From 2ca909d824a1a86437d081552e4cbe28ba79ff37 Mon Sep 17 00:00:00 2001 From: Tom Stellard Date: Wed, 6 May 2015 20:53:32 +0000 Subject: [PATCH] math: Add ldexp implementation Signed-off-by: Aaron Watry Tom Stellard: - Add denormal handling. - Share vectorization code with r600 implementation. Patch By: Aaron Watry llvm-svn: 236639 --- libclc/generic/lib/SOURCES | 1 + libclc/generic/lib/clcmacro.h | 4 + libclc/generic/lib/math/ldexp.cl | 138 ++++++++++++++++++++++++++++ libclc/{r600 => generic}/lib/math/ldexp.inc | 0 libclc/r600/lib/math/ldexp.cl | 2 +- 5 files changed, 144 insertions(+), 1 deletion(-) create mode 100644 libclc/generic/lib/math/ldexp.cl rename libclc/{r600 => generic}/lib/math/ldexp.inc (100%) diff --git a/libclc/generic/lib/SOURCES b/libclc/generic/lib/SOURCES index d21d6aa..8eca762 100644 --- a/libclc/generic/lib/SOURCES +++ b/libclc/generic/lib/SOURCES @@ -84,6 +84,7 @@ math/fmod.cl math/fract.cl math/half_sqrt.cl math/hypot.cl +math/ldexp.cl math/log10.cl math/log1p.cl math/mad.cl diff --git a/libclc/generic/lib/clcmacro.h b/libclc/generic/lib/clcmacro.h index 346adf2..9ef337b 100644 --- a/libclc/generic/lib/clcmacro.h +++ b/libclc/generic/lib/clcmacro.h @@ -115,6 +115,10 @@ _CLC_DEF _CLC_OVERLOAD RET_TYPE FUNCTION(ARG1_TYPE x, ARG2_TYPE y) { \ } \ _CLC_BINARY_VECTORIZE(_CLC_OVERLOAD _CLC_DEF, RET_TYPE, FUNCTION, ARG1_TYPE, ARG2_TYPE) +#define _CLC_DEFINE_BINARY_BUILTIN_WITH_SCALAR_SECOND_ARG(RET_TYPE, FUNCTION, BUILTIN, ARG1_TYPE, ARG2_TYPE) \ +_CLC_DEFINE_BINARY_BUILTIN(RET_TYPE, FUNCTION, BUILTIN, ARG1_TYPE, ARG2_TYPE) \ +_CLC_BINARY_VECTORIZE_SCALAR_SECOND_ARG(_CLC_OVERLOAD _CLC_DEF, RET_TYPE, FUNCTION, ARG1_TYPE, ARG2_TYPE) + #define _CLC_DEFINE_UNARY_BUILTIN(RET_TYPE, FUNCTION, BUILTIN, ARG1_TYPE) \ _CLC_DEF _CLC_OVERLOAD RET_TYPE FUNCTION(ARG1_TYPE x) { \ return BUILTIN(x); \ diff --git a/libclc/generic/lib/math/ldexp.cl b/libclc/generic/lib/math/ldexp.cl new file mode 100644 index 0000000..1802bf3 --- /dev/null +++ b/libclc/generic/lib/math/ldexp.cl @@ -0,0 +1,138 @@ +/* + * Copyright (c) 2014 Advanced Micro Devices, Inc. + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to deal + * in the Software without restriction, including without limitation the rights + * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell + * copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in + * all copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN + * THE SOFTWARE. + */ + +#include +#include "config.h" +#include "../clcmacro.h" +#include "math.h" + +_CLC_DEF _CLC_OVERLOAD float ldexp(float x, int n) { + + if (!__clc_fp32_subnormals_supported()) { + + // This treats subnormals as zeros + int i = as_int(x); + int e = (i >> 23) & 0xff; + int m = i & 0x007fffff; + int s = i & 0x80000000; + int v = add_sat(e, n); + v = clamp(v, 0, 0xff); + int mr = e == 0 | v == 0 | v == 0xff ? 0 : m; + int c = e == 0xff; + mr = c ? m : mr; + int er = c ? e : v; + er = e ? er : e; + return as_float( s | (er << 23) | mr ); + } + + /* supports denormal values */ + const int multiplier = 24; + float val_f; + uint val_ui; + uint sign; + int exponent; + val_ui = as_uint(x); + sign = val_ui & 0x80000000; + val_ui = val_ui & 0x7fffffff;/* remove the sign bit */ + int val_x = val_ui; + + exponent = val_ui >> 23; /* get the exponent */ + int dexp = exponent; + + /* denormal support */ + int fbh = 127 - (as_uint((float)(as_float(val_ui | 0x3f800000) - 1.0f)) >> 23); + int dexponent = 25 - fbh; + uint dval_ui = (( (val_ui << fbh) & 0x007fffff) | (dexponent << 23)); + int ex = dexponent + n - multiplier; + dexponent = ex; + uint val = sign | (ex << 23) | (dval_ui & 0x007fffff); + int ex1 = dexponent + multiplier; + ex1 = -ex1 +25; + dval_ui = (((dval_ui & 0x007fffff )| 0x800000) >> ex1); + dval_ui = dexponent > 0 ? val :dval_ui; + dval_ui = dexponent > 254 ? 0x7f800000 :dval_ui; /*overflow*/ + dval_ui = dexponent < -multiplier ? 0 : dval_ui; /*underflow*/ + dval_ui = dval_ui | sign; + val_f = as_float(dval_ui); + + exponent += n; + + val = sign | (exponent << 23) | (val_ui & 0x007fffff); + ex1 = exponent + multiplier; + ex1 = -ex1 +25; + val_ui = (((val_ui & 0x007fffff )| 0x800000) >> ex1); + val_ui = exponent > 0 ? val :val_ui; + val_ui = exponent > 254 ? 0x7f800000 :val_ui; /*overflow*/ + val_ui = exponent < -multiplier ? 0 : val_ui; /*underflow*/ + val_ui = val_ui | sign; + + val_ui = dexp == 0? dval_ui : val_ui; + val_f = as_float(val_ui); + + val_f = isnan(x) | isinf(x) | val_x == 0 ? x : val_f; + return val_f; +} + +// This defines all the ldexp(floatN, intN) variants. +_CLC_BINARY_VECTORIZE(_CLC_OVERLOAD _CLC_DEF, float, ldexp, float, int) + +#ifdef cl_khr_fp64 + +#pragma OPENCL EXTENSION cl_khr_fp64 : enable + +_CLC_DEF _CLC_OVERLOAD double ldexp(double x, int n) { + long l = as_ulong(x); + int e = (l >> 52) & 0x7ff; + long s = l & 0x8000000000000000; + + ulong ux = as_ulong(x * 0x1.0p+53); + int de = ((int)(ux >> 52) & 0x7ff) - 53; + int c = e == 0; + e = c ? de: e; + + ux = c ? ux : l; + + int v = e + n; + v = clamp(v, -0x7ff, 0x7ff); + + ux &= ~EXPBITS_DP64; + + double mr = as_double(ux | ((ulong)(v+53) << 52)); + mr = mr * 0x1.0p-53; + + mr = v > 0 ? as_double(ux | ((ulong)v << 52)) : mr; + + mr = v == 0x7ff ? as_double(s | PINFBITPATT_DP64) : mr; + mr = v < -53 ? as_double(s) : mr; + + mr = ((n == 0) | isinf(x) | (x == 0) ) ? x : mr; + return mr; +} + +// This defines all the ldexp(doubleN, intN) variants. +_CLC_BINARY_VECTORIZE(_CLC_OVERLOAD _CLC_DEF, double, ldexp, double, int) + +#endif + +// This defines all the ldexp(GENTYPE, int) variants +#define __CLC_BODY +#include diff --git a/libclc/r600/lib/math/ldexp.inc b/libclc/generic/lib/math/ldexp.inc similarity index 100% rename from libclc/r600/lib/math/ldexp.inc rename to libclc/generic/lib/math/ldexp.inc diff --git a/libclc/r600/lib/math/ldexp.cl b/libclc/r600/lib/math/ldexp.cl index 6742cb9..0461a53 100644 --- a/libclc/r600/lib/math/ldexp.cl +++ b/libclc/r600/lib/math/ldexp.cl @@ -34,5 +34,5 @@ _CLC_DEFINE_BINARY_BUILTIN(float, ldexp, __builtin_amdgpu_ldexpf, float, int); #endif // This defines all the ldexp(GENTYPE, int); -#define __CLC_BODY +#define __CLC_BODY <../../../generic/lib/math/ldexp.inc> #include -- 2.7.4