1 // Auto-generated file. Do not edit!
2 // Template: src/f32-igemm/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/igemm.h>
15
16
xnn_f32_igemm_minmax_ukernel_1x16__fma3_broadcast(size_t mr,size_t nc,size_t kc,size_t ks,const float ** restrict a,const float * restrict w,float * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const float * zero,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_igemm_minmax_ukernel_1x16__fma3_broadcast(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 size_t ks,
22 const float**restrict a,
23 const float*restrict w,
24 float*restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 size_t a_offset,
28 const float* zero,
29 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
30 {
31 assert(mr != 0);
32 assert(mr <= 1);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(float) == 0);
36 assert(ks != 0);
37 assert(ks % (1 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(float) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 float* c0 = c;
44
45 do {
46 __m256 vacc0x01234567 = _mm256_load_ps(w);
47 __m256 vacc0x89ABCDEF = _mm256_load_ps(w + 8);
48 w += 16;
49
50 size_t p = ks;
51 do {
52 const float* restrict a0 = a[0];
53 assert(a0 != NULL);
54 if XNN_UNPREDICTABLE(a0 != zero) {
55 a0 = (const float*) ((uintptr_t) a0 + a_offset);
56 }
57 a += 1;
58
59 size_t k = kc;
60 do {
61 const __m256 vb01234567 = _mm256_load_ps(w);
62 const __m256 vb89ABCDEF = _mm256_load_ps(w + 8);
63 w += 16;
64
65 const __m256 va0 = _mm256_broadcast_ss(a0);
66 a0 += 1;
67
68 vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567, vacc0x01234567);
69 vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEF, vacc0x89ABCDEF);
70 k -= sizeof(float);
71 } while (k != 0);
72 p -= 1 * sizeof(void*);
73 } while (p != 0);
74
75 const __m256 vmax = _mm256_broadcast_ps((const __m128*) params->sse.max);
76 vacc0x01234567 = _mm256_min_ps(vacc0x01234567, vmax);
77 vacc0x89ABCDEF = _mm256_min_ps(vacc0x89ABCDEF, vmax);
78
79 const __m256 vmin = _mm256_broadcast_ps((const __m128*) params->sse.min);
80 vacc0x01234567 = _mm256_max_ps(vacc0x01234567, vmin);
81 vacc0x89ABCDEF = _mm256_max_ps(vacc0x89ABCDEF, vmin);
82
83 if XNN_LIKELY(nc >= 16) {
84 _mm256_storeu_ps(c0, vacc0x01234567);
85 _mm256_storeu_ps(c0 + 8, vacc0x89ABCDEF);
86 c0 = (float*) ((uintptr_t) c0 + cn_stride);
87
88 a = (const float**restrict) ((uintptr_t) a - ks);
89 nc -= 16;
90 } else {
91 if (nc & 8) {
92 _mm256_storeu_ps(c0, vacc0x01234567);
93
94 vacc0x01234567 = vacc0x89ABCDEF;
95
96 c0 += 8;
97 }
98 __m128 vacc0x0123 = _mm256_castps256_ps128(vacc0x01234567);
99 if (nc & 4) {
100 _mm_storeu_ps(c0, vacc0x0123);
101
102 vacc0x0123 = _mm256_extractf128_ps(vacc0x01234567, 1);
103
104 c0 += 4;
105 }
106 if (nc & 2) {
107 _mm_storel_pi((__m64*) c0, vacc0x0123);
108
109 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
110
111 c0 += 2;
112 }
113 if (nc & 1) {
114 _mm_store_ss(c0, vacc0x0123);
115 }
116
117 nc = 0;
118 }
119 } while (nc != 0);
120 }
121