xref: /aosp_15_r20/external/XNNPACK/src/f32-vsqrt/neon-sqrt.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$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
9#include <assert.h>
10#include <math.h>
11
12#include <arm_neon.h>
13
14#include <xnnpack/common.h>
15#include <xnnpack/vunary.h>
16
17
18void xnn_f32_vsqrt_ukernel__neon_sqrt_x${BATCH_TILE}(
19    size_t n,
20    const float* x,
21    float* y,
22    const union xnn_f32_sqrt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23{
24  assert(n != 0);
25  assert(n % sizeof(float) == 0);
26
27  $if BATCH_TILE > 4:
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      $for N in range(0, BATCH_TILE, 4):
33        const float32x4_t vy${ABC[N:N+4]} = vsqrtq_f32(vx${ABC[N:N+4]});
34
35      $for N in range(0, BATCH_TILE, 4):
36        vst1q_f32(y, vy${ABC[N:N+4]}); y += 4;
37    }
38  for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
39    const float32x4_t vx = vld1q_f32(x); x += 4;
40    const float32x4_t vy = vsqrtq_f32(vx);
41    vst1q_f32(y, vy); y += 4;
42  }
43  if XNN_UNLIKELY(n != 0) {
44    const float32x4_t vx = vld1q_f32(x);
45
46    float32x2_t vy_lo = vsqrt_f32(vget_low_f32(vx));
47    const float32x2_t vx_hi = vget_high_f32(vx);
48    if (n & (2 * sizeof(float))) {
49      vst1_f32(y, vy_lo); y += 2;
50      vy_lo = vsqrt_f32(vx_hi);
51    }
52    if (n & (1 * sizeof(float))) {
53      vst1_lane_f32(y, vy_lo, 0);
54    }
55  }
56}
57