File size: 6,370 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
// Auto-generated file. Do not edit!
//   Template: src/f16-igemm/neonfp16arith-ld64.c.in
//   Generator: tools/xngen
//
// 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.


#include <assert.h>

#include <arm_neon.h>

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


void xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64(
    size_t mr,
    size_t nc,
    size_t kc,
    size_t ks,
    const void** restrict a,
    const void* restrict w,
    void* restrict c,
    size_t cm_stride,
    size_t cn_stride,
    size_t a_offset,
    const void* zero,
    const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(mr != 0);
  assert(mr <= 1);
  assert(nc != 0);
  assert(kc != 0);
  assert(kc % sizeof(uint16_t) == 0);
  assert(ks != 0);
  assert(ks % (1 * sizeof(void*)) == 0);
  assert(a_offset % sizeof(uint16_t) == 0);
  assert(a != NULL);
  assert(w != NULL);
  assert(c != NULL);

  uint16_t* c0 = (uint16_t*) c;

  do {
    float16x8_t vacc0x01234567 = vreinterpretq_f16_u16(vld1q_u16(w)); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
    float16x8_t vacc0x89ABCDEF = vreinterpretq_f16_u16(vld1q_u16(w)); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));

    size_t p = ks;
    do {
      const uint16_t* restrict a0 = (const uint16_t*) a[0];
      assert(a0 != NULL);
      if XNN_UNPREDICTABLE(a0 != zero) {
        a0 = (const uint16_t*) ((uintptr_t) a0 + a_offset);
      }
      a += 1;

      size_t k = kc;
      for (; k >= 4 * sizeof(uint16_t); k -= 4 * sizeof(uint16_t)) {
        const float16x4_t va0 = vreinterpret_f16_u16(vld1_u16(a0)); a0 += 4;

        const float16x8_t vb01234567c0 = vreinterpretq_f16_u16(vld1q_u16(w)); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
        const float16x8_t vb89ABCDEFc0 = vreinterpretq_f16_u16(vld1q_u16(w)); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));

        #if XNN_ARCH_ARM64
          vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c0, va0, 0);
          vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc0, va0, 0);
        #else
          vacc0x01234567 = vmlaq_lane_f16(vacc0x01234567, vb01234567c0, va0, 0);
          vacc0x89ABCDEF = vmlaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc0, va0, 0);
        #endif
        const float16x8_t vb01234567c1 = vreinterpretq_f16_u16(vld1q_u16(w)); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
        const float16x8_t vb89ABCDEFc1 = vreinterpretq_f16_u16(vld1q_u16(w)); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));

        #if XNN_ARCH_ARM64
          vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c1, va0, 1);
          vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc1, va0, 1);
        #else
          vacc0x01234567 = vmlaq_lane_f16(vacc0x01234567, vb01234567c1, va0, 1);
          vacc0x89ABCDEF = vmlaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc1, va0, 1);
        #endif
        const float16x8_t vb01234567c2 = vreinterpretq_f16_u16(vld1q_u16(w)); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
        const float16x8_t vb89ABCDEFc2 = vreinterpretq_f16_u16(vld1q_u16(w)); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));

        #if XNN_ARCH_ARM64
          vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c2, va0, 2);
          vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc2, va0, 2);
        #else
          vacc0x01234567 = vmlaq_lane_f16(vacc0x01234567, vb01234567c2, va0, 2);
          vacc0x89ABCDEF = vmlaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc2, va0, 2);
        #endif
        const float16x8_t vb01234567c3 = vreinterpretq_f16_u16(vld1q_u16(w)); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
        const float16x8_t vb89ABCDEFc3 = vreinterpretq_f16_u16(vld1q_u16(w)); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));

        #if XNN_ARCH_ARM64
          vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c3, va0, 3);
          vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc3, va0, 3);
        #else
          vacc0x01234567 = vmlaq_lane_f16(vacc0x01234567, vb01234567c3, va0, 3);
          vacc0x89ABCDEF = vmlaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc3, va0, 3);
        #endif
      }
      if XNN_UNLIKELY(k != 0) {
        do {
          const float16x8_t va0 = vreinterpretq_f16_u16(vld1q_dup_u16(a0)); a0 += 1;

          const float16x8_t vb01234567 = vreinterpretq_f16_u16(vld1q_u16(w)); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
          const float16x8_t vb89ABCDEF = vreinterpretq_f16_u16(vld1q_u16(w)); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));

          vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0, vb01234567);
          vacc0x89ABCDEF = vfmaq_f16(vacc0x89ABCDEF, va0, vb89ABCDEF);

          k -= sizeof(uint16_t);
        } while (k != 0);
      }
      p -= 1 * sizeof(void*);
    } while (p != 0);


    const float16x8_t vmin = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith.min));
    vacc0x01234567 = vmaxq_f16(vacc0x01234567, vmin);
    vacc0x89ABCDEF = vmaxq_f16(vacc0x89ABCDEF, vmin);

    const float16x8_t vmax = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith.max));
    vacc0x01234567 = vminq_f16(vacc0x01234567, vmax);
    vacc0x89ABCDEF = vminq_f16(vacc0x89ABCDEF, vmax);

    if XNN_LIKELY(nc >= 16) {
      vst1q_u16(c0, vreinterpretq_u16_f16(vacc0x01234567));
      vst1q_u16(c0 + 8, vreinterpretq_u16_f16(vacc0x89ABCDEF));
      c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);

      a = (const void**restrict) ((uintptr_t) a - ks);
      nc -= 16;
    } else {
      if (nc & 8) {
        vst1q_u16(c0, vreinterpretq_u16_f16(vacc0x01234567)); c0 += 8;

        vacc0x01234567 = vacc0x89ABCDEF;
      }
      float16x4_t vacc0x0123 = vget_low_f16(vacc0x01234567);
      if (nc & 4) {
        vst1_u16(c0, vreinterpret_u16_f16(vacc0x0123)); c0 += 4;

        vacc0x0123 = vget_high_f16(vacc0x01234567);
      }
      if (nc & 2) {
        vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vacc0x0123), 0); c0 += 2;

        vacc0x0123 = vext_f16(vacc0x0123, vacc0x0123, 2);
      }
      if (nc & 1) {
        vst1_lane_u16(c0, vreinterpret_u16_f16(vacc0x0123), 0);
      }

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