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_1x16__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_1x16__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 <= 1);
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
43 do {
44 __m256 vacc0x01234567 = _mm256_load_ps(acc + 0);
45 __m256 vacc0x89ABCDEF = _mm256_load_ps(acc + 8);
46 acc += 16;
47
48 size_t k = kc;
49 do {
50 const __m256 va0 = _mm256_broadcast_ss(a0);
51 a0 += 1;
52
53 const __m256 vb01234567 = _mm256_load_ps(w);
54 const __m256 vb89ABCDEF = _mm256_load_ps(w + 8);
55 w += 16;
56
57 vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567, vacc0x01234567);
58 vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEF, vacc0x89ABCDEF);
59
60 k -= sizeof(float);
61 } while (k != 0);
62
63 const __m256 vmin = _mm256_load_ps(params->avx.min);
64 vacc0x01234567 = _mm256_max_ps(vacc0x01234567, vmin);
65 vacc0x89ABCDEF = _mm256_max_ps(vacc0x89ABCDEF, vmin);
66
67 const __m256 vmax = _mm256_load_ps(params->avx.max);
68 vacc0x01234567 = _mm256_min_ps(vacc0x01234567, vmax);
69 vacc0x89ABCDEF = _mm256_min_ps(vacc0x89ABCDEF, vmax);
70
71 if XNN_LIKELY(nc >= 16) {
72 _mm256_storeu_ps(c0, vacc0x01234567);
73 _mm256_storeu_ps(c0 + 8, vacc0x89ABCDEF);
74 c0 = (float*) ((uintptr_t) c0 + cn_stride);
75
76 a0 = (const float*) ((uintptr_t) a0 - kc);
77
78 nc -= 16;
79 } else {
80 if (nc & 8) {
81 _mm256_storeu_ps(c0, vacc0x01234567);
82
83 vacc0x01234567 = vacc0x89ABCDEF;
84
85 c0 += 8;
86 }
87 __m128 vacc0x0123 = _mm256_castps256_ps128(vacc0x01234567);
88 if (nc & 4) {
89 _mm_storeu_ps(c0, vacc0x0123);
90
91 vacc0x0123 = _mm256_extractf128_ps(vacc0x01234567, 1);
92
93 c0 += 4;
94 }
95 if (nc & 2) {
96 _mm_storel_pi((__m64*) c0, vacc0x0123);
97
98 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
99
100 c0 += 2;
101 }
102 if (nc & 1) {
103 _mm_store_ss(c0, vacc0x0123);
104 }
105
106 nc = 0;
107 }
108 } while (nc != 0);
109 }
110