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$SIMD_TILE = BATCH_TILE // 8
9#include <assert.h>
10
11#include <immintrin.h>
12
13#include <xnnpack/common.h>
14#include <xnnpack/vunary.h>
15
16
17static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};
18
19void xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x${BATCH_TILE}(
20    size_t n,
21    const float* x,
22    float* y,
23    const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS(1)])
24{
25  assert(n % sizeof(float) == 0);
26
27  const __m256 vprescale = _mm256_broadcast_ps((const __m128*) params->sse.prescale);
28  const __m256 valpha = _mm256_broadcast_ps((const __m128*) params->sse.alpha);
29  const __m256 vbeta = _mm256_broadcast_ps((const __m128*) params->sse.beta);
30
31  const __m256 vsat_cutoff = _mm256_set1_ps(-0x1.154246p+4f);
32  const __m256 vmagic_bias = _mm256_set1_ps(0x1.800000p20f);
33  const __m256 vlog2e = _mm256_set1_ps(0x1.715476p+0f);
34  const __m256i vtable = _mm256_set_epi32(
35    0x3F7AC0C7, 0x3F7744FD, 0x3F75672A, 0x3F7504F3, 0x3F75FED7, 0x3F7837F0, 0x3F7B95C2, 0x3F800000);
36  const __m256 vminus_ln2 = _mm256_set1_ps(-0x1.62E43p-1f);
37  const __m256 vc4 = _mm256_set1_ps(0x1.5558ECp-5f);
38  const __m256 vc3 = _mm256_set1_ps(0x1.555C20p-3f);
39  const __m256 vc2 = _mm256_set1_ps(0x1.000000p-1f);
40
41  $if BATCH_TILE > 8:
42    for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) {
43      __m256 vx0 = _mm256_loadu_ps(x);
44      $for N in range(1, SIMD_TILE):
45        __m256 vx${N} = _mm256_loadu_ps(x + ${N * 8});
46      x += ${BATCH_TILE};
47
48      $for N in range(SIMD_TILE):
49        const __m256 vz${N} = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx${N}, vprescale));
50
51      $for N in range(SIMD_TILE):
52        __m256 vn${N} = _mm256_fmadd_ps(vz${N}, vlog2e, vmagic_bias);
53
54      $for N in range(SIMD_TILE):
55        const __m256i ven${N} = _mm256_slli_epi32(_mm256_castps_si256(vn${N}), 20);
56        const __m256i vl${N} = _mm256_permutevar8x32_epi32(vtable, _mm256_castps_si256(vn${N}));
57        vn${N} = _mm256_sub_ps(vn${N}, vmagic_bias);
58
59      $for N in range(SIMD_TILE):
60        __m256 vs${N} = _mm256_castsi256_ps(_mm256_add_epi32(vl${N}, ven${N}));
61        __m256 vt${N} = _mm256_fmadd_ps(vn${N}, vminus_ln2, vz${N});
62
63      $for N in range(SIMD_TILE):
64        __m256 vp${N} = _mm256_fmadd_ps(vc4, vt${N}, vc3);
65
66      $for N in range(SIMD_TILE):
67        vp${N} = _mm256_fmadd_ps(vp${N}, vt${N}, vc2);
68
69      $for N in range(SIMD_TILE):
70        vp${N} = _mm256_mul_ps(vp${N}, vt${N});
71        vt${N} = _mm256_mul_ps(vt${N}, vs${N});
72
73      $for N in range(SIMD_TILE):
74        vs${N} = _mm256_fmsub_ps(vs${N}, valpha, valpha);
75        vp${N} = _mm256_fmadd_ps(vp${N}, vt${N}, vt${N});
76
77      $for N in range(SIMD_TILE):
78        const __m256 ve${N} = _mm256_fmadd_ps(vp${N}, valpha, vs${N});
79        vx${N} = _mm256_mul_ps(vx${N}, vbeta);
80
81      $for N in range(SIMD_TILE):
82        const __m256 vy${N} = _mm256_blendv_ps(vx${N}, ve${N}, vx${N});
83
84      _mm256_storeu_ps(y, vy0);
85      $for N in range(1, SIMD_TILE):
86        _mm256_storeu_ps(y + ${N * 8}, vy${N});
87      y += ${BATCH_TILE};
88    }
89  for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
90    __m256 vx = _mm256_loadu_ps(x);
91    x += 8;
92
93    const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
94
95    __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
96    const __m256i ven = _mm256_slli_epi32(_mm256_castps_si256(vn), 20);
97    const __m256i vl = _mm256_permutevar8x32_epi32(vtable, _mm256_castps_si256(vn));
98    __m256 vs = _mm256_castsi256_ps(_mm256_add_epi32(vl, ven));
99    vn = _mm256_sub_ps(vn, vmagic_bias);
100
101    __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
102
103    __m256 vp = _mm256_fmadd_ps(vc4, vt, vc3);
104    vp = _mm256_fmadd_ps(vp, vt, vc2);
105    vp = _mm256_mul_ps(vp, vt);
106
107    vt = _mm256_mul_ps(vt, vs);
108    vs = _mm256_fmsub_ps(vs, valpha, valpha);
109    vp = _mm256_fmadd_ps(vp, vt, vt);
110    const __m256 ve = _mm256_fmadd_ps(vp, valpha, vs);
111
112    vx = _mm256_mul_ps(vx, vbeta);
113    const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
114
115    _mm256_storeu_ps(y, vy);
116    y += 8;
117  }
118  if XNN_UNLIKELY(n != 0) {
119    assert(n >= 1 * sizeof(float));
120    assert(n <= 7 * sizeof(float));
121    __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
122
123    __m256 vx = _mm256_maskload_ps(x, vmask);
124
125    const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
126
127    __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
128    const __m256i ven = _mm256_slli_epi32(_mm256_castps_si256(vn), 20);
129    const __m256i vl = _mm256_permutevar8x32_epi32(vtable, _mm256_castps_si256(vn));
130    __m256 vs = _mm256_castsi256_ps(_mm256_add_epi32(vl, ven));
131    vn = _mm256_sub_ps(vn, vmagic_bias);
132
133    __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
134
135    __m256 vp = _mm256_fmadd_ps(vc4, vt, vc3);
136    vp = _mm256_fmadd_ps(vp, vt, vc2);
137    vp = _mm256_mul_ps(vp, vt);
138
139    vt = _mm256_mul_ps(vt, vs);
140    vs = _mm256_fmsub_ps(vs, valpha, valpha);
141    vp = _mm256_fmadd_ps(vp, vt, vt);
142    const __m256 ve = _mm256_fmadd_ps(vp, valpha, vs);
143
144    vx = _mm256_mul_ps(vx, vbeta);
145    const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
146
147    // _mm256_maskstore_ps(y, vmask, vf) could be used here, but triggers msan failures (probably an msan bug).
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