1 // Auto-generated file. Do not edit!
2 // Template: src/f32-igemm/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/igemm.h>
15
16
xnn_f32_igemm_minmax_ukernel_3x16s4__fma3_broadcast(size_t mr,size_t nc,size_t kc,size_t ks,const float ** restrict a,const float * restrict w,float * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const float * zero,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_igemm_minmax_ukernel_3x16s4__fma3_broadcast(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 size_t ks,
22 const float**restrict a,
23 const float*restrict w,
24 float*restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 size_t a_offset,
28 const float* zero,
29 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 3);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(float) == 0);
36 assert(ks != 0);
37 assert(ks % (3 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(float) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 float* c0 = c;
44 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 c1 = c0;
47 }
48 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
49 if XNN_UNPREDICTABLE(mr <= 2) {
50 c2 = c1;
51 }
52
53 do {
54 __m256 vacc0x01234567 = _mm256_load_ps(w);
55 __m256 vacc0x89ABCDEF = _mm256_load_ps(w + 8);
56 __m256 vacc1x01234567 = vacc0x01234567;
57 __m256 vacc1x89ABCDEF = vacc0x89ABCDEF;
58 __m256 vacc2x01234567 = vacc0x01234567;
59 __m256 vacc2x89ABCDEF = vacc0x89ABCDEF;
60 w += 16;
61
62 size_t p = ks;
63 do {
64 const float* restrict a0 = a[0];
65 assert(a0 != NULL);
66 if XNN_UNPREDICTABLE(a0 != zero) {
67 a0 = (const float*) ((uintptr_t) a0 + a_offset);
68 }
69 const float* restrict a1 = a[1];
70 assert(a1 != NULL);
71 if XNN_UNPREDICTABLE(a1 != zero) {
72 a1 = (const float*) ((uintptr_t) a1 + a_offset);
73 }
74 const float* restrict a2 = a[2];
75 assert(a2 != NULL);
76 if XNN_UNPREDICTABLE(a2 != zero) {
77 a2 = (const float*) ((uintptr_t) a2 + a_offset);
78 }
79 a += 3;
80
81 size_t k = kc;
82 while (k >= 4 * sizeof(float)) {
83 __m256 va0 = _mm256_broadcast_ps((const __m128*) a0);
84 a0 += 4;
85 __m256 va1 = _mm256_broadcast_ps((const __m128*) a1);
86 a1 += 4;
87 __m256 va2 = _mm256_broadcast_ps((const __m128*) a2);
88 a2 += 4;
89
90
91 const __m256 vb01234567c0 = _mm256_load_ps(w + 0);
92 const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8);
93
94 vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c0, vacc0x01234567);
95 vacc1x01234567 = _mm256_fmadd_ps(va1, vb01234567c0, vacc1x01234567);
96 vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567c0, vacc2x01234567);
97 vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc0, vacc0x89ABCDEF);
98 vacc1x89ABCDEF = _mm256_fmadd_ps(va1, vb89ABCDEFc0, vacc1x89ABCDEF);
99 vacc2x89ABCDEF = _mm256_fmadd_ps(va2, vb89ABCDEFc0, vacc2x89ABCDEF);
100
101 va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
102 va1 = _mm256_permute_ps(va1, _MM_SHUFFLE(0, 3, 2, 1));
103 va2 = _mm256_permute_ps(va2, _MM_SHUFFLE(0, 3, 2, 1));
104
105 const __m256 vb01234567c1 = _mm256_load_ps(w + 16);
106 const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24);
107
108 vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c1, vacc0x01234567);
109 vacc1x01234567 = _mm256_fmadd_ps(va1, vb01234567c1, vacc1x01234567);
110 vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567c1, vacc2x01234567);
111 vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc1, vacc0x89ABCDEF);
112 vacc1x89ABCDEF = _mm256_fmadd_ps(va1, vb89ABCDEFc1, vacc1x89ABCDEF);
113 vacc2x89ABCDEF = _mm256_fmadd_ps(va2, vb89ABCDEFc1, vacc2x89ABCDEF);
114
115 va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
116 va1 = _mm256_permute_ps(va1, _MM_SHUFFLE(0, 3, 2, 1));
117 va2 = _mm256_permute_ps(va2, _MM_SHUFFLE(0, 3, 2, 1));
118
119 const __m256 vb01234567c2 = _mm256_load_ps(w + 32);
120 const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40);
121
122 vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c2, vacc0x01234567);
123 vacc1x01234567 = _mm256_fmadd_ps(va1, vb01234567c2, vacc1x01234567);
124 vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567c2, vacc2x01234567);
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
129 va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
130 va1 = _mm256_permute_ps(va1, _MM_SHUFFLE(0, 3, 2, 1));
131 va2 = _mm256_permute_ps(va2, _MM_SHUFFLE(0, 3, 2, 1));
132
133 const __m256 vb01234567c3 = _mm256_load_ps(w + 48);
134 const __m256 vb89ABCDEFc3 = _mm256_load_ps(w + 56);
135
136 vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c3, vacc0x01234567);
137 vacc1x01234567 = _mm256_fmadd_ps(va1, vb01234567c3, vacc1x01234567);
138 vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567c3, vacc2x01234567);
139 vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc3, vacc0x89ABCDEF);
140 vacc1x89ABCDEF = _mm256_fmadd_ps(va1, vb89ABCDEFc3, vacc1x89ABCDEF);
141 vacc2x89ABCDEF = _mm256_fmadd_ps(va2, vb89ABCDEFc3, vacc2x89ABCDEF);
142
143
144 w += 64;
145 k -= 4 * sizeof(float);
146 }
147 if XNN_UNLIKELY(k != 0) {
148 __m256 va0 = _mm256_broadcast_ps((const __m128*) a0);
149 a0 = (const float*) ((uintptr_t) a0 + k);
150 __m256 va1 = _mm256_broadcast_ps((const __m128*) a1);
151 a1 = (const float*) ((uintptr_t) a1 + k);
152 __m256 va2 = _mm256_broadcast_ps((const __m128*) a2);
153 a2 = (const float*) ((uintptr_t) a2 + k);
154
155 const __m256 vzero = _mm256_setzero_ps();
156
157 const __m256 vb01234567c0 = _mm256_load_ps(w + 0);
158 const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8);
159
160 vacc0x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb01234567c0, vzero, _CMP_NEQ_OQ)), vb01234567c0, vacc0x01234567);
161 vacc1x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb01234567c0, vzero, _CMP_NEQ_OQ)), vb01234567c0, vacc1x01234567);
162 vacc2x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb01234567c0, vzero, _CMP_NEQ_OQ)), vb01234567c0, vacc2x01234567);
163 vacc0x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb89ABCDEFc0, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc0, vacc0x89ABCDEF);
164 vacc1x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb89ABCDEFc0, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc0, vacc1x89ABCDEF);
165 vacc2x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb89ABCDEFc0, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc0, vacc2x89ABCDEF);
166
167 va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
168 va1 = _mm256_permute_ps(va1, _MM_SHUFFLE(0, 3, 2, 1));
169 va2 = _mm256_permute_ps(va2, _MM_SHUFFLE(0, 3, 2, 1));
170
171 const __m256 vb01234567c1 = _mm256_load_ps(w + 16);
172 const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24);
173
174 vacc0x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb01234567c1, vzero, _CMP_NEQ_OQ)), vb01234567c1, vacc0x01234567);
175 vacc1x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb01234567c1, vzero, _CMP_NEQ_OQ)), vb01234567c1, vacc1x01234567);
176 vacc2x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb01234567c1, vzero, _CMP_NEQ_OQ)), vb01234567c1, vacc2x01234567);
177 vacc0x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb89ABCDEFc1, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc1, vacc0x89ABCDEF);
178 vacc1x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb89ABCDEFc1, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc1, vacc1x89ABCDEF);
179 vacc2x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb89ABCDEFc1, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc1, vacc2x89ABCDEF);
180
181 va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
182 va1 = _mm256_permute_ps(va1, _MM_SHUFFLE(0, 3, 2, 1));
183 va2 = _mm256_permute_ps(va2, _MM_SHUFFLE(0, 3, 2, 1));
184
185 const __m256 vb01234567c2 = _mm256_load_ps(w + 32);
186 const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40);
187
188 vacc0x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb01234567c2, vzero, _CMP_NEQ_OQ)), vb01234567c2, vacc0x01234567);
189 vacc1x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb01234567c2, vzero, _CMP_NEQ_OQ)), vb01234567c2, vacc1x01234567);
190 vacc2x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb01234567c2, vzero, _CMP_NEQ_OQ)), vb01234567c2, vacc2x01234567);
191 vacc0x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb89ABCDEFc2, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc2, vacc0x89ABCDEF);
192 vacc1x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb89ABCDEFc2, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc2, vacc1x89ABCDEF);
193 vacc2x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb89ABCDEFc2, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc2, vacc2x89ABCDEF);
194
195 va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
196 va1 = _mm256_permute_ps(va1, _MM_SHUFFLE(0, 3, 2, 1));
197 va2 = _mm256_permute_ps(va2, _MM_SHUFFLE(0, 3, 2, 1));
198
199 const __m256 vb01234567c3 = _mm256_load_ps(w + 48);
200 const __m256 vb89ABCDEFc3 = _mm256_load_ps(w + 56);
201
202 vacc0x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb01234567c3, vzero, _CMP_NEQ_OQ)), vb01234567c3, vacc0x01234567);
203 vacc1x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb01234567c3, vzero, _CMP_NEQ_OQ)), vb01234567c3, vacc1x01234567);
204 vacc2x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb01234567c3, vzero, _CMP_NEQ_OQ)), vb01234567c3, vacc2x01234567);
205 vacc0x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb89ABCDEFc3, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc3, vacc0x89ABCDEF);
206 vacc1x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb89ABCDEFc3, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc3, vacc1x89ABCDEF);
207 vacc2x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb89ABCDEFc3, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc3, vacc2x89ABCDEF);
208
209
210 w += 64;
211 }
212 p -= 3 * sizeof(void*);
213 } while (p != 0);
214
215 const __m256 vmin = _mm256_load_ps(params->avx.min);
216 vacc0x01234567 = _mm256_max_ps(vacc0x01234567, vmin);
217 vacc1x01234567 = _mm256_max_ps(vacc1x01234567, vmin);
218 vacc2x01234567 = _mm256_max_ps(vacc2x01234567, vmin);
219 vacc0x89ABCDEF = _mm256_max_ps(vacc0x89ABCDEF, vmin);
220 vacc1x89ABCDEF = _mm256_max_ps(vacc1x89ABCDEF, vmin);
221 vacc2x89ABCDEF = _mm256_max_ps(vacc2x89ABCDEF, vmin);
222
223 const __m256 vmax = _mm256_load_ps(params->avx.max);
224 vacc0x01234567 = _mm256_min_ps(vacc0x01234567, vmax);
225 vacc1x01234567 = _mm256_min_ps(vacc1x01234567, vmax);
226 vacc2x01234567 = _mm256_min_ps(vacc2x01234567, vmax);
227 vacc0x89ABCDEF = _mm256_min_ps(vacc0x89ABCDEF, vmax);
228 vacc1x89ABCDEF = _mm256_min_ps(vacc1x89ABCDEF, vmax);
229 vacc2x89ABCDEF = _mm256_min_ps(vacc2x89ABCDEF, vmax);
230
231 if XNN_LIKELY(nc >= 16) {
232 _mm256_storeu_ps(c2, vacc2x01234567);
233 _mm256_storeu_ps(c2 + 8, vacc2x89ABCDEF);
234 c2 = (float*) ((uintptr_t) c2 + cn_stride);
235 _mm256_storeu_ps(c1, vacc1x01234567);
236 _mm256_storeu_ps(c1 + 8, vacc1x89ABCDEF);
237 c1 = (float*) ((uintptr_t) c1 + cn_stride);
238 _mm256_storeu_ps(c0, vacc0x01234567);
239 _mm256_storeu_ps(c0 + 8, vacc0x89ABCDEF);
240 c0 = (float*) ((uintptr_t) c0 + cn_stride);
241
242 a = (const float**restrict) ((uintptr_t) a - ks);
243 nc -= 16;
244 } else {
245 if (nc & 8) {
246 _mm256_storeu_ps(c2, vacc2x01234567);
247 _mm256_storeu_ps(c1, vacc1x01234567);
248 _mm256_storeu_ps(c0, vacc0x01234567);
249
250 vacc2x01234567 = vacc2x89ABCDEF;
251 vacc1x01234567 = vacc1x89ABCDEF;
252 vacc0x01234567 = vacc0x89ABCDEF;
253
254 c2 += 8;
255 c1 += 8;
256 c0 += 8;
257 }
258 __m128 vacc2x0123 = _mm256_castps256_ps128(vacc2x01234567);
259 __m128 vacc1x0123 = _mm256_castps256_ps128(vacc1x01234567);
260 __m128 vacc0x0123 = _mm256_castps256_ps128(vacc0x01234567);
261 if (nc & 4) {
262 _mm_storeu_ps(c2, vacc2x0123);
263 _mm_storeu_ps(c1, vacc1x0123);
264 _mm_storeu_ps(c0, vacc0x0123);
265
266 vacc2x0123 = _mm256_extractf128_ps(vacc2x01234567, 1);
267 vacc1x0123 = _mm256_extractf128_ps(vacc1x01234567, 1);
268 vacc0x0123 = _mm256_extractf128_ps(vacc0x01234567, 1);
269
270 c2 += 4;
271 c1 += 4;
272 c0 += 4;
273 }
274 if (nc & 2) {
275 _mm_storel_pi((__m64*) c2, vacc2x0123);
276 _mm_storel_pi((__m64*) c1, vacc1x0123);
277 _mm_storel_pi((__m64*) c0, vacc0x0123);
278
279 vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123);
280 vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
281 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
282
283 c2 += 2;
284 c1 += 2;
285 c0 += 2;
286 }
287 if (nc & 1) {
288 _mm_store_ss(c2, vacc2x0123);
289 _mm_store_ss(c1, vacc1x0123);
290 _mm_store_ss(c0, vacc0x0123);
291 }
292
293 nc = 0;
294 }
295 } while (nc != 0);
296 }
297