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_minmax_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 float * restrict acc,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_gemminc_minmax_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 float*restrict acc,
28 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
29 {
30 assert(mr != 0);
31 assert(mr <= 3);
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
55 do {
56 __m256 vacc0x01234567 = _mm256_load_ps(acc + 0);
57 __m256 vacc0x89ABCDEF = _mm256_load_ps(acc + 8);
58 __m256 vacc1x01234567 = _mm256_load_ps(acc + 16);
59 __m256 vacc1x89ABCDEF = _mm256_load_ps(acc + 24);
60 __m256 vacc2x01234567 = _mm256_load_ps(acc + 32);
61 __m256 vacc2x89ABCDEF = _mm256_load_ps(acc + 40);
62 acc += 48;
63
64 size_t k = kc;
65 do {
66 const __m256 va0 = _mm256_broadcast_ss(a0);
67 a0 += 1;
68 const __m256 va1 = _mm256_broadcast_ss(a1);
69 a1 += 1;
70 const __m256 va2 = _mm256_broadcast_ss(a2);
71 a2 += 1;
72
73 const __m256 vb01234567 = _mm256_load_ps(w);
74 const __m256 vb89ABCDEF = _mm256_load_ps(w + 8);
75 w += 16;
76
77 vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567, vacc0x01234567);
78 vacc1x01234567 = _mm256_fmadd_ps(va1, vb01234567, vacc1x01234567);
79 vacc2x01234567 = _mm256_fmadd_ps(va2, vb01234567, vacc2x01234567);
80 vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEF, vacc0x89ABCDEF);
81 vacc1x89ABCDEF = _mm256_fmadd_ps(va1, vb89ABCDEF, vacc1x89ABCDEF);
82 vacc2x89ABCDEF = _mm256_fmadd_ps(va2, vb89ABCDEF, vacc2x89ABCDEF);
83
84 k -= sizeof(float);
85 } while (k != 0);
86
87 const __m256 vmax = _mm256_broadcast_ps((const __m128*) params->sse.max);
88 vacc0x01234567 = _mm256_min_ps(vacc0x01234567, vmax);
89 vacc1x01234567 = _mm256_min_ps(vacc1x01234567, vmax);
90 vacc2x01234567 = _mm256_min_ps(vacc2x01234567, vmax);
91 vacc0x89ABCDEF = _mm256_min_ps(vacc0x89ABCDEF, vmax);
92 vacc1x89ABCDEF = _mm256_min_ps(vacc1x89ABCDEF, vmax);
93 vacc2x89ABCDEF = _mm256_min_ps(vacc2x89ABCDEF, vmax);
94
95 const __m256 vmin = _mm256_broadcast_ps((const __m128*) params->sse.min);
96 vacc0x01234567 = _mm256_max_ps(vacc0x01234567, vmin);
97 vacc1x01234567 = _mm256_max_ps(vacc1x01234567, vmin);
98 vacc2x01234567 = _mm256_max_ps(vacc2x01234567, vmin);
99 vacc0x89ABCDEF = _mm256_max_ps(vacc0x89ABCDEF, vmin);
100 vacc1x89ABCDEF = _mm256_max_ps(vacc1x89ABCDEF, vmin);
101 vacc2x89ABCDEF = _mm256_max_ps(vacc2x89ABCDEF, vmin);
102
103 if XNN_LIKELY(nc >= 16) {
104 _mm256_storeu_ps(c2, vacc2x01234567);
105 _mm256_storeu_ps(c2 + 8, vacc2x89ABCDEF);
106 c2 = (float*) ((uintptr_t) c2 + cn_stride);
107 _mm256_storeu_ps(c1, vacc1x01234567);
108 _mm256_storeu_ps(c1 + 8, vacc1x89ABCDEF);
109 c1 = (float*) ((uintptr_t) c1 + cn_stride);
110 _mm256_storeu_ps(c0, vacc0x01234567);
111 _mm256_storeu_ps(c0 + 8, vacc0x89ABCDEF);
112 c0 = (float*) ((uintptr_t) c0 + cn_stride);
113
114 a2 = (const float*) ((uintptr_t) a2 - kc);
115 a1 = (const float*) ((uintptr_t) a1 - kc);
116 a0 = (const float*) ((uintptr_t) a0 - kc);
117
118 nc -= 16;
119 } else {
120 if (nc & 8) {
121 _mm256_storeu_ps(c2, vacc2x01234567);
122 _mm256_storeu_ps(c1, vacc1x01234567);
123 _mm256_storeu_ps(c0, vacc0x01234567);
124
125 vacc2x01234567 = vacc2x89ABCDEF;
126 vacc1x01234567 = vacc1x89ABCDEF;
127 vacc0x01234567 = vacc0x89ABCDEF;
128
129 c2 += 8;
130 c1 += 8;
131 c0 += 8;
132 }
133 __m128 vacc2x0123 = _mm256_castps256_ps128(vacc2x01234567);
134 __m128 vacc1x0123 = _mm256_castps256_ps128(vacc1x01234567);
135 __m128 vacc0x0123 = _mm256_castps256_ps128(vacc0x01234567);
136 if (nc & 4) {
137 _mm_storeu_ps(c2, vacc2x0123);
138 _mm_storeu_ps(c1, vacc1x0123);
139 _mm_storeu_ps(c0, vacc0x0123);
140
141 vacc2x0123 = _mm256_extractf128_ps(vacc2x01234567, 1);
142 vacc1x0123 = _mm256_extractf128_ps(vacc1x01234567, 1);
143 vacc0x0123 = _mm256_extractf128_ps(vacc0x01234567, 1);
144
145 c2 += 4;
146 c1 += 4;
147 c0 += 4;
148 }
149 if (nc & 2) {
150 _mm_storel_pi((__m64*) c2, vacc2x0123);
151 _mm_storel_pi((__m64*) c1, vacc1x0123);
152 _mm_storel_pi((__m64*) c0, vacc0x0123);
153
154 vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123);
155 vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
156 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
157
158 c2 += 2;
159 c1 += 2;
160 c0 += 2;
161 }
162 if (nc & 1) {
163 _mm_store_ss(c2, vacc2x0123);
164 _mm_store_ss(c1, vacc1x0123);
165 _mm_store_ss(c0, vacc0x0123);
166 }
167
168 nc = 0;
169 }
170 } while (nc != 0);
171 }
172