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$assert BATCH_TILE % 8 == 0
7$assert BATCH_TILE >= 8
8$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
9#include <assert.h>
10
11#include <immintrin.h>
12
13#include <xnnpack/clamp.h>
14#include <xnnpack/common.h>
15
16
17static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};
18
19void xnn_f32_clamp_ukernel__avx_x${BATCH_TILE}(
20    size_t n,
21    const float* x,
22    float* y,
23    const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
24{
25  assert(n != 0);
26  assert(n % sizeof(float) == 0);
27  assert(x != NULL);
28  assert(y != NULL);
29
30  const __m256 vy_min = _mm256_broadcast_ps((const __m128*) params->sse.min);
31  const __m256 vy_max = _mm256_broadcast_ps((const __m128*) params->sse.max);
32
33  for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) {
34    __m256 vacc${ABC[0:8]} = _mm256_loadu_ps(x);
35    $for N in range(8, BATCH_TILE, 8):
36      __m256 vacc${ABC[N:N+8]} = _mm256_loadu_ps(x + ${N});
37    x += ${BATCH_TILE};
38
39    $for N in range(0, BATCH_TILE, 8):
40      vacc${ABC[N:N+8]} = _mm256_max_ps(vacc${ABC[N:N+8]}, vy_min);
41
42    $for N in range(0, BATCH_TILE, 8):
43      vacc${ABC[N:N+8]} = _mm256_min_ps(vacc${ABC[N:N+8]}, vy_max);
44
45    _mm256_storeu_ps(y, vacc${ABC[0:8]});
46    $for N in range(8, BATCH_TILE, 8):
47      _mm256_storeu_ps(y + ${N}, vacc${ABC[N:N+8]});
48    y += ${BATCH_TILE};
49  }
50  $if BATCH_TILE > 8:
51    for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
52      __m256 vacc = _mm256_loadu_ps(x);
53      x += 8;
54
55      vacc = _mm256_max_ps(vacc, vy_min);
56      vacc = _mm256_min_ps(vacc, vy_max);
57
58      _mm256_storeu_ps(y, vacc);
59      y += 8;
60    }
61  if XNN_UNLIKELY(n != 0) {
62    assert(n >= 1 * sizeof(float));
63    assert(n <= 7 * sizeof(float));
64    __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
65
66    __m256 vacc = _mm256_maskload_ps(x, vmask);
67    vacc = _mm256_max_ps(vacc, vy_min);
68    vacc = _mm256_min_ps(vacc, vy_max);
69
70    // _mm256_maskstore_ps(y, vmask, vacc) could be used here, but triggers msan failures (probably an msan bug).
71    __m128 vacc_lo = _mm256_castps256_ps128(vacc);
72    if (n & (4 * sizeof(float))) {
73      _mm_storeu_ps(y, vacc_lo);
74      vacc_lo = _mm256_extractf128_ps(vacc, 1);
75      y += 4;
76    }
77    if (n & (2 * sizeof(float))) {
78      _mm_storel_pi((__m64*) y, vacc_lo);
79      vacc_lo = _mm_movehl_ps(vacc_lo, vacc_lo);
80      y += 2;
81    }
82    if (n & (1 * sizeof(float))) {
83      _mm_store_ss(y, vacc_lo);
84    }
85  }
86}
87