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$SIMD_TILE = BATCH_TILE // 8
9#include <assert.h>
10
11#include <immintrin.h>
12
13#include <xnnpack/common.h>
14#include <xnnpack/vunary.h>
15
16
17static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};
18
19void xnn_f32_velu_ukernel__avx_rr2_lut4_p4_perm_x${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(0x1.8003F8p21f);
33  const __m256 vlog2e = _mm256_set1_ps(0x1.715476p+0f);
34  const __m256 vindex_mask = _mm256_castsi256_ps(_mm256_set1_epi32(0x3));
35  const __m256 vtable = _mm256_set_ps(
36    0x1.AE89FAp+0f, 0x1.6A09E6p+0f, 0x1.306FE0p+0f, 0x1.000000p+0f,
37    0x1.AE89FAp+0f, 0x1.6A09E6p+0f, 0x1.306FE0p+0f, 0x1.000000p+0f);
38  const __m256 vminus_ln2_hi = _mm256_set1_ps(-0x1.62E400p-1f);
39  const __m256 vminus_ln2_lo = _mm256_set1_ps(-0x1.7F7D1Cp-20f);
40  const __m256 vc4 = _mm256_set1_ps(0x1.554F9Ap-5f);
41  const __m256 vc3 = _mm256_set1_ps(0x1.557082p-3f);
42  const __m256 vc2 = _mm256_set1_ps(0x1.000002p-1f);
43  const __m256 vone = _mm256_set1_ps(1.0f);
44
45  $if BATCH_TILE > 8:
46    for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) {
47      __m256 vx0 = _mm256_loadu_ps(x);
48      $for N in range(1, SIMD_TILE):
49        __m256 vx${N} = _mm256_loadu_ps(x + ${N * 8});
50      x += ${BATCH_TILE};
51
52      $for N in range(SIMD_TILE):
53        const __m256 vz${N} = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx${N}, vprescale));
54
55      $for N in range(SIMD_TILE):
56        __m256 vn${N} = _mm256_add_ps(_mm256_mul_ps(vz${N}, vlog2e), vmagic_bias);
57
58      $for N in range(SIMD_TILE):
59        __m256 ven${N} = _mm256_andnot_ps(vindex_mask, vn${N});
60        const __m256 vl${N} = _mm256_permutevar_ps(vtable, _mm256_castps_si256(vn${N}));
61        const __m128 ven${N}_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(ven${N})), 21));
62
63      $for N in range(SIMD_TILE):
64        vn${N} = _mm256_sub_ps(vn${N}, vmagic_bias);
65        const __m128 ven${N}_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven${N}, 1)), 21));
66
67      $for N in range(SIMD_TILE):
68        __m256 vt${N} = _mm256_add_ps(_mm256_mul_ps(vn${N}, vminus_ln2_hi), vz${N});
69        ven${N} = _mm256_insertf128_ps(_mm256_castps128_ps256(ven${N}_lo), ven${N}_hi, 1);
70
71      $for N in range(SIMD_TILE):
72        vt${N} = _mm256_add_ps(_mm256_mul_ps(vn${N}, vminus_ln2_lo), vt${N});
73        __m256 vs${N} = _mm256_mul_ps(vl${N}, ven${N});
74
75      $for N in range(SIMD_TILE):
76        __m256 vp${N} = _mm256_add_ps(_mm256_mul_ps(vc4, vt${N}), vc3);
77
78      $for N in range(SIMD_TILE):
79        vp${N} = _mm256_add_ps(_mm256_mul_ps(vp${N}, vt${N}), vc2);
80
81      $for N in range(SIMD_TILE):
82        vp${N} = _mm256_mul_ps(vp${N}, vt${N});
83
84      $for N in range(SIMD_TILE):
85        vt${N} = _mm256_mul_ps(vt${N}, vs${N});
86        vs${N} = _mm256_sub_ps(vs${N}, vone);
87
88      $for N in range(SIMD_TILE):
89        vp${N} = _mm256_add_ps(_mm256_mul_ps(vp${N}, vt${N}), vt${N});
90
91      $for N in range(SIMD_TILE):
92        const __m256 ve${N} = _mm256_mul_ps(_mm256_add_ps(vp${N}, vs${N}), valpha);
93        vx${N} = _mm256_mul_ps(vx${N}, vbeta);
94
95      $for N in range(SIMD_TILE):
96        const __m256 vy${N} = _mm256_blendv_ps(vx${N}, ve${N}, vx${N});
97
98      _mm256_storeu_ps(y, vy0);
99      $for N in range(1, SIMD_TILE):
100        _mm256_storeu_ps(y + ${N * 8}, vy${N});
101      y += ${BATCH_TILE};
102    }
103  for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
104    __m256 vx = _mm256_loadu_ps(x);
105    x += 8;
106
107    const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
108
109    __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
110    __m256 ven = _mm256_andnot_ps(vindex_mask, vn);
111    const __m256 vl = _mm256_permutevar_ps(vtable, _mm256_castps_si256(vn));
112    const __m128 ven_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(ven)), 21));
113    vn = _mm256_sub_ps(vn, vmagic_bias);
114    const __m128 ven_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven, 1)), 21));
115
116    __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
117    ven = _mm256_insertf128_ps(_mm256_castps128_ps256(ven_lo), ven_hi, 1);
118    vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
119    __m256 vs = _mm256_mul_ps(vl, ven);
120
121    __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc4, vt), vc3);
122    vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
123    vp = _mm256_mul_ps(vp, vt);
124
125    vt = _mm256_mul_ps(vt, vs);
126    vs = _mm256_sub_ps(vs, vone);
127    vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vt);
128
129    const __m256 ve = _mm256_mul_ps(_mm256_add_ps(vp, vs), valpha);
130    vx = _mm256_mul_ps(vx, vbeta);
131    const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
132
133    _mm256_storeu_ps(y, vy);
134    y += 8;
135  }
136  if XNN_UNLIKELY(n != 0) {
137    assert(n >= 1 * sizeof(float));
138    assert(n <= 7 * sizeof(float));
139    __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
140
141    __m256 vx = _mm256_maskload_ps(x, vmask);
142
143    const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
144
145    __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
146    __m256 ven = _mm256_andnot_ps(vindex_mask, vn);
147    const __m256 vl = _mm256_permutevar_ps(vtable, _mm256_castps_si256(vn));
148    const __m128 ven_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(ven)), 21));
149    vn = _mm256_sub_ps(vn, vmagic_bias);
150    const __m128 ven_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven, 1)), 21));
151
152    __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
153    ven = _mm256_insertf128_ps(_mm256_castps128_ps256(ven_lo), ven_hi, 1);
154    vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
155    __m256 vs = _mm256_mul_ps(vl, ven);
156
157    __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc4, vt), vc3);
158    vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
159    vp = _mm256_mul_ps(vp, vt);
160
161    vt = _mm256_mul_ps(vt, vs);
162    vs = _mm256_sub_ps(vs, vone);
163    vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vt);
164
165    const __m256 ve = _mm256_mul_ps(_mm256_add_ps(vp, vs), valpha);
166    vx = _mm256_mul_ps(vx, vbeta);
167    const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
168
169    // _mm256_maskstore_ps(y, vmask, vf) could be used here, but triggers msan failures (probably an msan bug).
170    __m128 vy_lo = _mm256_castps256_ps128(vy);
171    if (n & (4 * sizeof(float))) {
172      _mm_storeu_ps(y, vy_lo);
173      vy_lo = _mm256_extractf128_ps(vy, 1);
174      y += 4;
175    }
176    if (n & (2 * sizeof(float))) {
177      _mm_storel_pi((__m64*) y, vy_lo);
178      vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
179      y += 2;
180    }
181    if (n & (1 * sizeof(float))) {
182      _mm_store_ss(y, vy_lo);
183    }
184  }
185}
186