• 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$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
9#include <assert.h>
10
11#include <xmmintrin.h>
12
13#include <xnnpack/clamp.h>
14#include <xnnpack/common.h>
15
16
17void xnn_f32_clamp_ukernel__sse_x${BATCH_TILE}(
18    size_t n,
19    const float* x,
20    float* y,
21    const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
22{
23  assert(n != 0);
24  assert(n % sizeof(float) == 0);
25  assert(x != NULL);
26  assert(y != NULL);
27
28  const __m128 vy_min = _mm_load_ps(params->sse.min);
29  const __m128 vy_max = _mm_load_ps(params->sse.max);
30
31  for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) {
32    __m128 vacc${ABC[0:4]} = _mm_loadu_ps(x);
33    $for N in range(4, BATCH_TILE, 4):
34      __m128 vacc${ABC[N:N+4]} = _mm_loadu_ps(x + ${N});
35    x += ${BATCH_TILE};
36
37    $for N in range(0, BATCH_TILE, 4):
38      vacc${ABC[N:N+4]} = _mm_max_ps(vacc${ABC[N:N+4]}, vy_min);
39
40    $for N in range(0, BATCH_TILE, 4):
41      vacc${ABC[N:N+4]} = _mm_min_ps(vacc${ABC[N:N+4]}, vy_max);
42
43    _mm_storeu_ps(y, vacc${ABC[0:4]});
44    $for N in range(4, BATCH_TILE, 4):
45      _mm_storeu_ps(y + ${N}, vacc${ABC[N:N+4]});
46    y += ${BATCH_TILE};
47  }
48  $if BATCH_TILE > 4:
49    for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
50      __m128 vacc = _mm_loadu_ps(x);
51      x += 4;
52
53      vacc = _mm_max_ps(vacc, vy_min);
54      vacc = _mm_min_ps(vacc, vy_max);
55
56      _mm_storeu_ps(y, vacc);
57      y += 4;
58    }
59  if XNN_UNLIKELY(n != 0) {
60    __m128 vacc = _mm_loadu_ps(x);
61    vacc = _mm_max_ps(vacc, vy_min);
62    vacc = _mm_min_ps(vacc, vy_max);
63
64    if (n & (2 * sizeof(float))) {
65      _mm_storel_pi((__m64*) y, vacc);
66      vacc = _mm_movehl_ps(vacc, vacc);
67      y += 2;
68    }
69    if (n & (1 * sizeof(float))) {
70      _mm_store_ss(y, vacc);
71    }
72  }
73}
74