• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
9#include <assert.h>
10#include <math.h>
11
12#include <arm_neon.h>
13
14#include <xnnpack/common.h>
15#include <xnnpack/vunary.h>
16
17
18void xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x${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_DISABLE_TSAN
23{
24  assert(n != 0);
25  assert(n % sizeof(float) == 0);
26
27  $if BATCH_TILE > 4:
28    const float32x4_t vhalf = vmovq_n_f32(0.5f);
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      $for N in range(0, BATCH_TILE, 4):
34        const float32x4_t vrsqrtx${ABC[N:N+4]} = vrsqrteq_f32(vx${ABC[N:N+4]});
35
36      $for N in range(0, BATCH_TILE, 4):
37        float32x4_t vsqrtx${ABC[N:N+4]} = vmulq_f32(vrsqrtx${ABC[N:N+4]}, vx${ABC[N:N+4]});
38        float32x4_t vhalfrsqrtx${ABC[N:N+4]} = vmulq_f32(vrsqrtx${ABC[N:N+4]}, vhalf);
39
40      $for N in range(0, BATCH_TILE, 4):
41        float32x4_t vresidual${ABC[N:N+4]} = vfmsq_f32(vhalf, vsqrtx${ABC[N:N+4]}, vhalfrsqrtx${ABC[N:N+4]});
42
43      $for N in range(0, BATCH_TILE, 4):
44        vhalfrsqrtx${ABC[N:N+4]} = vfmaq_f32(vhalfrsqrtx${ABC[N:N+4]}, vresidual${ABC[N:N+4]}, vhalfrsqrtx${ABC[N:N+4]});
45        vsqrtx${ABC[N:N+4]} = vfmaq_f32(vsqrtx${ABC[N:N+4]}, vresidual${ABC[N:N+4]}, vsqrtx${ABC[N:N+4]});
46
47      $for N in range(0, BATCH_TILE, 4):
48        vresidual${ABC[N:N+4]} = vfmsq_f32(vhalf, vsqrtx${ABC[N:N+4]}, vhalfrsqrtx${ABC[N:N+4]});
49
50      $for N in range(0, BATCH_TILE, 4):
51        vhalfrsqrtx${ABC[N:N+4]} = vfmaq_f32(vhalfrsqrtx${ABC[N:N+4]}, vresidual${ABC[N:N+4]}, vhalfrsqrtx${ABC[N:N+4]});
52        vsqrtx${ABC[N:N+4]} = vfmaq_f32(vsqrtx${ABC[N:N+4]}, vresidual${ABC[N:N+4]}, vsqrtx${ABC[N:N+4]});
53
54      $for N in range(0, BATCH_TILE, 4):
55        const float32x4_t vadjustment${ABC[N:N+4]} = vfmsq_f32(vx${ABC[N:N+4]}, vsqrtx${ABC[N:N+4]}, vsqrtx${ABC[N:N+4]});
56
57      $for N in range(0, BATCH_TILE, 4):
58        const float32x4_t vy${ABC[N:N+4]} = vfmaq_f32(vsqrtx${ABC[N:N+4]}, vhalfrsqrtx${ABC[N:N+4]}, vadjustment${ABC[N:N+4]});
59
60      $for N in range(0, BATCH_TILE, 4):
61        vst1q_f32(y, vy${ABC[N:N+4]}); y += 4;
62    }
63  if XNN_UNLIKELY(n != 0) {
64    do {
65      const float vx = *x++;
66      const float vy = sqrtf(vx);
67      *y++ = vy;
68      n -= sizeof(float);
69    } while (n != 0);
70  }
71}
72