xref: /aosp_15_r20/external/XNNPACK/src/f32-vrnd/vrndz-neon.c.in (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
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$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
9#include <assert.h>
10
11#include <arm_neon.h>
12
13#include <xnnpack/common.h>
14#include <xnnpack/math.h>
15#include <xnnpack/vunary.h>
16
17
18void xnn_f32_vrndz_ukernel__neon_x${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_OOB_READS
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  for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) {
29    $for N in range(0, BATCH_TILE, 4):
30      const float32x4_t vx${ABC[N:N+4]} = vld1q_f32(x); x += 4;
31
32    $for N in range(0, BATCH_TILE, 4):
33      const int32x4_t vintx${ABC[N:N+4]} = vcvtq_s32_f32(vx${ABC[N:N+4]});
34
35    $for N in range(0, BATCH_TILE, 4):
36      uint32x4_t vrndmask${ABC[N:N+4]} = vcaltq_f32(vx${ABC[N:N+4]}, vintegral_threshold);
37
38    $for N in range(0, BATCH_TILE, 4):
39      const float32x4_t vrndx${ABC[N:N+4]} = vcvtq_f32_s32(vintx${ABC[N:N+4]});
40
41    $for N in range(0, BATCH_TILE, 4):
42      vrndmask${ABC[N:N+4]} = vbicq_u32(vrndmask${ABC[N:N+4]}, vmovq_n_u32(UINT32_C(0x80000000)));
43
44    $for N in range(0, BATCH_TILE, 4):
45      const float32x4_t vy${ABC[N:N+4]} = vbslq_f32(vrndmask${ABC[N:N+4]}, vrndx${ABC[N:N+4]}, vx${ABC[N:N+4]});
46
47    $for N in range(0, BATCH_TILE, 4):
48      vst1q_f32(y, vy${ABC[N:N+4]}); y += 4;
49  }
50  $if BATCH_TILE > 4:
51    for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
52      const float32x4_t vx = vld1q_f32(x); x += 4;
53      const int32x4_t vintx = vcvtq_s32_f32(vx);
54      uint32x4_t vrndmask = vcaltq_f32(vx, vintegral_threshold);
55      const float32x4_t vrndx = vcvtq_f32_s32(vintx);
56      vrndmask = vbicq_u32(vrndmask, vmovq_n_u32(UINT32_C(0x80000000)));
57      const float32x4_t vy = vbslq_f32(vrndmask, vrndx, vx);
58      vst1q_f32(y, vy); y += 4;
59    }
60  if XNN_UNLIKELY(n != 0) {
61    const float32x4_t vx = vld1q_f32(x);
62    const int32x4_t vintx = vcvtq_s32_f32(vx);
63    uint32x4_t vrndmask = vcaltq_f32(vx, vintegral_threshold);
64    const float32x4_t vrndx = vcvtq_f32_s32(vintx);
65    vrndmask = vbicq_u32(vrndmask, vmovq_n_u32(UINT32_C(0x80000000)));
66    const float32x4_t vy = vbslq_f32(vrndmask, vrndx, vx);
67    float32x2_t vy_lo = vget_low_f32(vy);
68    if (n & (2 * sizeof(float))) {
69      vst1_f32(y, vy_lo); y += 2;
70      vy_lo = vget_high_f32(vy);
71    }
72    if (n & (1 * sizeof(float))) {
73      vst1_lane_f32(y, vy_lo, 0);
74    }
75  }
76}
77