xref: /aosp_15_r20/external/XNNPACK/src/f32-vsqrt/sse-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$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
9#include <assert.h>
10
11#include <xmmintrin.h>
12
13#include <xnnpack/common.h>
14#include <xnnpack/vunary.h>
15
16
17void xnn_f32_vsqrt_ukernel__sse_sqrt_x${BATCH_TILE}(
18    size_t n,
19    const float* x,
20    float* y,
21    const union xnn_f32_sqrt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
22{
23  assert(n != 0);
24  assert(n % sizeof(float) == 0);
25
26  $if BATCH_TILE > 4:
27    for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) {
28      const __m128 vx${ABC[0:4]} = _mm_loadu_ps(x);
29      $for N in range(4, BATCH_TILE, 4):
30        const __m128 vx${ABC[N:N+4]} = _mm_loadu_ps(x + ${N});
31      x += ${BATCH_TILE};
32
33      $for N in range(0, BATCH_TILE, 4):
34        const __m128 vy${ABC[N:N+4]} = _mm_sqrt_ps(vx${ABC[N:N+4]});
35
36      _mm_storeu_ps(y, vy${ABC[0:4]});
37      $for N in range(4, BATCH_TILE, 4):
38        _mm_storeu_ps(y + ${N}, vy${ABC[N:N+4]});
39      y += ${BATCH_TILE};
40    }
41  for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
42    const __m128 vx = _mm_loadu_ps(x);
43    x += 4;
44    const __m128 vy = _mm_sqrt_ps(vx);
45    _mm_storeu_ps(y, vy);
46    y += 4;
47  }
48  if XNN_UNLIKELY(n != 0) {
49    const __m128 vx = _mm_loadu_ps(x);
50    __m128 vy = _mm_sqrt_ps(vx);
51    if (n & (2 * sizeof(float))) {
52      _mm_storel_pi((__m64*) y, vy);
53      vy = _mm_movehl_ps(vy, vy);
54      y += 2;
55    }
56    if (n & (1 * sizeof(float))) {
57      _mm_store_ss(y, vy);
58    }
59  }
60}
61