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_1x8__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_1x8__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 w += 8;
44
45 size_t k = kc;
46 do {
47 const __m256 va0 = _mm256_broadcast_ss(a0);
48 a0 += 1;
49
50 const __m256 vb01234567 = _mm256_load_ps(w);
51 w += 8;
52
53 vacc0x01234567 = _mm256_add_ps(vacc0x01234567, _mm256_mul_ps(va0, vb01234567));
54
55 k -= sizeof(float);
56 } while (k != 0);
57
58 const __m256 vmax = _mm256_broadcast_ps((const __m128*) params->sse.max);
59 vacc0x01234567 = _mm256_min_ps(vacc0x01234567, vmax);
60
61 const __m256 vmin = _mm256_broadcast_ps((const __m128*) params->sse.min);
62 vacc0x01234567 = _mm256_max_ps(vacc0x01234567, vmin);
63
64 if XNN_LIKELY(nc >= 8) {
65 _mm256_storeu_ps(c0, vacc0x01234567);
66 c0 = (float*) ((uintptr_t) c0 + cn_stride);
67
68 a0 = (const float*) ((uintptr_t) a0 - kc);
69
70 nc -= 8;
71 } else {
72 __m128 vacc0x0123 = _mm256_castps256_ps128(vacc0x01234567);
73 if (nc & 4) {
74 _mm_storeu_ps(c0, vacc0x0123);
75
76 vacc0x0123 = _mm256_extractf128_ps(vacc0x01234567, 1);
77
78 c0 += 4;
79 }
80 if (nc & 2) {
81 _mm_storel_pi((__m64*) c0, vacc0x0123);
82
83 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
84
85 c0 += 2;
86 }
87 if (nc & 1) {
88 _mm_store_ss(c0, vacc0x0123);
89 }
90
91 nc = 0;
92 }
93 } while (nc != 0);
94 }
95