1 // Auto-generated file. Do not edit!
2 // Template: src/f32-gemm/avx-broadcast.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_8x8__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_8x8__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)])
29 {
30 assert(mr != 0);
31 assert(mr <= 8);
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 const float* a5 = (const float*) ((uintptr_t) a4 + a_stride);
67 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
68 if XNN_UNPREDICTABLE(mr < 6) {
69 a5 = a4;
70 c5 = c4;
71 }
72 const float* a6 = (const float*) ((uintptr_t) a5 + a_stride);
73 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
74 if XNN_UNPREDICTABLE(mr <= 6) {
75 a6 = a5;
76 c6 = c5;
77 }
78 const float* a7 = (const float*) ((uintptr_t) a6 + a_stride);
79 float* c7 = (float*) ((uintptr_t) c6 + cm_stride);
80 if XNN_UNPREDICTABLE(mr != 8) {
81 a7 = a6;
82 c7 = c6;
83 }
84
85 do {
86 __m256 vacc0x01234567 = _mm256_load_ps(acc + 0);
87 __m256 vacc1x01234567 = _mm256_load_ps(acc + 8);
88 __m256 vacc2x01234567 = _mm256_load_ps(acc + 16);
89 __m256 vacc3x01234567 = _mm256_load_ps(acc + 24);
90 __m256 vacc4x01234567 = _mm256_load_ps(acc + 32);
91 __m256 vacc5x01234567 = _mm256_load_ps(acc + 40);
92 __m256 vacc6x01234567 = _mm256_load_ps(acc + 48);
93 __m256 vacc7x01234567 = _mm256_load_ps(acc + 56);
94 acc += 64;
95
96 size_t k = kc;
97 do {
98 const __m256 va0 = _mm256_broadcast_ss(a0);
99 a0 += 1;
100 const __m256 va1 = _mm256_broadcast_ss(a1);
101 a1 += 1;
102 const __m256 va2 = _mm256_broadcast_ss(a2);
103 a2 += 1;
104 const __m256 va3 = _mm256_broadcast_ss(a3);
105 a3 += 1;
106 const __m256 va4 = _mm256_broadcast_ss(a4);
107 a4 += 1;
108 const __m256 va5 = _mm256_broadcast_ss(a5);
109 a5 += 1;
110 const __m256 va6 = _mm256_broadcast_ss(a6);
111 a6 += 1;
112 const __m256 va7 = _mm256_broadcast_ss(a7);
113 a7 += 1;
114
115 const __m256 vb01234567 = _mm256_load_ps(w);
116 w += 8;
117
118 vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567, vacc0x01234567);
119 vacc1x01234567 = _mm256_fmadd_ps(va1, vb01234567, vacc1x01234567);
120 vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567, vacc2x01234567);
121 vacc3x01234567 = _mm256_fmadd_ps(va3, vb01234567, vacc3x01234567);
122 vacc4x01234567 = _mm256_fmadd_ps(va4, vb01234567, vacc4x01234567);
123 vacc5x01234567 = _mm256_fmadd_ps(va5, vb01234567, vacc5x01234567);
124 vacc6x01234567 = _mm256_fmadd_ps(va6, vb01234567, vacc6x01234567);
125 vacc7x01234567 = _mm256_fmadd_ps(va7, vb01234567, vacc7x01234567);
126
127 k -= sizeof(float);
128 } while (k != 0);
129
130 const __m256 vmin = _mm256_load_ps(params->avx.min);
131 vacc0x01234567 = _mm256_max_ps(vacc0x01234567, vmin);
132 vacc1x01234567 = _mm256_max_ps(vacc1x01234567, vmin);
133 vacc2x01234567 = _mm256_max_ps(vacc2x01234567, vmin);
134 vacc3x01234567 = _mm256_max_ps(vacc3x01234567, vmin);
135 vacc4x01234567 = _mm256_max_ps(vacc4x01234567, vmin);
136 vacc5x01234567 = _mm256_max_ps(vacc5x01234567, vmin);
137 vacc6x01234567 = _mm256_max_ps(vacc6x01234567, vmin);
138 vacc7x01234567 = _mm256_max_ps(vacc7x01234567, vmin);
139
140 const __m256 vmax = _mm256_load_ps(params->avx.max);
141 vacc0x01234567 = _mm256_min_ps(vacc0x01234567, vmax);
142 vacc1x01234567 = _mm256_min_ps(vacc1x01234567, vmax);
143 vacc2x01234567 = _mm256_min_ps(vacc2x01234567, vmax);
144 vacc3x01234567 = _mm256_min_ps(vacc3x01234567, vmax);
145 vacc4x01234567 = _mm256_min_ps(vacc4x01234567, vmax);
146 vacc5x01234567 = _mm256_min_ps(vacc5x01234567, vmax);
147 vacc6x01234567 = _mm256_min_ps(vacc6x01234567, vmax);
148 vacc7x01234567 = _mm256_min_ps(vacc7x01234567, vmax);
149
150 if XNN_LIKELY(nc >= 8) {
151 _mm256_storeu_ps(c7, vacc7x01234567);
152 c7 = (float*) ((uintptr_t) c7 + cn_stride);
153 _mm256_storeu_ps(c6, vacc6x01234567);
154 c6 = (float*) ((uintptr_t) c6 + cn_stride);
155 _mm256_storeu_ps(c5, vacc5x01234567);
156 c5 = (float*) ((uintptr_t) c5 + cn_stride);
157 _mm256_storeu_ps(c4, vacc4x01234567);
158 c4 = (float*) ((uintptr_t) c4 + cn_stride);
159 _mm256_storeu_ps(c3, vacc3x01234567);
160 c3 = (float*) ((uintptr_t) c3 + cn_stride);
161 _mm256_storeu_ps(c2, vacc2x01234567);
162 c2 = (float*) ((uintptr_t) c2 + cn_stride);
163 _mm256_storeu_ps(c1, vacc1x01234567);
164 c1 = (float*) ((uintptr_t) c1 + cn_stride);
165 _mm256_storeu_ps(c0, vacc0x01234567);
166 c0 = (float*) ((uintptr_t) c0 + cn_stride);
167
168 a7 = (const float*) ((uintptr_t) a7 - kc);
169 a6 = (const float*) ((uintptr_t) a6 - kc);
170 a5 = (const float*) ((uintptr_t) a5 - kc);
171 a4 = (const float*) ((uintptr_t) a4 - kc);
172 a3 = (const float*) ((uintptr_t) a3 - kc);
173 a2 = (const float*) ((uintptr_t) a2 - kc);
174 a1 = (const float*) ((uintptr_t) a1 - kc);
175 a0 = (const float*) ((uintptr_t) a0 - kc);
176
177 nc -= 8;
178 } else {
179 __m128 vacc7x0123 = _mm256_castps256_ps128(vacc7x01234567);
180 __m128 vacc6x0123 = _mm256_castps256_ps128(vacc6x01234567);
181 __m128 vacc5x0123 = _mm256_castps256_ps128(vacc5x01234567);
182 __m128 vacc4x0123 = _mm256_castps256_ps128(vacc4x01234567);
183 __m128 vacc3x0123 = _mm256_castps256_ps128(vacc3x01234567);
184 __m128 vacc2x0123 = _mm256_castps256_ps128(vacc2x01234567);
185 __m128 vacc1x0123 = _mm256_castps256_ps128(vacc1x01234567);
186 __m128 vacc0x0123 = _mm256_castps256_ps128(vacc0x01234567);
187 if (nc & 4) {
188 _mm_storeu_ps(c7, vacc7x0123);
189 _mm_storeu_ps(c6, vacc6x0123);
190 _mm_storeu_ps(c5, vacc5x0123);
191 _mm_storeu_ps(c4, vacc4x0123);
192 _mm_storeu_ps(c3, vacc3x0123);
193 _mm_storeu_ps(c2, vacc2x0123);
194 _mm_storeu_ps(c1, vacc1x0123);
195 _mm_storeu_ps(c0, vacc0x0123);
196
197 vacc7x0123 = _mm256_extractf128_ps(vacc7x01234567, 1);
198 vacc6x0123 = _mm256_extractf128_ps(vacc6x01234567, 1);
199 vacc5x0123 = _mm256_extractf128_ps(vacc5x01234567, 1);
200 vacc4x0123 = _mm256_extractf128_ps(vacc4x01234567, 1);
201 vacc3x0123 = _mm256_extractf128_ps(vacc3x01234567, 1);
202 vacc2x0123 = _mm256_extractf128_ps(vacc2x01234567, 1);
203 vacc1x0123 = _mm256_extractf128_ps(vacc1x01234567, 1);
204 vacc0x0123 = _mm256_extractf128_ps(vacc0x01234567, 1);
205
206 c7 += 4;
207 c6 += 4;
208 c5 += 4;
209 c4 += 4;
210 c3 += 4;
211 c2 += 4;
212 c1 += 4;
213 c0 += 4;
214 }
215 if (nc & 2) {
216 _mm_storel_pi((__m64*) c7, vacc7x0123);
217 _mm_storel_pi((__m64*) c6, vacc6x0123);
218 _mm_storel_pi((__m64*) c5, vacc5x0123);
219 _mm_storel_pi((__m64*) c4, vacc4x0123);
220 _mm_storel_pi((__m64*) c3, vacc3x0123);
221 _mm_storel_pi((__m64*) c2, vacc2x0123);
222 _mm_storel_pi((__m64*) c1, vacc1x0123);
223 _mm_storel_pi((__m64*) c0, vacc0x0123);
224
225 vacc7x0123 = _mm_movehl_ps(vacc7x0123, vacc7x0123);
226 vacc6x0123 = _mm_movehl_ps(vacc6x0123, vacc6x0123);
227 vacc5x0123 = _mm_movehl_ps(vacc5x0123, vacc5x0123);
228 vacc4x0123 = _mm_movehl_ps(vacc4x0123, vacc4x0123);
229 vacc3x0123 = _mm_movehl_ps(vacc3x0123, vacc3x0123);
230 vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123);
231 vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
232 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
233
234 c7 += 2;
235 c6 += 2;
236 c5 += 2;
237 c4 += 2;
238 c3 += 2;
239 c2 += 2;
240 c1 += 2;
241 c0 += 2;
242 }
243 if (nc & 1) {
244 _mm_store_ss(c7, vacc7x0123);
245 _mm_store_ss(c6, vacc6x0123);
246 _mm_store_ss(c5, vacc5x0123);
247 _mm_store_ss(c4, vacc4x0123);
248 _mm_store_ss(c3, vacc3x0123);
249 _mm_store_ss(c2, vacc2x0123);
250 _mm_store_ss(c1, vacc1x0123);
251 _mm_store_ss(c0, vacc0x0123);
252 }
253
254 nc = 0;
255 }
256 } while (nc != 0);
257 }
258