File size: 5,634 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
// Copyright 2019 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.

$assert CHANNEL_TILE % 4 == 0
$assert KERNEL_TILE >= 2
$assert ACCUMULATORS >= 1
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <xmmintrin.h>

#include <xnnpack/dwconv.h>


void xnn_f32_dwconv_minmax_ukernel_${KERNEL_TILE}p${CHANNEL_TILE}c__sse${"" if ACCUMULATORS == 1 else "_acc%d" % ACCUMULATORS}(
    size_t channels,
    size_t output_width,
    const float** input,
    const float* weights,
    float* output,
    intptr_t input_stride,
    size_t output_increment,
    size_t input_offset,
    const float* zero,
    const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(channels != 0);
  assert(output_width != 0);

  const __m128 vmax = _mm_load_ps(params->sse.max);
  const __m128 vmin = _mm_load_ps(params->sse.min);
  do {
    $for K in range(KERNEL_TILE):
      const float* i${K} = input[${K}];
      assert(i${K} != NULL);
      if XNN_UNPREDICTABLE(i${K} != zero) {
        i${K} = (const float*) ((uintptr_t) i${K} + input_offset);
      }
    input = (const float**) ((uintptr_t) input + input_stride);

    size_t c = channels;
    const float* w = weights;
    for (; c >= ${CHANNEL_TILE}; c -= ${CHANNEL_TILE}) {
      __m128 vacc${ABC[0:4]}p0 = _mm_load_ps(w);
      $for C in range(4, CHANNEL_TILE, 4):
        __m128 vacc${ABC[C:C+4]}p0 = _mm_load_ps(w + ${C});

      $for K in range(KERNEL_TILE):

        const __m128 vi${K}x${ABC[0:4]} = _mm_loadu_ps(i${K});
        $for C in range(4, CHANNEL_TILE, 4):
          const __m128 vi${K}x${ABC[C:C+4]} = _mm_loadu_ps(i${K} + ${C});
        i${K} += ${CHANNEL_TILE};

        $for C in range(0, CHANNEL_TILE, 4):
          const __m128 vk${K}x${ABC[C:C+4]} = _mm_load_ps(w + ${(K + 1) * CHANNEL_TILE + C});
        $for C in range(0, CHANNEL_TILE, 4):
          $if 1 <= K < ACCUMULATORS:
            __m128 vacc${ABC[C:C+4]}p${K} = _mm_mul_ps(vi${K}x${ABC[C:C+4]}, vk${K}x${ABC[C:C+4]});
          $else:
            vacc${ABC[C:C+4]}p${K % ACCUMULATORS} = _mm_add_ps(vacc${ABC[C:C+4]}p${K % ACCUMULATORS}, _mm_mul_ps(vi${K}x${ABC[C:C+4]}, vk${K}x${ABC[C:C+4]}));

      w += ${(KERNEL_TILE + 1) * CHANNEL_TILE};

      $if ACCUMULATORS > 1:
        // Add up all accumulators to vacc${ABC[0:CHANNEL_TILE]}p0
        $ACC_SLICE = 1
        $while ACC_SLICE < ACCUMULATORS:
          $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
            $if A + ACC_SLICE < ACCUMULATORS:
              $for C in range(0, CHANNEL_TILE, 4):
                vacc${ABC[C:C+4]}p${A} = _mm_add_ps(vacc${ABC[C:C+4]}p${A}, vacc${ABC[C:C+4]}p${A + ACC_SLICE});
          $ACC_SLICE *= 2

      $for C in range(0, CHANNEL_TILE, 4):
        __m128 vacc${ABC[C:C+4]} = _mm_max_ps(vacc${ABC[C:C+4]}p0, vmin);
      $for C in range(0, CHANNEL_TILE, 4):
        vacc${ABC[C:C+4]} = _mm_min_ps(vacc${ABC[C:C+4]}, vmax);

      _mm_storeu_ps(output, vacc${ABC[0:4]});
      $for C in range(4, CHANNEL_TILE, 4):
        _mm_storeu_ps(output + ${C}, vacc${ABC[C:C+4]});
      output += ${CHANNEL_TILE};
    }
    $if CHANNEL_TILE > 4:
      for (; c >= 4; c -= 4) {
        __m128 vacc0123p0 = _mm_load_ps(w);
        $for K in range(KERNEL_TILE):

          const __m128 vi${K}x0123 = _mm_loadu_ps(i${K});
          i${K} += 4;

          const __m128 vk${K}x0123 = _mm_load_ps(w + ${(K + 1) * CHANNEL_TILE});
          $if 1 <= K < ACCUMULATORS:
            __m128 vacc0123p${K} = _mm_mul_ps(vi${K}x0123, vk${K}x0123);
          $else:
            vacc0123p${K % ACCUMULATORS} = _mm_add_ps(vacc0123p${K % ACCUMULATORS}, _mm_mul_ps(vi${K}x0123, vk${K}x0123));

        w += 4;

        $if ACCUMULATORS > 1:
          // Add up all accumulators to vacc${ABC[0:CHANNEL_TILE]}p0
          $ACC_SLICE = 1
          $while ACC_SLICE < ACCUMULATORS:
            $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
              $if A + ACC_SLICE < ACCUMULATORS:
                vacc0123p${A} = _mm_add_ps(vacc0123p${A}, vacc0123p${A + ACC_SLICE});
            $ACC_SLICE *= 2

        __m128 vacc0123 = _mm_max_ps(vacc0123p0, vmin);
        vacc0123 = _mm_min_ps(vacc0123, vmax);

        _mm_storeu_ps(output, vacc0123);
        output += 4;
      }
    if XNN_UNLIKELY(c != 0) {
      __m128 vacc0123p0 = _mm_load_ps(w);
      $for K in range(KERNEL_TILE):

        const __m128 vi${K}x0123 = _mm_loadu_ps(i${K});
        const __m128 vk${K}x0123 = _mm_load_ps(w + ${(K + 1) * CHANNEL_TILE});
        $if 1 <= K < ACCUMULATORS:
          __m128 vacc0123p${K} = _mm_mul_ps(vi${K}x0123, vk${K}x0123);
        $else:
          vacc0123p${K % ACCUMULATORS} = _mm_add_ps(vacc0123p${K % ACCUMULATORS}, _mm_mul_ps(vi${K}x0123, vk${K}x0123));

      $if ACCUMULATORS > 1:
        // Add up all accumulators to vacc${ABC[0:CHANNEL_TILE]}p0
        $ACC_SLICE = 1
        $while ACC_SLICE < ACCUMULATORS:
          $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
            $if A + ACC_SLICE < ACCUMULATORS:
              vacc0123p${A} = _mm_add_ps(vacc0123p${A}, vacc0123p${A + ACC_SLICE});
          $ACC_SLICE *= 2

      __m128 vacc0123 = _mm_max_ps(vacc0123p0, vmin);
      vacc0123 = _mm_min_ps(vacc0123, vmax);

      if (c & 2) {
        _mm_storel_pi((__m64*) output, vacc0123);
        vacc0123 = _mm_movehl_ps(vacc0123, vacc0123);
        output += 2;
      }
      if (c & 1) {
        _mm_store_ss(output, vacc0123);
        output += 1;
      }
    }

    output = (float*) ((uintptr_t) output + output_increment);
  } while (--output_width != 0);
}