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_3x16__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_3x16__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 <= 3);
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
53 do {
54 __m256 vacc0x01234567 = _mm256_load_ps(w + 0);
55 __m256 vacc0x89ABCDEF = _mm256_load_ps(w + 8);
56 __m256 vacc1x01234567 = vacc0x01234567;
57 __m256 vacc1x89ABCDEF = vacc0x89ABCDEF;
58 __m256 vacc2x01234567 = vacc0x01234567;
59 __m256 vacc2x89ABCDEF = vacc0x89ABCDEF;
60 w += 16;
61
62 size_t k = kc;
63 do {
64 const __m256 va0 = _mm256_broadcast_ss(a0);
65 a0 += 1;
66 const __m256 va1 = _mm256_broadcast_ss(a1);
67 a1 += 1;
68 const __m256 va2 = _mm256_broadcast_ss(a2);
69 a2 += 1;
70
71 const __m256 vb01234567 = _mm256_load_ps(w);
72 const __m256 vb89ABCDEF = _mm256_load_ps(w + 8);
73 w += 16;
74
75 vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567, vacc0x01234567);
76 vacc1x01234567 = _mm256_fmadd_ps(va1, vb01234567, vacc1x01234567);
77 vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567, vacc2x01234567);
78 vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEF, vacc0x89ABCDEF);
79 vacc1x89ABCDEF = _mm256_fmadd_ps(va1, vb89ABCDEF, vacc1x89ABCDEF);
80 vacc2x89ABCDEF = _mm256_fmadd_ps(va2, vb89ABCDEF, vacc2x89ABCDEF);
81
82 k -= sizeof(float);
83 } while (k != 0);
84
85 const __m256 vmax = _mm256_broadcast_ps((const __m128*) params->sse.max);
86 vacc0x01234567 = _mm256_min_ps(vacc0x01234567, vmax);
87 vacc1x01234567 = _mm256_min_ps(vacc1x01234567, vmax);
88 vacc2x01234567 = _mm256_min_ps(vacc2x01234567, vmax);
89 vacc0x89ABCDEF = _mm256_min_ps(vacc0x89ABCDEF, vmax);
90 vacc1x89ABCDEF = _mm256_min_ps(vacc1x89ABCDEF, vmax);
91 vacc2x89ABCDEF = _mm256_min_ps(vacc2x89ABCDEF, vmax);
92
93 const __m256 vmin = _mm256_broadcast_ps((const __m128*) params->sse.min);
94 vacc0x01234567 = _mm256_max_ps(vacc0x01234567, vmin);
95 vacc1x01234567 = _mm256_max_ps(vacc1x01234567, vmin);
96 vacc2x01234567 = _mm256_max_ps(vacc2x01234567, vmin);
97 vacc0x89ABCDEF = _mm256_max_ps(vacc0x89ABCDEF, vmin);
98 vacc1x89ABCDEF = _mm256_max_ps(vacc1x89ABCDEF, vmin);
99 vacc2x89ABCDEF = _mm256_max_ps(vacc2x89ABCDEF, vmin);
100
101 if XNN_LIKELY(nc >= 16) {
102 _mm256_storeu_ps(c2, vacc2x01234567);
103 _mm256_storeu_ps(c2 + 8, vacc2x89ABCDEF);
104 c2 = (float*) ((uintptr_t) c2 + cn_stride);
105 _mm256_storeu_ps(c1, vacc1x01234567);
106 _mm256_storeu_ps(c1 + 8, vacc1x89ABCDEF);
107 c1 = (float*) ((uintptr_t) c1 + cn_stride);
108 _mm256_storeu_ps(c0, vacc0x01234567);
109 _mm256_storeu_ps(c0 + 8, vacc0x89ABCDEF);
110 c0 = (float*) ((uintptr_t) c0 + cn_stride);
111
112 a2 = (const float*) ((uintptr_t) a2 - kc);
113 a1 = (const float*) ((uintptr_t) a1 - kc);
114 a0 = (const float*) ((uintptr_t) a0 - kc);
115
116 nc -= 16;
117 } else {
118 if (nc & 8) {
119 _mm256_storeu_ps(c2, vacc2x01234567);
120 _mm256_storeu_ps(c1, vacc1x01234567);
121 _mm256_storeu_ps(c0, vacc0x01234567);
122
123 vacc2x01234567 = vacc2x89ABCDEF;
124 vacc1x01234567 = vacc1x89ABCDEF;
125 vacc0x01234567 = vacc0x89ABCDEF;
126
127 c2 += 8;
128 c1 += 8;
129 c0 += 8;
130 }
131 __m128 vacc2x0123 = _mm256_castps256_ps128(vacc2x01234567);
132 __m128 vacc1x0123 = _mm256_castps256_ps128(vacc1x01234567);
133 __m128 vacc0x0123 = _mm256_castps256_ps128(vacc0x01234567);
134 if (nc & 4) {
135 _mm_storeu_ps(c2, vacc2x0123);
136 _mm_storeu_ps(c1, vacc1x0123);
137 _mm_storeu_ps(c0, vacc0x0123);
138
139 vacc2x0123 = _mm256_extractf128_ps(vacc2x01234567, 1);
140 vacc1x0123 = _mm256_extractf128_ps(vacc1x01234567, 1);
141 vacc0x0123 = _mm256_extractf128_ps(vacc0x01234567, 1);
142
143 c2 += 4;
144 c1 += 4;
145 c0 += 4;
146 }
147 if (nc & 2) {
148 _mm_storel_pi((__m64*) c2, vacc2x0123);
149 _mm_storel_pi((__m64*) c1, vacc1x0123);
150 _mm_storel_pi((__m64*) c0, vacc0x0123);
151
152 vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123);
153 vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
154 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
155
156 c2 += 2;
157 c1 += 2;
158 c0 += 2;
159 }
160 if (nc & 1) {
161 _mm_store_ss(c2, vacc2x0123);
162 _mm_store_ss(c1, vacc1x0123);
163 _mm_store_ss(c0, vacc0x0123);
164 }
165
166 nc = 0;
167 }
168 } while (nc != 0);
169 }
170