1));
124 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2, vb4567c1));
125 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3, vb4567c1));
126 vacc4x4567 = _mm_add_ps(vacc4x4567, _mm_mul_ps(va4, vb4567c1));
127
128 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
129 va1 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 3, 2, 1));
130 va2 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 3, 2, 1));
131 va3 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(0, 3, 2, 1));
132 va4 = _mm_shuffle_ps(va4, va4, _MM_SHUFFLE(0, 3, 2, 1));
133
134 const __m128 vb0123c2 = _mm_load_ps(w + 16);
135 const __m128 vb4567c2 = _mm_load_ps(w + 20);
136
137 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c2));
138 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123c2));
139 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123c2));
140 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3, vb0123c2));
141 vacc4x0123 = _mm_add_ps(vacc4x0123, _mm_mul_ps(va4, vb0123c2));
142 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567c2));
143 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1, vb4567c2));
144 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2, vb4567c2));
145 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3, vb4567c2));
146 vacc4x4567 = _mm_add_ps(vacc4x4567, _mm_mul_ps(va4, vb4567c2));
147
148 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
149 va1 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 3, 2, 1));
150 va2 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 3, 2, 1));
151 va3 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(0, 3, 2, 1));
152 va4 = _mm_shuffle_ps(va4, va4, _MM_SHUFFLE(0, 3, 2, 1));
153
154 const __m128 vb0123c3 = _mm_load_ps(w + 24);
155 const __m128 vb4567c3 = _mm_load_ps(w + 28);
156
157 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c3));
158 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123c3));
159 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123c3));
160 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3, vb0123c3));
161 vacc4x0123 = _mm_add_ps(vacc4x0123, _mm_mul_ps(va4, vb0123c3));
162 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567c3));
163 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1, vb4567c3));
164 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2, vb4567c3));
165 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3, vb4567c3));
166 vacc4x4567 = _mm_add_ps(vacc4x4567, _mm_mul_ps(va4, vb4567c3));
167
168
169 w += 32;
170 k -= 4 * sizeof(float);
171 }
172 if XNN_UNLIKELY(k != 0) {
173 __m128 va0 = _mm_loadu_ps(a0);
174 a0 = (const float*) ((uintptr_t) a0 + k);
175 __m128 va1 = _mm_loadu_ps(a1);
176 a1 = (const float*) ((uintptr_t) a1 + k);
177 __m128 va2 = _mm_loadu_ps(a2);
178 a2 = (const float*) ((uintptr_t) a2 + k);
179 __m128 va3 = _mm_loadu_ps(a3);
180 a3 = (const float*) ((uintptr_t) a3 + k);
181 __m128 va4 = _mm_loadu_ps(a4);
182 a4 = (const float*) ((uintptr_t) a4 + k);
183
184
185 const __m128 vb0123c0 = _mm_load_ps(w + 0);
186 const __m128 vb4567c0 = _mm_load_ps(w + 4);
187
188 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c0), va0), vb0123c0));
189 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c0), va1), vb0123c0));
190 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c0), va2), vb0123c0));
191 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c0), va3), vb0123c0));
192 vacc4x0123 = _mm_add_ps(vacc4x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c0), va4), vb0123c0));
193 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c0), va0), vb4567c0));
194 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c0), va1), vb4567c0));
195 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c0), va2), vb4567c0));
196 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c0), va3), vb4567c0));
197 vacc4x4567 = _mm_add_ps(vacc4x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c0), va4), vb4567c0));
198
199 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
200 va1 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 3, 2, 1));
201 va2 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 3, 2, 1));
202 va3 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(0, 3, 2, 1));
203 va4 = _mm_shuffle_ps(va4, va4, _MM_SHUFFLE(0, 3, 2, 1));
204
205 const __m128 vb0123c1 = _mm_load_ps(w + 8);
206 const __m128 vb4567c1 = _mm_load_ps(w + 12);
207
208 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c1), va0), vb0123c1));
209 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c1), va1), vb0123c1));
210 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c1), va2), vb0123c1));
211 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c1), va3), vb0123c1));
212 vacc4x0123 = _mm_add_ps(vacc4x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c1), va4), vb0123c1));
213 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c1), va0), vb4567c1));
214 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c1), va1), vb4567c1));
215 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c1), va2), vb4567c1));
216 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c1), va3), vb4567c1));
217 vacc4x4567 = _mm_add_ps(vacc4x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c1), va4), vb4567c1));
218
219 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
220 va1 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 3, 2, 1));
221 va2 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 3, 2, 1));
222 va3 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(0, 3, 2, 1));
223 va4 = _mm_shuffle_ps(va4, va4, _MM_SHUFFLE(0, 3, 2, 1));
224
225 const __m128 vb0123c2 = _mm_load_ps(w + 16);
226 const __m128 vb4567c2 = _mm_load_ps(w + 20);
227
228 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c2), va0), vb0123c2));
229 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c2), va1), vb0123c2));
230 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c2), va2), vb0123c2));
231 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c2), va3), vb0123c2));
232 vacc4x0123 = _mm_add_ps(vacc4x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c2), va4), vb0123c2));
233 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c2), va0), vb4567c2));
234 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c2), va1), vb4567c2));
235 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c2), va2), vb4567c2));
236 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c2), va3), vb4567c2));
237 vacc4x4567 = _mm_add_ps(vacc4x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c2), va4), vb4567c2));
238
239 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
240 va1 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 3, 2, 1));
241 va2 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 3, 2, 1));
242 va3 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(0, 3, 2, 1));
243 va4 = _mm_shuffle_ps(va4, va4, _MM_SHUFFLE(0, 3, 2, 1));
244
245 const __m128 vb0123c3 = _mm_load_ps(w + 24);
246 const __m128 vb4567c3 = _mm_load_ps(w + 28);
247
248 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c3), va0), vb0123c3));
249 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c3), va1), vb0123c3));
250 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c3), va2), vb0123c3));
251 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c3), va3), vb0123c3));
252 vacc4x0123 = _mm_add_ps(vacc4x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c3), va4), vb0123c3));
253 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c3), va0), vb4567c3));
254 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c3), va1), vb4567c3));
255 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c3), va2), vb4567c3));
256 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c3), va3), vb4567c3));
257 vacc4x4567 = _mm_add_ps(vacc4x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c3), va4), vb4567c3));
258
259
260 w += 32;
261 }
262
263 const __m128 vmax = _mm_load_ps(params->sse.max);
264 vacc0x0123 = _mm_min_ps(vacc0x0123, vmax);
265 vacc1x0123 = _mm_min_ps(vacc1x0123, vmax);
266 vacc2x0123 = _mm_min_ps(vacc2x0123, vmax);
267 vacc3x0123 = _mm_min_ps(vacc3x0123, vmax);
268 vacc4x0123 = _mm_min_ps(vacc4x0123, vmax);
269 vacc0x4567 = _mm_min_ps(vacc0x4567, vmax);
270 vacc1x4567 = _mm_min_ps(vacc1x4567, vmax);
271 vacc2x4567 = _mm_min_ps(vacc2x4567, vmax);
272 vacc3x4567 = _mm_min_ps(vacc3x4567, vmax);
273 vacc4x4567 = _mm_min_ps(vacc4x4567, vmax);
274
275 const __m128 vmin = _mm_load_ps(params->sse.min);
276 vacc0x0123 = _mm_max_ps(vacc0x0123, vmin);
277 vacc1x0123 = _mm_max_ps(vacc1x0123, vmin);
278 vacc2x0123 = _mm_max_ps(vacc2x0123, vmin);
279 vacc3x0123 = _mm_max_ps(vacc3x0123, vmin);
280 vacc4x0123 = _mm_max_ps(vacc4x0123, vmin);
281 vacc0x4567 = _mm_max_ps(vacc0x4567, vmin);
282 vacc1x4567 = _mm_max_ps(vacc1x4567, vmin);
283 vacc2x4567 = _mm_max_ps(vacc2x4567, vmin);
284 vacc3x4567 = _mm_max_ps(vacc3x4567, vmin);
285 vacc4x4567 = _mm_max_ps(vacc4x4567, vmin);
286
287 if XNN_LIKELY(nc >= 8) {
288 _mm_storeu_ps(c4, vacc4x0123);
289 _mm_storeu_ps(c4 + 4, vacc4x4567);
290 c4 = (float*) ((uintptr_t) c4 + cn_stride);
291 _mm_storeu_ps(c3, vacc3x0123);
292 _mm_storeu_ps(c3 + 4, vacc3x4567);
293 c3 = (float*) ((uintptr_t) c3 + cn_stride);
294 _mm_storeu_ps(c2, vacc2x0123);
295 _mm_storeu_ps(c2 + 4, vacc2x4567);
296 c2 = (float*) ((uintptr_t) c2 + cn_stride);
297 _mm_storeu_ps(c1, vacc1x0123);
298 _mm_storeu_ps(c1 + 4, vacc1x4567);
299 c1 = (float*) ((uintptr_t) c1 + cn_stride);
300 _mm_storeu_ps(c0, vacc0x0123);
301 _mm_storeu_ps(c0 + 4, vacc0x4567);
302 c0 = (float*) ((uintptr_t) c0 + cn_stride);
303
304 a4 = (const float*) ((uintptr_t) a4 - kc);
305 a3 = (const float*) ((uintptr_t) a3 - kc);
306 a2 = (const float*) ((uintptr_t) a2 - kc);
307 a1 = (const float*) ((uintptr_t) a1 - kc);
308 a0 = (const float*) ((uintptr_t) a0 - kc);
309
310 nc -= 8;
311 } else {
312 if (nc & 4) {
313 _mm_storeu_ps(c4, vacc4x0123);
314 _mm_storeu_ps(c3, vacc3x0123);
315 _mm_storeu_ps(c2, vacc2x0123);
316 _mm_storeu_ps(c1, vacc1x0123);
317 _mm_storeu_ps(c0, vacc0x0123);
318
319 vacc4x0123 = vacc4x4567;
320 vacc3x0123 = vacc3x4567;
321 vacc2x0123 = vacc2x4567;
322 vacc1x0123 = vacc1x4567;
323 vacc0x0123 = vacc0x4567;
324
325 c4 += 4;
326 c3 += 4;
327 c2 += 4;
328 c1 += 4;
329 c0 += 4;
330 }
331 if (nc & 2) {
332 _mm_storel_pi((__m64*) c4, vacc4x0123);
333 _mm_storel_pi((__m64*) c3, vacc3x0123);
334 _mm_storel_pi((__m64*) c2, vacc2x0123);
335 _mm_storel_pi((__m64*) c1, vacc1x0123);
336 _mm_storel_pi((__m64*) c0, vacc0x0123);
337
338 vacc4x0123 = _mm_movehl_ps(vacc4x0123, vacc4x0123);
339 vacc3x0123 = _mm_movehl_ps(vacc3x0123, vacc3x0123);
340 vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123);
341 vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
342 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
343
344 c4 += 2;
345 c3 += 2;
346 c2 += 2;
347 c1 += 2;
348 c0 += 2;
349 }
350 if (nc & 1) {
351 _mm_store_ss(c4, vacc4x0123);
352 _mm_store_ss(c3, vacc3x0123);
353 _mm_store_ss(c2, vacc2x0123);
354 _mm_store_ss(c1, vacc1x0123);
355 _mm_store_ss(c0, vacc0x0123);
356 }
357
358 nc = 0;
359 }
360 } while (nc != 0);
361 }
362