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_3x16s4__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_3x16s4__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 <= 3);
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
55 do {
56 __m256 vacc0x01234567 = _mm256_load_ps(acc + 0);
57 __m256 vacc0x89ABCDEF = _mm256_load_ps(acc + 8);
58 __m256 vacc1x01234567 = _mm256_load_ps(acc + 16);
59 __m256 vacc1x89ABCDEF = _mm256_load_ps(acc + 24);
60 __m256 vacc2x01234567 = _mm256_load_ps(acc + 32);
61 __m256 vacc2x89ABCDEF = _mm256_load_ps(acc + 40);
62 acc += 48;
63
64 size_t k = kc;
65 while (k >= 4 * sizeof(float)) {
66 __m256 va0 = _mm256_broadcast_ps((const __m128*) a0);
67 a0 += 4;
68 __m256 va1 = _mm256_broadcast_ps((const __m128*) a1);
69 a1 += 4;
70 __m256 va2 = _mm256_broadcast_ps((const __m128*) a2);
71 a2 += 4;
72
73
74 const __m256 vb01234567c0 = _mm256_load_ps(w + 0);
75 const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8);
76
77 vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c0, vacc0x01234567);
78 vacc1x01234567 = _mm256_fmadd_ps(va1, vb01234567c0, vacc1x01234567);
79 vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567c0, vacc2x01234567);
80 vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc0, vacc0x89ABCDEF);
81 vacc1x89ABCDEF = _mm256_fmadd_ps(va1, vb89ABCDEFc0, vacc1x89ABCDEF);
82 vacc2x89ABCDEF = _mm256_fmadd_ps(va2, vb89ABCDEFc0, vacc2x89ABCDEF);
83
84 va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
85 va1 = _mm256_permute_ps(va1, _MM_SHUFFLE(0, 3, 2, 1));
86 va2 = _mm256_permute_ps(va2, _MM_SHUFFLE(0, 3, 2, 1));
87
88 const __m256 vb01234567c1 = _mm256_load_ps(w + 16);
89 const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24);
90
91 vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c1, vacc0x01234567);
92 vacc1x01234567 = _mm256_fmadd_ps(va1, vb01234567c1, vacc1x01234567);
93 vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567c1, vacc2x01234567);
94 vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc1, vacc0x89ABCDEF);
95 vacc1x89ABCDEF = _mm256_fmadd_ps(va1, vb89ABCDEFc1, vacc1x89ABCDEF);
96 vacc2x89ABCDEF = _mm256_fmadd_ps(va2, vb89ABCDEFc1, vacc2x89ABCDEF);
97
98 va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
99 va1 = _mm256_permute_ps(va1, _MM_SHUFFLE(0, 3, 2, 1));
100 va2 = _mm256_permute_ps(va2, _MM_SHUFFLE(0, 3, 2, 1));
101
102 const __m256 vb01234567c2 = _mm256_load_ps(w + 32);
103 const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40);
104
105 vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c2, vacc0x01234567);
106 vacc1x01234567 = _mm256_fmadd_ps(va1, vb01234567c2, vacc1x01234567);
107 vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567c2, vacc2x01234567);
108 vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc2, vacc0x89ABCDEF);
109 vacc1x89ABCDEF = _mm256_fmadd_ps(va1, vb89ABCDEFc2, vacc1x89ABCDEF);
110 vacc2x89ABCDEF = _mm256_fmadd_ps(va2, vb89ABCDEFc2, vacc2x89ABCDEF);
111
112 va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
113 va1 = _mm256_permute_ps(va1, _MM_SHUFFLE(0, 3, 2, 1));
114 va2 = _mm256_permute_ps(va2, _MM_SHUFFLE(0, 3, 2, 1));
115
116 const __m256 vb01234567c3 = _mm256_load_ps(w + 48);
117 const __m256 vb89ABCDEFc3 = _mm256_load_ps(w + 56);
118
119 vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c3, vacc0x01234567);
120 vacc1x01234567 = _mm256_fmadd_ps(va1, vb01234567c3, vacc1x01234567);
121 vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567c3, vacc2x01234567);
122 vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc3, vacc0x89ABCDEF);
123 vacc1x89ABCDEF = _mm256_fmadd_ps(va1, vb89ABCDEFc3, vacc1x89ABCDEF);
124 vacc2x89ABCDEF = _mm256_fmadd_ps(va2, vb89ABCDEFc3, vacc2x89ABCDEF);
125
126
127 w += 64;
128 k -= 4 * sizeof(float);
129 }
130 if XNN_UNLIKELY(k != 0) {
131 __m256 va0 = _mm256_broadcast_ps((const __m128*) a0);
132 a0 = (const float*) ((uintptr_t) a0 + k);
133 __m256 va1 = _mm256_broadcast_ps((const __m128*) a1);
134 a1 = (const float*) ((uintptr_t) a1 + k);
135 __m256 va2 = _mm256_broadcast_ps((const __m128*) a2);
136 a2 = (const float*) ((uintptr_t) a2 + k);
137
138 const __m256 vzero = _mm256_setzero_ps();
139
140 const __m256 vb01234567c0 = _mm256_load_ps(w + 0);
141 const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8);
142
143 vacc0x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb01234567c0, vzero, _CMP_NEQ_OQ)), vb01234567c0, vacc0x01234567);
144 vacc1x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb01234567c0, vzero, _CMP_NEQ_OQ)), vb01234567c0, vacc1x01234567);
145 vacc2x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb01234567c0, vzero, _CMP_NEQ_OQ)), vb01234567c0, vacc2x01234567);
146 vacc0x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb89ABCDEFc0, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc0, vacc0x89ABCDEF);
147 vacc1x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb89ABCDEFc0, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc0, vacc1x89ABCDEF);
148 vacc2x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb89ABCDEFc0, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc0, vacc2x89ABCDEF);
149
150 va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
151 va1 = _mm256_permute_ps(va1, _MM_SHUFFLE(0, 3, 2, 1));
152 va2 = _mm256_permute_ps(va2, _MM_SHUFFLE(0, 3, 2, 1));
153
154 const __m256 vb01234567c1 = _mm256_load_ps(w + 16);
155 const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24);
156
157 vacc0x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb01234567c1, vzero, _CMP_NEQ_OQ)), vb01234567c1, vacc0x01234567);
158 vacc1x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb01234567c1, vzero, _CMP_NEQ_OQ)), vb01234567c1, vacc1x01234567);
159 vacc2x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb01234567c1, vzero, _CMP_NEQ_OQ)), vb01234567c1, vacc2x01234567);
160 vacc0x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb89ABCDEFc1, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc1, vacc0x89ABCDEF);
161 vacc1x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb89ABCDEFc1, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc1, vacc1x89ABCDEF);
162 vacc2x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb89ABCDEFc1, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc1, vacc2x89ABCDEF);
163
164 va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
165 va1 = _mm256_permute_ps(va1, _MM_SHUFFLE(0, 3, 2, 1));
166 va2 = _mm256_permute_ps(va2, _MM_SHUFFLE(0, 3, 2, 1));
167
168 const __m256 vb01234567c2 = _mm256_load_ps(w + 32);
169 const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40);
170
171 vacc0x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb01234567c2, vzero, _CMP_NEQ_OQ)), vb01234567c2, vacc0x01234567);
172 vacc1x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb01234567c2, vzero, _CMP_NEQ_OQ)), vb01234567c2, vacc1x01234567);
173 vacc2x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb01234567c2, vzero, _CMP_NEQ_OQ)), vb01234567c2, vacc2x01234567);
174 vacc0x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb89ABCDEFc2, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc2, vacc0x89ABCDEF);
175 vacc1x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb89ABCDEFc2, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc2, vacc1x89ABCDEF);
176 vacc2x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb89ABCDEFc2, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc2, vacc2x89ABCDEF);
177
178 va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
179 va1 = _mm256_permute_ps(va1, _MM_SHUFFLE(0, 3, 2, 1));
180 va2 = _mm256_permute_ps(va2, _MM_SHUFFLE(0, 3, 2, 1));
181
182 const __m256 vb01234567c3 = _mm256_load_ps(w + 48);
183 const __m256 vb89ABCDEFc3 = _mm256_load_ps(w + 56);
184
185 vacc0x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb01234567c3, vzero, _CMP_NEQ_OQ)), vb01234567c3, vacc0x01234567);
186 vacc1x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb01234567c3, vzero, _CMP_NEQ_OQ)), vb01234567c3, vacc1x01234567);
187 vacc2x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb01234567c3, vzero, _CMP_NEQ_OQ)), vb01234567c3, vacc2x01234567);
188 vacc0x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb89ABCDEFc3, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc3, vacc0x89ABCDEF);
189 vacc1x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb89ABCDEFc3, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc3, vacc1x89ABCDEF);
190 vacc2x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb89ABCDEFc3, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc3, vacc2x89ABCDEF);
191
192
193 w += 64;
194 }
195
196 const __m256 vmin = _mm256_load_ps(params->avx.min);
197 vacc0x01234567 = _mm256_max_ps(vacc0x01234567, vmin);
198 vacc1x01234567 = _mm256_max_ps(vacc1x01234567, vmin);
199 vacc2x01234567 = _mm256_max_ps(vacc2x01234567, vmin);
200 vacc0x89ABCDEF = _mm256_max_ps(vacc0x89ABCDEF, vmin);
201 vacc1x89ABCDEF = _mm256_max_ps(vacc1x89ABCDEF, vmin);
202 vacc2x89ABCDEF = _mm256_max_ps(vacc2x89ABCDEF, vmin);
203
204 const __m256 vmax = _mm256_load_ps(params->avx.max);
205 vacc0x01234567 = _mm256_min_ps(vacc0x01234567, vmax);
206 vacc1x01234567 = _mm256_min_ps(vacc1x01234567, vmax);
207 vacc2x01234567 = _mm256_min_ps(vacc2x01234567, vmax);
208 vacc0x89ABCDEF = _mm256_min_ps(vacc0x89ABCDEF, vmax);
209 vacc1x89ABCDEF = _mm256_min_ps(vacc1x89ABCDEF, vmax);
210 vacc2x89ABCDEF = _mm256_min_ps(vacc2x89ABCDEF, vmax);
211
212 if XNN_LIKELY(nc >= 16) {
213 _mm256_storeu_ps(c2, vacc2x01234567);
214 _mm256_storeu_ps(c2 + 8, vacc2x89ABCDEF);
215 c2 = (float*) ((uintptr_t) c2 + cn_stride);
216 _mm256_storeu_ps(c1, vacc1x01234567);
217 _mm256_storeu_ps(c1 + 8, vacc1x89ABCDEF);
218 c1 = (float*) ((uintptr_t) c1 + cn_stride);
219 _mm256_storeu_ps(c0, vacc0x01234567);
220 _mm256_storeu_ps(c0 + 8, vacc0x89ABCDEF);
221 c0 = (float*) ((uintptr_t) c0 + cn_stride);
222
223 a2 = (const float*) ((uintptr_t) a2 - kc);
224 a1 = (const float*) ((uintptr_t) a1 - kc);
225 a0 = (const float*) ((uintptr_t) a0 - kc);
226
227 nc -= 16;
228 } else {
229 if (nc & 8) {
230 _mm256_storeu_ps(c2, vacc2x01234567);
231 _mm256_storeu_ps(c1, vacc1x01234567);
232 _mm256_storeu_ps(c0, vacc0x01234567);
233
234 vacc2x01234567 = vacc2x89ABCDEF;
235 vacc1x01234567 = vacc1x89ABCDEF;
236 vacc0x01234567 = vacc0x89ABCDEF;
237
238 c2 += 8;
239 c1 += 8;
240 c0 += 8;
241 }
242 __m128 vacc2x0123 = _mm256_castps256_ps128(vacc2x01234567);
243 __m128 vacc1x0123 = _mm256_castps256_ps128(vacc1x01234567);
244 __m128 vacc0x0123 = _mm256_castps256_ps128(vacc0x01234567);
245 if (nc & 4) {
246 _mm_storeu_ps(c2, vacc2x0123);
247 _mm_storeu_ps(c1, vacc1x0123);
248 _mm_storeu_ps(c0, vacc0x0123);
249
250 vacc2x0123 = _mm256_extractf128_ps(vacc2x01234567, 1);
251 vacc1x0123 = _mm256_extractf128_ps(vacc1x01234567, 1);
252 vacc0x0123 = _mm256_extractf128_ps(vacc0x01234567, 1);
253
254 c2 += 4;
255 c1 += 4;
256 c0 += 4;
257 }
258 if (nc & 2) {
259 _mm_storel_pi((__m64*) c2, vacc2x0123);
260 _mm_storel_pi((__m64*) c1, vacc1x0123);
261 _mm_storel_pi((__m64*) c0, vacc0x0123);
262
263 vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123);
264 vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
265 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
266
267 c2 += 2;
268 c1 += 2;
269 c0 += 2;
270 }
271 if (nc & 1) {
272 _mm_store_ss(c2, vacc2x0123);
273 _mm_store_ss(c1, vacc1x0123);
274 _mm_store_ss(c0, vacc0x0123);
275 }
276
277 nc = 0;
278 }
279 } while (nc != 0);
280 }
281