1 // Auto-generated file. Do not edit!
2 // Template: src/f32-igemm/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/igemm.h>
15
16
xnn_f32_igemm_minmax_ukernel_5x16__avx_broadcast(size_t mr,size_t nc,size_t kc,size_t ks,const float ** restrict a,const float * restrict w,float * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const float * zero,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_igemm_minmax_ukernel_5x16__avx_broadcast(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 size_t ks,
22 const float**restrict a,
23 const float*restrict w,
24 float*restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 size_t a_offset,
28 const float* zero,
29 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
30 {
31 assert(mr != 0);
32 assert(mr <= 5);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(float) == 0);
36 assert(ks != 0);
37 assert(ks % (5 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(float) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 float* c0 = c;
44 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 c1 = c0;
47 }
48 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
49 if XNN_UNPREDICTABLE(mr <= 2) {
50 c2 = c1;
51 }
52 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
53 if XNN_UNPREDICTABLE(mr < 4) {
54 c3 = c2;
55 }
56 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
57 if XNN_UNPREDICTABLE(mr <= 4) {
58 c4 = c3;
59 }
60
61 do {
62 __m256 vacc0x01234567 = _mm256_load_ps(w);
63 __m256 vacc0x89ABCDEF = _mm256_load_ps(w + 8);
64 __m256 vacc1x01234567 = vacc0x01234567;
65 __m256 vacc1x89ABCDEF = vacc0x89ABCDEF;
66 __m256 vacc2x01234567 = vacc0x01234567;
67 __m256 vacc2x89ABCDEF = vacc0x89ABCDEF;
68 __m256 vacc3x01234567 = vacc0x01234567;
69 __m256 vacc3x89ABCDEF = vacc0x89ABCDEF;
70 __m256 vacc4x01234567 = vacc0x01234567;
71 __m256 vacc4x89ABCDEF = vacc0x89ABCDEF;
72 w += 16;
73
74 size_t p = ks;
75 do {
76 const float* restrict a0 = a[0];
77 assert(a0 != NULL);
78 if XNN_UNPREDICTABLE(a0 != zero) {
79 a0 = (const float*) ((uintptr_t) a0 + a_offset);
80 }
81 const float* restrict a1 = a[1];
82 assert(a1 != NULL);
83 if XNN_UNPREDICTABLE(a1 != zero) {
84 a1 = (const float*) ((uintptr_t) a1 + a_offset);
85 }
86 const float* restrict a2 = a[2];
87 assert(a2 != NULL);
88 if XNN_UNPREDICTABLE(a2 != zero) {
89 a2 = (const float*) ((uintptr_t) a2 + a_offset);
90 }
91 const float* restrict a3 = a[3];
92 assert(a3 != NULL);
93 if XNN_UNPREDICTABLE(a3 != zero) {
94 a3 = (const float*) ((uintptr_t) a3 + a_offset);
95 }
96 const float* restrict a4 = a[4];
97 assert(a4 != NULL);
98 if XNN_UNPREDICTABLE(a4 != zero) {
99 a4 = (const float*) ((uintptr_t) a4 + a_offset);
100 }
101 a += 5;
102
103 size_t k = kc;
104 do {
105 const __m256 vb01234567 = _mm256_load_ps(w);
106 const __m256 vb89ABCDEF = _mm256_load_ps(w + 8);
107 w += 16;
108
109 const __m256 va0 = _mm256_broadcast_ss(a0);
110 a0 += 1;
111 const __m256 va1 = _mm256_broadcast_ss(a1);
112 a1 += 1;
113 const __m256 va2 = _mm256_broadcast_ss(a2);
114 a2 += 1;
115 const __m256 va3 = _mm256_broadcast_ss(a3);
116 a3 += 1;
117 const __m256 va4 = _mm256_broadcast_ss(a4);
118 a4 += 1;
119
120 vacc0x01234567 = _mm256_add_ps(vacc0x01234567, _mm256_mul_ps(va0, vb01234567));
121 vacc0x89ABCDEF = _mm256_add_ps(vacc0x89ABCDEF, _mm256_mul_ps(va0, vb89ABCDEF));
122 vacc1x01234567 = _mm256_add_ps(vacc1x01234567, _mm256_mul_ps(va1, vb01234567));
123 vacc1x89ABCDEF = _mm256_add_ps(vacc1x89ABCDEF, _mm256_mul_ps(va1, vb89ABCDEF));
124 vacc2x01234567 = _mm256_add_ps(vacc2x01234567, _mm256_mul_ps(va2, vb01234567));
125 vacc2x89ABCDEF = _mm256_add_ps(vacc2x89ABCDEF, _mm256_mul_ps(va2, vb89ABCDEF));
126 vacc3x01234567 = _mm256_add_ps(vacc3x01234567, _mm256_mul_ps(va3, vb01234567));
127 vacc3x89ABCDEF = _mm256_add_ps(vacc3x89ABCDEF, _mm256_mul_ps(va3, vb89ABCDEF));
128 vacc4x01234567 = _mm256_add_ps(vacc4x01234567, _mm256_mul_ps(va4, vb01234567));
129 vacc4x89ABCDEF = _mm256_add_ps(vacc4x89ABCDEF, _mm256_mul_ps(va4, vb89ABCDEF));
130 k -= sizeof(float);
131 } while (k != 0);
132 p -= 5 * sizeof(void*);
133 } while (p != 0);
134
135 const __m256 vmax = _mm256_broadcast_ps((const __m128*) params->sse.max);
136 vacc0x01234567 = _mm256_min_ps(vacc0x01234567, vmax);
137 vacc1x01234567 = _mm256_min_ps(vacc1x01234567, vmax);
138 vacc2x01234567 = _mm256_min_ps(vacc2x01234567, vmax);
139 vacc3x01234567 = _mm256_min_ps(vacc3x01234567, vmax);
140 vacc4x01234567 = _mm256_min_ps(vacc4x01234567, vmax);
141 vacc0x89ABCDEF = _mm256_min_ps(vacc0x89ABCDEF, vmax);
142 vacc1x89ABCDEF = _mm256_min_ps(vacc1x89ABCDEF, vmax);
143 vacc2x89ABCDEF = _mm256_min_ps(vacc2x89ABCDEF, vmax);
144 vacc3x89ABCDEF = _mm256_min_ps(vacc3x89ABCDEF, vmax);
145 vacc4x89ABCDEF = _mm256_min_ps(vacc4x89ABCDEF, vmax);
146
147 const __m256 vmin = _mm256_broadcast_ps((const __m128*) params->sse.min);
148 vacc0x01234567 = _mm256_max_ps(vacc0x01234567, vmin);
149 vacc1x01234567 = _mm256_max_ps(vacc1x01234567, vmin);
150 vacc2x01234567 = _mm256_max_ps(vacc2x01234567, vmin);
151 vacc3x01234567 = _mm256_max_ps(vacc3x01234567, vmin);
152 vacc4x01234567 = _mm256_max_ps(vacc4x01234567, vmin);
153 vacc0x89ABCDEF = _mm256_max_ps(vacc0x89ABCDEF, vmin);
154 vacc1x89ABCDEF = _mm256_max_ps(vacc1x89ABCDEF, vmin);
155 vacc2x89ABCDEF = _mm256_max_ps(vacc2x89ABCDEF, vmin);
156 vacc3x89ABCDEF = _mm256_max_ps(vacc3x89ABCDEF, vmin);
157 vacc4x89ABCDEF = _mm256_max_ps(vacc4x89ABCDEF, vmin);
158
159 if XNN_LIKELY(nc >= 16) {
160 _mm256_storeu_ps(c4, vacc4x01234567);
161 _mm256_storeu_ps(c4 + 8, vacc4x89ABCDEF);
162 c4 = (float*) ((uintptr_t) c4 + cn_stride);
163 _mm256_storeu_ps(c3, vacc3x01234567);
164 _mm256_storeu_ps(c3 + 8, vacc3x89ABCDEF);
165 c3 = (float*) ((uintptr_t) c3 + cn_stride);
166 _mm256_storeu_ps(c2, vacc2x01234567);
167 _mm256_storeu_ps(c2 + 8, vacc2x89ABCDEF);
168 c2 = (float*) ((uintptr_t) c2 + cn_stride);
169 _mm256_storeu_ps(c1, vacc1x01234567);
170 _mm256_storeu_ps(c1 + 8, vacc1x89ABCDEF);
171 c1 = (float*) ((uintptr_t) c1 + cn_stride);
172 _mm256_storeu_ps(c0, vacc0x01234567);
173 _mm256_storeu_ps(c0 + 8, vacc0x89ABCDEF);
174 c0 = (float*) ((uintptr_t) c0 + cn_stride);
175
176 a = (const float**restrict) ((uintptr_t) a - ks);
177 nc -= 16;
178 } else {
179 if (nc & 8) {
180 _mm256_storeu_ps(c4, vacc4x01234567);
181 _mm256_storeu_ps(c3, vacc3x01234567);
182 _mm256_storeu_ps(c2, vacc2x01234567);
183 _mm256_storeu_ps(c1, vacc1x01234567);
184 _mm256_storeu_ps(c0, vacc0x01234567);
185
186 vacc4x01234567 = vacc4x89ABCDEF;
187 vacc3x01234567 = vacc3x89ABCDEF;
188 vacc2x01234567 = vacc2x89ABCDEF;
189 vacc1x01234567 = vacc1x89ABCDEF;
190 vacc0x01234567 = vacc0x89ABCDEF;
191
192 c4 += 8;
193 c3 += 8;
194 c2 += 8;
195 c1 += 8;
196 c0 += 8;
197 }
198 __m128 vacc4x0123 = _mm256_castps256_ps128(vacc4x01234567);
199 __m128 vacc3x0123 = _mm256_castps256_ps128(vacc3x01234567);
200 __m128 vacc2x0123 = _mm256_castps256_ps128(vacc2x01234567);
201 __m128 vacc1x0123 = _mm256_castps256_ps128(vacc1x01234567);
202 __m128 vacc0x0123 = _mm256_castps256_ps128(vacc0x01234567);
203 if (nc & 4) {
204 _mm_storeu_ps(c4, vacc4x0123);
205 _mm_storeu_ps(c3, vacc3x0123);
206 _mm_storeu_ps(c2, vacc2x0123);
207 _mm_storeu_ps(c1, vacc1x0123);
208 _mm_storeu_ps(c0, vacc0x0123);
209
210 vacc4x0123 = _mm256_extractf128_ps(vacc4x01234567, 1);
211 vacc3x0123 = _mm256_extractf128_ps(vacc3x01234567, 1);
212 vacc2x0123 = _mm256_extractf128_ps(vacc2x01234567, 1);
213 vacc1x0123 = _mm256_extractf128_ps(vacc1x01234567, 1);
214 vacc0x0123 = _mm256_extractf128_ps(vacc0x01234567, 1);
215
216 c4 += 4;
217 c3 += 4;
218 c2 += 4;
219 c1 += 4;
220 c0 += 4;
221 }
222 if (nc & 2) {
223 _mm_storel_pi((__m64*) c4, vacc4x0123);
224 _mm_storel_pi((__m64*) c3, vacc3x0123);
225 _mm_storel_pi((__m64*) c2, vacc2x0123);
226 _mm_storel_pi((__m64*) c1, vacc1x0123);
227 _mm_storel_pi((__m64*) c0, vacc0x0123);
228
229 vacc4x0123 = _mm_movehl_ps(vacc4x0123, vacc4x0123);
230 vacc3x0123 = _mm_movehl_ps(vacc3x0123, vacc3x0123);
231 vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123);
232 vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
233 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
234
235 c4 += 2;
236 c3 += 2;
237 c2 += 2;
238 c1 += 2;
239 c0 += 2;
240 }
241 if (nc & 1) {
242 _mm_store_ss(c4, vacc4x0123);
243 _mm_store_ss(c3, vacc3x0123);
244 _mm_store_ss(c2, vacc2x0123);
245 _mm_store_ss(c1, vacc1x0123);
246 _mm_store_ss(c0, vacc0x0123);
247 }
248
249 nc = 0;
250 }
251 } while (nc != 0);
252 }
253