File size: 2,511 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
// Auto-generated file. Do not edit!
//   Template: src/f16-f32acc-rsum/neonfp16.c.in
//   Generator: tools/xngen
//
// Copyright 2023 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/common.h>
#include <xnnpack/reduce.h>


void xnn_f16_f32acc_rsum_ukernel__neonfp16_x16_acc2(
    size_t batch,
    const void* input,
    void* output,
    const union xnn_f16_f32acc_scale_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(batch != 0);
  assert(batch % sizeof(uint16_t) == 0);
  assert(input != NULL);
  assert(output != NULL);

  const uint16_t* i = (const uint16_t*) input;
  uint16_t* o = (uint16_t*) output;
  float32x4_t vacc0 = vmovq_n_f32(0.0f);
  float32x4_t vacc1 = vmovq_n_f32(0.0f);
  for (; batch >= 16 * sizeof(uint16_t); batch -= 16 * sizeof(uint16_t)) {
    const float16x8_t vh01 = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
    const float16x8_t vh23 = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;

    const float32x4_t vt0 = vcvt_f32_f16(vget_low_f16(vh01));
    const float32x4_t vt1 = vcvt_f32_f16(vget_high_f16(vh01));
    const float32x4_t vt2 = vcvt_f32_f16(vget_low_f16(vh23));
    const float32x4_t vt3 = vcvt_f32_f16(vget_high_f16(vh23));

    vacc0 = vaddq_f32(vacc0, vt0);
    vacc1 = vaddq_f32(vacc1, vt1);
    vacc0 = vaddq_f32(vacc0, vt2);
    vacc1 = vaddq_f32(vacc1, vt3);
  }
  vacc0 = vaddq_f32(vacc0, vacc1);
  for (; batch >= 4 * sizeof(uint16_t); batch -= 4 * sizeof(uint16_t)) {
    const float16x4_t vh = vreinterpret_f16_u16(vld1_u16(i)); i += 4;
    const float32x4_t vt = vcvt_f32_f16(vh);
    vacc0 = vaddq_f32(vacc0, vt);
  }
  const float32x2_t vscale = vld1_dup_f32(&params->scalar.scale);
  float32x2_t vacc = vadd_f32(vget_low_f32(vacc0), vget_high_f32(vacc0));
  if XNN_UNLIKELY(batch & (2 * sizeof(uint16_t))) {
    const float16x4_t vh = vreinterpret_f16_u32(vld1_dup_u32((const void*) i)); i += 2;
    const float32x4_t vt = vcvt_f32_f16(vh);
    vacc = vadd_f32(vacc, vget_low_f32(vt));
  }
  vacc = vpadd_f32(vacc, vacc);
  if XNN_UNLIKELY(batch & (1 * sizeof(uint16_t))) {
    const float16x4_t vh = vreinterpret_f16_u16(vld1_dup_u16(i));
    const float32x4_t vt = vcvt_f32_f16(vh);
    vacc = vadd_f32(vacc, vget_low_f32(vt));
  }
  vacc = vmul_f32(vacc, vscale);
  const float16x4_t vout = vcvt_f16_f32(vcombine_f32(vacc, vacc));
  vst1_lane_u16(o, vreinterpret_u16_f16(vout), 0);
}