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_minmax_ukernel_1x16__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 union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_gemm_minmax_ukernel_1x16__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 union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29 assert(mr != 0);
30 assert(mr <= 1);
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
41 do {
42 __m256 vacc0x01234567 = _mm256_load_ps(w + 0);
43 __m256 vacc0x89ABCDEF = _mm256_load_ps(w + 8);
44 w += 16;
45
46 size_t k = kc;
47 do {
48 const __m256 va0 = _mm256_broadcast_ss(a0);
49 a0 += 1;
50
51 const __m256 vb01234567 = _mm256_load_ps(w);
52 const __m256 vb89ABCDEF = _mm256_load_ps(w + 8);
53 w += 16;
54
55 vacc0x01234567 = _mm256_add_ps(vacc0x01234567, _mm256_mul_ps(va0, vb01234567));
56 vacc0x89ABCDEF = _mm256_add_ps(vacc0x89ABCDEF, _mm256_mul_ps(va0, vb89ABCDEF));
57
58 k -= sizeof(float);
59 } while (k != 0);
60
61 const __m256 vmax = _mm256_broadcast_ps((const __m128*) params->sse.max);
62 vacc0x01234567 = _mm256_min_ps(vacc0x01234567, vmax);
63 vacc0x89ABCDEF = _mm256_min_ps(vacc0x89ABCDEF, vmax);
64
65 const __m256 vmin = _mm256_broadcast_ps((const __m128*) params->sse.min);
66 vacc0x01234567 = _mm256_max_ps(vacc0x01234567, vmin);
67 vacc0x89ABCDEF = _mm256_max_ps(vacc0x89ABCDEF, vmin);
68
69 if XNN_LIKELY(nc >= 16) {
70 _mm256_storeu_ps(c0, vacc0x01234567);
71 _mm256_storeu_ps(c0 + 8, vacc0x89ABCDEF);
72 c0 = (float*) ((uintptr_t) c0 + cn_stride);
73
74 a0 = (const float*) ((uintptr_t) a0 - kc);
75
76 nc -= 16;
77 } else {
78 if (nc & 8) {
79 _mm256_storeu_ps(c0, vacc0x01234567);
80
81 vacc0x01234567 = vacc0x89ABCDEF;
82
83 c0 += 8;
84 }
85 __m128 vacc0x0123 = _mm256_castps256_ps128(vacc0x01234567);
86 if (nc & 4) {
87 _mm_storeu_ps(c0, vacc0x0123);
88
89 vacc0x0123 = _mm256_extractf128_ps(vacc0x01234567, 1);
90
91 c0 += 4;
92 }
93 if (nc & 2) {
94 _mm_storel_pi((__m64*) c0, vacc0x0123);
95
96 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
97
98 c0 += 2;
99 }
100 if (nc & 1) {
101 _mm_store_ss(c0, vacc0x0123);
102 }
103
104 nc = 0;
105 }
106 } while (nc != 0);
107 }
108