Hugging Face's logo

Androidonnxfork
/
test

Model card Files Community
5
test
File size: 5,865 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
 
// 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 NR % 8 == 0
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"

#include <assert.h>

#include <arm_neon.h>

#include <xnnpack/common.h>

#include <xnnpack/gemm.h>
#include <xnnpack/intrinsics-polyfill.h>


void xnn_f16_gemm${"inc" if INC else ""}_minmax_ukernel_${MR}x${NR}__neonfp16arith_ld64(
    size_t mr,
    size_t nc,
    size_t kc,
    const void* restrict a,
    size_t a_stride,
    const void* restrict w,
    void* restrict c,
    size_t cm_stride,
    size_t cn_stride,
    $if INC:
      const void* restrict acc,
    const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(mr != 0);
  assert(mr <= ${MR});
  assert(nc != 0);
  assert(kc != 0);
  assert(kc % sizeof(uint16_t) == 0);
  assert(a != NULL);
  assert(w != NULL);
  assert(c != NULL);
  $if INC:
    assert(acc != NULL);

  const uint16_t* a0 = (const uint16_t*) a;
  uint16_t* c0 = (uint16_t*) c;
  $for M in range(1, MR):
    const uint16_t* a${M} = (const uint16_t*) ((uintptr_t) a${M-1} + a_stride);
    uint16_t* c${M} = (uint16_t*) ((uintptr_t) c${M-1} + cm_stride);
    $if M % 2 == 0:
      if XNN_UNPREDICTABLE(mr <= ${M}) {
        a${M} = a${M-1};
        c${M} = c${M-1};
      }
    $elif M + 1 == MR:
      if XNN_UNPREDICTABLE(mr != ${M+1}) {
        a${M} = a${M-1};
        c${M} = c${M-1};
      }
    $else:
      if XNN_UNPREDICTABLE(mr < ${M+1}) {
        a${M} = a${M-1};
        c${M} = c${M-1};
      }

  do {
    $if INC:
      $for M in range(MR):
        $for N in range(0, NR, 8):
          float16x8_t vacc${M}x${ABC[N:N+8]} = vreinterpretq_f16_u16(vld1q_u16(acc)); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
    $else:
      $for N in range(0, NR, 8):
        float16x8_t vacc0x${ABC[N:N+8]} = vreinterpretq_f16_u16(vld1q_u16(w)); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
      $for M in range(1, MR):
        $for N in range(0, NR, 8):
          float16x8_t vacc${M}x${ABC[N:N+8]} = vacc0x${ABC[N:N+8]};

    size_t k = kc;
    while (k >= 4 * sizeof(uint16_t)) {
      $for M in range(MR):
        const float16x4_t va${M} = vreinterpret_f16_u16(vld1_u16(a${M})); a${M} += 4;

      $for L in range(4):
        $for N in range(0, NR, 8):
          const float16x8_t vb${ABC[N:N+8]}c${L} = vreinterpretq_f16_u16(vld1q_u16(w)); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));

        #if XNN_ARCH_ARM64
          $for N in range(0, NR, 8):
            $for M in range(MR):
              vacc${M}x${ABC[N:N+8]} = vfmaq_lane_f16(vacc${M}x${ABC[N:N+8]}, vb${ABC[N:N+8]}c${L}, va${M}, ${L});
        #else
          $for N in range(0, NR, 8):
            $for M in range(MR):
              vacc${M}x${ABC[N:N+8]} = vmlaq_lane_f16(vacc${M}x${ABC[N:N+8]}, vb${ABC[N:N+8]}c${L}, va${M}, ${L});
        #endif

      k -= 4 * sizeof(uint16_t);
    }
    if XNN_UNLIKELY(k != 0) {
      do {
        $for M in range(MR):
          const float16x8_t va${M} = vreinterpretq_f16_u16(vld1q_dup_u16(a${M})); a${M} += 1;

        $for N in range(0, NR, 8):
          const float16x8_t vb${ABC[N:N+8]} = vreinterpretq_f16_u16(vld1q_u16(w)); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));

        $for N in range(0, NR, 8):
          $for M in range(MR):
            vacc${M}x${ABC[N:N+8]} = vfmaq_f16(vacc${M}x${ABC[N:N+8]}, va${M}, vb${ABC[N:N+8]});

        k -= sizeof(uint16_t);
      } while (k != 0);
    }


    const float16x8_t vmin = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith.min));
    $for N in range(0, NR, 8):
      $for M in range(MR):
        vacc${M}x${ABC[N:N+8]} = vmaxq_f16(vacc${M}x${ABC[N:N+8]}, vmin);

    const float16x8_t vmax = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith.max));
    $for N in range(0, NR, 8):
      $for M in range(MR):
        vacc${M}x${ABC[N:N+8]} = vminq_f16(vacc${M}x${ABC[N:N+8]}, vmax);

    if XNN_LIKELY(nc >= ${NR}) {
      $for M in range(MR):
        vst1q_u16(c${M}, vreinterpretq_u16_f16(vacc${M}x${ABC[0:8]}));
        $for N in range(8, NR, 8):
          vst1q_u16(c${M} + ${N}, vreinterpretq_u16_f16(vacc${M}x${ABC[N:N+8]}));
        c${M} = (uint16_t*) ((uintptr_t) c${M} + cn_stride);

      $for M in range(MR):
        a${M} = (const uint16_t*) ((uintptr_t) a${M} - kc);

      nc -= ${NR};
    } else {
      $for LOG2N in reversed(range(NR.bit_length())):
        $if NR != 1 << LOG2N:
          if (nc & ${1 << LOG2N}) {
            $if LOG2N >= 3:
              $for N in range(0, 1 << LOG2N, 8):
                $for M in range(MR):
                  vst1q_u16(c${M}, vreinterpretq_u16_f16(vacc${M}x${ABC[N:N+8]})); c${M} += 8;

              $for M in range(MR):
                $for N in range(0, NR - (1 << LOG2N), 8):
                  vacc${M}x${ABC[N:N+8]} = vacc${M}x${ABC[N + (1 << LOG2N):N + (1 << LOG2N)+8]};
            $elif LOG2N == 2:
              $for M in range(MR):
                vst1_u16(c${M}, vreinterpret_u16_f16(vacc${M}x${ABC[0:4]})); c${M} += 4;

              $for M in range(MR):
                vacc${M}x${ABC[0:4]} = vget_high_f16(vacc${M}x${ABC[0:8]});
            $elif LOG2N == 1:
              $for M in range(MR):
                vst1_lane_u32((void*) c${M}, vreinterpret_u32_f16(vacc${M}x${ABC[0:4]}), 0); c${M} += 2;

              $for M in range(MR):
                vacc${M}x${ABC[0:4]} = vext_f16(vacc${M}x${ABC[0:4]}, vacc${M}x${ABC[0:4]}, 2);
            $elif LOG2N == 0:
              $for M in range(MR):
                vst1_lane_u16(c${M}, vreinterpret_u16_f16(vacc${M}x${ABC[0:4]}), 0);
          }
        $if LOG2N == 3:
          $for M in range(MR):
            float16x4_t vacc${M}x${ABC[0:4]} = vget_low_f16(vacc${M}x${ABC[0:8]});

      nc = 0;
    }
  } while (nc != 0);
}