File size: 8,989 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
// Auto-generated file. Do not edit!
//   Template: src/f32-dwconv2d-chw/3x3p1-ssse3.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 <tmmintrin.h>

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


void xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4_acc2(
    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 __m128 vmask = _mm_load_ps((const float*) params->sse_stride1.mask);
  const __m128 vmax = _mm_load_ps(params->sse_stride1.max);
  const __m128 vmin = _mm_load_ps(params->sse_stride1.min);

  const __m128 vbias = _mm_load1_ps(weights);
  const __m128 vk00 = _mm_load1_ps(weights + 1);
  const __m128 vk01 = _mm_load1_ps(weights + 2);
  const __m128 vk02 = _mm_load1_ps(weights + 3);
  const __m128 vk10 = _mm_load1_ps(weights + 4);
  const __m128 vk11 = _mm_load1_ps(weights + 5);
  const __m128 vk12 = _mm_load1_ps(weights + 6);
  const __m128 vk20 = _mm_load1_ps(weights + 7);
  const __m128 vk21 = _mm_load1_ps(weights + 8);
  const __m128 vk22 = _mm_load1_ps(weights + 9);

  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);

  float* o0 = output;
  float* o1 = (float*) ((uintptr_t) o0 + 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;
    }

    __m128 vi0x0123 = _mm_setzero_ps();
    __m128 vi1x0123 = _mm_setzero_ps();
    __m128 vi2x0123 = _mm_setzero_ps();
    __m128 vi3x0123 = _mm_setzero_ps();

    __m128 vi0x4567 = _mm_loadu_ps(i0);
    i0 += 4;
    __m128 vi1x4567 = _mm_loadu_ps(i1);
    i1 += 4;
    __m128 vi2x4567 = _mm_loadu_ps(i2);
    i2 += 4;
    __m128 vi3x4567 = _mm_loadu_ps(i3);
    i3 += 4;

    size_t w = input_width;
    for (; w > 4 * sizeof(float); w -= 4 * sizeof(float)) {
      const __m128 vi0x89AB = _mm_loadu_ps(i0);
      i0 += 4;
      const __m128 vi1x89AB = _mm_loadu_ps(i1);
      i1 += 4;
      const __m128 vi2x89AB = _mm_loadu_ps(i2);
      i2 += 4;
      const __m128 vi3x89AB = _mm_loadu_ps(i3);
      i3 += 4;

      __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x4567, vk01));
      __m128 vo1p0 = _mm_add_ps(vbias, _mm_mul_ps(vi1x4567, vk01));
      __m128 vo0p1 = _mm_mul_ps(vi1x4567, vk11);
      __m128 vo1p1 = _mm_mul_ps(vi2x4567, vk11);
      vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x4567, vk21));
      vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x4567, vk21));

      const __m128 vi0x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x4567), _mm_castps_si128(vi0x0123), 12));
      const __m128 vi1x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x4567), _mm_castps_si128(vi1x0123), 12));
      const __m128 vi2x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x4567), _mm_castps_si128(vi2x0123), 12));
      const __m128 vi3x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi3x4567), _mm_castps_si128(vi3x0123), 12));

      vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi0x3456, vk00));
      vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi1x3456, vk00));
      vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x3456, vk10));
      vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x3456, vk10));
      vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi2x3456, vk20));
      vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi3x3456, vk20));

      vi0x0123 = vi0x4567;
      vi1x0123 = vi1x4567;
      vi2x0123 = vi2x4567;
      vi3x0123 = vi3x4567;

      const __m128 vi0x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x89AB), _mm_castps_si128(vi0x4567), 4));
      const __m128 vi1x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x89AB), _mm_castps_si128(vi1x4567), 4));
      const __m128 vi2x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x89AB), _mm_castps_si128(vi2x4567), 4));
      const __m128 vi3x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi3x89AB), _mm_castps_si128(vi3x4567), 4));

      vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x5678, vk02));
      vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x5678, vk02));
      vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi1x5678, vk12));
      vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi2x5678, vk12));
      vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x5678, vk22));
      vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x5678, vk22));

      vi0x4567 = vi0x89AB;
      vi1x4567 = vi1x89AB;
      vi2x4567 = vi2x89AB;
      vi3x4567 = vi3x89AB;

      vo0p0 = _mm_add_ps(vo0p0, vo0p1);
      vo1p0 = _mm_add_ps(vo1p0, vo1p1);

      __m128 vo0 = _mm_max_ps(vo0p0, vmin);
      __m128 vo1 = _mm_max_ps(vo1p0, vmin);

      vo0 = _mm_min_ps(vo0, vmax);
      vo1 = _mm_min_ps(vo1, vmax);

      _mm_storeu_ps(o1, vo1);
      o1 += 4;
      _mm_storeu_ps(o0, vo0);
      o0 += 4;
    }
    // Always process the last block of 1..4 pixels.
    assert(w >= 1 * sizeof(float));
    assert(w <= 4 * sizeof(float));
    {
      vi0x4567 = _mm_and_ps(vmask, vi0x4567);
      vi1x4567 = _mm_and_ps(vmask, vi1x4567);
      vi2x4567 = _mm_and_ps(vmask, vi2x4567);
      vi3x4567 = _mm_and_ps(vmask, vi3x4567);

      __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x4567, vk01));
      __m128 vo1p0 = _mm_add_ps(vbias, _mm_mul_ps(vi1x4567, vk01));
      __m128 vo0p1 = _mm_mul_ps(vi1x4567, vk11);
      __m128 vo1p1 = _mm_mul_ps(vi2x4567, vk11);
      vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x4567, vk21));
      vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x4567, vk21));

      const __m128 vi0x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x4567), _mm_castps_si128(vi0x0123), 12));
      const __m128 vi1x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x4567), _mm_castps_si128(vi1x0123), 12));
      const __m128 vi2x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x4567), _mm_castps_si128(vi2x0123), 12));
      const __m128 vi3x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi3x4567), _mm_castps_si128(vi3x0123), 12));

      vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi0x3456, vk00));
      vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi1x3456, vk00));
      vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x3456, vk10));
      vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x3456, vk10));
      vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi2x3456, vk20));
      vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi3x3456, vk20));

      const __m128i vzero = _mm_setzero_si128();
      const __m128 vi0x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi0x4567), 4));
      const __m128 vi1x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi1x4567), 4));
      const __m128 vi2x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi2x4567), 4));
      const __m128 vi3x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi3x4567), 4));

      vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x5678, vk02));
      vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x5678, vk02));
      vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi1x5678, vk12));
      vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi2x5678, vk12));
      vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x5678, vk22));
      vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x5678, vk22));

      vo0p0 = _mm_add_ps(vo0p0, vo0p1);
      vo1p0 = _mm_add_ps(vo1p0, vo1p1);

      __m128 vo0 = _mm_max_ps(vo0p0, vmin);
      __m128 vo1 = _mm_max_ps(vo1p0, vmin);

      vo0 = _mm_min_ps(vo0, vmax);
      vo1 = _mm_min_ps(vo1, vmax);

      if XNN_LIKELY(w == 4 * sizeof(float)) {
        _mm_storeu_ps(o1, vo1);
        o1 += 4;
        _mm_storeu_ps(o0, vo0);
        o0 += 4;
      } else {
        if (w & (2 * sizeof(float))) {
          _mm_storel_pi((__m64*) o1, vo1);
          o1 += 2;
          _mm_storel_pi((__m64*) o0, vo0);
          o0 += 2;

          vo0 = _mm_movehl_ps(vo0, vo0);
          vo1 = _mm_movehl_ps(vo1, vo1);
        }
        if (w & (1 * sizeof(float))) {
          _mm_store_ss(o1, vo1);
          o1 += 1;
          _mm_store_ss(o0, vo0);
          o0 += 1;
        }
      }
    }

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

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

    output_height = doz(output_height, 2);
  } while (output_height != 0);
}