File size: 6,547 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
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
// 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/gavgpool.h>
#include <xnnpack/math.h>


void xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_arm_c4(
    size_t rows,
    size_t channels,
    const float* input,
    size_t input_stride,
    const float* zero,
    float* buffer,
    float* output,
    const union xnn_f32_scaleminmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(rows > 7);
  assert(channels != 0);

  const float* i0 = input;
  const float* i1 = (const float*) ((uintptr_t) i0 + input_stride);
  const float* i2 = (const float*) ((uintptr_t) i1 + input_stride);
  const float* i3 = (const float*) ((uintptr_t) i2 + input_stride);
  const float* i4 = (const float*) ((uintptr_t) i3 + input_stride);
  const float* i5 = (const float*) ((uintptr_t) i4 + input_stride);
  const float* i6 = (const float*) ((uintptr_t) i5 + input_stride);
  const size_t packed_channels = round_up_po2(channels, 4);
  const size_t input_increment = 7 * input_stride - packed_channels * sizeof(float);

  float* b = buffer;
  for (size_t c = 0; c < channels; c += 4) {
    const v128_t vi0 = wasm_v128_load(i0);
    i0 += 4;
    const v128_t vi1 = wasm_v128_load(i1);
    i1 += 4;
    const v128_t vi2 = wasm_v128_load(i2);
    i2 += 4;
    const v128_t vi3 = wasm_v128_load(i3);
    i3 += 4;
    const v128_t vi4 = wasm_v128_load(i4);
    i4 += 4;
    const v128_t vi5 = wasm_v128_load(i5);
    i5 += 4;
    const v128_t vi6 = wasm_v128_load(i6);
    i6 += 4;

    const v128_t vsum01 = wasm_f32x4_add(vi0, vi1);
    const v128_t vsum23 = wasm_f32x4_add(vi2, vi3);
    const v128_t vsum45 = wasm_f32x4_add(vi4, vi5);

    const v128_t vsum016 = wasm_f32x4_add(vsum01, vi6);
    const v128_t vsum2345 = wasm_f32x4_add(vsum23, vsum45);

    const v128_t vsum = wasm_f32x4_add(vsum016, vsum2345);

    wasm_v128_store(b, vsum);
    b += 4;
  }
  for (rows -= 7; rows > 7; rows -= 7) {
    b = buffer;

    i0 = (const float*) ((uintptr_t) i0 + input_increment);
    i1 = (const float*) ((uintptr_t) i1 + input_increment);
    i2 = (const float*) ((uintptr_t) i2 + input_increment);
    i3 = (const float*) ((uintptr_t) i3 + input_increment);
    i4 = (const float*) ((uintptr_t) i4 + input_increment);
    i5 = (const float*) ((uintptr_t) i5 + input_increment);
    i6 = (const float*) ((uintptr_t) i6 + input_increment);

    for (size_t c = 0; c < channels; c += 4) {
      const v128_t vi0 = wasm_v128_load(i0);
      i0 += 4;
      const v128_t vi1 = wasm_v128_load(i1);
      i1 += 4;
      const v128_t vi2 = wasm_v128_load(i2);
      i2 += 4;
      const v128_t vi3 = wasm_v128_load(i3);
      i3 += 4;
      const v128_t vi4 = wasm_v128_load(i4);
      i4 += 4;
      const v128_t vi5 = wasm_v128_load(i5);
      i5 += 4;
      const v128_t vi6 = wasm_v128_load(i6);
      i6 += 4;
      const v128_t vacc = wasm_v128_load(b);

      const v128_t vsum01 = wasm_f32x4_add(vi0, vi1);
      const v128_t vsum23 = wasm_f32x4_add(vi2, vi3);
      const v128_t vsum45 = wasm_f32x4_add(vi4, vi5);
      const v128_t vsum6a = wasm_f32x4_add(vi6, vacc);

      const v128_t vsum0123 = wasm_f32x4_add(vsum01, vsum23);
      const v128_t vsum456a = wasm_f32x4_add(vsum45, vsum6a);

      const v128_t vsum = wasm_f32x4_add(vsum0123, vsum456a);

      wasm_v128_store(b, vsum); b += 4;
    }
  }

  i0 = (const float*) ((uintptr_t) i0 + input_increment);
  i1 = (const float*) ((uintptr_t) i1 + input_increment);
  if (rows < 2) {
    i1 = zero;
  }
  i2 = (const float*) ((uintptr_t) i2 + input_increment);
  if (rows <= 2) {
    i2 = zero;
  }
  i3 = (const float*) ((uintptr_t) i3 + input_increment);
  if (rows < 4) {
    i3 = zero;
  }
  i4 = (const float*) ((uintptr_t) i4 + input_increment);
  if (rows <= 4) {
    i4 = zero;
  }
  i5 = (const float*) ((uintptr_t) i5 + input_increment);
  if (rows < 6) {
    i5 = zero;
  }
  i6 = (const float*) ((uintptr_t) i6 + input_increment);
  if (rows <= 6) {
    i6 = zero;
  }
  const v128_t vscale = wasm_v128_load32_splat(&params->scalar.scale);
  const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
  const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);

  b = buffer;
  while (channels >= 4) {
    const v128_t vi0 = wasm_v128_load(i0);
    i0 += 4;
    const v128_t vi1 = wasm_v128_load(i1);
    i1 += 4;
    const v128_t vi2 = wasm_v128_load(i2);
    i2 += 4;
    const v128_t vi3 = wasm_v128_load(i3);
    i3 += 4;
    const v128_t vi4 = wasm_v128_load(i4);
    i4 += 4;
    const v128_t vi5 = wasm_v128_load(i5);
    i5 += 4;
    const v128_t vi6 = wasm_v128_load(i6);
    i6 += 4;
    const v128_t vacc = wasm_v128_load(b);
    b += 4;

    const v128_t vsum01 = wasm_f32x4_add(vi0, vi1);
    const v128_t vsum23 = wasm_f32x4_add(vi2, vi3);
    const v128_t vsum45 = wasm_f32x4_add(vi4, vi5);
    const v128_t vsum6a = wasm_f32x4_add(vi6, vacc);

    const v128_t vsum0123 = wasm_f32x4_add(vsum01, vsum23);
    const v128_t vsum456a = wasm_f32x4_add(vsum45, vsum6a);

    const v128_t vsum = wasm_f32x4_add(vsum0123, vsum456a);

    v128_t vout = wasm_f32x4_mul(vsum, vscale);
    vout = wasm_f32x4_max(vout, vmin);
    vout = wasm_f32x4_min(vout, vmax);

    wasm_v128_store(output, vout);
    output += 4;

    channels -= 4;
  }
  if (channels != 0) {
    const v128_t vi0 = wasm_v128_load(i0);
    const v128_t vi1 = wasm_v128_load(i1);
    const v128_t vi2 = wasm_v128_load(i2);
    const v128_t vi3 = wasm_v128_load(i3);
    const v128_t vi4 = wasm_v128_load(i4);
    const v128_t vi5 = wasm_v128_load(i5);
    const v128_t vi6 = wasm_v128_load(i6);
    const v128_t vacc = wasm_v128_load(b);

    const v128_t vsum01 = wasm_f32x4_add(vi0, vi1);
    const v128_t vsum23 = wasm_f32x4_add(vi2, vi3);
    const v128_t vsum45 = wasm_f32x4_add(vi4, vi5);
    const v128_t vsum6a = wasm_f32x4_add(vi6, vacc);

    const v128_t vsum0123 = wasm_f32x4_add(vsum01, vsum23);
    const v128_t vsum456a = wasm_f32x4_add(vsum45, vsum6a);

    const v128_t vsum = wasm_f32x4_add(vsum0123, vsum456a);

    v128_t vout = wasm_f32x4_mul(vsum, vscale);
    vout = wasm_f32x4_max(vout, vmin);
    vout = wasm_f32x4_min(vout, vmax);

    if (channels & 2) {
      wasm_v128_store64_lane(output, vout, 0);
      vout = wasm_v64x2_shuffle(vout, vout, 1, 1);
      output += 2;
    }
    if (channels & 1) {
      wasm_v128_store32_lane(output, vout, 0);
      output += 1;
    }
  }
}