File size: 3,979 Bytes
8b7c501
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
// Auto-generated file. Do not edit!
//   Template: src/f32-gemm/wasmsimd-splat.c.in
//   Generator: tools/xngen
//
// Copyright 2020 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.

#include <assert.h>

#include <wasm_simd128.h>

#include <xnnpack/gemm.h>


void xnn_f32_gemm_minmax_ukernel_1x8__wasmsimd_arm_splat(
    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)])
{
  assert(mr != 0);
  assert(mr <= 1);
  assert(nc != 0);
  assert(kc != 0);
  assert(kc % sizeof(float) == 0);
  assert(a != NULL);
  assert(w != NULL);
  assert(c != NULL);

  const float* a0 = a;
  float* c0 = c;

  const v128_t vmin = wasm_v128_load64_splat(params->wasmsimd.min);
  const v128_t vmax = wasm_v128_load64_splat(params->wasmsimd.max);
  do {
    v128_t vacc0x0123 = wasm_v128_load(w + 0);
    v128_t vacc0x4567 = wasm_v128_load(w + 4);
    w += 8;

    size_t k = kc;
    while (k >= 4 * sizeof(float)) {
      const v128_t va0 = wasm_v128_load(a0);
      a0 += 4;

      const v128_t va0c0 = wasm_v32x4_shuffle(va0, va0, 0, 0, 0, 0);

      const v128_t vb0123c0 = wasm_v128_load(w + 0);
      const v128_t vb4567c0 = wasm_v128_load(w + 4);

      vacc0x0123 = wasm_f32x4_add(wasm_f32x4_mul(va0c0, vb0123c0), vacc0x0123);
      vacc0x4567 = wasm_f32x4_add(wasm_f32x4_mul(va0c0, vb4567c0), vacc0x4567);
      const v128_t va0c1 = wasm_v32x4_shuffle(va0, va0, 1, 1, 1, 1);

      const v128_t vb0123c1 = wasm_v128_load(w + 8);
      const v128_t vb4567c1 = wasm_v128_load(w + 12);

      vacc0x0123 = wasm_f32x4_add(wasm_f32x4_mul(va0c1, vb0123c1), vacc0x0123);
      vacc0x4567 = wasm_f32x4_add(wasm_f32x4_mul(va0c1, vb4567c1), vacc0x4567);
      const v128_t va0c2 = wasm_v32x4_shuffle(va0, va0, 2, 2, 2, 2);

      const v128_t vb0123c2 = wasm_v128_load(w + 16);
      const v128_t vb4567c2 = wasm_v128_load(w + 20);

      vacc0x0123 = wasm_f32x4_add(wasm_f32x4_mul(va0c2, vb0123c2), vacc0x0123);
      vacc0x4567 = wasm_f32x4_add(wasm_f32x4_mul(va0c2, vb4567c2), vacc0x4567);
      const v128_t va0c3 = wasm_v32x4_shuffle(va0, va0, 3, 3, 3, 3);

      const v128_t vb0123c3 = wasm_v128_load(w + 24);
      const v128_t vb4567c3 = wasm_v128_load(w + 28);

      vacc0x0123 = wasm_f32x4_add(wasm_f32x4_mul(va0c3, vb0123c3), vacc0x0123);
      vacc0x4567 = wasm_f32x4_add(wasm_f32x4_mul(va0c3, vb4567c3), vacc0x4567);
      w += 32;
      k -= 4 * sizeof(float);
    }
    if XNN_UNLIKELY(k != 0) {
      do {
        const v128_t va0 = wasm_v128_load32_splat(a0);
        a0 += 1;

        const v128_t vb0123 = wasm_v128_load(w);
        const v128_t vb4567 = wasm_v128_load(w + 4);
        w += 8;

        vacc0x0123 = wasm_f32x4_add(wasm_f32x4_mul(va0, vb0123), vacc0x0123);
        vacc0x4567 = wasm_f32x4_add(wasm_f32x4_mul(va0, vb4567), vacc0x4567);

        k -= sizeof(float);
      } while (k != 0);
    }

    vacc0x0123 = wasm_f32x4_max(vmin, vacc0x0123);
    vacc0x4567 = wasm_f32x4_max(vmin, vacc0x4567);

    vacc0x0123 = wasm_f32x4_min(vmax, vacc0x0123);
    vacc0x4567 = wasm_f32x4_min(vmax, vacc0x4567);

    if XNN_LIKELY(nc >= 8) {
      wasm_v128_store(c0, vacc0x0123);
      wasm_v128_store(c0 + 4, vacc0x4567);
      c0 = (float*) ((uintptr_t) c0 + cn_stride);

      a0 = (const float*) ((uintptr_t) a0 - kc);

      nc -= 8;
    } else {
      if (nc & 4) {
        wasm_v128_store(c0, vacc0x0123);

        vacc0x0123 = vacc0x4567;

        c0 += 4;
      }
      if (nc & 2) {
        wasm_v128_store64_lane(c0, vacc0x0123, 0);

        vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);

        c0 += 2;
      }
      if (nc & 1) {
        wasm_v128_store32_lane(c0, vacc0x0123, 0);
      }

      nc = 0;
    }
  } while (nc != 0);
}