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_5x16__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_5x16__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 do {
80 const __m256 va0 = _mm256_broadcast_ss(a0);
81 a0 += 1;
82 const __m256 va1 = _mm256_broadcast_ss(a1);
83 a1 += 1;
84 const __m256 va2 = _mm256_broadcast_ss(a2);
85 a2 += 1;
86 const __m256 va3 = _mm256_broadcast_ss(a3);
87 a3 += 1;
88 const __m256 va4 = _mm256_broadcast_ss(a4);
89 a4 += 1;
90
91 const __m256 vb01234567 = _mm256_load_ps(w);
92 const __m256 vb89ABCDEF = _mm256_load_ps(w + 8);
93 w += 16;
94
95 vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567, vacc0x01234567);
96 vacc1x01234567 = _mm256_fmadd_ps(va1, vb01234567, vacc1x01234567);
97 vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567, vacc2x01234567);
98 vacc3x01234567 = _mm256_fmadd_ps(va3, vb01234567, vacc3x01234567);
99 vacc4x01234567 = _mm256_fmadd_ps(va4, vb01234567, vacc4x01234567);
100 vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEF, vacc0x89ABCDEF);
101 vacc1x89ABCDEF = _mm256_fmadd_ps(va1, vb89ABCDEF, vacc1x89ABCDEF);
102 vacc2x89ABCDEF = _mm256_fmadd_ps(va2, vb89ABCDEF, vacc2x89ABCDEF);
103 vacc3x89ABCDEF = _mm256_fmadd_ps(va3, vb89ABCDEF, vacc3x89ABCDEF);
104 vacc4x89ABCDEF = _mm256_fmadd_ps(va4, vb89ABCDEF, vacc4x89ABCDEF);
105
106 k -= sizeof(float);
107 } while (k != 0);
108
109 const __m256 vmax = _mm256_broadcast_ps((const __m128*) params->sse.max);
110 vacc0x01234567 = _mm256_min_ps(vacc0x01234567, vmax);
111 vacc1x01234567 = _mm256_min_ps(vacc1x01234567, vmax);
112 vacc2x01234567 = _mm256_min_ps(vacc2x01234567, vmax);
113 vacc3x01234567 = _mm256_min_ps(vacc3x01234567, vmax);
114 vacc4x01234567 = _mm256_min_ps(vacc4x01234567, vmax);
115 vacc0x89ABCDEF = _mm256_min_ps(vacc0x89ABCDEF, vmax);
116 vacc1x89ABCDEF = _mm256_min_ps(vacc1x89ABCDEF, vmax);
117 vacc2x89ABCDEF = _mm256_min_ps(vacc2x89ABCDEF, vmax);
118 vacc3x89ABCDEF = _mm256_min_ps(vacc3x89ABCDEF, vmax);
119 vacc4x89ABCDEF = _mm256_min_ps(vacc4x89ABCDEF, vmax);
120
121 const __m256 vmin = _mm256_broadcast_ps((const __m128*) params->sse.min);
122 vacc0x01234567 = _mm256_max_ps(vacc0x01234567, vmin);
123 vacc1x01234567 = _mm256_max_ps(vacc1x01234567, vmin);
124 vacc2x01234567 = _mm256_max_ps(vacc2x01234567, vmin);
125 vacc3x01234567 = _mm256_max_ps(vacc3x01234567, vmin);
126 vacc4x01234567 = _mm256_max_ps(vacc4x01234567, vmin);
127 vacc0x89ABCDEF = _mm256_max_ps(vacc0x89ABCDEF, vmin);
128 vacc1x89ABCDEF = _mm256_max_ps(vacc1x89ABCDEF, vmin);
129 vacc2x89ABCDEF = _mm256_max_ps(vacc2x89ABCDEF, vmin);
130 vacc3x89ABCDEF = _mm256_max_ps(vacc3x89ABCDEF, vmin);
131 vacc4x89ABCDEF = _mm256_max_ps(vacc4x89ABCDEF, vmin);
132
133 if XNN_LIKELY(nc >= 16) {
134 _mm256_storeu_ps(c4, vacc4x01234567);
135 _mm256_storeu_ps(c4 + 8, vacc4x89ABCDEF);
136 c4 = (float*) ((uintptr_t) c4 + cn_stride);
137 _mm256_storeu_ps(c3, vacc3x01234567);
138 _mm256_storeu_ps(c3 + 8, vacc3x89ABCDEF);
139 c3 = (float*) ((uintptr_t) c3 + cn_stride);
140 _mm256_storeu_ps(c2, vacc2x01234567);
141 _mm256_storeu_ps(c2 + 8, vacc2x89ABCDEF);
142 c2 = (float*) ((uintptr_t) c2 + cn_stride);
143 _mm256_storeu_ps(c1, vacc1x01234567);
144 _mm256_storeu_ps(c1 + 8, vacc1x89ABCDEF);
145 c1 = (float*) ((uintptr_t) c1 + cn_stride);
146 _mm256_storeu_ps(c0, vacc0x01234567);
147 _mm256_storeu_ps(c0 + 8, vacc0x89ABCDEF);
148 c0 = (float*) ((uintptr_t) c0 + cn_stride);
149
150 a4 = (const float*) ((uintptr_t) a4 - kc);
151 a3 = (const float*) ((uintptr_t) a3 - kc);
152 a2 = (const float*) ((uintptr_t) a2 - kc);
153 a1 = (const float*) ((uintptr_t) a1 - kc);
154 a0 = (const float*) ((uintptr_t) a0 - kc);
155
156 nc -= 16;
157 } else {
158 if (nc & 8) {
159 _mm256_storeu_ps(c4, vacc4x01234567);
160 _mm256_storeu_ps(c3, vacc3x01234567);
161 _mm256_storeu_ps(c2, vacc2x01234567);
162 _mm256_storeu_ps(c1, vacc1x01234567);
163 _mm256_storeu_ps(c0, vacc0x01234567);
164
165 vacc4x01234567 = vacc4x89ABCDEF;
166 vacc3x01234567 = vacc3x89ABCDEF;
167 vacc2x01234567 = vacc2x89ABCDEF;
168 vacc1x01234567 = vacc1x89ABCDEF;
169 vacc0x01234567 = vacc0x89ABCDEF;
170
171 c4 += 8;
172 c3 += 8;
173 c2 += 8;
174 c1 += 8;
175 c0 += 8;
176 }
177 __m128 vacc4x0123 = _mm256_castps256_ps128(vacc4x01234567);
178 __m128 vacc3x0123 = _mm256_castps256_ps128(vacc3x01234567);
179 __m128 vacc2x0123 = _mm256_castps256_ps128(vacc2x01234567);
180 __m128 vacc1x0123 = _mm256_castps256_ps128(vacc1x01234567);
181 __m128 vacc0x0123 = _mm256_castps256_ps128(vacc0x01234567);
182 if (nc & 4) {
183 _mm_storeu_ps(c4, vacc4x0123);
184 _mm_storeu_ps(c3, vacc3x0123);
185 _mm_storeu_ps(c2, vacc2x0123);
186 _mm_storeu_ps(c1, vacc1x0123);
187 _mm_storeu_ps(c0, vacc0x0123);
188
189 vacc4x0123 = _mm256_extractf128_ps(vacc4x01234567, 1);
190 vacc3x0123 = _mm256_extractf128_ps(vacc3x01234567, 1);
191 vacc2x0123 = _mm256_extractf128_ps(vacc2x01234567, 1);
192 vacc1x0123 = _mm256_extractf128_ps(vacc1x01234567, 1);
193 vacc0x0123 = _mm256_extractf128_ps(vacc0x01234567, 1);
194
195 c4 += 4;
196 c3 += 4;
197 c2 += 4;
198 c1 += 4;
199 c0 += 4;
200 }
201 if (nc & 2) {
202 _mm_storel_pi((__m64*) c4, vacc4x0123);
203 _mm_storel_pi((__m64*) c3, vacc3x0123);
204 _mm_storel_pi((__m64*) c2, vacc2x0123);
205 _mm_storel_pi((__m64*) c1, vacc1x0123);
206 _mm_storel_pi((__m64*) c0, vacc0x0123);
207
208 vacc4x0123 = _mm_movehl_ps(vacc4x0123, vacc4x0123);
209 vacc3x0123 = _mm_movehl_ps(vacc3x0123, vacc3x0123);
210 vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123);
211 vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
212 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
213
214 c4 += 2;
215 c3 += 2;
216 c2 += 2;
217 c1 += 2;
218 c0 += 2;
219 }
220 if (nc & 1) {
221 _mm_store_ss(c4, vacc4x0123);
222 _mm_store_ss(c3, vacc3x0123);
223 _mm_store_ss(c2, vacc2x0123);
224 _mm_store_ss(c1, vacc1x0123);
225 _mm_store_ss(c0, vacc0x0123);
226 }
227
228 nc = 0;
229 }
230 } while (nc != 0);
231 }
232