• 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$assert BATCH_TILE % 8 == 0
7$assert BATCH_TILE >= 8
8$SIMD_TILE = BATCH_TILE // 8
9#include <assert.h>
10
11#include <immintrin.h>
12
13#include <xnnpack/common.h>
14#include <xnnpack/vunary.h>
15
16
17void xnn_f32_velu_ukernel__avx2_rr1_p6_x${BATCH_TILE}(
18    size_t n,
19    const float* x,
20    float* y,
21    const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS(1)])
22{
23  assert(n % sizeof(float) == 0);
24
25  const __m256 vprescale = _mm256_load_ps(params->avx2_rr1_p6.prescale);
26  const __m256 valpha = _mm256_load_ps(params->avx2_rr1_p6.alpha);
27  const __m256 vbeta = _mm256_load_ps(params->avx2_rr1_p6.beta);
28  const __m256 vsat_cutoff = _mm256_load_ps(params->avx2_rr1_p6.sat_cutoff);
29  const __m256 vmagic_bias = _mm256_load_ps(params->avx2_rr1_p6.magic_bias);
30  const __m256 vlog2e = _mm256_load_ps(params->avx2_rr1_p6.log2e);
31  const __m256 vminus_ln2 = _mm256_load_ps(params->avx2_rr1_p6.minus_ln2);
32  const __m256 vc6 = _mm256_load_ps(params->avx2_rr1_p6.c6);
33  const __m256 vc5 = _mm256_load_ps(params->avx2_rr1_p6.c5);
34  const __m256 vc4 = _mm256_load_ps(params->avx2_rr1_p6.c4);
35  const __m256 vc3 = _mm256_load_ps(params->avx2_rr1_p6.c3);
36  const __m256 vc2 = _mm256_load_ps(params->avx2_rr1_p6.c2);
37
38  $if BATCH_TILE > 8:
39    for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) {
40      __m256 vx0 = _mm256_loadu_ps(x);
41      $for N in range(1, SIMD_TILE):
42        __m256 vx${N} = _mm256_loadu_ps(x + ${N * 8});
43      x += ${BATCH_TILE};
44
45      $for N in range(SIMD_TILE):
46        const __m256 vz${N} = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx${N}, vprescale));
47
48      $for N in range(SIMD_TILE):
49        __m256 vn${N} = _mm256_fmadd_ps(vz${N}, vlog2e, vmagic_bias);
50
51      $for N in range(SIMD_TILE):
52        __m256 vs${N} = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn${N}), 23));
53        vn${N} = _mm256_sub_ps(vn${N}, vmagic_bias);
54
55      $for N in range(SIMD_TILE):
56        __m256 vt${N} = _mm256_fmadd_ps(vn${N}, vminus_ln2, vz${N});
57
58      $for N in range(SIMD_TILE):
59        __m256 vp${N} = _mm256_fmadd_ps(vc6, vt${N}, vc5);
60
61      $for N in range(SIMD_TILE):
62        vp${N} = _mm256_fmadd_ps(vp${N}, vt${N}, vc4);
63
64      $for N in range(SIMD_TILE):
65        vp${N} = _mm256_fmadd_ps(vp${N}, vt${N}, vc3);
66
67      $for N in range(SIMD_TILE):
68        vp${N} = _mm256_fmadd_ps(vp${N}, vt${N}, vc2);
69
70      $for N in range(SIMD_TILE):
71        vp${N} = _mm256_mul_ps(vp${N}, vt${N});
72        vt${N} = _mm256_mul_ps(vt${N}, vs${N});
73
74      $for N in range(SIMD_TILE):
75        vs${N} = _mm256_fmsub_ps(vs${N}, valpha, valpha);
76        vp${N} = _mm256_fmadd_ps(vp${N}, vt${N}, vt${N});
77
78      $for N in range(SIMD_TILE):
79        const __m256 ve${N} = _mm256_fmadd_ps(vp${N}, valpha, vs${N});
80        vx${N} = _mm256_mul_ps(vx${N}, vbeta);
81
82      $for N in range(SIMD_TILE):
83        const __m256 vy${N} = _mm256_blendv_ps(vx${N}, ve${N}, vx${N});
84
85      _mm256_storeu_ps(y, vy0);
86      $for N in range(1, SIMD_TILE):
87        _mm256_storeu_ps(y + ${N * 8}, vy${N});
88      y += ${BATCH_TILE};
89    }
90  for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
91    __m256 vx = _mm256_loadu_ps(x);
92    x += 8;
93
94    const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
95
96    __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
97    __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
98    vn = _mm256_sub_ps(vn, vmagic_bias);
99
100    __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
101
102    __m256 vp = _mm256_fmadd_ps(vc6, vt, vc5);
103    vp = _mm256_fmadd_ps(vp, vt, vc4);
104    vp = _mm256_fmadd_ps(vp, vt, vc3);
105    vp = _mm256_fmadd_ps(vp, vt, vc2);
106    vp = _mm256_mul_ps(vp, vt);
107
108    vt = _mm256_mul_ps(vt, vs);
109    vs = _mm256_fmsub_ps(vs, valpha, valpha);
110    vp = _mm256_fmadd_ps(vp, vt, vt);
111    const __m256 ve = _mm256_fmadd_ps(vp, valpha, vs);
112
113    vx = _mm256_mul_ps(vx, vbeta);
114    const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
115
116    _mm256_storeu_ps(y, vy);
117    y += 8;
118  }
119  if XNN_UNLIKELY(n != 0) {
120    assert(n >= 1 * sizeof(float));
121    assert(n <= 7 * sizeof(float));
122    const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &params->avx2_rr1_p6.mask_table[7] - n));
123
124    __m256 vx = _mm256_maskload_ps(x, vmask);
125
126    const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
127
128    __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
129    __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
130    vn = _mm256_sub_ps(vn, vmagic_bias);
131
132    __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
133
134    __m256 vp = _mm256_fmadd_ps(vc6, vt, vc5);
135    vp = _mm256_fmadd_ps(vp, vt, vc4);
136    vp = _mm256_fmadd_ps(vp, vt, vc3);
137    vp = _mm256_fmadd_ps(vp, vt, vc2);
138    vp = _mm256_mul_ps(vp, vt);
139
140    vt = _mm256_mul_ps(vt, vs);
141    vs = _mm256_fmsub_ps(vs, valpha, valpha);
142    vp = _mm256_fmadd_ps(vp, vt, vt);
143    const __m256 ve = _mm256_fmadd_ps(vp, valpha, vs);
144
145    vx = _mm256_mul_ps(vx, vbeta);
146    const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
147
148    __m128 vy_lo = _mm256_castps256_ps128(vy);
149    if (n & (4 * sizeof(float))) {
150      _mm_storeu_ps(y, vy_lo);
151      vy_lo = _mm256_extractf128_ps(vy, 1);
152      y += 4;
153    }
154    if (n & (2 * sizeof(float))) {
155      _mm_storel_pi((__m64*) y, vy_lo);
156      vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
157      y += 2;
158    }
159    if (n & (1 * sizeof(float))) {
160      _mm_store_ss(y, vy_lo);
161    }
162  }
163}
164