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$SIMD_TILE = BATCH_TILE // 8
9#include <assert.h>
10
11#include <immintrin.h>
12
13#include <xnnpack/common.h>
14#include <xnnpack/vunary.h>
15
16
17static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};
18
19void xnn_f32_velu_ukernel__avx2_rr1_p6_x${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(0x1.8000FEp23f);
33  const __m256 vlog2e = _mm256_set1_ps(0x1.715476p+0f);
34  const __m256 vminus_ln2 = _mm256_set1_ps(-0x1.62E43p-1f);
35  const __m256 vc6 = _mm256_set1_ps(0x1.6b7338p-10f);
36  const __m256 vc5 = _mm256_set1_ps(0x1.12278Ep-7f);
37  const __m256 vc4 = _mm256_set1_ps(0x1.555716p-5f);
38  const __m256 vc3 = _mm256_set1_ps(0x1.5554B0p-3f);
39  const __m256 vc2 = _mm256_set1_ps(0x1.FFFFFEp-2f);
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      $for N in range(SIMD_TILE):
55        __m256 vs${N} = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn${N}), 23));
56        vn${N} = _mm256_sub_ps(vn${N}, vmagic_bias);
57
58      $for N in range(SIMD_TILE):
59        __m256 vt${N} = _mm256_fmadd_ps(vn${N}, vminus_ln2, vz${N});
60
61      $for N in range(SIMD_TILE):
62        __m256 vp${N} = _mm256_fmadd_ps(vc6, vt${N}, vc5);
63
64      $for N in range(SIMD_TILE):
65        vp${N} = _mm256_fmadd_ps(vp${N}, vt${N}, vc4);
66
67      $for N in range(SIMD_TILE):
68        vp${N} = _mm256_fmadd_ps(vp${N}, vt${N}, vc3);
69
70      $for N in range(SIMD_TILE):
71        vp${N} = _mm256_fmadd_ps(vp${N}, vt${N}, vc2);
72
73      $for N in range(SIMD_TILE):
74        vp${N} = _mm256_mul_ps(vp${N}, vt${N});
75        vt${N} = _mm256_mul_ps(vt${N}, vs${N});
76
77      $for N in range(SIMD_TILE):
78        vs${N} = _mm256_fmsub_ps(vs${N}, valpha, valpha);
79        vp${N} = _mm256_fmadd_ps(vp${N}, vt${N}, vt${N});
80
81      $for N in range(SIMD_TILE):
82        const __m256 ve${N} = _mm256_fmadd_ps(vp${N}, valpha, vs${N});
83        vx${N} = _mm256_mul_ps(vx${N}, vbeta);
84
85      $for N in range(SIMD_TILE):
86        const __m256 vy${N} = _mm256_blendv_ps(vx${N}, ve${N}, vx${N});
87
88      _mm256_storeu_ps(y, vy0);
89      $for N in range(1, SIMD_TILE):
90        _mm256_storeu_ps(y + ${N * 8}, vy${N});
91      y += ${BATCH_TILE};
92    }
93  for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
94    __m256 vx = _mm256_loadu_ps(x);
95    x += 8;
96
97    const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
98
99    __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
100    __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
101    vn = _mm256_sub_ps(vn, vmagic_bias);
102
103    __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
104
105    __m256 vp = _mm256_fmadd_ps(vc6, vt, vc5);
106    vp = _mm256_fmadd_ps(vp, vt, vc4);
107    vp = _mm256_fmadd_ps(vp, vt, vc3);
108    vp = _mm256_fmadd_ps(vp, vt, vc2);
109    vp = _mm256_mul_ps(vp, vt);
110
111    vt = _mm256_mul_ps(vt, vs);
112    vs = _mm256_fmsub_ps(vs, valpha, valpha);
113    vp = _mm256_fmadd_ps(vp, vt, vt);
114    const __m256 ve = _mm256_fmadd_ps(vp, valpha, vs);
115
116    vx = _mm256_mul_ps(vx, vbeta);
117    const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
118
119    _mm256_storeu_ps(y, vy);
120    y += 8;
121  }
122  if XNN_UNLIKELY(n != 0) {
123    assert(n >= 1 * sizeof(float));
124    assert(n <= 7 * sizeof(float));
125    __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
126
127    __m256 vx = _mm256_maskload_ps(x, vmask);
128
129    const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
130
131    __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
132    __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
133    vn = _mm256_sub_ps(vn, vmagic_bias);
134
135    __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
136
137    __m256 vp = _mm256_fmadd_ps(vc6, vt, vc5);
138    vp = _mm256_fmadd_ps(vp, vt, vc4);
139    vp = _mm256_fmadd_ps(vp, vt, vc3);
140    vp = _mm256_fmadd_ps(vp, vt, vc2);
141    vp = _mm256_mul_ps(vp, vt);
142
143    vt = _mm256_mul_ps(vt, vs);
144    vs = _mm256_fmsub_ps(vs, valpha, valpha);
145    vp = _mm256_fmadd_ps(vp, vt, vt);
146    const __m256 ve = _mm256_fmadd_ps(vp, valpha, vs);
147
148    vx = _mm256_mul_ps(vx, vbeta);
149    const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
150
151    // _mm256_maskstore_ps(y, vmask, vf) could be used here, but triggers msan failures (probably an msan bug).
152    __m128 vy_lo = _mm256_castps256_ps128(vy);
153    if (n & (4 * sizeof(float))) {
154      _mm_storeu_ps(y, vy_lo);
155      vy_lo = _mm256_extractf128_ps(vy, 1);
156      y += 4;
157    }
158    if (n & (2 * sizeof(float))) {
159      _mm_storel_pi((__m64*) y, vy_lo);
160      vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
161      y += 2;
162    }
163    if (n & (1 * sizeof(float))) {
164      _mm_store_ss(y, vy_lo);
165    }
166  }
167}
168