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_ukernel_5x16__avx_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_output_params params[restrict static1])17 void xnn_f32_gemminc_ukernel_5x16__avx_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_output_params params[restrict static 1])
29 {
30 assert(mr != 0);
31 assert(mr <= 5);
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
67 do {
68 __m256 vacc0x01234567 = _mm256_load_ps(acc + 0);
69 __m256 vacc0x89ABCDEF = _mm256_load_ps(acc + 8);
70 __m256 vacc1x01234567 = _mm256_load_ps(acc + 16);
71 __m256 vacc1x89ABCDEF = _mm256_load_ps(acc + 24);
72 __m256 vacc2x01234567 = _mm256_load_ps(acc + 32);
73 __m256 vacc2x89ABCDEF = _mm256_load_ps(acc + 40);
74 __m256 vacc3x01234567 = _mm256_load_ps(acc + 48);
75 __m256 vacc3x89ABCDEF = _mm256_load_ps(acc + 56);
76 __m256 vacc4x01234567 = _mm256_load_ps(acc + 64);
77 __m256 vacc4x89ABCDEF = _mm256_load_ps(acc + 72);
78 acc += 80;
79
80 size_t k = kc;
81 do {
82 const __m256 va0 = _mm256_broadcast_ss(a0);
83 a0 += 1;
84 const __m256 va1 = _mm256_broadcast_ss(a1);
85 a1 += 1;
86 const __m256 va2 = _mm256_broadcast_ss(a2);
87 a2 += 1;
88 const __m256 va3 = _mm256_broadcast_ss(a3);
89 a3 += 1;
90 const __m256 va4 = _mm256_broadcast_ss(a4);
91 a4 += 1;
92
93 const __m256 vb01234567 = _mm256_load_ps(w);
94 const __m256 vb89ABCDEF = _mm256_load_ps(w + 8);
95 w += 16;
96
97 vacc0x01234567 = _mm256_add_ps(vacc0x01234567, _mm256_mul_ps(va0, vb01234567));
98 vacc1x01234567 = _mm256_add_ps(vacc1x01234567, _mm256_mul_ps(va1, vb01234567));
99 vacc2x01234567 = _mm256_add_ps(vacc2x01234567, _mm256_mul_ps(va2, vb01234567));
100 vacc3x01234567 = _mm256_add_ps(vacc3x01234567, _mm256_mul_ps(va3, vb01234567));
101 vacc4x01234567 = _mm256_add_ps(vacc4x01234567, _mm256_mul_ps(va4, vb01234567));
102 vacc0x89ABCDEF = _mm256_add_ps(vacc0x89ABCDEF, _mm256_mul_ps(va0, vb89ABCDEF));
103 vacc1x89ABCDEF = _mm256_add_ps(vacc1x89ABCDEF, _mm256_mul_ps(va1, vb89ABCDEF));
104 vacc2x89ABCDEF = _mm256_add_ps(vacc2x89ABCDEF, _mm256_mul_ps(va2, vb89ABCDEF));
105 vacc3x89ABCDEF = _mm256_add_ps(vacc3x89ABCDEF, _mm256_mul_ps(va3, vb89ABCDEF));
106 vacc4x89ABCDEF = _mm256_add_ps(vacc4x89ABCDEF, _mm256_mul_ps(va4, vb89ABCDEF));
107
108 k -= sizeof(float);
109 } while (k != 0);
110
111 const __m256 vmax = _mm256_broadcast_ps((const __m128*) params->sse.max);
112 vacc0x01234567 = _mm256_min_ps(vacc0x01234567, vmax);
113 vacc1x01234567 = _mm256_min_ps(vacc1x01234567, vmax);
114 vacc2x01234567 = _mm256_min_ps(vacc2x01234567, vmax);
115 vacc3x01234567 = _mm256_min_ps(vacc3x01234567, vmax);
116 vacc4x01234567 = _mm256_min_ps(vacc4x01234567, vmax);
117 vacc0x89ABCDEF = _mm256_min_ps(vacc0x89ABCDEF, vmax);
118 vacc1x89ABCDEF = _mm256_min_ps(vacc1x89ABCDEF, vmax);
119 vacc2x89ABCDEF = _mm256_min_ps(vacc2x89ABCDEF, vmax);
120 vacc3x89ABCDEF = _mm256_min_ps(vacc3x89ABCDEF, vmax);
121 vacc4x89ABCDEF = _mm256_min_ps(vacc4x89ABCDEF, vmax);
122
123 const __m256 vmin = _mm256_broadcast_ps((const __m128*) params->sse.min);
124 vacc0x01234567 = _mm256_max_ps(vacc0x01234567, vmin);
125 vacc1x01234567 = _mm256_max_ps(vacc1x01234567, vmin);
126 vacc2x01234567 = _mm256_max_ps(vacc2x01234567, vmin);
127 vacc3x01234567 = _mm256_max_ps(vacc3x01234567, vmin);
128 vacc4x01234567 = _mm256_max_ps(vacc4x01234567, vmin);
129 vacc0x89ABCDEF = _mm256_max_ps(vacc0x89ABCDEF, vmin);
130 vacc1x89ABCDEF = _mm256_max_ps(vacc1x89ABCDEF, vmin);
131 vacc2x89ABCDEF = _mm256_max_ps(vacc2x89ABCDEF, vmin);
132 vacc3x89ABCDEF = _mm256_max_ps(vacc3x89ABCDEF, vmin);
133 vacc4x89ABCDEF = _mm256_max_ps(vacc4x89ABCDEF, vmin);
134
135 if XNN_LIKELY(nc >= 16) {
136 _mm256_storeu_ps(c4, vacc4x01234567);
137 _mm256_storeu_ps(c4 + 8, vacc4x89ABCDEF);
138 c4 = (float*) ((uintptr_t) c4 + cn_stride);
139 _mm256_storeu_ps(c3, vacc3x01234567);
140 _mm256_storeu_ps(c3 + 8, vacc3x89ABCDEF);
141 c3 = (float*) ((uintptr_t) c3 + cn_stride);
142 _mm256_storeu_ps(c2, vacc2x01234567);
143 _mm256_storeu_ps(c2 + 8, vacc2x89ABCDEF);
144 c2 = (float*) ((uintptr_t) c2 + cn_stride);
145 _mm256_storeu_ps(c1, vacc1x01234567);
146 _mm256_storeu_ps(c1 + 8, vacc1x89ABCDEF);
147 c1 = (float*) ((uintptr_t) c1 + cn_stride);
148 _mm256_storeu_ps(c0, vacc0x01234567);
149 _mm256_storeu_ps(c0 + 8, vacc0x89ABCDEF);
150 c0 = (float*) ((uintptr_t) c0 + cn_stride);
151
152 a4 = (const float*) ((uintptr_t) a4 - kc);
153 a3 = (const float*) ((uintptr_t) a3 - kc);
154 a2 = (const float*) ((uintptr_t) a2 - kc);
155 a1 = (const float*) ((uintptr_t) a1 - kc);
156 a0 = (const float*) ((uintptr_t) a0 - kc);
157
158 nc -= 16;
159 } else {
160 if (nc & 8) {
161 _mm256_storeu_ps(c4, vacc4x01234567);
162 _mm256_storeu_ps(c3, vacc3x01234567);
163 _mm256_storeu_ps(c2, vacc2x01234567);
164 _mm256_storeu_ps(c1, vacc1x01234567);
165 _mm256_storeu_ps(c0, vacc0x01234567);
166
167 vacc4x01234567 = vacc4x89ABCDEF;
168 vacc3x01234567 = vacc3x89ABCDEF;
169 vacc2x01234567 = vacc2x89ABCDEF;
170 vacc1x01234567 = vacc1x89ABCDEF;
171 vacc0x01234567 = vacc0x89ABCDEF;
172
173 c4 += 8;
174 c3 += 8;
175 c2 += 8;
176 c1 += 8;
177 c0 += 8;
178 }
179 __m128 vacc4x0123 = _mm256_castps256_ps128(vacc4x01234567);
180 __m128 vacc3x0123 = _mm256_castps256_ps128(vacc3x01234567);
181 __m128 vacc2x0123 = _mm256_castps256_ps128(vacc2x01234567);
182 __m128 vacc1x0123 = _mm256_castps256_ps128(vacc1x01234567);
183 __m128 vacc0x0123 = _mm256_castps256_ps128(vacc0x01234567);
184 if (nc & 4) {
185 _mm_storeu_ps(c4, vacc4x0123);
186 _mm_storeu_ps(c3, vacc3x0123);
187 _mm_storeu_ps(c2, vacc2x0123);
188 _mm_storeu_ps(c1, vacc1x0123);
189 _mm_storeu_ps(c0, vacc0x0123);
190
191 vacc4x0123 = _mm256_extractf128_ps(vacc4x01234567, 1);
192 vacc3x0123 = _mm256_extractf128_ps(vacc3x01234567, 1);
193 vacc2x0123 = _mm256_extractf128_ps(vacc2x01234567, 1);
194 vacc1x0123 = _mm256_extractf128_ps(vacc1x01234567, 1);
195 vacc0x0123 = _mm256_extractf128_ps(vacc0x01234567, 1);
196
197 c4 += 4;
198 c3 += 4;
199 c2 += 4;
200 c1 += 4;
201 c0 += 4;
202 }
203 if (nc & 2) {
204 _mm_storel_pi((__m64*) c4, vacc4x0123);
205 _mm_storel_pi((__m64*) c3, vacc3x0123);
206 _mm_storel_pi((__m64*) c2, vacc2x0123);
207 _mm_storel_pi((__m64*) c1, vacc1x0123);
208 _mm_storel_pi((__m64*) c0, vacc0x0123);
209
210 vacc4x0123 = _mm_movehl_ps(vacc4x0123, vacc4x0123);
211 vacc3x0123 = _mm_movehl_ps(vacc3x0123, vacc3x0123);
212 vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123);
213 vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
214 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
215
216 c4 += 2;
217 c3 += 2;
218 c2 += 2;
219 c1 += 2;
220 c0 += 2;
221 }
222 if (nc & 1) {
223 _mm_store_ss(c4, vacc4x0123);
224 _mm_store_ss(c3, vacc3x0123);
225 _mm_store_ss(c2, vacc2x0123);
226 _mm_store_ss(c1, vacc1x0123);
227 _mm_store_ss(c0, vacc0x0123);
228 }
229
230 nc = 0;
231 }
232 } while (nc != 0);
233 }
234