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$assert BATCH_TILE % 16 == 0
7$assert BATCH_TILE >= 16
8$SIMD_TILE = BATCH_TILE // 16
9#include <assert.h>
10
11#include <immintrin.h>
12
13#include <xnnpack/common.h>
14#include <xnnpack/intrinsics-polyfill.h>
15#include <xnnpack/vunary.h>
16
17
18void xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x${BATCH_TILE}(
19    size_t n,
20    const float* x,
21    float* y,
22    const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS(1)])
23{
24  assert(n != 0);
25  assert(n % sizeof(float) == 0);
26
27  const __m512 vprescale = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.prescale));
28  const __m512 valpha = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.alpha));
29  const __m512 vbeta = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.beta));
30
31  const __m512 vsat_cutoff = _mm512_set1_ps(-0x1.154246p+4f);
32  const __m512 vmagic_bias = _mm512_set1_ps(0x1.800000p19f);
33  const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f);
34  const __m512i vtable = _mm512_set_epi32(
35    0x3F7D257D, 0x3F7AC0C7, 0x3F78CCDF, 0x3F7744FD, 0x3F76248C, 0x3F75672A, 0x3F7508A4, 0x3F7504F3,
36    0x3F75583F, 0x3F75FED7, 0x3F76F532, 0x3F7837F0, 0x3F79C3D3, 0x3F7B95C2, 0x3F7DAAC3, 0x3F800000);
37  const __m512 vminus_ln2 = _mm512_set1_ps(-0x1.62E43p-1f);
38  const __m512 vc3 = _mm512_set1_ps(0x1.55561Cp-3f);
39  const __m512 vc2 = _mm512_set1_ps(0x1.0001ECp-1f);
40
41  $if BATCH_TILE > 16:
42    for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) {
43      __m512 vx0 = _mm512_loadu_ps(x);
44      $for N in range(1, SIMD_TILE):
45        __m512 vx${N} = _mm512_loadu_ps(x + ${N * 16});
46      x += ${BATCH_TILE};
47
48      $for N in range(SIMD_TILE):
49        const __m512 vz${N} = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx${N}, vprescale));
50
51      $for N in range(SIMD_TILE):
52        __m512 vn${N} = _mm512_fmadd_ps(vz${N}, vlog2e, vmagic_bias);
53
54      $for N in range(SIMD_TILE):
55        const __m512i ven${N} = _mm512_slli_epi32(_mm512_castps_si512(vn${N}), 19);
56        const __m512i vl${N} = _mm512_permutexvar_epi32(_mm512_castps_si512(vn${N}), vtable);
57
58      $for N in range(SIMD_TILE):
59        __m512 vs${N} = _mm512_castsi512_ps(_mm512_add_epi32(vl${N}, ven${N}));
60        vn${N} = _mm512_sub_ps(vn${N}, vmagic_bias);
61
62      $for N in range(SIMD_TILE):
63        __m512 vt${N} = _mm512_fmadd_ps(vn${N}, vminus_ln2, vz${N});
64
65      $for N in range(SIMD_TILE):
66        __m512 vp${N} = _mm512_fmadd_ps(vc3, vt${N}, vc2);
67
68      $for N in range(SIMD_TILE):
69        vp${N} = _mm512_mul_ps(vp${N}, vt${N});
70        vt${N} = _mm512_mul_ps(vt${N}, vs${N});
71
72      $for N in range(SIMD_TILE):
73        vs${N} = _mm512_fmsub_ps(vs${N}, valpha, valpha);
74
75      $for N in range(SIMD_TILE):
76        vp${N} = _mm512_fmadd_ps(vp${N}, vt${N}, vt${N});
77
78      const __m512 vzero = _mm512_setzero_ps();
79      $for N in range(SIMD_TILE):
80        __m512 vy${N} = _mm512_fmadd_ps(vp${N}, valpha, vs${N});
81        const __mmask16 vsign${N} = _mm512_cmp_ps_mask(vx${N}, vzero, _CMP_NLT_US);
82
83      $for N in range(SIMD_TILE):
84        vy${N} = _mm512_mask_mul_ps(vy${N}, vsign${N}, vx${N}, vbeta);
85
86      _mm512_storeu_ps(y, vy0);
87      $for N in range(1, SIMD_TILE):
88        _mm512_storeu_ps(y + ${N * 16}, vy${N});
89      y += ${BATCH_TILE};
90    }
91  for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
92    __m512 vx = _mm512_loadu_ps(x);
93    x += 16;
94
95    const __m512 vz = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx, vprescale));
96    const __mmask16 vsign = _mm512_cmp_ps_mask(vx, _mm512_setzero_ps(), _CMP_NLT_US);
97
98    __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
99    const __m512i ven = _mm512_slli_epi32(_mm512_castps_si512(vn), 19);
100    const __m512i vl = _mm512_permutexvar_epi32(_mm512_castps_si512(vn), vtable);
101    __m512 vs = _mm512_castsi512_ps(_mm512_add_epi32(vl, ven));
102    vn = _mm512_sub_ps(vn, vmagic_bias);
103
104    __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
105
106    __m512 vp = _mm512_fmadd_ps(vc3, vt, vc2);
107    vp = _mm512_mul_ps(vp, vt);
108
109    vt = _mm512_mul_ps(vt, vs);
110    vs = _mm512_fmsub_ps(vs, valpha, valpha);
111    vp = _mm512_fmadd_ps(vp, vt, vt);
112    __m512 vy = _mm512_fmadd_ps(vp, valpha, vs);
113
114    vy = _mm512_mask_mul_ps(vy, vsign, vx, vbeta);
115
116    _mm512_storeu_ps(y, vy);
117    y += 16;
118  }
119  if XNN_UNLIKELY(n != 0) {
120    assert(n >= 1 * sizeof(float));
121    assert(n <= 15 * sizeof(float));
122    // Prepare mask for valid 32-bit elements (depends on n).
123    n >>= 2 /* log2(sizeof(float)) */;
124    const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
125
126    __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
127
128    const __m512 vz = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx, vprescale));
129    const __mmask16 vsign = _mm512_cmp_ps_mask(vx, _mm512_setzero_ps(), _CMP_NLT_US);
130
131    __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
132    const __m512i ven = _mm512_slli_epi32(_mm512_castps_si512(vn), 19);
133    const __m512i vl = _mm512_permutexvar_epi32(_mm512_castps_si512(vn), vtable);
134    __m512 vs = _mm512_castsi512_ps(_mm512_add_epi32(vl, ven));
135    vn = _mm512_sub_ps(vn, vmagic_bias);
136
137    __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
138
139    __m512 vp = _mm512_fmadd_ps(vc3, vt, vc2);
140    vp = _mm512_mul_ps(vp, vt);
141
142    vt = _mm512_mul_ps(vt, vs);
143    vs = _mm512_fmsub_ps(vs, valpha, valpha);
144    vp = _mm512_fmadd_ps(vp, vt, vt);
145    __m512 vy = _mm512_fmadd_ps(vp, valpha, vs);
146
147    vy = _mm512_mask_mul_ps(vy, vsign, vx, vbeta);
148
149    _mm512_mask_storeu_ps(y, vmask, vy);
150  }
151}
152