Hugging Face's logo

Androidonnxfork
/
test

Model card Files Community
5
test
File size: 6,361 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 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.

$assert CHANNEL_TILE % 8 == 0
$assert KERNEL_TILE >= 2
$assert ACCUMULATORS >= 1
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <arm_neon.h>

#include <xnnpack/dwconv.h>


void xnn_f16_dwconv_minmax_ukernel_${KERNEL_TILE}p${CHANNEL_TILE}c__neonfp16arith${"" if ACCUMULATORS == 1 else "_acc%d" % ACCUMULATORS}(
    size_t channels,
    size_t output_width,
    const void** input,
    const void* weights,
    void* output_ptr,
    intptr_t input_stride,
    size_t output_increment,
    size_t input_offset,
    const void* zero,
    const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(channels != 0);
  assert(output_width != 0);

  uint16_t* output = (uint16_t*) output_ptr;
  const float16x8_t vmin = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith.min));
  const float16x8_t vmax = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith.max));
  do {
    $for K in range(KERNEL_TILE):
      const uint16_t* i${K} = (const uint16_t*) input[${K}];
      assert(i${K} != NULL);
      if XNN_UNPREDICTABLE(i${K} != (const uint16_t*) zero) {
        i${K} = (const uint16_t*) ((uintptr_t) i${K} + input_offset);
      }

    input = (const void**) ((uintptr_t) input + input_stride);

    size_t c = channels;
    const uint16_t* w = (const uint16_t*) weights;
    for (; c >= ${CHANNEL_TILE}; c -= ${CHANNEL_TILE}) {
      $for C in range(0, CHANNEL_TILE, 8):
        float16x8_t vacc${ABC[C:C+8]}p0 = vreinterpretq_f16_u16(vld1q_u16(w)); w += 8;

      $for K in range(KERNEL_TILE):

        $for C in range(0, CHANNEL_TILE, 8):
          const float16x8_t vi${K}x${ABC[C:C+8]} = vreinterpretq_f16_u16(vld1q_u16(i${K})); i${K} += 8;
        $for C in range(0, CHANNEL_TILE, 8):
          const float16x8_t vk${K}x${ABC[C:C+8]} = vreinterpretq_f16_u16(vld1q_u16(w)); w += 8;
        $for C in range(0, CHANNEL_TILE, 8):
          $if 1 <= K < ACCUMULATORS:
            float16x8_t vacc${ABC[C:C+8]}p${K} = vmulq_f16(vi${K}x${ABC[C:C+8]}, vk${K}x${ABC[C:C+8]});
          $else:
            vacc${ABC[C:C+8]}p${K % ACCUMULATORS} = vfmaq_f16(vacc${ABC[C:C+8]}p${K % ACCUMULATORS}, vi${K}x${ABC[C:C+8]}, vk${K}x${ABC[C:C+8]});

      $if ACCUMULATORS > 1:
        // Add up all accumulators to vacc${ABC[0:CHANNEL_TILE]}p0
        $ACC_STEP = 1
        $while ACC_STEP < ACCUMULATORS:
          $for A in range(0, ACCUMULATORS, ACC_STEP * 2):
            $if A + ACC_STEP < ACCUMULATORS:
              $for C in range(0, CHANNEL_TILE, 8):
                vacc${ABC[C:C+8]}p${A} = vaddq_f16(vacc${ABC[C:C+8]}p${A}, vacc${ABC[C:C+8]}p${A + ACC_STEP});
          $ACC_STEP *= 2

      $for C in range(0, CHANNEL_TILE, 8):
        float16x8_t vacc${ABC[C:C+8]} = vmaxq_f16(vacc${ABC[C:C+8]}p0, vmin);
      $for C in range(0, CHANNEL_TILE, 8):
        vacc${ABC[C:C+8]} = vminq_f16(vacc${ABC[C:C+8]}, vmax);

      $for C in range(0, CHANNEL_TILE, 8):
        vst1q_u16(output, vreinterpretq_u16_f16(vacc${ABC[C:C+8]})); output += 8;
    }
    $if CHANNEL_TILE > 8:
      for (; c >= 8; c -= 8) {
        float16x8_t vacc01234567p0 = vreinterpretq_f16_u16(vld1q_u16(w)); w += 8;

        $for K in range(KERNEL_TILE):

          const float16x8_t vi${K}x01234567 = vreinterpretq_f16_u16(vld1q_u16(i${K})); i${K} += 8;
          const float16x8_t vk${K}x01234567 = vreinterpretq_f16_u16(vld1q_u16(w + ${(K + 1) * CHANNEL_TILE - 8}));
          $if 1 <= K < ACCUMULATORS:
            float16x8_t vacc01234567p${K} = vmulq_f16(vi${K}x01234567, vk${K}x01234567);
          $else:
            vacc01234567p${K % ACCUMULATORS} = vfmaq_f16(vacc01234567p${K % ACCUMULATORS}, vi${K}x01234567, vk${K}x01234567);

        $if ACCUMULATORS > 1:
          // Add up all accumulators to vacc01234567p0
          $ACC_STEP = 1
          $while ACC_STEP < ACCUMULATORS:
            $for A in range(0, ACCUMULATORS, ACC_STEP * 2):
              $if A + ACC_STEP < ACCUMULATORS:
                vacc01234567p${A} = vaddq_f16(vacc01234567p${A}, vacc01234567p${A + ACC_STEP});
            $ACC_STEP *= 2

        float16x8_t vacc01234567 = vmaxq_f16(vacc01234567p0, vmin);
        vacc01234567 = vminq_f16(vacc01234567, vmax);

        vst1q_u16(output, vreinterpretq_u16_f16(vacc01234567)); output += 8;
      }
    if XNN_UNLIKELY(c != 0) {
      $if CHANNEL_TILE == 8:
        float16x8_t vacc01234567p0 = vreinterpretq_f16_u16(vld1q_u16(w)); w += 8;
      $else:
        float16x8_t vacc01234567p0 = vreinterpretq_f16_u16(vld1q_u16(w));

      $for K in range(KERNEL_TILE):

        const float16x8_t vi${K}x01234567 = vreinterpretq_f16_u16(vld1q_u16(i${K}));
        $if CHANNEL_TILE == 8:
          const float16x8_t vk${K}x01234567 = vreinterpretq_f16_u16(vld1q_u16(w)); w += 8;
        $else:
          const float16x8_t vk${K}x01234567 = vreinterpretq_f16_u16(vld1q_u16(w + ${(K + 1) * CHANNEL_TILE}));
        $if 1 <= K < ACCUMULATORS:
          float16x8_t vacc01234567p${K} = vmulq_f16(vi${K}x01234567, vk${K}x01234567);
        $else:
          vacc01234567p${K % ACCUMULATORS} = vfmaq_f16(vacc01234567p${K % ACCUMULATORS}, vi${K}x01234567, vk${K}x01234567);

      $if ACCUMULATORS > 1:
        // Add up all accumulators to vacc01234567p0
        $ACC_STEP = 1
        $while ACC_STEP < ACCUMULATORS:
          $for A in range(0, ACCUMULATORS, ACC_STEP * 2):
            $if A + ACC_STEP < ACCUMULATORS:
              vacc01234567p${A} = vaddq_f16(vacc01234567p${A}, vacc01234567p${A + ACC_STEP});
          $ACC_STEP *= 2

      float16x8_t vacc01234567 = vmaxq_f16(vacc01234567p0, vmin);
      vacc01234567 = vminq_f16(vacc01234567, vmax);

      float16x4_t vacc0123 = vget_low_f16(vacc01234567);
      if (c & 4) {
        vst1_u16(output, vreinterpret_u16_f16(vacc0123)); output += 4;
        vacc0123 = vget_high_f16(vacc01234567);
      }
      if (c & 2) {
        vst1_lane_u32((void*) output, vreinterpret_u32_f16(vacc0123), 0); output += 2;
        vacc0123 = vext_f16(vacc0123, vacc0123, 2);
      }
      if (c & 1) {
        vst1_lane_u16(output, vreinterpret_u16_f16(vacc0123), 0); output += 1;
      }
    }

    output = (uint16_t*) ((uintptr_t) output + output_increment);
  } while (--output_width != 0);
}