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
|
/*========================== begin_copyright_notice ============================
Copyright (C) 2017-2021 Intel Corporation
SPDX-License-Identifier: MIT
============================= end_copyright_notice ===========================*/
#include "../include/BiF_Definitions.cl"
#include "../../Headers/spirv.h"
#include "../IMF/FP32/pow_s_la.cl"
#include "../IMF/FP32/pow_s_la_noLUT.cl"
#include "../IMF/FP32/pow_s_prev.cl"
#if defined(cl_khr_fp64)
#include "../IMF/FP64/pow_d_la.cl"
#include "../IMF/FP64/pow_d_la_noLUT.cl"
#endif // defined(cl_khr_fp64)
INLINE float __attribute__((overloadable)) __spirv_ocl_pow( float x, float y )
{
if(BIF_FLAG_CTRL_GET(FastRelaxedMath) && (!BIF_FLAG_CTRL_GET(APIRS)))
{
// Undefined for x = 0 and y = 0.
// Undefined for x < 0 and noninteger y.
// For x >= 0, or x < 0 and even y, derived implementations implement this as:
// exp2(y * log2(x)).
// For x < 0 and odd y, derived implementations implement this as:
// -exp2(y * log2(fabs(x)).
//
// This expansion is technically undefined when x == 0, since
// log2(x) is undefined, however our native log2 returns -inf
// in this case. Since exp2( y * -inf ) is zero for finite y,
// we'll end up with zero, hence the "correct" results.
float pr = __spirv_ocl_fabs( x );
// TBD: Which is faster?
// Note that USC has a pattern match optimization to turn
// log-mul-exp into pow. Additionally, there are some specific
// LLVM optimizations for pow. So, preferring pow for now.
#if 0
pr = __spirv_ocl_log2( pr );
pr = y * pr;
pr = __spirv_ocl_exp2( pr );
#else
pr = __spirv_ocl_native_powr( pr, y );
#endif
// Check for a positive x by checking the sign bit as an integer,
// not a float. This correctly handles x = -0.0f. Arguably, we
// don't have to do this since -cl-fast-relaxed-math implies
// -cl-no-signed-zeros, but it's easy enough to do.
int iy = (int)y;
float nr = -pr;
nr = ( iy & 1 ) ? nr : pr; // positive result for even y, else negative result
float result = ( as_int(x) >= 0 ) ? pr : nr;// positive result for positive x, else unchanged
return result;
}
else if (BIF_FLAG_CTRL_GET(UseHighAccuracyMath))
{
return __ocl_svml_powf_noLUT(x, y);
}
else
{
bool precisionWA =
( ( as_uint(y) == 0x39D0A3C4 ) | ( as_uint(y) == 0x3ada1700 ) | ( as_uint(y) == 0x3b81ef38 ) | ( as_uint(y) == 0x3b2434e1 ) | ( as_uint(y) == 0x3D7319F7 ) ) &
( ( as_uint(x) >= 0x3f7ff65f ) & ( as_uint(x) < 0x3F800000 ) );
if ( precisionWA )
{
return as_float(0x3F7FFFFF);
}
// Previous version of pow builtin is called here, because new one introduced some critical functional regressions.
// TODO: The target is to call '__ocl_svml_powf' here.
return __ocl_svml_px_powf1(x, y);
}
}
GENERATE_SPIRV_OCL_VECTOR_FUNCTIONS_2ARGS_VV_LOOP( pow, float, float, float, f32, f32 )
#if defined(cl_khr_fp64)
INLINE double __attribute__((overloadable)) __spirv_ocl_pow( double x, double y )
{
double result;
if (BIF_FLAG_CTRL_GET(UseHighAccuracyMath)) {
result = __ocl_svml_pow_noLUT(x, y);
} else {
result = __ocl_svml_pow(x, y);
}
return result;
}
GENERATE_SPIRV_OCL_VECTOR_FUNCTIONS_2ARGS_VV_LOOP( pow, double, double, double, f64, f64 )
#endif // defined(cl_khr_fp64)
#if defined(cl_khr_fp16)
INLINE half __attribute__((overloadable)) __spirv_ocl_pow( half x, half y )
{
return __spirv_ocl_pow((float)x, (float)y );
}
GENERATE_SPIRV_OCL_VECTOR_FUNCTIONS_2ARGS_VV_LOOP( pow, half, half, half, f16, f16 )
#endif // defined(cl_khr_fp16)
#if defined(IGC_SPV_INTEL_bfloat16_arithmetic)
INLINE bfloat __attribute__((overloadable)) __spirv_ocl_pow( bfloat x, bfloat y )
{
return __spirv_ocl_pow((float)x, (float)y);
}
GENERATE_SPIRV_OCL_VECTOR_FUNCTIONS_2ARGS_VV_LOOP( pow, bfloat, bfloat, bfloat, , )
#endif // defined(IGC_SPV_INTEL_bfloat16_arithmetic)
|
