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$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
9#include <assert.h>
10
11#include <xmmintrin.h>
12
13#include <xnnpack/vunary.h>
14#include <xnnpack/common.h>
15
16
17void xnn_f32_relu_ukernel__sse_x${BATCH_TILE}(
18    size_t n,
19    const float* x,
20    float* y,
21    const union xnn_f32_relu_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 vzero = _mm_setzero_ps();
29
30  for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) {
31    __m128 vacc${ABC[0:4]} = _mm_loadu_ps(x);
32    $for N in range(4, BATCH_TILE, 4):
33      __m128 vacc${ABC[N:N+4]} = _mm_loadu_ps(x + ${N});
34    x += ${BATCH_TILE};
35
36    $for N in range(0, BATCH_TILE, 4):
37      vacc${ABC[N:N+4]} = _mm_max_ps(vacc${ABC[N:N+4]}, vzero);
38
39    _mm_storeu_ps(y, vacc${ABC[0:4]});
40    $for N in range(4, BATCH_TILE, 4):
41      _mm_storeu_ps(y + ${N}, vacc${ABC[N:N+4]});
42    y += ${BATCH_TILE};
43  }
44  $if BATCH_TILE > 4:
45    for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
46      __m128 vacc = _mm_loadu_ps(x);
47      x += 4;
48
49      vacc = _mm_max_ps(vacc, vzero);
50
51      _mm_storeu_ps(y, vacc);
52      y += 4;
53    }
54  if XNN_UNLIKELY(n != 0) {
55    __m128 vacc = _mm_loadu_ps(x);
56    vacc = _mm_max_ps(vacc, vzero);
57
58    if (n & (2 * sizeof(float))) {
59      _mm_storel_pi((__m64*) y, vacc);
60      vacc = _mm_movehl_ps(vacc, vacc);
61      y += 2;
62    }
63    if (n & (1 * sizeof(float))) {
64      _mm_store_ss(y, vacc);
65    }
66  }
67}
68