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_6x8__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_6x8__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 <= 6);
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 const float* a5 = (const float*) ((uintptr_t) a4 + a_stride);
67 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
68 if XNN_UNPREDICTABLE(mr != 6) {
69 a5 = a4;
70 c5 = c4;
71 }
72
73 do {
74 __m256 vacc0x01234567 = _mm256_load_ps(acc + 0);
75 __m256 vacc1x01234567 = _mm256_load_ps(acc + 8);
76 __m256 vacc2x01234567 = _mm256_load_ps(acc + 16);
77 __m256 vacc3x01234567 = _mm256_load_ps(acc + 24);
78 __m256 vacc4x01234567 = _mm256_load_ps(acc + 32);
79 __m256 vacc5x01234567 = _mm256_load_ps(acc + 40);
80 acc += 48;
81
82 size_t k = kc;
83 do {
84 const __m256 va0 = _mm256_broadcast_ss(a0);
85 a0 += 1;
86 const __m256 va1 = _mm256_broadcast_ss(a1);
87 a1 += 1;
88 const __m256 va2 = _mm256_broadcast_ss(a2);
89 a2 += 1;
90 const __m256 va3 = _mm256_broadcast_ss(a3);
91 a3 += 1;
92 const __m256 va4 = _mm256_broadcast_ss(a4);
93 a4 += 1;
94 const __m256 va5 = _mm256_broadcast_ss(a5);
95 a5 += 1;
96
97 const __m256 vb01234567 = _mm256_load_ps(w);
98 w += 8;
99
100 vacc0x01234567 = _mm256_add_ps(vacc0x01234567, _mm256_mul_ps(va0, vb01234567));
101 vacc1x01234567 = _mm256_add_ps(vacc1x01234567, _mm256_mul_ps(va1, vb01234567));
102 vacc2x01234567 = _mm256_add_ps(vacc2x01234567, _mm256_mul_ps(va2, vb01234567));
103 vacc3x01234567 = _mm256_add_ps(vacc3x01234567, _mm256_mul_ps(va3, vb01234567));
104 vacc4x01234567 = _mm256_add_ps(vacc4x01234567, _mm256_mul_ps(va4, vb01234567));
105 vacc5x01234567 = _mm256_add_ps(vacc5x01234567, _mm256_mul_ps(va5, vb01234567));
106
107 k -= sizeof(float);
108 } while (k != 0);
109
110 const __m256 vmax = _mm256_broadcast_ps((const __m128*) params->sse.max);
111 vacc0x01234567 = _mm256_min_ps(vacc0x01234567, vmax);
112 vacc1x01234567 = _mm256_min_ps(vacc1x01234567, vmax);
113 vacc2x01234567 = _mm256_min_ps(vacc2x01234567, vmax);
114 vacc3x01234567 = _mm256_min_ps(vacc3x01234567, vmax);
115 vacc4x01234567 = _mm256_min_ps(vacc4x01234567, vmax);
116 vacc5x01234567 = _mm256_min_ps(vacc5x01234567, vmax);
117
118 const __m256 vmin = _mm256_broadcast_ps((const __m128*) params->sse.min);
119 vacc0x01234567 = _mm256_max_ps(vacc0x01234567, vmin);
120 vacc1x01234567 = _mm256_max_ps(vacc1x01234567, vmin);
121 vacc2x01234567 = _mm256_max_ps(vacc2x01234567, vmin);
122 vacc3x01234567 = _mm256_max_ps(vacc3x01234567, vmin);
123 vacc4x01234567 = _mm256_max_ps(vacc4x01234567, vmin);
124 vacc5x01234567 = _mm256_max_ps(vacc5x01234567, vmin);
125
126 if XNN_LIKELY(nc >= 8) {
127 _mm256_storeu_ps(c5, vacc5x01234567);
128 c5 = (float*) ((uintptr_t) c5 + cn_stride);
129 _mm256_storeu_ps(c4, vacc4x01234567);
130 c4 = (float*) ((uintptr_t) c4 + cn_stride);
131 _mm256_storeu_ps(c3, vacc3x01234567);
132 c3 = (float*) ((uintptr_t) c3 + cn_stride);
133 _mm256_storeu_ps(c2, vacc2x01234567);
134 c2 = (float*) ((uintptr_t) c2 + cn_stride);
135 _mm256_storeu_ps(c1, vacc1x01234567);
136 c1 = (float*) ((uintptr_t) c1 + cn_stride);
137 _mm256_storeu_ps(c0, vacc0x01234567);
138 c0 = (float*) ((uintptr_t) c0 + cn_stride);
139
140 a5 = (const float*) ((uintptr_t) a5 - kc);
141 a4 = (const float*) ((uintptr_t) a4 - kc);
142 a3 = (const float*) ((uintptr_t) a3 - kc);
143 a2 = (const float*) ((uintptr_t) a2 - kc);
144 a1 = (const float*) ((uintptr_t) a1 - kc);
145 a0 = (const float*) ((uintptr_t) a0 - kc);
146
147 nc -= 8;
148 } else {
149 __m128 vacc5x0123 = _mm256_castps256_ps128(vacc5x01234567);
150 __m128 vacc4x0123 = _mm256_castps256_ps128(vacc4x01234567);
151 __m128 vacc3x0123 = _mm256_castps256_ps128(vacc3x01234567);
152 __m128 vacc2x0123 = _mm256_castps256_ps128(vacc2x01234567);
153 __m128 vacc1x0123 = _mm256_castps256_ps128(vacc1x01234567);
154 __m128 vacc0x0123 = _mm256_castps256_ps128(vacc0x01234567);
155 if (nc & 4) {
156 _mm_storeu_ps(c5, vacc5x0123);
157 _mm_storeu_ps(c4, vacc4x0123);
158 _mm_storeu_ps(c3, vacc3x0123);
159 _mm_storeu_ps(c2, vacc2x0123);
160 _mm_storeu_ps(c1, vacc1x0123);
161 _mm_storeu_ps(c0, vacc0x0123);
162
163 vacc5x0123 = _mm256_extractf128_ps(vacc5x01234567, 1);
164 vacc4x0123 = _mm256_extractf128_ps(vacc4x01234567, 1);
165 vacc3x0123 = _mm256_extractf128_ps(vacc3x01234567, 1);
166 vacc2x0123 = _mm256_extractf128_ps(vacc2x01234567, 1);
167 vacc1x0123 = _mm256_extractf128_ps(vacc1x01234567, 1);
168 vacc0x0123 = _mm256_extractf128_ps(vacc0x01234567, 1);
169
170 c5 += 4;
171 c4 += 4;
172 c3 += 4;
173 c2 += 4;
174 c1 += 4;
175 c0 += 4;
176 }
177 if (nc & 2) {
178 _mm_storel_pi((__m64*) c5, vacc5x0123);
179 _mm_storel_pi((__m64*) c4, vacc4x0123);
180 _mm_storel_pi((__m64*) c3, vacc3x0123);
181 _mm_storel_pi((__m64*) c2, vacc2x0123);
182 _mm_storel_pi((__m64*) c1, vacc1x0123);
183 _mm_storel_pi((__m64*) c0, vacc0x0123);
184
185 vacc5x0123 = _mm_movehl_ps(vacc5x0123, vacc5x0123);
186 vacc4x0123 = _mm_movehl_ps(vacc4x0123, vacc4x0123);
187 vacc3x0123 = _mm_movehl_ps(vacc3x0123, vacc3x0123);
188 vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123);
189 vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
190 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
191
192 c5 += 2;
193 c4 += 2;
194 c3 += 2;
195 c2 += 2;
196 c1 += 2;
197 c0 += 2;
198 }
199 if (nc & 1) {
200 _mm_store_ss(c5, vacc5x0123);
201 _mm_store_ss(c4, vacc4x0123);
202 _mm_store_ss(c3, vacc3x0123);
203 _mm_store_ss(c2, vacc2x0123);
204 _mm_store_ss(c1, vacc1x0123);
205 _mm_store_ss(c0, vacc0x0123);
206 }
207
208 nc = 0;
209 }
210 } while (nc != 0);
211 }
212