libclc: Rewrite log implementation as gentype inc file (#187537)
Follow the ordinary gentype conventions for the log implementation, instead of using a plain header. This doesn't quite yet enable vectorization, due to how the table is currently indexed. This should make it easier for targets to selectively overload the function for a subset of types.
This commit is contained in:
Matt Arsenault
GitHub
parent
441790b31f
commit
a5de509e4e
@ -6,9 +6,19 @@
|
||||
//
|
||||
//===----------------------------------------------------------------------===//
|
||||
|
||||
#include "clc/clc_convert.h"
|
||||
#include "clc/float/definitions.h"
|
||||
#include "clc/internal/clc.h"
|
||||
#include "clc/math/clc_ep.h"
|
||||
#include "clc/math/clc_fabs.h"
|
||||
#include "clc/math/clc_fma.h"
|
||||
#include "clc/math/clc_frexp.h"
|
||||
#include "clc/math/clc_ldexp.h"
|
||||
#include "clc/math/clc_log2.h"
|
||||
#include "clc/math/clc_mad.h"
|
||||
#include "clc/math/math.h"
|
||||
#include "clc/relational/clc_isinf.h"
|
||||
#include "clc/relational/clc_isnan.h"
|
||||
|
||||
/*
|
||||
*log(x) = log2(x) * (1/log2(e))
|
||||
|
||||
@ -6,21 +6,25 @@
|
||||
//
|
||||
//===----------------------------------------------------------------------===//
|
||||
|
||||
#include "clc/clc_convert.h"
|
||||
#include "clc/float/definitions.h"
|
||||
#include "clc/internal/clc.h"
|
||||
#include "clc/math/clc_ep.h"
|
||||
#include "clc/math/clc_fabs.h"
|
||||
#include "clc/math/clc_fma.h"
|
||||
#include "clc/math/clc_frexp.h"
|
||||
#include "clc/math/clc_ldexp.h"
|
||||
#include "clc/math/clc_mad.h"
|
||||
#include "clc/math/math.h"
|
||||
#include "clc/math/tables.h"
|
||||
|
||||
#ifdef cl_khr_fp64
|
||||
#pragma OPENCL EXTENSION cl_khr_fp64 : enable
|
||||
#endif // cl_khr_fp64
|
||||
|
||||
#ifdef cl_khr_fp16
|
||||
#pragma OPENCL EXTENSION cl_khr_fp16 : enable
|
||||
#endif // cl_khr_fp16
|
||||
#include "clc/relational/clc_isinf.h"
|
||||
#include "clc/relational/clc_isnan.h"
|
||||
|
||||
#define COMPILING_LOG10
|
||||
#include "clc_log_base.h"
|
||||
#define __CLC_BODY "clc_log_base.inc"
|
||||
#include "clc/math/gentype.inc"
|
||||
#undef COMPILING_LOG10
|
||||
|
||||
#define __CLC_FUNCTION __clc_log10
|
||||
#define __CLC_BODY "clc/shared/unary_def_scalarize.inc"
|
||||
#define __CLC_BODY "clc/shared/unary_def_scalarize_loop.inc"
|
||||
#include "clc/math/gentype.inc"
|
||||
|
||||
@ -6,21 +6,25 @@
|
||||
//
|
||||
//===----------------------------------------------------------------------===//
|
||||
|
||||
#include "clc/clc_convert.h"
|
||||
#include "clc/float/definitions.h"
|
||||
#include "clc/internal/clc.h"
|
||||
#include "clc/math/clc_ep.h"
|
||||
#include "clc/math/clc_fabs.h"
|
||||
#include "clc/math/clc_fma.h"
|
||||
#include "clc/math/clc_frexp.h"
|
||||
#include "clc/math/clc_ldexp.h"
|
||||
#include "clc/math/clc_mad.h"
|
||||
#include "clc/math/math.h"
|
||||
#include "clc/math/tables.h"
|
||||
|
||||
#ifdef cl_khr_fp64
|
||||
#pragma OPENCL EXTENSION cl_khr_fp64 : enable
|
||||
#endif // cl_khr_fp64
|
||||
|
||||
#ifdef cl_khr_fp16
|
||||
#pragma OPENCL EXTENSION cl_khr_fp16 : enable
|
||||
#endif // cl_khr_fp16
|
||||
#include "clc/relational/clc_isinf.h"
|
||||
#include "clc/relational/clc_isnan.h"
|
||||
|
||||
#define COMPILING_LOG2
|
||||
#include "clc_log_base.h"
|
||||
#define __CLC_BODY "clc_log_base.inc"
|
||||
#include "clc/math/gentype.inc"
|
||||
#undef COMPILING_LOG2
|
||||
|
||||
#define __CLC_FUNCTION __clc_log2
|
||||
#define __CLC_BODY "clc/shared/unary_def_scalarize.inc"
|
||||
#define __CLC_BODY "clc/shared/unary_def_scalarize_loop.inc"
|
||||
#include "clc/math/gentype.inc"
|
||||
|
||||
@ -6,17 +6,6 @@
|
||||
//
|
||||
//===----------------------------------------------------------------------===//
|
||||
|
||||
#include "clc/float/definitions.h"
|
||||
#include "clc/math/clc_ep.h"
|
||||
#include "clc/math/clc_fabs.h"
|
||||
#include "clc/math/clc_fma.h"
|
||||
#include "clc/math/clc_frexp.h"
|
||||
#include "clc/math/clc_ldexp.h"
|
||||
#include "clc/math/clc_mad.h"
|
||||
#include "clc/math/math.h"
|
||||
#include "clc/relational/clc_isinf.h"
|
||||
#include "clc/relational/clc_isnan.h"
|
||||
|
||||
/*
|
||||
Algorithm:
|
||||
|
||||
@ -72,54 +61,57 @@
|
||||
|
||||
*/
|
||||
|
||||
_CLC_OVERLOAD _CLC_DEF float
|
||||
#ifdef __CLC_SCALAR
|
||||
|
||||
#if __CLC_FPSIZE == 32
|
||||
_CLC_OVERLOAD _CLC_DEF __CLC_FLOATN
|
||||
#if defined(COMPILING_LOG2)
|
||||
__clc_log2(float x)
|
||||
__clc_log2(__CLC_FLOATN x)
|
||||
#elif defined(COMPILING_LOG10)
|
||||
__clc_log10(float x)
|
||||
__clc_log10(__CLC_FLOATN x)
|
||||
#else
|
||||
__clc_log(float x)
|
||||
__clc_log(__CLC_FLOATN x)
|
||||
#endif
|
||||
{
|
||||
|
||||
#if defined(COMPILING_LOG2)
|
||||
const float LOG2E = 0x1.715476p+0f; // 1.4426950408889634
|
||||
const float LOG2E_HEAD = 0x1.700000p+0f; // 1.4375
|
||||
const float LOG2E_TAIL = 0x1.547652p-8f; // 0.00519504072
|
||||
const __CLC_FLOATN LOG2E = 0x1.715476p+0f; // 1.4426950408889634
|
||||
const __CLC_FLOATN LOG2E_HEAD = 0x1.700000p+0f; // 1.4375
|
||||
const __CLC_FLOATN LOG2E_TAIL = 0x1.547652p-8f; // 0.00519504072
|
||||
#elif defined(COMPILING_LOG10)
|
||||
const float LOG10E = 0x1.bcb7b2p-2f; // 0.43429448190325182
|
||||
const float LOG10E_HEAD = 0x1.bc0000p-2f; // 0.43359375
|
||||
const float LOG10E_TAIL = 0x1.6f62a4p-11f; // 0.0007007319
|
||||
const float LOG10_2_HEAD = 0x1.340000p-2f; // 0.30078125
|
||||
const float LOG10_2_TAIL = 0x1.04d426p-12f; // 0.000248745637
|
||||
const __CLC_FLOATN LOG10E = 0x1.bcb7b2p-2f; // 0.43429448190325182
|
||||
const __CLC_FLOATN LOG10E_HEAD = 0x1.bc0000p-2f; // 0.43359375
|
||||
const __CLC_FLOATN LOG10E_TAIL = 0x1.6f62a4p-11f; // 0.0007007319
|
||||
const __CLC_FLOATN LOG10_2_HEAD = 0x1.340000p-2f; // 0.30078125
|
||||
const __CLC_FLOATN LOG10_2_TAIL = 0x1.04d426p-12f; // 0.000248745637
|
||||
#else
|
||||
const float LOG2_HEAD = 0x1.62e000p-1f; // 0.693115234
|
||||
const float LOG2_TAIL = 0x1.0bfbe8p-15f; // 0.0000319461833
|
||||
const __CLC_FLOATN LOG2_HEAD = 0x1.62e000p-1f; // 0.693115234
|
||||
const __CLC_FLOATN LOG2_TAIL = 0x1.0bfbe8p-15f; // 0.0000319461833
|
||||
#endif
|
||||
|
||||
uint xi = __clc_as_uint(x);
|
||||
uint ax = xi & EXSIGNBIT_SP32;
|
||||
|
||||
// Calculations for |x-1| < 2^-4
|
||||
float r = x - 1.0f;
|
||||
__CLC_FLOATN r = x - 1.0f;
|
||||
int near1 = __clc_fabs(r) < 0x1.0p-4f;
|
||||
float u2 = MATH_DIVIDE(r, 2.0f + r);
|
||||
float corr = u2 * r;
|
||||
float u = u2 + u2;
|
||||
float v = u * u;
|
||||
float znear1, z1, z2;
|
||||
__CLC_FLOATN u2 = MATH_DIVIDE(r, 2.0f + r);
|
||||
__CLC_FLOATN corr = u2 * r;
|
||||
__CLC_FLOATN u = u2 + u2;
|
||||
__CLC_FLOATN v = u * u;
|
||||
__CLC_FLOATN znear1, z1, z2;
|
||||
|
||||
// 2/(5 * 2^5), 2/(3 * 2^3)
|
||||
z2 = __clc_mad(u, __clc_mad(v, 0x1.99999ap-7f, 0x1.555556p-4f) * v, -corr);
|
||||
|
||||
#if defined(COMPILING_LOG2)
|
||||
z1 = __clc_as_float(__clc_as_int(r) & 0xffff0000);
|
||||
z1 = __CLC_AS_FLOATN(__CLC_AS_INTN(r) & 0xffff0000);
|
||||
z2 = z2 + (r - z1);
|
||||
znear1 = __clc_mad(
|
||||
z1, LOG2E_HEAD,
|
||||
__clc_mad(z2, LOG2E_HEAD, __clc_mad(z1, LOG2E_TAIL, z2 * LOG2E_TAIL)));
|
||||
#elif defined(COMPILING_LOG10)
|
||||
z1 = __clc_as_float(__clc_as_int(r) & 0xffff0000);
|
||||
z1 = __CLC_AS_FLOATN(__clc_as_int(r) & 0xffff0000);
|
||||
z2 = z2 + (r - z1);
|
||||
znear1 = __clc_mad(
|
||||
z1, LOG10E_HEAD,
|
||||
@ -132,24 +124,24 @@ __clc_log(float x)
|
||||
int m = (int)(xi >> EXPSHIFTBITS_SP32) - EXPBIAS_SP32;
|
||||
|
||||
// Normalize subnormal
|
||||
uint xis = __clc_as_uint(__clc_as_float(xi | 0x3f800000) - 1.0f);
|
||||
uint xis = __CLC_AS_UINTN(__CLC_AS_FLOATN(xi | 0x3f800000) - 1.0f);
|
||||
int ms = (int)(xis >> EXPSHIFTBITS_SP32) - 253;
|
||||
int c = m == -127;
|
||||
m = c ? ms : m;
|
||||
uint xin = c ? xis : xi;
|
||||
|
||||
float mf = (float)m;
|
||||
__CLC_FLOATN mf = (__CLC_FLOATN)m;
|
||||
uint indx = (xin & 0x007f0000) + ((xin & 0x00008000) << 1);
|
||||
|
||||
// F - Y
|
||||
float f = __clc_as_float(0x3f000000 | indx) -
|
||||
__clc_as_float(0x3f000000 | (xin & MANTBITS_SP32));
|
||||
__CLC_FLOATN f = __CLC_AS_FLOATN(0x3f000000 | indx) -
|
||||
__CLC_AS_FLOATN(0x3f000000 | (xin & MANTBITS_SP32));
|
||||
|
||||
indx = indx >> 16;
|
||||
r = f * __CLC_USE_TABLE(log_inv_tbl, indx);
|
||||
|
||||
// 1/3, 1/2
|
||||
float poly = __clc_mad(__clc_mad(r, 0x1.555556p-2f, 0.5f), r * r, r);
|
||||
__CLC_FLOATN poly = __clc_mad(__clc_mad(r, 0x1.555556p-2f, 0.5f), r * r, r);
|
||||
|
||||
#if defined(COMPILING_LOG2)
|
||||
float2 tv = __CLC_USE_TABLE(log2_tbl, indx);
|
||||
@ -165,63 +157,63 @@ __clc_log(float x)
|
||||
z2 = __clc_mad(mf, LOG2_TAIL, -poly) + tv.s1;
|
||||
#endif
|
||||
|
||||
float z = z1 + z2;
|
||||
__CLC_FLOATN z = z1 + z2;
|
||||
z = near1 ? znear1 : z;
|
||||
|
||||
// Corner cases
|
||||
z = ax >= PINFBITPATT_SP32 ? x : z;
|
||||
z = xi != ax ? __clc_as_float(QNANBITPATT_SP32) : z;
|
||||
z = ax == 0 ? __clc_as_float(NINFBITPATT_SP32) : z;
|
||||
z = xi != ax ? __CLC_AS_FLOATN(QNANBITPATT_SP32) : z;
|
||||
z = ax == 0 ? __CLC_AS_FLOATN(NINFBITPATT_SP32) : z;
|
||||
|
||||
return z;
|
||||
}
|
||||
|
||||
#ifdef cl_khr_fp64
|
||||
#elif __CLC_FPSIZE == 64
|
||||
|
||||
_CLC_OVERLOAD _CLC_DEF double
|
||||
_CLC_OVERLOAD _CLC_DEF __CLC_DOUBLEN
|
||||
#if defined(COMPILING_LOG2)
|
||||
__clc_log2(double a)
|
||||
__clc_log2(__CLC_DOUBLEN a)
|
||||
#elif defined(COMPILING_LOG10)
|
||||
__clc_log10(double a)
|
||||
__clc_log10(__CLC_DOUBLEN a)
|
||||
#else
|
||||
__clc_log(double a)
|
||||
__clc_log(__CLC_DOUBLEN a)
|
||||
#endif
|
||||
{
|
||||
int a_exp;
|
||||
double m = __clc_frexp(a, &a_exp);
|
||||
int b = m < (2.0 / 3.0);
|
||||
__CLC_INTN a_exp;
|
||||
__CLC_DOUBLEN m = __clc_frexp(a, &a_exp);
|
||||
__CLC_INTN b = m < (2.0 / 3.0);
|
||||
m = __clc_ldexp(m, b);
|
||||
int e = a_exp - b;
|
||||
__CLC_INTN e = a_exp - b;
|
||||
|
||||
__clc_ep_pair_double x = __clc_ep_div(m - 1.0, __clc_ep_fast_add(1.0, m));
|
||||
double s = x.hi * x.hi;
|
||||
double p = __clc_mad(s, __clc_mad(s, __clc_mad(s,
|
||||
__CLC_EP_PAIR x = __clc_ep_div(m - 1.0, __clc_ep_fast_add(1.0, m));
|
||||
__CLC_DOUBLEN s = x.hi * x.hi;
|
||||
__CLC_DOUBLEN p = __clc_mad(s, __clc_mad(s, __clc_mad(s,
|
||||
__clc_mad(s, __clc_mad(s, __clc_mad(s, 0x1.3ab76bf559e2bp-3, 0x1.385386b47b09ap-3),
|
||||
0x1.7474dd7f4df2ep-3), 0x1.c71c016291751p-3),
|
||||
0x1.249249b27acf1p-2), 0x1.99999998ef7b6p-2), 0x1.5555555555780p-1);
|
||||
__clc_ep_pair_double r =
|
||||
__CLC_EP_PAIR r =
|
||||
__clc_ep_fast_add(__clc_ep_ldexp(x, 1), s * x.hi * p);
|
||||
|
||||
#if defined COMPILING_LOG2
|
||||
r = __clc_ep_add(
|
||||
(double)e,
|
||||
__CLC_CONVERT_DOUBLEN(e),
|
||||
__clc_ep_mul(
|
||||
__clc_ep_make_pair(0x1.71547652b82fep+0, 0x1.777d0ffda0d24p-56), r));
|
||||
#elif defined COMPILING_LOG10
|
||||
r = __clc_ep_add(
|
||||
__clc_ep_mul(
|
||||
__clc_ep_make_pair(0x1.34413509f79ffp-2, -0x1.9dc1da994fd21p-59),
|
||||
(double)e),
|
||||
__CLC_CONVERT_DOUBLEN(e)),
|
||||
__clc_ep_mul(
|
||||
__clc_ep_make_pair(0x1.bcb7b1526e50ep-2, 0x1.95355baaafad3p-57), r));
|
||||
#else
|
||||
r = __clc_ep_add(__clc_ep_mul(__clc_ep_make_pair(0x1.62e42fefa39efp-1,
|
||||
0x1.abc9e3b39803fp-56),
|
||||
(double)e),
|
||||
__CLC_CONVERT_DOUBLEN(e)),
|
||||
r);
|
||||
#endif
|
||||
|
||||
double ret = r.hi;
|
||||
__CLC_DOUBLEN ret = r.hi;
|
||||
|
||||
ret = __clc_isinf(a) ? a : ret;
|
||||
ret = a < 0.0 ? DBL_NAN : ret;
|
||||
@ -230,23 +222,22 @@ __clc_log(double a)
|
||||
return ret;
|
||||
}
|
||||
|
||||
#endif // cl_khr_fp64
|
||||
#elif __CLC_FPSIZE == 16
|
||||
|
||||
#ifdef cl_khr_fp16
|
||||
|
||||
_CLC_OVERLOAD _CLC_DEF half
|
||||
_CLC_OVERLOAD _CLC_DEF __CLC_HALFN
|
||||
#if defined(COMPILING_LOG2)
|
||||
__clc_log2(half x) {
|
||||
return (half)__clc_log2((float)x);
|
||||
__clc_log2(__CLC_HALFN x) {
|
||||
return __CLC_CONVERT_HALFN(__clc_log2(__CLC_CONVERT_FLOATN(x)));
|
||||
}
|
||||
#elif defined(COMPILING_LOG10)
|
||||
__clc_log10(half x) {
|
||||
return (half)__clc_log10((float)x);
|
||||
__clc_log10(__CLC_HALFN x) {
|
||||
return __CLC_CONVERT_HALFN(__clc_log10(__CLC_CONVERT_FLOATN(x)));
|
||||
}
|
||||
#else
|
||||
__clc_log(half x) {
|
||||
return (half)__clc_log((float)x);
|
||||
__clc_log(__CLC_HALFN x) {
|
||||
return __CLC_CONVERT_HALFN(__clc_log(__CLC_CONVERT_FLOATN(x));
|
||||
}
|
||||
#endif
|
||||
|
||||
#endif // cl_khr_fp16
|
||||
#endif // __CLC_FPSIZE
|
||||
#endif // __CLC_SCALAR
|
||||
Loading…
x
Reference in New Issue
Block a user