1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
|
/*========================== begin_copyright_notice ============================
Copyright (C) 2024 Intel Corporation
SPDX-License-Identifier: MIT
============================= end_copyright_notice ===========================*/
#include "../imf.h"
#pragma OPENCL FP_CONTRACT OFF
static __constant union {
unsigned int w;
float f;
int i;
} __sexpm1_ep_c4 = {0x3c0b4a77u};
static __constant union {
unsigned int w;
float f;
int i;
} __sexpm1_ep_c3 = {0x3d2e1d2cu};
static __constant union {
unsigned int w;
float f;
int i;
} __sexpm1_ep_c2 = {0x3e2aa0ccu};
static __constant union {
unsigned int w;
float f;
int i;
} __sexpm1_ep_c1 = {0x3efff2b6u};
static __constant union {
unsigned int w;
float f;
int i;
} __sexpm1_ep_c0 = {0x35bb1a40u};
static __constant union {
unsigned int w;
float f;
int i;
} __sexpm1_ep_fL2E = {0x3FB8AA3Bu};
static __constant union {
unsigned int w;
float f;
int i;
} __sexpm1_ep_fShifter = {0x4b40007fu};
// log(2) high, low
static __constant union {
unsigned int w;
float f;
int i;
} __sexpm1_ep_NL2H = {0xbf317218u};
static __constant union {
unsigned int w;
float f;
int i;
} __sexpm1_ep_NL2L = {0x3102E308u};
__attribute__((always_inline)) inline int
__ocl_svml_internal_sexpm1_ep(float *a, float *pres) {
int nRet = 0;
float xin = *a;
// float expm1f(float xin)
union {
unsigned int w;
float f;
int i;
} xf, fN, xL2E, fS, xa;
union {
unsigned int w;
float f;
int i;
} T, sc, res;
float R, poly, Th;
xf.f = xin;
xL2E.f =
SPIRV_OCL_BUILTIN(fma, _f32_f32_f32, )(xf.f, __sexpm1_ep_fL2E.f, 0.0f);
fN.f = SPIRV_OCL_BUILTIN(trunc, _f32, )(xL2E.f);
fS.f = __sexpm1_ep_fShifter.f + fN.f;
// reduced argument
R = SPIRV_OCL_BUILTIN(fma, _f32_f32_f32, )(fN.f, __sexpm1_ep_NL2H.f, xf.f);
// R = SP_FMA(fN.f, _VSTATIC(NL2L).f, R);
// 2^N
T.w = fS.w << 23;
// e^R - 1
poly = SPIRV_OCL_BUILTIN(fma, _f32_f32_f32, )(__sexpm1_ep_c4.f, R,
__sexpm1_ep_c3.f);
poly = SPIRV_OCL_BUILTIN(fma, _f32_f32_f32, )(poly, R, __sexpm1_ep_c2.f);
poly = SPIRV_OCL_BUILTIN(fma, _f32_f32_f32, )(poly, R, __sexpm1_ep_c1.f);
poly = SPIRV_OCL_BUILTIN(fma, _f32_f32_f32, )(poly, R, __sexpm1_ep_c0.f);
poly = SPIRV_OCL_BUILTIN(fma, _f32_f32_f32, )(poly, R, R);
// maxabs(T,-1), minabs(T,-1)
Th = T.f - 1.0f;
// 2^N*poly + 2^N - 1
res.f = SPIRV_OCL_BUILTIN(fma, _f32_f32_f32, )(T.f, poly, Th);
// ensure expm1(-0)=-0
xa.f = SPIRV_OCL_BUILTIN(fabs, _f32, )(xf.f);
res.w |= (xf.w ^ xa.w);
if (SPIRV_OCL_BUILTIN(fabs, _f32, )(xf.f) <= 87.0f) {
*pres = res.f;
return nRet;
}
// special case and overflow path
if (xf.f < 0.0f) {
*pres = -1.0f;
return nRet;
}
if (!(xf.f < 128.0f)) {
// +Inf or NaN?
xa.w = xf.w & 0x7fffffff;
if (xa.w > 0x7f800000) {
*pres = xf.f + res.f;
return nRet;
}
// overflow
res.w = 0x7f800000 - 1;
res.f = res.f * res.f; // to set OF flag
//
nRet = 3;
{
*pres = res.f;
return nRet;
}
}
// at or near overflow
// 2^(N-64), N=(int)dN
T.w = (fS.w - 64) << 23;
res.f = SPIRV_OCL_BUILTIN(fma, _f32_f32_f32, )(T.f, poly, T.f);
// final scaling
sc.w = 0x5f800000u;
res.f *= sc.f;
// determine if overflow
if (res.w == 0x7f800000)
nRet = 3;
*pres = res.f;
return nRet;
}
float __ocl_svml_expm1f_ep(float x) {
float r;
__ocl_svml_internal_sexpm1_ep(&x, &r);
return r;
}
|
