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_ukernel_7x8__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_output_params params[restrict static1])17 void xnn_f32_gemm_ukernel_7x8__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_output_params params[restrict static 1])
28 {
29 assert(mr != 0);
30 assert(mr <= 7);
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
77 do {
78 __m256 vacc0x01234567 = _mm256_load_ps(w + 0);
79 __m256 vacc1x01234567 = vacc0x01234567;
80 __m256 vacc2x01234567 = vacc0x01234567;
81 __m256 vacc3x01234567 = vacc0x01234567;
82 __m256 vacc4x01234567 = vacc0x01234567;
83 __m256 vacc5x01234567 = vacc0x01234567;
84 __m256 vacc6x01234567 = vacc0x01234567;
85 w += 8;
86
87 size_t k = kc;
88 do {
89 const __m256 va0 = _mm256_broadcast_ss(a0);
90 a0 += 1;
91 const __m256 va1 = _mm256_broadcast_ss(a1);
92 a1 += 1;
93 const __m256 va2 = _mm256_broadcast_ss(a2);
94 a2 += 1;
95 const __m256 va3 = _mm256_broadcast_ss(a3);
96 a3 += 1;
97 const __m256 va4 = _mm256_broadcast_ss(a4);
98 a4 += 1;
99 const __m256 va5 = _mm256_broadcast_ss(a5);
100 a5 += 1;
101 const __m256 va6 = _mm256_broadcast_ss(a6);
102 a6 += 1;
103
104 const __m256 vb01234567 = _mm256_load_ps(w);
105 w += 8;
106
107 vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567, vacc0x01234567);
108 vacc1x01234567 = _mm256_fmadd_ps(va1, vb01234567, vacc1x01234567);
109 vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567, vacc2x01234567);
110 vacc3x01234567 = _mm256_fmadd_ps(va3, vb01234567, vacc3x01234567);
111 vacc4x01234567 = _mm256_fmadd_ps(va4, vb01234567, vacc4x01234567);
112 vacc5x01234567 = _mm256_fmadd_ps(va5, vb01234567, vacc5x01234567);
113 vacc6x01234567 = _mm256_fmadd_ps(va6, vb01234567, vacc6x01234567);
114
115 k -= sizeof(float);
116 } while (k != 0);
117
118 const __m256 vmax = _mm256_broadcast_ps((const __m128*) params->sse.max);
119 vacc0x01234567 = _mm256_min_ps(vacc0x01234567, vmax);
120 vacc1x01234567 = _mm256_min_ps(vacc1x01234567, vmax);
121 vacc2x01234567 = _mm256_min_ps(vacc2x01234567, vmax);
122 vacc3x01234567 = _mm256_min_ps(vacc3x01234567, vmax);
123 vacc4x01234567 = _mm256_min_ps(vacc4x01234567, vmax);
124 vacc5x01234567 = _mm256_min_ps(vacc5x01234567, vmax);
125 vacc6x01234567 = _mm256_min_ps(vacc6x01234567, vmax);
126
127 const __m256 vmin = _mm256_broadcast_ps((const __m128*) params->sse.min);
128 vacc0x01234567 = _mm256_max_ps(vacc0x01234567, vmin);
129 vacc1x01234567 = _mm256_max_ps(vacc1x01234567, vmin);
130 vacc2x01234567 = _mm256_max_ps(vacc2x01234567, vmin);
131 vacc3x01234567 = _mm256_max_ps(vacc3x01234567, vmin);
132 vacc4x01234567 = _mm256_max_ps(vacc4x01234567, vmin);
133 vacc5x01234567 = _mm256_max_ps(vacc5x01234567, vmin);
134 vacc6x01234567 = _mm256_max_ps(vacc6x01234567, vmin);
135
136 if XNN_LIKELY(nc >= 8) {
137 _mm256_storeu_ps(c6, vacc6x01234567);
138 c6 = (float*) ((uintptr_t) c6 + cn_stride);
139 _mm256_storeu_ps(c5, vacc5x01234567);
140 c5 = (float*) ((uintptr_t) c5 + cn_stride);
141 _mm256_storeu_ps(c4, vacc4x01234567);
142 c4 = (float*) ((uintptr_t) c4 + cn_stride);
143 _mm256_storeu_ps(c3, vacc3x01234567);
144 c3 = (float*) ((uintptr_t) c3 + cn_stride);
145 _mm256_storeu_ps(c2, vacc2x01234567);
146 c2 = (float*) ((uintptr_t) c2 + cn_stride);
147 _mm256_storeu_ps(c1, vacc1x01234567);
148 c1 = (float*) ((uintptr_t) c1 + cn_stride);
149 _mm256_storeu_ps(c0, vacc0x01234567);
150 c0 = (float*) ((uintptr_t) c0 + cn_stride);
151
152 a6 = (const float*) ((uintptr_t) a6 - kc);
153 a5 = (const float*) ((uintptr_t) a5 - kc);
154 a4 = (const float*) ((uintptr_t) a4 - kc);
155 a3 = (const float*) ((uintptr_t) a3 - kc);
156 a2 = (const float*) ((uintptr_t) a2 - kc);
157 a1 = (const float*) ((uintptr_t) a1 - kc);
158 a0 = (const float*) ((uintptr_t) a0 - kc);
159
160 nc -= 8;
161 } else {
162 __m128 vacc6x0123 = _mm256_castps256_ps128(vacc6x01234567);
163 __m128 vacc5x0123 = _mm256_castps256_ps128(vacc5x01234567);
164 __m128 vacc4x0123 = _mm256_castps256_ps128(vacc4x01234567);
165 __m128 vacc3x0123 = _mm256_castps256_ps128(vacc3x01234567);
166 __m128 vacc2x0123 = _mm256_castps256_ps128(vacc2x01234567);
167 __m128 vacc1x0123 = _mm256_castps256_ps128(vacc1x01234567);
168 __m128 vacc0x0123 = _mm256_castps256_ps128(vacc0x01234567);
169 if (nc & 4) {
170 _mm_storeu_ps(c6, vacc6x0123);
171 _mm_storeu_ps(c5, vacc5x0123);
172 _mm_storeu_ps(c4, vacc4x0123);
173 _mm_storeu_ps(c3, vacc3x0123);
174 _mm_storeu_ps(c2, vacc2x0123);
175 _mm_storeu_ps(c1, vacc1x0123);
176 _mm_storeu_ps(c0, vacc0x0123);
177
178 vacc6x0123 = _mm256_extractf128_ps(vacc6x01234567, 1);
179 vacc5x0123 = _mm256_extractf128_ps(vacc5x01234567, 1);
180 vacc4x0123 = _mm256_extractf128_ps(vacc4x01234567, 1);
181 vacc3x0123 = _mm256_extractf128_ps(vacc3x01234567, 1);
182 vacc2x0123 = _mm256_extractf128_ps(vacc2x01234567, 1);
183 vacc1x0123 = _mm256_extractf128_ps(vacc1x01234567, 1);
184 vacc0x0123 = _mm256_extractf128_ps(vacc0x01234567, 1);
185
186 c6 += 4;
187 c5 += 4;
188 c4 += 4;
189 c3 += 4;
190 c2 += 4;
191 c1 += 4;
192 c0 += 4;
193 }
194 if (nc & 2) {
195 _mm_storel_pi((__m64*) c6, vacc6x0123);
196 _mm_storel_pi((__m64*) c5, vacc5x0123);
197 _mm_storel_pi((__m64*) c4, vacc4x0123);
198 _mm_storel_pi((__m64*) c3, vacc3x0123);
199 _mm_storel_pi((__m64*) c2, vacc2x0123);
200 _mm_storel_pi((__m64*) c1, vacc1x0123);
201 _mm_storel_pi((__m64*) c0, vacc0x0123);
202
203 vacc6x0123 = _mm_movehl_ps(vacc6x0123, vacc6x0123);
204 vacc5x0123 = _mm_movehl_ps(vacc5x0123, vacc5x0123);
205 vacc4x0123 = _mm_movehl_ps(vacc4x0123, vacc4x0123);
206 vacc3x0123 = _mm_movehl_ps(vacc3x0123, vacc3x0123);
207 vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123);
208 vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
209 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
210
211 c6 += 2;
212 c5 += 2;
213 c4 += 2;
214 c3 += 2;
215 c2 += 2;
216 c1 += 2;
217 c0 += 2;
218 }
219 if (nc & 1) {
220 _mm_store_ss(c6, vacc6x0123);
221 _mm_store_ss(c5, vacc5x0123);
222 _mm_store_ss(c4, vacc4x0123);
223 _mm_store_ss(c3, vacc3x0123);
224 _mm_store_ss(c2, vacc2x0123);
225 _mm_store_ss(c1, vacc1x0123);
226 _mm_store_ss(c0, vacc0x0123);
227 }
228
229 nc = 0;
230 }
231 } while (nc != 0);
232 }
233