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_5x16s4__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_5x16s4__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 <= 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 __m256 vacc0x01234567 = _mm256_load_ps(acc + 0);
69 __m256 vacc0x89ABCDEF = _mm256_load_ps(acc + 8);
70 __m256 vacc1x01234567 = _mm256_load_ps(acc + 16);
71 __m256 vacc1x89ABCDEF = _mm256_load_ps(acc + 24);
72 __m256 vacc2x01234567 = _mm256_load_ps(acc + 32);
73 __m256 vacc2x89ABCDEF = _mm256_load_ps(acc + 40);
74 __m256 vacc3x01234567 = _mm256_load_ps(acc + 48);
75 __m256 vacc3x89ABCDEF = _mm256_load_ps(acc + 56);
76 __m256 vacc4x01234567 = _mm256_load_ps(acc + 64);
77 __m256 vacc4x89ABCDEF = _mm256_load_ps(acc + 72);
78 acc += 80;
79
80 size_t k = kc;
81 while (k >= 4 * sizeof(float)) {
82 __m256 va0 = _mm256_broadcast_ps((const __m128*) a0);
83 a0 += 4;
84 __m256 va1 = _mm256_broadcast_ps((const __m128*) a1);
85 a1 += 4;
86 __m256 va2 = _mm256_broadcast_ps((const __m128*) a2);
87 a2 += 4;
88 __m256 va3 = _mm256_broadcast_ps((const __m128*) a3);
89 a3 += 4;
90 __m256 va4 = _mm256_broadcast_ps((const __m128*) a4);
91 a4 += 4;
92
93
94 const __m256 vb01234567c0 = _mm256_load_ps(w + 0);
95 const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8);
96
97 vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c0, vacc0x01234567);
98 vacc1x01234567 = _mm256_fmadd_ps(va1, vb01234567c0, vacc1x01234567);
99 vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567c0, vacc2x01234567);
100 vacc3x01234567 = _mm256_fmadd_ps(va3, vb01234567c0, vacc3x01234567);
101 vacc4x01234567 = _mm256_fmadd_ps(va4, vb01234567c0, vacc4x01234567);
102 vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc0, vacc0x89ABCDEF);
103 vacc1x89ABCDEF = _mm256_fmadd_ps(va1, vb89ABCDEFc0, vacc1x89ABCDEF);
104 vacc2x89ABCDEF = _mm256_fmadd_ps(va2, vb89ABCDEFc0, vacc2x89ABCDEF);
105 vacc3x89ABCDEF = _mm256_fmadd_ps(va3, vb89ABCDEFc0, vacc3x89ABCDEF);
106 vacc4x89ABCDEF = _mm256_fmadd_ps(va4, vb89ABCDEFc0, vacc4x89ABCDEF);
107
108 va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
109 va1 = _mm256_permute_ps(va1, _MM_SHUFFLE(0, 3, 2, 1));
110 va2 = _mm256_permute_ps(va2, _MM_SHUFFLE(0, 3, 2, 1));
111 va3 = _mm256_permute_ps(va3, _MM_SHUFFLE(0, 3, 2, 1));
112 va4 = _mm256_permute_ps(va4, _MM_SHUFFLE(0, 3, 2, 1));
113
114 const __m256 vb01234567c1 = _mm256_load_ps(w + 16);
115 const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24);
116
117 vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c1, vacc0x01234567);
118 vacc1x01234567 = _mm256_fmadd_ps(va1, vb01234567c1, vacc1x01234567);
119 vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567c1, vacc2x01234567);
120 vacc3x01234567 = _mm256_fmadd_ps(va3, vb01234567c1, vacc3x01234567);
121 vacc4x01234567 = _mm256_fmadd_ps(va4, vb01234567c1, vacc4x01234567);
122 vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc1, vacc0x89ABCDEF);
123 vacc1x89ABCDEF = _mm256_fmadd_ps(va1, vb89ABCDEFc1, vacc1x89ABCDEF);
124 vacc2x89ABCDEF = _mm256_fmadd_ps(va2, vb89ABCDEFc1, vacc2x89ABCDEF);
125 vacc3x89ABCDEF = _mm256_fmadd_ps(va3, vb89ABCDEFc1, vacc3x89ABCDEF);
126 vacc4x89ABCDEF = _mm256_fmadd_ps(va4, vb89ABCDEFc1, vacc4x89ABCDEF);
127
128 va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
129 va1 = _mm256_permute_ps(va1, _MM_SHUFFLE(0, 3, 2, 1));
130 va2 = _mm256_permute_ps(va2, _MM_SHUFFLE(0, 3, 2, 1));
131 va3 = _mm256_permute_ps(va3, _MM_SHUFFLE(0, 3, 2, 1));
132 va4 = _mm256_permute_ps(va4, _MM_SHUFFLE(0, 3, 2, 1));
133
134 const __m256 vb01234567c2 = _mm256_load_ps(w + 32);
135 const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40);
136
137 vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c2, vacc0x01234567);
138 vacc1x01234567 = _mm256_fmadd_ps(va1, vb01234567c2, vacc1x01234567);
139 vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567c2, vacc2x01234567);
140 vacc3x01234567 = _mm256_fmadd_ps(va3, vb01234567c2, vacc3x01234567);
141 vacc4x01234567 = _mm256_fmadd_ps(va4, vb01234567c2, vacc4x01234567);
142 vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc2, vacc0x89ABCDEF);
143 vacc1x89ABCDEF = _mm256_fmadd_ps(va1, vb89ABCDEFc2, vacc1x89ABCDEF);
144 vacc2x89ABCDEF = _mm256_fmadd_ps(va2, vb89ABCDEFc2, vacc2x89ABCDEF);
145 vacc3x89ABCDEF = _mm256_fmadd_ps(va3, vb89ABCDEFc2, vacc3x89ABCDEF);
146 vacc4x89ABCDEF = _mm256_fmadd_ps(va4, vb89ABCDEFc2, vacc4x89ABCDEF);
147
148 va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
149 va1 = _mm256_permute_ps(va1, _MM_SHUFFLE(0, 3, 2, 1));
150 va2 = _mm256_permute_ps(va2, _MM_SHUFFLE(0, 3, 2, 1));
151 va3 = _mm256_permute_ps(va3, _MM_SHUFFLE(0, 3, 2, 1));
152 va4 = _mm256_permute_ps(va4, _MM_SHUFFLE(0, 3, 2, 1));
153
154 const __m256 vb01234567c3 = _mm256_load_ps(w + 48);
155 const __m256 vb89ABCDEFc3 = _mm256_load_ps(w + 56);
156
157 vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c3, vacc0x01234567);
158 vacc1x01234567 = _mm256_fmadd_ps(va1, vb01234567c3, vacc1x01234567);
159 vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567c3, vacc2x01234567);
160 vacc3x01234567 = _mm256_fmadd_ps(va3, vb01234567c3, vacc3x01234567);
161 vacc4x01234567 = _mm256_fmadd_ps(va4, vb01234567c3, vacc4x01234567);
162 vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc3, vacc0x89ABCDEF);
163 vacc1x89ABCDEF = _mm256_fmadd_ps(va1, vb89ABCDEFc3, vacc1x89ABCDEF);
164 vacc2x89ABCDEF = _mm256_fmadd_ps(va2, vb89ABCDEFc3, vacc2x89ABCDEF);
165 vacc3x89ABCDEF = _mm256_fmadd_ps(va3, vb89ABCDEFc3, vacc3x89ABCDEF);
166 vacc4x89ABCDEF = _mm256_fmadd_ps(va4, vb89ABCDEFc3, vacc4x89ABCDEF);
167
168
169 w += 64;
170 k -= 4 * sizeof(float);
171 }
172 if XNN_UNLIKELY(k != 0) {
173 __m256 va0 = _mm256_broadcast_ps((const __m128*) a0);
174 a0 = (const float*) ((uintptr_t) a0 + k);
175 __m256 va1 = _mm256_broadcast_ps((const __m128*) a1);
176 a1 = (const float*) ((uintptr_t) a1 + k);
177 __m256 va2 = _mm256_broadcast_ps((const __m128*) a2);
178 a2 = (const float*) ((uintptr_t) a2 + k);
179 __m256 va3 = _mm256_broadcast_ps((const __m128*) a3);
180 a3 = (const float*) ((uintptr_t) a3 + k);
181 __m256 va4 = _mm256_broadcast_ps((const __m128*) a4);
182 a4 = (const float*) ((uintptr_t) a4 + k);
183
184 const __m256 vzero = _mm256_setzero_ps();
185
186 const __m256 vb01234567c0 = _mm256_load_ps(w + 0);
187 const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8);
188
189 vacc0x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb01234567c0, vzero, _CMP_NEQ_OQ)), vb01234567c0, vacc0x01234567);
190 vacc1x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb01234567c0, vzero, _CMP_NEQ_OQ)), vb01234567c0, vacc1x01234567);
191 vacc2x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb01234567c0, vzero, _CMP_NEQ_OQ)), vb01234567c0, vacc2x01234567);
192 vacc3x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va3, _mm256_cmp_ps(vb01234567c0, vzero, _CMP_NEQ_OQ)), vb01234567c0, vacc3x01234567);
193 vacc4x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va4, _mm256_cmp_ps(vb01234567c0, vzero, _CMP_NEQ_OQ)), vb01234567c0, vacc4x01234567);
194 vacc0x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb89ABCDEFc0, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc0, vacc0x89ABCDEF);
195 vacc1x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb89ABCDEFc0, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc0, vacc1x89ABCDEF);
196 vacc2x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb89ABCDEFc0, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc0, vacc2x89ABCDEF);
197 vacc3x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va3, _mm256_cmp_ps(vb89ABCDEFc0, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc0, vacc3x89ABCDEF);
198 vacc4x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va4, _mm256_cmp_ps(vb89ABCDEFc0, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc0, vacc4x89ABCDEF);
199
200 va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
201 va1 = _mm256_permute_ps(va1, _MM_SHUFFLE(0, 3, 2, 1));
202 va2 = _mm256_permute_ps(va2, _MM_SHUFFLE(0, 3, 2, 1));
203 va3 = _mm256_permute_ps(va3, _MM_SHUFFLE(0, 3, 2, 1));
204 va4 = _mm256_permute_ps(va4, _MM_SHUFFLE(0, 3, 2, 1));
205
206 const __m256 vb01234567c1 = _mm256_load_ps(w + 16);
207 const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24);
208
209 vacc0x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb01234567c1, vzero, _CMP_NEQ_OQ)), vb01234567c1, vacc0x01234567);
210 vacc1x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb01234567c1, vzero, _CMP_NEQ_OQ)), vb01234567c1, vacc1x01234567);
211 vacc2x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb01234567c1, vzero, _CMP_NEQ_OQ)), vb01234567c1, vacc2x01234567);
212 vacc3x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va3, _mm256_cmp_ps(vb01234567c1, vzero, _CMP_NEQ_OQ)), vb01234567c1, vacc3x01234567);
213 vacc4x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va4, _mm256_cmp_ps(vb01234567c1, vzero, _CMP_NEQ_OQ)), vb01234567c1, vacc4x01234567);
214 vacc0x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb89ABCDEFc1, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc1, vacc0x89ABCDEF);
215 vacc1x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb89ABCDEFc1, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc1, vacc1x89ABCDEF);
216 vacc2x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb89ABCDEFc1, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc1, vacc2x89ABCDEF);
217 vacc3x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va3, _mm256_cmp_ps(vb89ABCDEFc1, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc1, vacc3x89ABCDEF);
218 vacc4x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va4, _mm256_cmp_ps(vb89ABCDEFc1, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc1, vacc4x89ABCDEF);
219
220 va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
221 va1 = _mm256_permute_ps(va1, _MM_SHUFFLE(0, 3, 2, 1));
222 va2 = _mm256_permute_ps(va2, _MM_SHUFFLE(0, 3, 2, 1));
223 va3 = _mm256_permute_ps(va3, _MM_SHUFFLE(0, 3, 2, 1));
224 va4 = _mm256_permute_ps(va4, _MM_SHUFFLE(0, 3, 2, 1));
225
226 const __m256 vb01234567c2 = _mm256_load_ps(w + 32);
227 const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40);
228
229 vacc0x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb01234567c2, vzero, _CMP_NEQ_OQ)), vb01234567c2, vacc0x01234567);
230 vacc1x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb01234567c2, vzero, _CMP_NEQ_OQ)), vb01234567c2, vacc1x01234567);
231 vacc2x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb01234567c2, vzero, _CMP_NEQ_OQ)), vb01234567c2, vacc2x01234567);
232 vacc3x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va3, _mm256_cmp_ps(vb01234567c2, vzero, _CMP_NEQ_OQ)), vb01234567c2, vacc3x01234567);
233 vacc4x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va4, _mm256_cmp_ps(vb01234567c2, vzero, _CMP_NEQ_OQ)), vb01234567c2, vacc4x01234567);
234 vacc0x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb89ABCDEFc2, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc2, vacc0x89ABCDEF);
235 vacc1x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb89ABCDEFc2, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc2, vacc1x89ABCDEF);
236 vacc2x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb89ABCDEFc2, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc2, vacc2x89ABCDEF);
237 vacc3x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va3, _mm256_cmp_ps(vb89ABCDEFc2, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc2, vacc3x89ABCDEF);
238 vacc4x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va4, _mm256_cmp_ps(vb89ABCDEFc2, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc2, vacc4x89ABCDEF);
239
240 va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
241 va1 = _mm256_permute_ps(va1, _MM_SHUFFLE(0, 3, 2, 1));
242 va2 = _mm256_permute_ps(va2, _MM_SHUFFLE(0, 3, 2, 1));
243 va3 = _mm256_permute_ps(va3, _MM_SHUFFLE(0, 3, 2, 1));
244 va4 = _mm256_permute_ps(va4, _MM_SHUFFLE(0, 3, 2, 1));
245
246 const __m256 vb01234567c3 = _mm256_load_ps(w + 48);
247 const __m256 vb89ABCDEFc3 = _mm256_load_ps(w + 56);
248
249 vacc0x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb01234567c3, vzero, _CMP_NEQ_OQ)), vb01234567c3, vacc0x01234567);
250 vacc1x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb01234567c3, vzero, _CMP_NEQ_OQ)), vb01234567c3, vacc1x01234567);
251 vacc2x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb01234567c3, vzero, _CMP_NEQ_OQ)), vb01234567c3, vacc2x01234567);
252 vacc3x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va3, _mm256_cmp_ps(vb01234567c3, vzero, _CMP_NEQ_OQ)), vb01234567c3, vacc3x01234567);
253 vacc4x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va4, _mm256_cmp_ps(vb01234567c3, vzero, _CMP_NEQ_OQ)), vb01234567c3, vacc4x01234567);
254 vacc0x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb89ABCDEFc3, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc3, vacc0x89ABCDEF);
255 vacc1x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb89ABCDEFc3, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc3, vacc1x89ABCDEF);
256 vacc2x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb89ABCDEFc3, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc3, vacc2x89ABCDEF);
257 vacc3x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va3, _mm256_cmp_ps(vb89ABCDEFc3, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc3, vacc3x89ABCDEF);
258 vacc4x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va4, _mm256_cmp_ps(vb89ABCDEFc3, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc3, vacc4x89ABCDEF);
259
260
261 w += 64;
262 }
263
264 const __m256 vmin = _mm256_load_ps(params->avx.min);
265 vacc0x01234567 = _mm256_max_ps(vacc0x01234567, vmin);
266 vacc1x01234567 = _mm256_max_ps(vacc1x01234567, vmin);
267 vacc2x01234567 = _mm256_max_ps(vacc2x01234567, vmin);
268 vacc3x01234567 = _mm256_max_ps(vacc3x01234567, vmin);
269 vacc4x01234567 = _mm256_max_ps(vacc4x01234567, vmin);
270 vacc0x89ABCDEF = _mm256_max_ps(vacc0x89ABCDEF, vmin);
271 vacc1x89ABCDEF = _mm256_max_ps(vacc1x89ABCDEF, vmin);
272 vacc2x89ABCDEF = _mm256_max_ps(vacc2x89ABCDEF, vmin);
273 vacc3x89ABCDEF = _mm256_max_ps(vacc3x89ABCDEF, vmin);
274 vacc4x89ABCDEF = _mm256_max_ps(vacc4x89ABCDEF, vmin);
275
276 const __m256 vmax = _mm256_load_ps(params->avx.max);
277 vacc0x01234567 = _mm256_min_ps(vacc0x01234567, vmax);
278 vacc1x01234567 = _mm256_min_ps(vacc1x01234567, vmax);
279 vacc2x01234567 = _mm256_min_ps(vacc2x01234567, vmax);
280 vacc3x01234567 = _mm256_min_ps(vacc3x01234567, vmax);
281 vacc4x01234567 = _mm256_min_ps(vacc4x01234567, vmax);
282 vacc0x89ABCDEF = _mm256_min_ps(vacc0x89ABCDEF, vmax);
283 vacc1x89ABCDEF = _mm256_min_ps(vacc1x89ABCDEF, vmax);
284 vacc2x89ABCDEF = _mm256_min_ps(vacc2x89ABCDEF, vmax);
285 vacc3x89ABCDEF = _mm256_min_ps(vacc3x89ABCDEF, vmax);
286 vacc4x89ABCDEF = _mm256_min_ps(vacc4x89ABCDEF, vmax);
287
288 if XNN_LIKELY(nc >= 16) {
289 _mm256_storeu_ps(c4, vacc4x01234567);
290 _mm256_storeu_ps(c4 + 8, vacc4x89ABCDEF);
291 c4 = (float*) ((uintptr_t) c4 + cn_stride);
292 _mm256_storeu_ps(c3, vacc3x01234567);
293 _mm256_storeu_ps(c3 + 8, vacc3x89ABCDEF);
294 c3 = (float*) ((uintptr_t) c3 + cn_stride);
295 _mm256_storeu_ps(c2, vacc2x01234567);
296 _mm256_storeu_ps(c2 + 8, vacc2x89ABCDEF);
297 c2 = (float*) ((uintptr_t) c2 + cn_stride);
298 _mm256_storeu_ps(c1, vacc1x01234567);
299 _mm256_storeu_ps(c1 + 8, vacc1x89ABCDEF);
300 c1 = (float*) ((uintptr_t) c1 + cn_stride);
301 _mm256_storeu_ps(c0, vacc0x01234567);
302 _mm256_storeu_ps(c0 + 8, vacc0x89ABCDEF);
303 c0 = (float*) ((uintptr_t) c0 + cn_stride);
304
305 a4 = (const float*) ((uintptr_t) a4 - kc);
306 a3 = (const float*) ((uintptr_t) a3 - kc);
307 a2 = (const float*) ((uintptr_t) a2 - kc);
308 a1 = (const float*) ((uintptr_t) a1 - kc);
309 a0 = (const float*) ((uintptr_t) a0 - kc);
310
311 nc -= 16;
312 } else {
313 if (nc & 8) {
314 _mm256_storeu_ps(c4, vacc4x01234567);
315 _mm256_storeu_ps(c3, vacc3x01234567);
316 _mm256_storeu_ps(c2, vacc2x01234567);
317 _mm256_storeu_ps(c1, vacc1x01234567);
318 _mm256_storeu_ps(c0, vacc0x01234567);
319
320 vacc4x01234567 = vacc4x89ABCDEF;
321 vacc3x01234567 = vacc3x89ABCDEF;
322 vacc2x01234567 = vacc2x89ABCDEF;
323 vacc1x01234567 = vacc1x89ABCDEF;
324 vacc0x01234567 = vacc0x89ABCDEF;
325
326 c4 += 8;
327 c3 += 8;
328 c2 += 8;
329 c1 += 8;
330 c0 += 8;
331 }
332 __m128 vacc4x0123 = _mm256_castps256_ps128(vacc4x01234567);
333 __m128 vacc3x0123 = _mm256_castps256_ps128(vacc3x01234567);
334 __m128 vacc2x0123 = _mm256_castps256_ps128(vacc2x01234567);
335 __m128 vacc1x0123 = _mm256_castps256_ps128(vacc1x01234567);
336 __m128 vacc0x0123 = _mm256_castps256_ps128(vacc0x01234567);
337 if (nc & 4) {
338 _mm_storeu_ps(c4, vacc4x0123);
339 _mm_storeu_ps(c3, vacc3x0123);
340 _mm_storeu_ps(c2, vacc2x0123);
341 _mm_storeu_ps(c1, vacc1x0123);
342 _mm_storeu_ps(c0, vacc0x0123);
343
344 vacc4x0123 = _mm256_extractf128_ps(vacc4x01234567, 1);
345 vacc3x0123 = _mm256_extractf128_ps(vacc3x01234567, 1);
346 vacc2x0123 = _mm256_extractf128_ps(vacc2x01234567, 1);
347 vacc1x0123 = _mm256_extractf128_ps(vacc1x01234567, 1);
348 vacc0x0123 = _mm256_extractf128_ps(vacc0x01234567, 1);
349
350 c4 += 4;
351 c3 += 4;
352 c2 += 4;
353 c1 += 4;
354 c0 += 4;
355 }
356 if (nc & 2) {
357 _mm_storel_pi((__m64*) c4, vacc4x0123);
358 _mm_storel_pi((__m64*) c3, vacc3x0123);
359 _mm_storel_pi((__m64*) c2, vacc2x0123);
360 _mm_storel_pi((__m64*) c1, vacc1x0123);
361 _mm_storel_pi((__m64*) c0, vacc0x0123);
362
363 vacc4x0123 = _mm_movehl_ps(vacc4x0123, vacc4x0123);
364 vacc3x0123 = _mm_movehl_ps(vacc3x0123, vacc3x0123);
365 vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123);
366 vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
367 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
368
369 c4 += 2;
370 c3 += 2;
371 c2 += 2;
372 c1 += 2;
373 c0 += 2;
374 }
375 if (nc & 1) {
376 _mm_store_ss(c4, vacc4x0123);
377 _mm_store_ss(c3, vacc3x0123);
378 _mm_store_ss(c2, vacc2x0123);
379 _mm_store_ss(c1, vacc1x0123);
380 _mm_store_ss(c0, vacc0x0123);
381 }
382
383 nc = 0;
384 }
385 } while (nc != 0);
386 }
387