File size: 5,693 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
// 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 CHANNEL_TILE >= 8
$assert ROW_TILE >= 1
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <arm_neon.h>

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


void xnn_f16_vmulcaddc_minmax_ukernel_c${CHANNEL_TILE}__neonfp16arith_${ROW_TILE}x(
    size_t rows,
    size_t channels,
    const void* restrict input,
    size_t input_stride,
    const void* restrict weights,
    void* restrict output,
    size_t output_stride,
    const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(rows != 0);
  assert(channels != 0);
  assert(channels % sizeof(uint16_t) == 0);

  const uint16_t* i0 = (const uint16_t*) input;
  uint16_t* o0 = (uint16_t*) output;
  $for M in range(1, ROW_TILE):
    const uint16_t* i${M} = (const uint16_t*) ((uintptr_t) i${M-1} + input_stride);
    uint16_t* o${M} = (uint16_t*) ((uintptr_t) o${M-1} + output_stride);

  const size_t input_increment = input_stride * ${ROW_TILE} - channels;
  const size_t output_increment = output_stride * ${ROW_TILE} - channels;

  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 M in range(1, ROW_TILE):
      $if M % 2 == 0:
        if XNN_UNPREDICTABLE(rows <= ${M}) {
          i${M} = i${M-1};
          o${M} = o${M-1};
        }
      $else:
        if XNN_UNPREDICTABLE(rows < ${M+1}) {
          i${M} = i${M-1};
          o${M} = o${M-1};
        }

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

        $for M in range(ROW_TILE):
          $for C in range(0, CHANNEL_TILE, 8):
            float16x8_t vacc${M}x${ABC[C:C+8]} = vreinterpretq_f16_u16(vld1q_u16(i${M})); i${M} += 8;

        $for C in range(0, CHANNEL_TILE, 8):
          const float16x8_t vbias${ABC[C:C+8]} = vreinterpretq_f16_u16(vld1q_u16(w)); w += 8;

        $for M in range(ROW_TILE):
          $for C in range(0, CHANNEL_TILE, 8):
            vacc${M}x${ABC[C:C+8]} = vfmaq_f16(vbias${ABC[C:C+8]}, vscale${ABC[C:C+8]}, vacc${M}x${ABC[C:C+8]});

        $for M in range(ROW_TILE):
          $for C in range(0, CHANNEL_TILE, 8):
            vacc${M}x${ABC[C:C+8]} = vmaxq_f16(vacc${M}x${ABC[C:C+8]}, vmin);

        $for M in range(ROW_TILE):
          $for C in range(0, CHANNEL_TILE, 8):
            vacc${M}x${ABC[C:C+8]} = vminq_f16(vacc${M}x${ABC[C:C+8]}, vmax);

        $for M in range(ROW_TILE):
          $for C in range(0, CHANNEL_TILE, 8):
            vst1q_u16(o${M}, vreinterpretq_u16_f16(vacc${M}x${ABC[C:C+8]})); o${M} += 8;
      }
    for (; c >= 8 * sizeof(uint16_t); c -= 8 * sizeof(uint16_t)) {
      const float16x8_t vscale01234567 = vreinterpretq_f16_u16(vld1q_u16(w));

      $for M in range(ROW_TILE):
        float16x8_t vacc${M}x01234567 = vreinterpretq_f16_u16(vld1q_u16(i${M})); i${M} += 8;

      const float16x8_t vbias01234567 = vreinterpretq_f16_u16(vld1q_u16(w + ${CHANNEL_TILE}));
      w += ${8 if CHANNEL_TILE > 8 else CHANNEL_TILE * 2};

      $for M in range(ROW_TILE):
        vacc${M}x01234567 = vfmaq_f16(vbias01234567, vscale01234567, vacc${M}x01234567);

      $for M in range(ROW_TILE):
        vacc${M}x01234567 = vmaxq_f16(vacc${M}x01234567, vmin);

      $for M in range(ROW_TILE):
        vacc${M}x01234567 = vminq_f16(vacc${M}x01234567, vmax);

      $for M in range(ROW_TILE):
        vst1q_u16(o${M}, vreinterpretq_u16_f16(vacc${M}x01234567)); o${M} += 8;
    }
    if XNN_UNLIKELY(c != 0) {
      const float16x8_t vscale01234567 = vreinterpretq_f16_u16(vld1q_u16(w));

      $for M in range(ROW_TILE):
        float16x8_t vacc${M}x01234567 = vreinterpretq_f16_u16(vld1q_u16(i${M})); i${M} = (const uint16_t*) ((uintptr_t) i${M} + c);

      const float16x8_t vbias01234567 = vreinterpretq_f16_u16(vld1q_u16(w + ${CHANNEL_TILE}));

      $for M in range(ROW_TILE):
        vacc${M}x01234567 = vfmaq_f16(vbias01234567, vscale01234567, vacc${M}x01234567);

      $for M in range(ROW_TILE):
        vacc${M}x01234567 = vmaxq_f16(vacc${M}x01234567, vmin);

      $for M in range(ROW_TILE):
        vacc${M}x01234567 = vminq_f16(vacc${M}x01234567, vmax);

      $for M in range(ROW_TILE):
        float16x4_t vacc${M}x0123 = vget_low_f16(vacc${M}x01234567);
      if (c & (4 * sizeof(uint16_t))) {
        $for M in range(ROW_TILE):
          vst1_u16(o${M}, vreinterpret_u16_f16(vacc${M}x0123)); o${M} += 4;

        $for M in range(ROW_TILE):
          vacc${M}x0123 = vget_high_f16(vacc${M}x01234567);
      }
      if (c & (2 * sizeof(uint16_t))) {
        $for M in range(ROW_TILE):
          vst1_lane_u32((void*) o${M}, vreinterpret_u32_f16(vacc${M}x0123), 0); o${M} += 2;

        $for M in range(ROW_TILE):
          vacc${M}x0123 = vext_f16(vacc${M}x0123, vacc${M}x0123, 2);
      }
      if (c & (1 * sizeof(uint16_t))) {
        $for M in range(ROW_TILE):
          vst1_lane_u16(o${M}, vreinterpret_u16_f16(vacc${M}x0123), 0); o${M} += 1;
      }
    }
    $for M in range(ROW_TILE):
      i${M} = (const uint16_t*) ((uintptr_t) i${M} + input_increment);
      o${M} = (uint16_t*) ((uintptr_t) o${M} + output_increment);
    rows = doz(rows, ${ROW_TILE});
  } while (rows != 0);
}