1// Copyright 2020 Google LLC
2//
3// This source code is licensed under the BSD-style license found in the
4// LICENSE file in the root directory of this source tree.
5
6$assert BATCH_TILE % 4 == 0
7$assert BATCH_TILE >= 4
8$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
9#include <assert.h>
10
11#include <arm_neon.h>
12
13#include <xnnpack/common.h>
14#include <xnnpack/math.h>
15#include <xnnpack/vunary.h>
16
17
18void xnn_f32_vrndd_ukernel__neon_x${BATCH_TILE}(
19    size_t n,
20    const float* x,
21    float* y,
22    const union xnn_f32_rnd_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
23{
24  assert(n != 0);
25  assert(n % sizeof(float) == 0);
26
27  const float32x4_t vintegral_threshold = vreinterpretq_f32_u32(vmovq_n_u32(UINT32_C(0x4B000000)));
28  const uint32x4_t vone = vreinterpretq_u32_f32(vmovq_n_f32(1.0f));
29  for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) {
30    $for N in range(0, BATCH_TILE, 4):
31      const float32x4_t vx${ABC[N:N+4]} = vld1q_f32(x); x += 4;
32
33    $for N in range(0, BATCH_TILE, 4):
34      const int32x4_t vintx${ABC[N:N+4]} = vcvtq_s32_f32(vx${ABC[N:N+4]});
35
36    $for N in range(0, BATCH_TILE, 4):
37      uint32x4_t vrndmask${ABC[N:N+4]} = vcaltq_f32(vx${ABC[N:N+4]}, vintegral_threshold);
38
39    $for N in range(0, BATCH_TILE, 4):
40      const float32x4_t vprerndx${ABC[N:N+4]} = vcvtq_f32_s32(vintx${ABC[N:N+4]});
41
42    $for N in range(0, BATCH_TILE, 4):
43      vrndmask${ABC[N:N+4]} = vbicq_u32(vrndmask${ABC[N:N+4]}, vmovq_n_u32(UINT32_C(0x80000000)));
44
45    $for N in range(0, BATCH_TILE, 4):
46      const float32x4_t vrndx${ABC[N:N+4]} = vbslq_f32(vrndmask${ABC[N:N+4]}, vprerndx${ABC[N:N+4]}, vx${ABC[N:N+4]});
47
48    $for N in range(0, BATCH_TILE, 4):
49      const uint32x4_t vadjmask${ABC[N:N+4]} = vcgtq_f32(vrndx${ABC[N:N+4]}, vx${ABC[N:N+4]});
50
51    $for N in range(0, BATCH_TILE, 4):
52      const float32x4_t vadjrndx${ABC[N:N+4]} = vreinterpretq_f32_u32(vandq_u32(vadjmask${ABC[N:N+4]}, vone));
53
54    $for N in range(0, BATCH_TILE, 4):
55      const float32x4_t vy${ABC[N:N+4]} = vsubq_f32(vrndx${ABC[N:N+4]}, vadjrndx${ABC[N:N+4]});
56
57    $for N in range(0, BATCH_TILE, 4):
58      vst1q_f32(y, vy${ABC[N:N+4]}); y += 4;
59  }
60  $if BATCH_TILE > 4:
61    for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
62      const float32x4_t vx = vld1q_f32(x); x += 4;
63      const int32x4_t vintx = vcvtq_s32_f32(vx);
64      uint32x4_t vrndmask = vcaltq_f32(vx, vintegral_threshold);
65      const float32x4_t vprerndx = vcvtq_f32_s32(vintx);
66      vrndmask = vbicq_u32(vrndmask, vmovq_n_u32(UINT32_C(0x80000000)));
67      const float32x4_t vrndx = vbslq_f32(vrndmask, vprerndx, vx);
68      const uint32x4_t vadjmask = vcgtq_f32(vrndx, vx);
69      const float32x4_t vadjrndx = vreinterpretq_f32_u32(vandq_u32(vadjmask, vone));
70      const float32x4_t vy = vsubq_f32(vrndx, vadjrndx);
71      vst1q_f32(y, vy); y += 4;
72    }
73  if XNN_UNLIKELY(n != 0) {
74    const float32x4_t vx = vld1q_f32(x);
75    const int32x4_t vintx = vcvtq_s32_f32(vx);
76    uint32x4_t vrndmask = vcaltq_f32(vx, vintegral_threshold);
77    const float32x4_t vprerndx = vcvtq_f32_s32(vintx);
78    vrndmask = vbicq_u32(vrndmask, vmovq_n_u32(UINT32_C(0x80000000)));
79    const float32x4_t vrndx = vbslq_f32(vrndmask, vprerndx, vx);
80    const uint32x4_t vadjmask = vcgtq_f32(vrndx, vx);
81    const float32x4_t vadjrndx = vreinterpretq_f32_u32(vandq_u32(vadjmask, vone));
82    const float32x4_t vy = vsubq_f32(vrndx, vadjrndx);
83    float32x2_t vy_lo = vget_low_f32(vy);
84    if (n & (2 * sizeof(float))) {
85      vst1_f32(y, vy_lo); y += 2;
86      vy_lo = vget_high_f32(vy);
87    }
88    if (n & (1 * sizeof(float))) {
89      vst1_lane_f32(y, vy_lo, 0);
90    }
91  }
92}
93