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_gemm_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 union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_gemm_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 union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29   assert(mr != 0);
30   assert(mr <= 5);
31   assert(nc != 0);
32   assert(kc != 0);
33   assert(kc % sizeof(float) == 0);
34   assert(a != NULL);
35   assert(w != NULL);
36   assert(c != NULL);
37 
38   const float* a0 = a;
39   float* c0 = c;
40   const float* a1 = (const float*) ((uintptr_t) a0 + a_stride);
41   float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
42   if XNN_UNPREDICTABLE(mr < 2) {
43     a1 = a0;
44     c1 = c0;
45   }
46   const float* a2 = (const float*) ((uintptr_t) a1 + a_stride);
47   float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
48   if XNN_UNPREDICTABLE(mr <= 2) {
49     a2 = a1;
50     c2 = c1;
51   }
52   const float* a3 = (const float*) ((uintptr_t) a2 + a_stride);
53   float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
54   if XNN_UNPREDICTABLE(mr < 4) {
55     a3 = a2;
56     c3 = c2;
57   }
58   const float* a4 = (const float*) ((uintptr_t) a3 + a_stride);
59   float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
60   if XNN_UNPREDICTABLE(mr <= 4) {
61     a4 = a3;
62     c4 = c3;
63   }
64 
65   do {
66     __m256 vacc0x01234567 = _mm256_load_ps(w + 0);
67     __m256 vacc0x89ABCDEF = _mm256_load_ps(w + 8);
68     __m256 vacc1x01234567 = vacc0x01234567;
69     __m256 vacc1x89ABCDEF = vacc0x89ABCDEF;
70     __m256 vacc2x01234567 = vacc0x01234567;
71     __m256 vacc2x89ABCDEF = vacc0x89ABCDEF;
72     __m256 vacc3x01234567 = vacc0x01234567;
73     __m256 vacc3x89ABCDEF = vacc0x89ABCDEF;
74     __m256 vacc4x01234567 = vacc0x01234567;
75     __m256 vacc4x89ABCDEF = vacc0x89ABCDEF;
76     w += 16;
77 
78     size_t k = kc;
79     while (k >= 4 * sizeof(float)) {
80       __m256 va0 = _mm256_broadcast_ps((const __m128*) a0);
81       a0 += 4;
82       __m256 va1 = _mm256_broadcast_ps((const __m128*) a1);
83       a1 += 4;
84       __m256 va2 = _mm256_broadcast_ps((const __m128*) a2);
85       a2 += 4;
86       __m256 va3 = _mm256_broadcast_ps((const __m128*) a3);
87       a3 += 4;
88       __m256 va4 = _mm256_broadcast_ps((const __m128*) a4);
89       a4 += 4;
90 
91 
92       const __m256 vb01234567c0 = _mm256_load_ps(w + 0);
93       const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8);
94 
95       vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c0, vacc0x01234567);
96       vacc1x01234567 = _mm256_fmadd_ps(va1, vb01234567c0, vacc1x01234567);
97       vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567c0, vacc2x01234567);
98       vacc3x01234567 = _mm256_fmadd_ps(va3, vb01234567c0, vacc3x01234567);
99       vacc4x01234567 = _mm256_fmadd_ps(va4, vb01234567c0, vacc4x01234567);
100       vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc0, vacc0x89ABCDEF);
101       vacc1x89ABCDEF = _mm256_fmadd_ps(va1, vb89ABCDEFc0, vacc1x89ABCDEF);
102       vacc2x89ABCDEF = _mm256_fmadd_ps(va2, vb89ABCDEFc0, vacc2x89ABCDEF);
103       vacc3x89ABCDEF = _mm256_fmadd_ps(va3, vb89ABCDEFc0, vacc3x89ABCDEF);
104       vacc4x89ABCDEF = _mm256_fmadd_ps(va4, vb89ABCDEFc0, vacc4x89ABCDEF);
105 
106       va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
107       va1 = _mm256_permute_ps(va1, _MM_SHUFFLE(0, 3, 2, 1));
108       va2 = _mm256_permute_ps(va2, _MM_SHUFFLE(0, 3, 2, 1));
109       va3 = _mm256_permute_ps(va3, _MM_SHUFFLE(0, 3, 2, 1));
110       va4 = _mm256_permute_ps(va4, _MM_SHUFFLE(0, 3, 2, 1));
111 
112       const __m256 vb01234567c1 = _mm256_load_ps(w + 16);
113       const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24);
114 
115       vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c1, vacc0x01234567);
116       vacc1x01234567 = _mm256_fmadd_ps(va1, vb01234567c1, vacc1x01234567);
117       vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567c1, vacc2x01234567);
118       vacc3x01234567 = _mm256_fmadd_ps(va3, vb01234567c1, vacc3x01234567);
119       vacc4x01234567 = _mm256_fmadd_ps(va4, vb01234567c1, vacc4x01234567);
120       vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc1, vacc0x89ABCDEF);
121       vacc1x89ABCDEF = _mm256_fmadd_ps(va1, vb89ABCDEFc1, vacc1x89ABCDEF);
122       vacc2x89ABCDEF = _mm256_fmadd_ps(va2, vb89ABCDEFc1, vacc2x89ABCDEF);
123       vacc3x89ABCDEF = _mm256_fmadd_ps(va3, vb89ABCDEFc1, vacc3x89ABCDEF);
124       vacc4x89ABCDEF = _mm256_fmadd_ps(va4, vb89ABCDEFc1, vacc4x89ABCDEF);
125 
126       va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
127       va1 = _mm256_permute_ps(va1, _MM_SHUFFLE(0, 3, 2, 1));
128       va2 = _mm256_permute_ps(va2, _MM_SHUFFLE(0, 3, 2, 1));
129       va3 = _mm256_permute_ps(va3, _MM_SHUFFLE(0, 3, 2, 1));
130       va4 = _mm256_permute_ps(va4, _MM_SHUFFLE(0, 3, 2, 1));
131 
132       const __m256 vb01234567c2 = _mm256_load_ps(w + 32);
133       const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40);
134 
135       vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c2, vacc0x01234567);
136       vacc1x01234567 = _mm256_fmadd_ps(va1, vb01234567c2, vacc1x01234567);
137       vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567c2, vacc2x01234567);
138       vacc3x01234567 = _mm256_fmadd_ps(va3, vb01234567c2, vacc3x01234567);
139       vacc4x01234567 = _mm256_fmadd_ps(va4, vb01234567c2, vacc4x01234567);
140       vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc2, vacc0x89ABCDEF);
141       vacc1x89ABCDEF = _mm256_fmadd_ps(va1, vb89ABCDEFc2, vacc1x89ABCDEF);
142       vacc2x89ABCDEF = _mm256_fmadd_ps(va2, vb89ABCDEFc2, vacc2x89ABCDEF);
143       vacc3x89ABCDEF = _mm256_fmadd_ps(va3, vb89ABCDEFc2, vacc3x89ABCDEF);
144       vacc4x89ABCDEF = _mm256_fmadd_ps(va4, vb89ABCDEFc2, vacc4x89ABCDEF);
145 
146       va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
147       va1 = _mm256_permute_ps(va1, _MM_SHUFFLE(0, 3, 2, 1));
148       va2 = _mm256_permute_ps(va2, _MM_SHUFFLE(0, 3, 2, 1));
149       va3 = _mm256_permute_ps(va3, _MM_SHUFFLE(0, 3, 2, 1));
150       va4 = _mm256_permute_ps(va4, _MM_SHUFFLE(0, 3, 2, 1));
151 
152       const __m256 vb01234567c3 = _mm256_load_ps(w + 48);
153       const __m256 vb89ABCDEFc3 = _mm256_load_ps(w + 56);
154 
155       vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c3, vacc0x01234567);
156       vacc1x01234567 = _mm256_fmadd_ps(va1, vb01234567c3, vacc1x01234567);
157       vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567c3, vacc2x01234567);
158       vacc3x01234567 = _mm256_fmadd_ps(va3, vb01234567c3, vacc3x01234567);
159       vacc4x01234567 = _mm256_fmadd_ps(va4, vb01234567c3, vacc4x01234567);
160       vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc3, vacc0x89ABCDEF);
161       vacc1x89ABCDEF = _mm256_fmadd_ps(va1, vb89ABCDEFc3, vacc1x89ABCDEF);
162       vacc2x89ABCDEF = _mm256_fmadd_ps(va2, vb89ABCDEFc3, vacc2x89ABCDEF);
163       vacc3x89ABCDEF = _mm256_fmadd_ps(va3, vb89ABCDEFc3, vacc3x89ABCDEF);
164       vacc4x89ABCDEF = _mm256_fmadd_ps(va4, vb89ABCDEFc3, vacc4x89ABCDEF);
165 
166 
167       w += 64;
168       k -= 4 * sizeof(float);
169     }
170     if XNN_UNLIKELY(k != 0) {
171       do {
172         const __m256 va0 = _mm256_broadcast_ss(a0);
173         a0 += 1;
174         const __m256 va1 = _mm256_broadcast_ss(a1);
175         a1 += 1;
176         const __m256 va2 = _mm256_broadcast_ss(a2);
177         a2 += 1;
178         const __m256 va3 = _mm256_broadcast_ss(a3);
179         a3 += 1;
180         const __m256 va4 = _mm256_broadcast_ss(a4);
181         a4 += 1;
182 
183         const __m256 vb01234567 = _mm256_load_ps(w);
184         const __m256 vb89ABCDEF = _mm256_load_ps(w + 8);
185         w += 16;
186 
187         vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567, vacc0x01234567);
188         vacc1x01234567 = _mm256_fmadd_ps(va1, vb01234567, vacc1x01234567);
189         vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567, vacc2x01234567);
190         vacc3x01234567 = _mm256_fmadd_ps(va3, vb01234567, vacc3x01234567);
191         vacc4x01234567 = _mm256_fmadd_ps(va4, vb01234567, vacc4x01234567);
192         vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEF, vacc0x89ABCDEF);
193         vacc1x89ABCDEF = _mm256_fmadd_ps(va1, vb89ABCDEF, vacc1x89ABCDEF);
194         vacc2x89ABCDEF = _mm256_fmadd_ps(va2, vb89ABCDEF, vacc2x89ABCDEF);
195         vacc3x89ABCDEF = _mm256_fmadd_ps(va3, vb89ABCDEF, vacc3x89ABCDEF);
196         vacc4x89ABCDEF = _mm256_fmadd_ps(va4, vb89ABCDEF, vacc4x89ABCDEF);
197 
198         k -= sizeof(float);
199       } while (k != 0);
200     }
201 
202     const __m256 vmax = _mm256_broadcast_ps((const __m128*) params->sse.max);
203     vacc0x01234567 = _mm256_min_ps(vacc0x01234567, vmax);
204     vacc1x01234567 = _mm256_min_ps(vacc1x01234567, vmax);
205     vacc2x01234567 = _mm256_min_ps(vacc2x01234567, vmax);
206     vacc3x01234567 = _mm256_min_ps(vacc3x01234567, vmax);
207     vacc4x01234567 = _mm256_min_ps(vacc4x01234567, vmax);
208     vacc0x89ABCDEF = _mm256_min_ps(vacc0x89ABCDEF, vmax);
209     vacc1x89ABCDEF = _mm256_min_ps(vacc1x89ABCDEF, vmax);
210     vacc2x89ABCDEF = _mm256_min_ps(vacc2x89ABCDEF, vmax);
211     vacc3x89ABCDEF = _mm256_min_ps(vacc3x89ABCDEF, vmax);
212     vacc4x89ABCDEF = _mm256_min_ps(vacc4x89ABCDEF, vmax);
213 
214     const __m256 vmin = _mm256_broadcast_ps((const __m128*) params->sse.min);
215     vacc0x01234567 = _mm256_max_ps(vacc0x01234567, vmin);
216     vacc1x01234567 = _mm256_max_ps(vacc1x01234567, vmin);
217     vacc2x01234567 = _mm256_max_ps(vacc2x01234567, vmin);
218     vacc3x01234567 = _mm256_max_ps(vacc3x01234567, vmin);
219     vacc4x01234567 = _mm256_max_ps(vacc4x01234567, vmin);
220     vacc0x89ABCDEF = _mm256_max_ps(vacc0x89ABCDEF, vmin);
221     vacc1x89ABCDEF = _mm256_max_ps(vacc1x89ABCDEF, vmin);
222     vacc2x89ABCDEF = _mm256_max_ps(vacc2x89ABCDEF, vmin);
223     vacc3x89ABCDEF = _mm256_max_ps(vacc3x89ABCDEF, vmin);
224     vacc4x89ABCDEF = _mm256_max_ps(vacc4x89ABCDEF, vmin);
225 
226     if XNN_LIKELY(nc >= 16) {
227       _mm256_storeu_ps(c4, vacc4x01234567);
228       _mm256_storeu_ps(c4 + 8, vacc4x89ABCDEF);
229       c4 = (float*) ((uintptr_t) c4 + cn_stride);
230       _mm256_storeu_ps(c3, vacc3x01234567);
231       _mm256_storeu_ps(c3 + 8, vacc3x89ABCDEF);
232       c3 = (float*) ((uintptr_t) c3 + cn_stride);
233       _mm256_storeu_ps(c2, vacc2x01234567);
234       _mm256_storeu_ps(c2 + 8, vacc2x89ABCDEF);
235       c2 = (float*) ((uintptr_t) c2 + cn_stride);
236       _mm256_storeu_ps(c1, vacc1x01234567);
237       _mm256_storeu_ps(c1 + 8, vacc1x89ABCDEF);
238       c1 = (float*) ((uintptr_t) c1 + cn_stride);
239       _mm256_storeu_ps(c0, vacc0x01234567);
240       _mm256_storeu_ps(c0 + 8, vacc0x89ABCDEF);
241       c0 = (float*) ((uintptr_t) c0 + cn_stride);
242 
243       a4 = (const float*) ((uintptr_t) a4 - kc);
244       a3 = (const float*) ((uintptr_t) a3 - kc);
245       a2 = (const float*) ((uintptr_t) a2 - kc);
246       a1 = (const float*) ((uintptr_t) a1 - kc);
247       a0 = (const float*) ((uintptr_t) a0 - kc);
248 
249       nc -= 16;
250     } else {
251       if (nc & 8) {
252         _mm256_storeu_ps(c4, vacc4x01234567);
253         _mm256_storeu_ps(c3, vacc3x01234567);
254         _mm256_storeu_ps(c2, vacc2x01234567);
255         _mm256_storeu_ps(c1, vacc1x01234567);
256         _mm256_storeu_ps(c0, vacc0x01234567);
257 
258         vacc4x01234567 = vacc4x89ABCDEF;
259         vacc3x01234567 = vacc3x89ABCDEF;
260         vacc2x01234567 = vacc2x89ABCDEF;
261         vacc1x01234567 = vacc1x89ABCDEF;
262         vacc0x01234567 = vacc0x89ABCDEF;
263 
264         c4 += 8;
265         c3 += 8;
266         c2 += 8;
267         c1 += 8;
268         c0 += 8;
269       }
270       __m128 vacc4x0123 = _mm256_castps256_ps128(vacc4x01234567);
271       __m128 vacc3x0123 = _mm256_castps256_ps128(vacc3x01234567);
272       __m128 vacc2x0123 = _mm256_castps256_ps128(vacc2x01234567);
273       __m128 vacc1x0123 = _mm256_castps256_ps128(vacc1x01234567);
274       __m128 vacc0x0123 = _mm256_castps256_ps128(vacc0x01234567);
275       if (nc & 4) {
276         _mm_storeu_ps(c4, vacc4x0123);
277         _mm_storeu_ps(c3, vacc3x0123);
278         _mm_storeu_ps(c2, vacc2x0123);
279         _mm_storeu_ps(c1, vacc1x0123);
280         _mm_storeu_ps(c0, vacc0x0123);
281 
282         vacc4x0123 = _mm256_extractf128_ps(vacc4x01234567, 1);
283         vacc3x0123 = _mm256_extractf128_ps(vacc3x01234567, 1);
284         vacc2x0123 = _mm256_extractf128_ps(vacc2x01234567, 1);
285         vacc1x0123 = _mm256_extractf128_ps(vacc1x01234567, 1);
286         vacc0x0123 = _mm256_extractf128_ps(vacc0x01234567, 1);
287 
288         c4 += 4;
289         c3 += 4;
290         c2 += 4;
291         c1 += 4;
292         c0 += 4;
293       }
294       if (nc & 2) {
295         _mm_storel_pi((__m64*) c4, vacc4x0123);
296         _mm_storel_pi((__m64*) c3, vacc3x0123);
297         _mm_storel_pi((__m64*) c2, vacc2x0123);
298         _mm_storel_pi((__m64*) c1, vacc1x0123);
299         _mm_storel_pi((__m64*) c0, vacc0x0123);
300 
301         vacc4x0123 = _mm_movehl_ps(vacc4x0123, vacc4x0123);
302         vacc3x0123 = _mm_movehl_ps(vacc3x0123, vacc3x0123);
303         vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123);
304         vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
305         vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
306 
307         c4 += 2;
308         c3 += 2;
309         c2 += 2;
310         c1 += 2;
311         c0 += 2;
312       }
313       if (nc & 1) {
314         _mm_store_ss(c4, vacc4x0123);
315         _mm_store_ss(c3, vacc3x0123);
316         _mm_store_ss(c2, vacc2x0123);
317         _mm_store_ss(c1, vacc1x0123);
318         _mm_store_ss(c0, vacc0x0123);
319       }
320 
321       nc = 0;
322     }
323   } while (nc != 0);
324 }
325