Author: Fraser Cormack Date: 2025-04-29T16:41:07+01:00 New Revision: ea688c031e99c48a35e0c764d0224c7eed559b12
URL: https://github.com/llvm/llvm-project/commit/ea688c031e99c48a35e0c764d0224c7eed559b12 DIFF: https://github.com/llvm/llvm-project/commit/ea688c031e99c48a35e0c764d0224c7eed559b12.diff LOG: [libclc] Move fdim to CLC library; simplify (#137811) This commit moves the fdim builtin to the CLC library. It simultaneously simplifies the codegen, unifying it between scalar and vector and avoiding bithacking for vector types. Added: libclc/clc/include/clc/math/clc_fdim.h libclc/clc/lib/generic/math/clc_fdim.cl libclc/clc/lib/generic/math/clc_fdim.inc Modified: libclc/clc/include/clc/math/gentype.inc libclc/clc/lib/generic/SOURCES libclc/generic/lib/math/fdim.cl Removed: libclc/generic/lib/math/fdim.inc ################################################################################ diff --git a/libclc/clc/include/clc/math/clc_fdim.h b/libclc/clc/include/clc/math/clc_fdim.h new file mode 100644 index 0000000000000..4477d1ecb7d31 --- /dev/null +++ b/libclc/clc/include/clc/math/clc_fdim.h @@ -0,0 +1,20 @@ +//===----------------------------------------------------------------------===// +// +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. +// See https://llvm.org/LICENSE.txt for license information. +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception +// +//===----------------------------------------------------------------------===// + +#ifndef __CLC_MATH_CLC_FDIM_H__ +#define __CLC_MATH_CLC_FDIM_H__ + +#define __CLC_BODY <clc/shared/binary_decl.inc> +#define __CLC_FUNCTION __clc_fdim + +#include <clc/math/gentype.inc> + +#undef __CLC_BODY +#undef __CLC_FUNCTION + +#endif // __CLC_MATH_CLC_FDIM_H__ diff --git a/libclc/clc/include/clc/math/gentype.inc b/libclc/clc/include/clc/math/gentype.inc index 0c7dd5b220b9c..196d2e96922eb 100644 --- a/libclc/clc/include/clc/math/gentype.inc +++ b/libclc/clc/include/clc/math/gentype.inc @@ -57,6 +57,8 @@ #define __CLC_CONVERT_UINTN __CLC_XCONCAT(__clc_convert_, __CLC_UINTN) #define __CLC_CONVERT_ULONGN __CLC_XCONCAT(__clc_convert_, __CLC_ULONGN) +#define __CLC_CONVERT_BIT_INTN __CLC_XCONCAT(__clc_convert_, __CLC_BIT_INTN) + // See definitions of __CLC_S_GENTYPE/__CLC_U_GENTYPE below, which depend on the // specific size of floating-point type. These are the signed and unsigned // integers of the same bitwidth and element count as the GENTYPE. They match @@ -329,6 +331,8 @@ #undef __CLC_CONVERT_UINTN #undef __CLC_CONVERT_ULONGN +#undef __CLC_CONVERT_BIT_INTN + #undef __CLC_ULONGN #undef __CLC_UINTN #undef __CLC_USHORTN diff --git a/libclc/clc/lib/generic/SOURCES b/libclc/clc/lib/generic/SOURCES index dafc69e1d83a4..a69a0530a0ac0 100644 --- a/libclc/clc/lib/generic/SOURCES +++ b/libclc/clc/lib/generic/SOURCES @@ -41,6 +41,7 @@ math/clc_exp2.cl math/clc_expm1.cl math/clc_exp_helper.cl math/clc_fabs.cl +math/clc_fdim.cl math/clc_fma.cl math/clc_fmax.cl math/clc_fmin.cl diff --git a/libclc/clc/lib/generic/math/clc_fdim.cl b/libclc/clc/lib/generic/math/clc_fdim.cl new file mode 100644 index 0000000000000..4e76448cde88f --- /dev/null +++ b/libclc/clc/lib/generic/math/clc_fdim.cl @@ -0,0 +1,18 @@ +//===----------------------------------------------------------------------===// +// +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. +// See https://llvm.org/LICENSE.txt for license information. +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception +// +//===----------------------------------------------------------------------===// + +#include <clc/clc_convert.h> +#include <clc/float/definitions.h> +#include <clc/internal/clc.h> +#include <clc/math/clc_fmax.h> +#include <clc/math/math.h> +#include <clc/relational/clc_isnan.h> +#include <clc/relational/clc_select.h> + +#define __CLC_BODY <clc_fdim.inc> +#include <clc/math/gentype.inc> diff --git a/libclc/clc/lib/generic/math/clc_fdim.inc b/libclc/clc/lib/generic/math/clc_fdim.inc new file mode 100644 index 0000000000000..712818bb0a305 --- /dev/null +++ b/libclc/clc/lib/generic/math/clc_fdim.inc @@ -0,0 +1,15 @@ +//===----------------------------------------------------------------------===// +// +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. +// See https://llvm.org/LICENSE.txt for license information. +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception +// +//===----------------------------------------------------------------------===// + +_CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __clc_fdim(__CLC_GENTYPE x, + __CLC_GENTYPE y) { + return __clc_select( + __builtin_elementwise_max(x - y, (__CLC_GENTYPE)__CLC_FP_LIT(0.0)), + __CLC_GENTYPE_NAN, + __CLC_CONVERT_BIT_INTN(__clc_isnan(x) || __clc_isnan(y))); +} diff --git a/libclc/generic/lib/math/fdim.cl b/libclc/generic/lib/math/fdim.cl index 6c336094e7f5c..dc078d42fd591 100644 --- a/libclc/generic/lib/math/fdim.cl +++ b/libclc/generic/lib/math/fdim.cl @@ -7,7 +7,8 @@ //===----------------------------------------------------------------------===// #include <clc/clc.h> -#include <clc/math/math.h> +#include <clc/math/clc_fdim.h> -#define __CLC_BODY <fdim.inc> +#define FUNCTION fdim +#define __CLC_BODY <clc/shared/binary_def.inc> #include <clc/math/gentype.inc> diff --git a/libclc/generic/lib/math/fdim.inc b/libclc/generic/lib/math/fdim.inc deleted file mode 100644 index ad22b7e9b4990..0000000000000 --- a/libclc/generic/lib/math/fdim.inc +++ /dev/null @@ -1,82 +0,0 @@ -//===----------------------------------------------------------------------===// -// -// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. -// See https://llvm.org/LICENSE.txt for license information. -// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception -// -//===----------------------------------------------------------------------===// - -#if __CLC_FPSIZE == 32 -#ifdef __CLC_SCALAR -_CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE fdim(__CLC_GENTYPE x, __CLC_GENTYPE y) { - if (__builtin_isnan(x) || __builtin_isnan(y)) - return as_float(QNANBITPATT_SP32); - return fmax(x - y, 0.0f); -} -#define __CLC_FDIM_VEC(width) \ -_CLC_OVERLOAD _CLC_DEF float##width fdim(float##width x, float##width y) { \ - /* Determine if x or y is NaN. */ \ - /* Vector true is -1, i.e. all-bits-set, and NaN==NaN is false. */ \ - /* If either is NaN, then ~((x==x) & (y==y)) will be 0 (e.g. ~(-1)), as will n. */ \ - int##width n = ~((x == x) & (y == y)) & QNANBITPATT_SP32; \ - /* Calculate x-y if x>y, otherwise positive 0, again taking */ \ - /* advantage of vector true being all-bits-set. */ \ - int##width r = (x > y) & as_int##width(x - y); \ - return as_float##width(n | r); \ -} -__CLC_FDIM_VEC(2) -__CLC_FDIM_VEC(3) -__CLC_FDIM_VEC(4) -__CLC_FDIM_VEC(8) -__CLC_FDIM_VEC(16) -#undef __CLC_FDIM_VEC -#endif -#endif - -#if __CLC_FPSIZE == 64 -#ifdef __CLC_SCALAR -_CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE fdim(__CLC_GENTYPE x, private __CLC_GENTYPE y) { - long n = -(isnan(x) | isnan(y)) & QNANBITPATT_DP64; - long r = -(x > y) & as_long(x - y); - return as_double(n | r); -} -#define __CLC_FDIM_VEC(width) \ -_CLC_OVERLOAD _CLC_DEF double##width fdim(double##width x, double##width y) { \ - /* See comment in float implementation for explanation. */ \ - long##width n = ~((x == x) & (y == y)) & QNANBITPATT_DP64; \ - long##width r = (x > y) & as_long##width(x - y); \ - return as_double##width(n | r); \ -} -__CLC_FDIM_VEC(2) -__CLC_FDIM_VEC(3) -__CLC_FDIM_VEC(4) -__CLC_FDIM_VEC(8) -__CLC_FDIM_VEC(16) -#undef __CLC_FDIM_VEC -#endif -#endif - -#if __CLC_FPSIZE == 16 -#ifdef __CLC_SCALAR -#define QNANBITPATT_FP16 ((short)0x7e00) -_CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE fdim(__CLC_GENTYPE x, - private __CLC_GENTYPE y) { - short n = -(isnan(x) | isnan(y)) & QNANBITPATT_FP16; - short r = -(x > y) & as_short(x - y); - return as_half((short)(n | r)); -} -#define __CLC_FDIM_VEC(width) \ - _CLC_OVERLOAD _CLC_DEF half##width fdim(half##width x, half##width y) { \ - /* See comment in float implementation for explanation. */ \ - short##width n = ~((x == x) & (y == y)) & QNANBITPATT_FP16; \ - short##width r = (x > y) & as_short##width(x - y); \ - return as_half##width(n | r); \ - } -__CLC_FDIM_VEC(2) -__CLC_FDIM_VEC(3) -__CLC_FDIM_VEC(4) -__CLC_FDIM_VEC(8) -__CLC_FDIM_VEC(16) -#undef __CLC_FDIM_VEC -#endif -#endif _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
