File size: 13,813 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
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
// Auto-generated file. Do not edit!
//   Template: src/f32-dwconv2d-chw/3x3p1-neon.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 <arm_neon.h>

#include <xnnpack/dwconv.h>
#include <xnnpack/math.h>


void xnn_f32_dwconv2d_chw_ukernel_3x3p1__aarch64_neonfma_4x4(
    size_t input_height,
    size_t input_width,
    const float* input,
    const float* weights,
    const float* zero,
    float* output,
    uint32_t padding_top,
    const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(input_height != 0);
  assert(input_width != 0);
  assert(input_width % sizeof(float) == 0);
  assert(padding_top == 1);

  const uint32x4_t vmask = vld1q_u32(params->neon_stride1.mask);
  const float32x4_t vmax = vld1q_dup_f32(&params->neon_stride1.max);
  const float32x4_t vmin = vld1q_dup_f32(&params->neon_stride1.min);

  const float32x4_t vw0123 = vld1q_f32(weights);
  const float32x4_t vw4567 = vld1q_f32(weights + 4);
  const float32x2_t vw89 = vld1_f32(weights + 8);

  const size_t input_decrement = round_up_po2(input_width, 4 * sizeof(float));

  const float* i0 = zero;
  const float* i1 = input;
  const float* i2 = (const float*) ((uintptr_t) i1 + input_width);
  const float* i3 = (const float*) ((uintptr_t) i2 + input_width);
  const float* i4 = (const float*) ((uintptr_t) i3 + input_width);
  const float* i5 = (const float*) ((uintptr_t) i4 + input_width);

  float* o0 = output;
  float* o1 = (float*) ((uintptr_t) o0 + input_width);
  float* o2 = (float*) ((uintptr_t) o1 + input_width);
  float* o3 = (float*) ((uintptr_t) o2 + input_width);

  size_t output_height = input_height;
  do {
    if XNN_UNPREDICTABLE(output_height < 2) {
      i2 = zero;
      o1 = o0;
    }
    if XNN_UNPREDICTABLE(output_height < 3) {
      i3 = zero;
      o2 = o1;
    }
    if XNN_UNPREDICTABLE(output_height < 4) {
      i4 = zero;
      o3 = o2;
    }
    if XNN_UNPREDICTABLE(output_height < 5) {
      i5 = zero;
    }

    float32x4_t vi0x0123 = vmovq_n_f32(0.0f);
    float32x4_t vi1x0123 = vmovq_n_f32(0.0f);
    float32x4_t vi2x0123 = vmovq_n_f32(0.0f);
    float32x4_t vi3x0123 = vmovq_n_f32(0.0f);
    float32x4_t vi4x0123 = vmovq_n_f32(0.0f);
    float32x4_t vi5x0123 = vmovq_n_f32(0.0f);

    float32x4_t vi0x4567 = vld1q_f32(i0); i0 += 4;
    float32x4_t vi1x4567 = vld1q_f32(i1); i1 += 4;
    float32x4_t vi2x4567 = vld1q_f32(i2); i2 += 4;
    float32x4_t vi3x4567 = vld1q_f32(i3); i3 += 4;
    float32x4_t vi4x4567 = vld1q_f32(i4); i4 += 4;
    float32x4_t vi5x4567 = vld1q_f32(i5); i5 += 4;

    size_t w = input_width;
    for (; w > 4 * sizeof(float); w -= 4 * sizeof(float)) {
      float32x4_t vo0p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
      float32x4_t vo1p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
      float32x4_t vo2p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
      float32x4_t vo3p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);

      const float32x4_t vi0x89AB = vld1q_f32(i0); i0 += 4;
      const float32x4_t vi1x89AB = vld1q_f32(i1); i1 += 4;
      const float32x4_t vi2x89AB = vld1q_f32(i2); i2 += 4;
      const float32x4_t vi3x89AB = vld1q_f32(i3); i3 += 4;
      const float32x4_t vi4x89AB = vld1q_f32(i4); i4 += 4;
      const float32x4_t vi5x89AB = vld1q_f32(i5); i5 += 4;

      vo0p0 = vfmaq_lane_f32(vo0p0, vi0x4567, vget_high_f32(vw0123), 0);
      vo1p0 = vfmaq_lane_f32(vo1p0, vi1x4567, vget_high_f32(vw0123), 0);
      vo2p0 = vfmaq_lane_f32(vo2p0, vi2x4567, vget_high_f32(vw0123), 0);
      vo3p0 = vfmaq_lane_f32(vo3p0, vi3x4567, vget_high_f32(vw0123), 0);

      vo0p0 = vfmaq_lane_f32(vo0p0, vi1x4567, vget_low_f32(vw4567), 1);
      vo1p0 = vfmaq_lane_f32(vo1p0, vi2x4567, vget_low_f32(vw4567), 1);
      vo2p0 = vfmaq_lane_f32(vo2p0, vi3x4567, vget_low_f32(vw4567), 1);
      vo3p0 = vfmaq_lane_f32(vo3p0, vi4x4567, vget_low_f32(vw4567), 1);

      vo0p0 = vfmaq_lane_f32(vo0p0, vi2x4567, vw89, 0);
      vo1p0 = vfmaq_lane_f32(vo1p0, vi3x4567, vw89, 0);
      vo2p0 = vfmaq_lane_f32(vo2p0, vi4x4567, vw89, 0);
      vo3p0 = vfmaq_lane_f32(vo3p0, vi5x4567, vw89, 0);

      const float32x4_t vi0x3456 = vextq_f32(vi0x0123, vi0x4567, 3);
      const float32x4_t vi1x3456 = vextq_f32(vi1x0123, vi1x4567, 3);
      const float32x4_t vi2x3456 = vextq_f32(vi2x0123, vi2x4567, 3);
      const float32x4_t vi3x3456 = vextq_f32(vi3x0123, vi3x4567, 3);
      const float32x4_t vi4x3456 = vextq_f32(vi4x0123, vi4x4567, 3);
      const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3);

      vo0p0 = vfmaq_lane_f32(vo0p0, vi0x3456, vget_low_f32(vw0123), 1);
      vo1p0 = vfmaq_lane_f32(vo1p0, vi1x3456, vget_low_f32(vw0123), 1);
      vo2p0 = vfmaq_lane_f32(vo2p0, vi2x3456, vget_low_f32(vw0123), 1);
      vo3p0 = vfmaq_lane_f32(vo3p0, vi3x3456, vget_low_f32(vw0123), 1);

      vo0p0 = vfmaq_lane_f32(vo0p0, vi1x3456, vget_low_f32(vw4567), 0);
      vo1p0 = vfmaq_lane_f32(vo1p0, vi2x3456, vget_low_f32(vw4567), 0);
      vo2p0 = vfmaq_lane_f32(vo2p0, vi3x3456, vget_low_f32(vw4567), 0);
      vo3p0 = vfmaq_lane_f32(vo3p0, vi4x3456, vget_low_f32(vw4567), 0);

      vo0p0 = vfmaq_lane_f32(vo0p0, vi2x3456, vget_high_f32(vw4567), 1);
      vo1p0 = vfmaq_lane_f32(vo1p0, vi3x3456, vget_high_f32(vw4567), 1);
      vo2p0 = vfmaq_lane_f32(vo2p0, vi4x3456, vget_high_f32(vw4567), 1);
      vo3p0 = vfmaq_lane_f32(vo3p0, vi5x3456, vget_high_f32(vw4567), 1);

      vi0x0123 = vi0x4567;
      vi1x0123 = vi1x4567;
      vi2x0123 = vi2x4567;
      vi3x0123 = vi3x4567;
      vi4x0123 = vi4x4567;
      vi5x0123 = vi5x4567;

      const float32x4_t vi0x5678 = vextq_f32(vi0x4567, vi0x89AB, 1);
      const float32x4_t vi1x5678 = vextq_f32(vi1x4567, vi1x89AB, 1);
      const float32x4_t vi2x5678 = vextq_f32(vi2x4567, vi2x89AB, 1);
      const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vi3x89AB, 1);
      const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vi4x89AB, 1);
      const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vi5x89AB, 1);

      vo0p0 = vfmaq_lane_f32(vo0p0, vi0x5678, vget_high_f32(vw0123), 1);
      vo1p0 = vfmaq_lane_f32(vo1p0, vi1x5678, vget_high_f32(vw0123), 1);
      vo2p0 = vfmaq_lane_f32(vo2p0, vi2x5678, vget_high_f32(vw0123), 1);
      vo3p0 = vfmaq_lane_f32(vo3p0, vi3x5678, vget_high_f32(vw0123), 1);

      vo0p0 = vfmaq_lane_f32(vo0p0, vi1x5678, vget_high_f32(vw4567), 0);
      vo1p0 = vfmaq_lane_f32(vo1p0, vi2x5678, vget_high_f32(vw4567), 0);
      vo2p0 = vfmaq_lane_f32(vo2p0, vi3x5678, vget_high_f32(vw4567), 0);
      vo3p0 = vfmaq_lane_f32(vo3p0, vi4x5678, vget_high_f32(vw4567), 0);

      vo0p0 = vfmaq_lane_f32(vo0p0, vi2x5678, vw89, 1);
      vo1p0 = vfmaq_lane_f32(vo1p0, vi3x5678, vw89, 1);
      vo2p0 = vfmaq_lane_f32(vo2p0, vi4x5678, vw89, 1);
      vo3p0 = vfmaq_lane_f32(vo3p0, vi5x5678, vw89, 1);

      vi0x4567 = vi0x89AB;
      vi1x4567 = vi1x89AB;
      vi2x4567 = vi2x89AB;
      vi3x4567 = vi3x89AB;
      vi4x4567 = vi4x89AB;
      vi5x4567 = vi5x89AB;


      float32x4_t vo0 = vmaxq_f32(vo0p0, vmin);
      float32x4_t vo1 = vmaxq_f32(vo1p0, vmin);
      float32x4_t vo2 = vmaxq_f32(vo2p0, vmin);
      float32x4_t vo3 = vmaxq_f32(vo3p0, vmin);

      vo0 = vminq_f32(vo0, vmax);
      vo1 = vminq_f32(vo1, vmax);
      vo2 = vminq_f32(vo2, vmax);
      vo3 = vminq_f32(vo3, vmax);

      vst1q_f32(o3, vo3); o3 += 4;
      vst1q_f32(o2, vo2); o2 += 4;
      vst1q_f32(o1, vo1); o1 += 4;
      vst1q_f32(o0, vo0); o0 += 4;
    }
    // Always process the last block of 1..4 pixels.
    assert(w >= 1 * sizeof(float));
    assert(w <= 4 * sizeof(float));
    {
      float32x4_t vo0p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
      float32x4_t vo1p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
      float32x4_t vo2p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);
      float32x4_t vo3p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0);

      vi0x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi0x4567)));
      vi1x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi1x4567)));
      vi2x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi2x4567)));
      vi3x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi3x4567)));
      vi4x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi4x4567)));
      vi5x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi5x4567)));

      vo0p0 = vfmaq_lane_f32(vo0p0, vi0x4567, vget_high_f32(vw0123), 0);
      vo1p0 = vfmaq_lane_f32(vo1p0, vi1x4567, vget_high_f32(vw0123), 0);
      vo2p0 = vfmaq_lane_f32(vo2p0, vi2x4567, vget_high_f32(vw0123), 0);
      vo3p0 = vfmaq_lane_f32(vo3p0, vi3x4567, vget_high_f32(vw0123), 0);

      vo0p0 = vfmaq_lane_f32(vo0p0, vi1x4567, vget_low_f32(vw4567), 1);
      vo1p0 = vfmaq_lane_f32(vo1p0, vi2x4567, vget_low_f32(vw4567), 1);
      vo2p0 = vfmaq_lane_f32(vo2p0, vi3x4567, vget_low_f32(vw4567), 1);
      vo3p0 = vfmaq_lane_f32(vo3p0, vi4x4567, vget_low_f32(vw4567), 1);

      vo0p0 = vfmaq_lane_f32(vo0p0, vi2x4567, vw89, 0);
      vo1p0 = vfmaq_lane_f32(vo1p0, vi3x4567, vw89, 0);
      vo2p0 = vfmaq_lane_f32(vo2p0, vi4x4567, vw89, 0);
      vo3p0 = vfmaq_lane_f32(vo3p0, vi5x4567, vw89, 0);

      const float32x4_t vi0x3456 = vextq_f32(vi0x0123, vi0x4567, 3);
      const float32x4_t vi1x3456 = vextq_f32(vi1x0123, vi1x4567, 3);
      const float32x4_t vi2x3456 = vextq_f32(vi2x0123, vi2x4567, 3);
      const float32x4_t vi3x3456 = vextq_f32(vi3x0123, vi3x4567, 3);
      const float32x4_t vi4x3456 = vextq_f32(vi4x0123, vi4x4567, 3);
      const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3);

      vo0p0 = vfmaq_lane_f32(vo0p0, vi0x3456, vget_low_f32(vw0123), 1);
      vo1p0 = vfmaq_lane_f32(vo1p0, vi1x3456, vget_low_f32(vw0123), 1);
      vo2p0 = vfmaq_lane_f32(vo2p0, vi2x3456, vget_low_f32(vw0123), 1);
      vo3p0 = vfmaq_lane_f32(vo3p0, vi3x3456, vget_low_f32(vw0123), 1);

      vo0p0 = vfmaq_lane_f32(vo0p0, vi1x3456, vget_low_f32(vw4567), 0);
      vo1p0 = vfmaq_lane_f32(vo1p0, vi2x3456, vget_low_f32(vw4567), 0);
      vo2p0 = vfmaq_lane_f32(vo2p0, vi3x3456, vget_low_f32(vw4567), 0);
      vo3p0 = vfmaq_lane_f32(vo3p0, vi4x3456, vget_low_f32(vw4567), 0);

      vo0p0 = vfmaq_lane_f32(vo0p0, vi2x3456, vget_high_f32(vw4567), 1);
      vo1p0 = vfmaq_lane_f32(vo1p0, vi3x3456, vget_high_f32(vw4567), 1);
      vo2p0 = vfmaq_lane_f32(vo2p0, vi4x3456, vget_high_f32(vw4567), 1);
      vo3p0 = vfmaq_lane_f32(vo3p0, vi5x3456, vget_high_f32(vw4567), 1);

      const float32x4_t vzero = vmovq_n_f32(0.0f);
      const float32x4_t vi0x5678 = vextq_f32(vi0x4567, vzero, 1);
      const float32x4_t vi1x5678 = vextq_f32(vi1x4567, vzero, 1);
      const float32x4_t vi2x5678 = vextq_f32(vi2x4567, vzero, 1);
      const float32x4_t vi3x5678 = vextq_f32(vi3x4567, vzero, 1);
      const float32x4_t vi4x5678 = vextq_f32(vi4x4567, vzero, 1);
      const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vzero, 1);

      vo0p0 = vfmaq_lane_f32(vo0p0, vi0x5678, vget_high_f32(vw0123), 1);
      vo1p0 = vfmaq_lane_f32(vo1p0, vi1x5678, vget_high_f32(vw0123), 1);
      vo2p0 = vfmaq_lane_f32(vo2p0, vi2x5678, vget_high_f32(vw0123), 1);
      vo3p0 = vfmaq_lane_f32(vo3p0, vi3x5678, vget_high_f32(vw0123), 1);

      vo0p0 = vfmaq_lane_f32(vo0p0, vi1x5678, vget_high_f32(vw4567), 0);
      vo1p0 = vfmaq_lane_f32(vo1p0, vi2x5678, vget_high_f32(vw4567), 0);
      vo2p0 = vfmaq_lane_f32(vo2p0, vi3x5678, vget_high_f32(vw4567), 0);
      vo3p0 = vfmaq_lane_f32(vo3p0, vi4x5678, vget_high_f32(vw4567), 0);

      vo0p0 = vfmaq_lane_f32(vo0p0, vi2x5678, vw89, 1);
      vo1p0 = vfmaq_lane_f32(vo1p0, vi3x5678, vw89, 1);
      vo2p0 = vfmaq_lane_f32(vo2p0, vi4x5678, vw89, 1);
      vo3p0 = vfmaq_lane_f32(vo3p0, vi5x5678, vw89, 1);


      float32x4_t vo0 = vmaxq_f32(vo0p0, vmin);
      float32x4_t vo1 = vmaxq_f32(vo1p0, vmin);
      float32x4_t vo2 = vmaxq_f32(vo2p0, vmin);
      float32x4_t vo3 = vmaxq_f32(vo3p0, vmin);

      vo0 = vminq_f32(vo0, vmax);
      vo1 = vminq_f32(vo1, vmax);
      vo2 = vminq_f32(vo2, vmax);
      vo3 = vminq_f32(vo3, vmax);

      if XNN_LIKELY(w == 4 * sizeof(float)) {
        vst1q_f32(o3, vo3); o3 += 4;
        vst1q_f32(o2, vo2); o2 += 4;
        vst1q_f32(o1, vo1); o1 += 4;
        vst1q_f32(o0, vo0); o0 += 4;
      } else {
        float32x2_t vo0_lo = vget_low_f32(vo0);
        float32x2_t vo1_lo = vget_low_f32(vo1);
        float32x2_t vo2_lo = vget_low_f32(vo2);
        float32x2_t vo3_lo = vget_low_f32(vo3);
        if (w & (2 * sizeof(float))) {
          vst1_f32(o3, vo3_lo); o3 += 2;
          vst1_f32(o2, vo2_lo); o2 += 2;
          vst1_f32(o1, vo1_lo); o1 += 2;
          vst1_f32(o0, vo0_lo); o0 += 2;

          vo0_lo = vget_high_f32(vo0);
          vo1_lo = vget_high_f32(vo1);
          vo2_lo = vget_high_f32(vo2);
          vo3_lo = vget_high_f32(vo3);
        }
        if (w & (1 * sizeof(float))) {
          vst1_lane_f32(o3, vo3_lo, 0); o3 += 1;
          vst1_lane_f32(o2, vo2_lo, 0); o2 += 1;
          vst1_lane_f32(o1, vo1_lo, 0); o1 += 1;
          vst1_lane_f32(o0, vo0_lo, 0); o0 += 1;
        }
      }
    }

    i0 = (const float*) ((uintptr_t) i4 - input_decrement);
    i1 = (const float*) ((uintptr_t) i5 - input_decrement);
    i2 = (const float*) ((uintptr_t) i1 + input_width);
    i3 = (const float*) ((uintptr_t) i2 + input_width);
    i4 = (const float*) ((uintptr_t) i3 + input_width);
    i5 = (const float*) ((uintptr_t) i4 + input_width);

    o0 = o3;
    o1 = (float*) ((uintptr_t) o0 + input_width);
    o2 = (float*) ((uintptr_t) o1 + input_width);
    o3 = (float*) ((uintptr_t) o2 + input_width);

    output_height = doz(output_height, 4);
  } while (output_height != 0);
}