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_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 union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_gemm_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 union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29 assert(mr != 0);
30 assert(mr <= 3);
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
53 do {
54 __m256 vacc0x01234567 = _mm256_load_ps(w + 0);
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 k = kc;
63 while (k >= 4 * sizeof(float)) {
64 __m256 va0 = _mm256_broadcast_ps((const __m128*) a0);
65 a0 += 4;
66 __m256 va1 = _mm256_broadcast_ps((const __m128*) a1);
67 a1 += 4;
68 __m256 va2 = _mm256_broadcast_ps((const __m128*) a2);
69 a2 += 4;
70
71
72 const __m256 vb01234567c0 = _mm256_load_ps(w + 0);
73 const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8);
74
75 vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c0, vacc0x01234567);
76 vacc1x01234567 = _mm256_fmadd_ps(va1, vb01234567c0, vacc1x01234567);
77 vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567c0, vacc2x01234567);
78 vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc0, vacc0x89ABCDEF);
79 vacc1x89ABCDEF = _mm256_fmadd_ps(va1, vb89ABCDEFc0, vacc1x89ABCDEF);
80 vacc2x89ABCDEF = _mm256_fmadd_ps(va2, vb89ABCDEFc0, vacc2x89ABCDEF);
81
82 va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
83 va1 = _mm256_permute_ps(va1, _MM_SHUFFLE(0, 3, 2, 1));
84 va2 = _mm256_permute_ps(va2, _MM_SHUFFLE(0, 3, 2, 1));
85
86 const __m256 vb01234567c1 = _mm256_load_ps(w + 16);
87 const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24);
88
89 vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c1, vacc0x01234567);
90 vacc1x01234567 = _mm256_fmadd_ps(va1, vb01234567c1, vacc1x01234567);
91 vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567c1, vacc2x01234567);
92 vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc1, vacc0x89ABCDEF);
93 vacc1x89ABCDEF = _mm256_fmadd_ps(va1, vb89ABCDEFc1, vacc1x89ABCDEF);
94 vacc2x89ABCDEF = _mm256_fmadd_ps(va2, vb89ABCDEFc1, vacc2x89ABCDEF);
95
96 va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
97 va1 = _mm256_permute_ps(va1, _MM_SHUFFLE(0, 3, 2, 1));
98 va2 = _mm256_permute_ps(va2, _MM_SHUFFLE(0, 3, 2, 1));
99
100 const __m256 vb01234567c2 = _mm256_load_ps(w + 32);
101 const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40);
102
103 vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c2, vacc0x01234567);
104 vacc1x01234567 = _mm256_fmadd_ps(va1, vb01234567c2, vacc1x01234567);
105 vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567c2, vacc2x01234567);
106 vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc2, vacc0x89ABCDEF);
107 vacc1x89ABCDEF = _mm256_fmadd_ps(va1, vb89ABCDEFc2, vacc1x89ABCDEF);
108 vacc2x89ABCDEF = _mm256_fmadd_ps(va2, vb89ABCDEFc2, vacc2x89ABCDEF);
109
110 va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
111 va1 = _mm256_permute_ps(va1, _MM_SHUFFLE(0, 3, 2, 1));
112 va2 = _mm256_permute_ps(va2, _MM_SHUFFLE(0, 3, 2, 1));
113
114 const __m256 vb01234567c3 = _mm256_load_ps(w + 48);
115 const __m256 vb89ABCDEFc3 = _mm256_load_ps(w + 56);
116
117 vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c3, vacc0x01234567);
118 vacc1x01234567 = _mm256_fmadd_ps(va1, vb01234567c3, vacc1x01234567);
119 vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567c3, vacc2x01234567);
120 vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc3, vacc0x89ABCDEF);
121 vacc1x89ABCDEF = _mm256_fmadd_ps(va1, vb89ABCDEFc3, vacc1x89ABCDEF);
122 vacc2x89ABCDEF = _mm256_fmadd_ps(va2, vb89ABCDEFc3, vacc2x89ABCDEF);
123
124
125 w += 64;
126 k -= 4 * sizeof(float);
127 }
128 if XNN_UNLIKELY(k != 0) {
129 __m256 va0 = _mm256_broadcast_ps((const __m128*) a0);
130 a0 = (const float*) ((uintptr_t) a0 + k);
131 __m256 va1 = _mm256_broadcast_ps((const __m128*) a1);
132 a1 = (const float*) ((uintptr_t) a1 + k);
133 __m256 va2 = _mm256_broadcast_ps((const __m128*) a2);
134 a2 = (const float*) ((uintptr_t) a2 + k);
135
136 const __m256 vzero = _mm256_setzero_ps();
137
138 const __m256 vb01234567c0 = _mm256_load_ps(w + 0);
139 const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8);
140
141 vacc0x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb01234567c0, vzero, _CMP_NEQ_OQ)), vb01234567c0, vacc0x01234567);
142 vacc1x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb01234567c0, vzero, _CMP_NEQ_OQ)), vb01234567c0, vacc1x01234567);
143 vacc2x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb01234567c0, vzero, _CMP_NEQ_OQ)), vb01234567c0, vacc2x01234567);
144 vacc0x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb89ABCDEFc0, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc0, vacc0x89ABCDEF);
145 vacc1x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb89ABCDEFc0, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc0, vacc1x89ABCDEF);
146 vacc2x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb89ABCDEFc0, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc0, vacc2x89ABCDEF);
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
152 const __m256 vb01234567c1 = _mm256_load_ps(w + 16);
153 const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24);
154
155 vacc0x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb01234567c1, vzero, _CMP_NEQ_OQ)), vb01234567c1, vacc0x01234567);
156 vacc1x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb01234567c1, vzero, _CMP_NEQ_OQ)), vb01234567c1, vacc1x01234567);
157 vacc2x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb01234567c1, vzero, _CMP_NEQ_OQ)), vb01234567c1, vacc2x01234567);
158 vacc0x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb89ABCDEFc1, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc1, vacc0x89ABCDEF);
159 vacc1x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb89ABCDEFc1, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc1, vacc1x89ABCDEF);
160 vacc2x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb89ABCDEFc1, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc1, vacc2x89ABCDEF);
161
162 va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
163 va1 = _mm256_permute_ps(va1, _MM_SHUFFLE(0, 3, 2, 1));
164 va2 = _mm256_permute_ps(va2, _MM_SHUFFLE(0, 3, 2, 1));
165
166 const __m256 vb01234567c2 = _mm256_load_ps(w + 32);
167 const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40);
168
169 vacc0x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb01234567c2, vzero, _CMP_NEQ_OQ)), vb01234567c2, vacc0x01234567);
170 vacc1x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb01234567c2, vzero, _CMP_NEQ_OQ)), vb01234567c2, vacc1x01234567);
171 vacc2x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb01234567c2, vzero, _CMP_NEQ_OQ)), vb01234567c2, vacc2x01234567);
172 vacc0x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb89ABCDEFc2, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc2, vacc0x89ABCDEF);
173 vacc1x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb89ABCDEFc2, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc2, vacc1x89ABCDEF);
174 vacc2x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb89ABCDEFc2, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc2, vacc2x89ABCDEF);
175
176 va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
177 va1 = _mm256_permute_ps(va1, _MM_SHUFFLE(0, 3, 2, 1));
178 va2 = _mm256_permute_ps(va2, _MM_SHUFFLE(0, 3, 2, 1));
179
180 const __m256 vb01234567c3 = _mm256_load_ps(w + 48);
181 const __m256 vb89ABCDEFc3 = _mm256_load_ps(w + 56);
182
183 vacc0x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb01234567c3, vzero, _CMP_NEQ_OQ)), vb01234567c3, vacc0x01234567);
184 vacc1x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb01234567c3, vzero, _CMP_NEQ_OQ)), vb01234567c3, vacc1x01234567);
185 vacc2x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb01234567c3, vzero, _CMP_NEQ_OQ)), vb01234567c3, vacc2x01234567);
186 vacc0x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb89ABCDEFc3, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc3, vacc0x89ABCDEF);
187 vacc1x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va1, _mm256_cmp_ps(vb89ABCDEFc3, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc3, vacc1x89ABCDEF);
188 vacc2x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va2, _mm256_cmp_ps(vb89ABCDEFc3, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc3, vacc2x89ABCDEF);
189
190
191 w += 64;
192 }
193
194 const __m256 vmin = _mm256_load_ps(params->avx.min);
195 vacc0x01234567 = _mm256_max_ps(vacc0x01234567, vmin);
196 vacc1x01234567 = _mm256_max_ps(vacc1x01234567, vmin);
197 vacc2x01234567 = _mm256_max_ps(vacc2x01234567, vmin);
198 vacc0x89ABCDEF = _mm256_max_ps(vacc0x89ABCDEF, vmin);
199 vacc1x89ABCDEF = _mm256_max_ps(vacc1x89ABCDEF, vmin);
200 vacc2x89ABCDEF = _mm256_max_ps(vacc2x89ABCDEF, vmin);
201
202 const __m256 vmax = _mm256_load_ps(params->avx.max);
203 vacc0x01234567 = _mm256_min_ps(vacc0x01234567, vmax);
204 vacc1x01234567 = _mm256_min_ps(vacc1x01234567, vmax);
205 vacc2x01234567 = _mm256_min_ps(vacc2x01234567, vmax);
206 vacc0x89ABCDEF = _mm256_min_ps(vacc0x89ABCDEF, vmax);
207 vacc1x89ABCDEF = _mm256_min_ps(vacc1x89ABCDEF, vmax);
208 vacc2x89ABCDEF = _mm256_min_ps(vacc2x89ABCDEF, vmax);
209
210 if XNN_LIKELY(nc >= 16) {
211 _mm256_storeu_ps(c2, vacc2x01234567);
212 _mm256_storeu_ps(c2 + 8, vacc2x89ABCDEF);
213 c2 = (float*) ((uintptr_t) c2 + cn_stride);
214 _mm256_storeu_ps(c1, vacc1x01234567);
215 _mm256_storeu_ps(c1 + 8, vacc1x89ABCDEF);
216 c1 = (float*) ((uintptr_t) c1 + cn_stride);
217 _mm256_storeu_ps(c0, vacc0x01234567);
218 _mm256_storeu_ps(c0 + 8, vacc0x89ABCDEF);
219 c0 = (float*) ((uintptr_t) c0 + cn_stride);
220
221 a2 = (const float*) ((uintptr_t) a2 - kc);
222 a1 = (const float*) ((uintptr_t) a1 - kc);
223 a0 = (const float*) ((uintptr_t) a0 - kc);
224
225 nc -= 16;
226 } else {
227 if (nc & 8) {
228 _mm256_storeu_ps(c2, vacc2x01234567);
229 _mm256_storeu_ps(c1, vacc1x01234567);
230 _mm256_storeu_ps(c0, vacc0x01234567);
231
232 vacc2x01234567 = vacc2x89ABCDEF;
233 vacc1x01234567 = vacc1x89ABCDEF;
234 vacc0x01234567 = vacc0x89ABCDEF;
235
236 c2 += 8;
237 c1 += 8;
238 c0 += 8;
239 }
240 __m128 vacc2x0123 = _mm256_castps256_ps128(vacc2x01234567);
241 __m128 vacc1x0123 = _mm256_castps256_ps128(vacc1x01234567);
242 __m128 vacc0x0123 = _mm256_castps256_ps128(vacc0x01234567);
243 if (nc & 4) {
244 _mm_storeu_ps(c2, vacc2x0123);
245 _mm_storeu_ps(c1, vacc1x0123);
246 _mm_storeu_ps(c0, vacc0x0123);
247
248 vacc2x0123 = _mm256_extractf128_ps(vacc2x01234567, 1);
249 vacc1x0123 = _mm256_extractf128_ps(vacc1x01234567, 1);
250 vacc0x0123 = _mm256_extractf128_ps(vacc0x01234567, 1);
251
252 c2 += 4;
253 c1 += 4;
254 c0 += 4;
255 }
256 if (nc & 2) {
257 _mm_storel_pi((__m64*) c2, vacc2x0123);
258 _mm_storel_pi((__m64*) c1, vacc1x0123);
259 _mm_storel_pi((__m64*) c0, vacc0x0123);
260
261 vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123);
262 vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
263 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
264
265 c2 += 2;
266 c1 += 2;
267 c0 += 2;
268 }
269 if (nc & 1) {
270 _mm_store_ss(c2, vacc2x0123);
271 _mm_store_ss(c1, vacc1x0123);
272 _mm_store_ss(c0, vacc0x0123);
273 }
274
275 nc = 0;
276 }
277 } while (nc != 0);
278 }
279