blob: 21df21a325b0058ab8be4162542b5181e499afe0 [file] [log] [blame]
XNNPACK Teamb455b122019-09-27 18:10:33 -07001// Auto-generated file. Do not edit!
2// Template: src/f32-gemm/psimd-loadsplat.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 <psimd.h>
13
14#include <xnnpack/gemm.h>
15
16
17void xnn_f32_gemm_ukernel_1x8__psimd_loadsplat(
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,
Marat Dukhaneb09a6b2020-04-08 17:34:32 -070027 const union xnn_f32_minmax_params params[restrict static 1])
XNNPACK Teamb455b122019-09-27 18:10:33 -070028{
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 psimd_f32 vacc0x0123 = psimd_load_f32(w + 0);
43 psimd_f32 vacc0x4567 = psimd_load_f32(w + 4);
44 w += 8;
45
46 size_t k = kc;
47 do {
48 const psimd_f32 va0 = psimd_load_splat_f32(a0);
49 a0 += 1;
50
51 const psimd_f32 vb0123 = psimd_load_f32(w);
52 const psimd_f32 vb4567 = psimd_load_f32(w + 4);
53 w += 8;
54
55 vacc0x0123 = psimd_qfma_f32(vacc0x0123, va0, vb0123);
56 vacc0x4567 = psimd_qfma_f32(vacc0x4567, va0, vb4567);
57
58 k -= sizeof(float);
59 } while (k != 0);
60
61 const psimd_f32 vmax = psimd_load_splat_f32(&params->scalar.max);
62 vacc0x0123 = psimd_min_f32(vacc0x0123, vmax);
63 vacc0x4567 = psimd_min_f32(vacc0x4567, vmax);
64
65 const psimd_f32 vmin = psimd_load_splat_f32(&params->scalar.min);
66 vacc0x0123 = psimd_max_f32(vacc0x0123, vmin);
67 vacc0x4567 = psimd_max_f32(vacc0x4567, vmin);
68
69 if XNN_LIKELY(nc >= 8) {
70 psimd_store_f32(c0, vacc0x0123);
71 psimd_store_f32(c0 + 4, vacc0x4567);
72 c0 = (float*) ((uintptr_t) c0 + cn_stride);
73
74 a0 = (const float*) ((uintptr_t) a0 - kc);
75
76 nc -= 8;
77 } else {
78 if (nc & 4) {
79 psimd_store_f32(c0, vacc0x0123);
80
81 vacc0x0123 = vacc0x4567;
82
83 c0 += 4;
84 }
85 if (nc & 2) {
86 psimd_store2_f32(c0, vacc0x0123);
87
88 vacc0x0123 = psimd_concat_hi_f32(vacc0x0123, vacc0x0123);
89
90 c0 += 2;
91 }
92 if (nc & 1) {
93 psimd_store1_f32(c0, vacc0x0123);
94 }
95
96 nc = 0;
97 }
98 } while (nc != 0);
99}