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_gemm_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 union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_gemm_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 union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29 assert(mr != 0);
30 assert(mr <= 8);
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 const float* a5 = (const float*) ((uintptr_t) a4 + a_stride);
65 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
66 if XNN_UNPREDICTABLE(mr < 6) {
67 a5 = a4;
68 c5 = c4;
69 }
70 const float* a6 = (const float*) ((uintptr_t) a5 + a_stride);
71 float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
72 if XNN_UNPREDICTABLE(mr <= 6) {
73 a6 = a5;
74 c6 = c5;
75 }
76 const float* a7 = (const float*) ((uintptr_t) a6 + a_stride);
77 float* c7 = (float*) ((uintptr_t) c6 + cm_stride);
78 if XNN_UNPREDICTABLE(mr != 8) {
79 a7 = a6;
80 c7 = c6;
81 }
82
83 do {
84 __m256 vacc0x01234567 = _mm256_load_ps(w + 0);
85 __m256 vacc1x01234567 = vacc0x01234567;
86 __m256 vacc2x01234567 = vacc0x01234567;
87 __m256 vacc3x01234567 = vacc0x01234567;
88 __m256 vacc4x01234567 = vacc0x01234567;
89 __m256 vacc5x01234567 = vacc0x01234567;
90 __m256 vacc6x01234567 = vacc0x01234567;
91 __m256 vacc7x01234567 = vacc0x01234567;
92 w += 8;
93
94 size_t k = kc;
95 do {
96 const __m256 va0 = _mm256_broadcast_ss(a0);
97 a0 += 1;
98 const __m256 va1 = _mm256_broadcast_ss(a1);
99 a1 += 1;
100 const __m256 va2 = _mm256_broadcast_ss(a2);
101 a2 += 1;
102 const __m256 va3 = _mm256_broadcast_ss(a3);
103 a3 += 1;
104 const __m256 va4 = _mm256_broadcast_ss(a4);
105 a4 += 1;
106 const __m256 va5 = _mm256_broadcast_ss(a5);
107 a5 += 1;
108 const __m256 va6 = _mm256_broadcast_ss(a6);
109 a6 += 1;
110 const __m256 va7 = _mm256_broadcast_ss(a7);
111 a7 += 1;
112
113 const __m256 vb01234567 = _mm256_load_ps(w);
114 w += 8;
115
116 vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567, vacc0x01234567);
117 vacc1x01234567 = _mm256_fmadd_ps(va1, vb01234567, vacc1x01234567);
118 vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567, vacc2x01234567);
119 vacc3x01234567 = _mm256_fmadd_ps(va3, vb01234567, vacc3x01234567);
120 vacc4x01234567 = _mm256_fmadd_ps(va4, vb01234567, vacc4x01234567);
121 vacc5x01234567 = _mm256_fmadd_ps(va5, vb01234567, vacc5x01234567);
122 vacc6x01234567 = _mm256_fmadd_ps(va6, vb01234567, vacc6x01234567);
123 vacc7x01234567 = _mm256_fmadd_ps(va7, vb01234567, vacc7x01234567);
124
125 k -= sizeof(float);
126 } while (k != 0);
127
128 const __m256 vmax = _mm256_broadcast_ps((const __m128*) params->sse.max);
129 vacc0x01234567 = _mm256_min_ps(vacc0x01234567, vmax);
130 vacc1x01234567 = _mm256_min_ps(vacc1x01234567, vmax);
131 vacc2x01234567 = _mm256_min_ps(vacc2x01234567, vmax);
132 vacc3x01234567 = _mm256_min_ps(vacc3x01234567, vmax);
133 vacc4x01234567 = _mm256_min_ps(vacc4x01234567, vmax);
134 vacc5x01234567 = _mm256_min_ps(vacc5x01234567, vmax);
135 vacc6x01234567 = _mm256_min_ps(vacc6x01234567, vmax);
136 vacc7x01234567 = _mm256_min_ps(vacc7x01234567, vmax);
137
138 const __m256 vmin = _mm256_broadcast_ps((const __m128*) params->sse.min);
139 vacc0x01234567 = _mm256_max_ps(vacc0x01234567, vmin);
140 vacc1x01234567 = _mm256_max_ps(vacc1x01234567, vmin);
141 vacc2x01234567 = _mm256_max_ps(vacc2x01234567, vmin);
142 vacc3x01234567 = _mm256_max_ps(vacc3x01234567, vmin);
143 vacc4x01234567 = _mm256_max_ps(vacc4x01234567, vmin);
144 vacc5x01234567 = _mm256_max_ps(vacc5x01234567, vmin);
145 vacc6x01234567 = _mm256_max_ps(vacc6x01234567, vmin);
146 vacc7x01234567 = _mm256_max_ps(vacc7x01234567, vmin);
147
148 if XNN_LIKELY(nc >= 8) {
149 _mm256_storeu_ps(c7, vacc7x01234567);
150 c7 = (float*) ((uintptr_t) c7 + cn_stride);
151 _mm256_storeu_ps(c6, vacc6x01234567);
152 c6 = (float*) ((uintptr_t) c6 + cn_stride);
153 _mm256_storeu_ps(c5, vacc5x01234567);
154 c5 = (float*) ((uintptr_t) c5 + cn_stride);
155 _mm256_storeu_ps(c4, vacc4x01234567);
156 c4 = (float*) ((uintptr_t) c4 + cn_stride);
157 _mm256_storeu_ps(c3, vacc3x01234567);
158 c3 = (float*) ((uintptr_t) c3 + cn_stride);
159 _mm256_storeu_ps(c2, vacc2x01234567);
160 c2 = (float*) ((uintptr_t) c2 + cn_stride);
161 _mm256_storeu_ps(c1, vacc1x01234567);
162 c1 = (float*) ((uintptr_t) c1 + cn_stride);
163 _mm256_storeu_ps(c0, vacc0x01234567);
164 c0 = (float*) ((uintptr_t) c0 + cn_stride);
165
166 a7 = (const float*) ((uintptr_t) a7 - kc);
167 a6 = (const float*) ((uintptr_t) a6 - kc);
168 a5 = (const float*) ((uintptr_t) a5 - kc);
169 a4 = (const float*) ((uintptr_t) a4 - kc);
170 a3 = (const float*) ((uintptr_t) a3 - kc);
171 a2 = (const float*) ((uintptr_t) a2 - kc);
172 a1 = (const float*) ((uintptr_t) a1 - kc);
173 a0 = (const float*) ((uintptr_t) a0 - kc);
174
175 nc -= 8;
176 } else {
177 __m128 vacc7x0123 = _mm256_castps256_ps128(vacc7x01234567);
178 __m128 vacc6x0123 = _mm256_castps256_ps128(vacc6x01234567);
179 __m128 vacc5x0123 = _mm256_castps256_ps128(vacc5x01234567);
180 __m128 vacc4x0123 = _mm256_castps256_ps128(vacc4x01234567);
181 __m128 vacc3x0123 = _mm256_castps256_ps128(vacc3x01234567);
182 __m128 vacc2x0123 = _mm256_castps256_ps128(vacc2x01234567);
183 __m128 vacc1x0123 = _mm256_castps256_ps128(vacc1x01234567);
184 __m128 vacc0x0123 = _mm256_castps256_ps128(vacc0x01234567);
185 if (nc & 4) {
186 _mm_storeu_ps(c7, vacc7x0123);
187 _mm_storeu_ps(c6, vacc6x0123);
188 _mm_storeu_ps(c5, vacc5x0123);
189 _mm_storeu_ps(c4, vacc4x0123);
190 _mm_storeu_ps(c3, vacc3x0123);
191 _mm_storeu_ps(c2, vacc2x0123);
192 _mm_storeu_ps(c1, vacc1x0123);
193 _mm_storeu_ps(c0, vacc0x0123);
194
195 vacc7x0123 = _mm256_extractf128_ps(vacc7x01234567, 1);
196 vacc6x0123 = _mm256_extractf128_ps(vacc6x01234567, 1);
197 vacc5x0123 = _mm256_extractf128_ps(vacc5x01234567, 1);
198 vacc4x0123 = _mm256_extractf128_ps(vacc4x01234567, 1);
199 vacc3x0123 = _mm256_extractf128_ps(vacc3x01234567, 1);
200 vacc2x0123 = _mm256_extractf128_ps(vacc2x01234567, 1);
201 vacc1x0123 = _mm256_extractf128_ps(vacc1x01234567, 1);
202 vacc0x0123 = _mm256_extractf128_ps(vacc0x01234567, 1);
203
204 c7 += 4;
205 c6 += 4;
206 c5 += 4;
207 c4 += 4;
208 c3 += 4;
209 c2 += 4;
210 c1 += 4;
211 c0 += 4;
212 }
213 if (nc & 2) {
214 _mm_storel_pi((__m64*) c7, vacc7x0123);
215 _mm_storel_pi((__m64*) c6, vacc6x0123);
216 _mm_storel_pi((__m64*) c5, vacc5x0123);
217 _mm_storel_pi((__m64*) c4, vacc4x0123);
218 _mm_storel_pi((__m64*) c3, vacc3x0123);
219 _mm_storel_pi((__m64*) c2, vacc2x0123);
220 _mm_storel_pi((__m64*) c1, vacc1x0123);
221 _mm_storel_pi((__m64*) c0, vacc0x0123);
222
223 vacc7x0123 = _mm_movehl_ps(vacc7x0123, vacc7x0123);
224 vacc6x0123 = _mm_movehl_ps(vacc6x0123, vacc6x0123);
225 vacc5x0123 = _mm_movehl_ps(vacc5x0123, vacc5x0123);
226 vacc4x0123 = _mm_movehl_ps(vacc4x0123, vacc4x0123);
227 vacc3x0123 = _mm_movehl_ps(vacc3x0123, vacc3x0123);
228 vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123);
229 vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
230 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
231
232 c7 += 2;
233 c6 += 2;
234 c5 += 2;
235 c4 += 2;
236 c3 += 2;
237 c2 += 2;
238 c1 += 2;
239 c0 += 2;
240 }
241 if (nc & 1) {
242 _mm_store_ss(c7, vacc7x0123);
243 _mm_store_ss(c6, vacc6x0123);
244 _mm_store_ss(c5, vacc5x0123);
245 _mm_store_ss(c4, vacc4x0123);
246 _mm_store_ss(c3, vacc3x0123);
247 _mm_store_ss(c2, vacc2x0123);
248 _mm_store_ss(c1, vacc1x0123);
249 _mm_store_ss(c0, vacc0x0123);
250 }
251
252 nc = 0;
253 }
254 } while (nc != 0);
255 }
256