Author: Matt Arsenault Date: 2026-03-24T10:58:59Z New Revision: aeba5d62e5352aa3ed2f0d5a6a563ac873dde716
URL: https://github.com/llvm/llvm-project/commit/aeba5d62e5352aa3ed2f0d5a6a563ac873dde716 DIFF: https://github.com/llvm/llvm-project/commit/aeba5d62e5352aa3ed2f0d5a6a563ac873dde716.diff LOG: libclc: Update cosh (#188214) This was originally ported from rocm device libs in 9cb070f96a8a9af5f513ffba0a8eed362623f216. Merge in more recent changes. Added: Modified: libclc/clc/lib/generic/math/clc_cosh.cl libclc/clc/lib/generic/math/clc_cosh.inc Removed: ################################################################################ diff --git a/libclc/clc/lib/generic/math/clc_cosh.cl b/libclc/clc/lib/generic/math/clc_cosh.cl index 1a376685eeae9..5dd426068e6b4 100644 --- a/libclc/clc/lib/generic/math/clc_cosh.cl +++ b/libclc/clc/lib/generic/math/clc_cosh.cl @@ -8,17 +8,11 @@ #include "clc/clc_convert.h" #include "clc/float/definitions.h" -#include "clc/internal/clc.h" #include "clc/math/clc_copysign.h" -#include "clc/math/clc_exp.h" +#include "clc/math/clc_cosh.h" +#include "clc/math/clc_ep.h" +#include "clc/math/clc_exp2_fast.h" #include "clc/math/clc_fabs.h" -#include "clc/math/clc_fma.h" -#include "clc/math/clc_mad.h" -#include "clc/math/math.h" -#include "clc/math/tables.h" -#include "clc/relational/clc_isinf.h" -#include "clc/relational/clc_isnan.h" -#include "clc/shared/clc_min.h" #define __CLC_BODY "clc_cosh.inc" #include "clc/math/gentype.inc" diff --git a/libclc/clc/lib/generic/math/clc_cosh.inc b/libclc/clc/lib/generic/math/clc_cosh.inc index a2be83fcf0860..47783ebe0a63e 100644 --- a/libclc/clc/lib/generic/math/clc_cosh.inc +++ b/libclc/clc/lib/generic/math/clc_cosh.inc @@ -8,191 +8,38 @@ #if __CLC_FPSIZE == 32 -_CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __clc_cosh(__CLC_GENTYPE x) { - // After dealing with special cases the computation is split into regions as - // follows. abs(x) >= max_cosh_arg: cosh(x) = sign(x)*Inf abs(x) >= - // small_threshold: cosh(x) = sign(x)*exp(abs(x))/2 computed using the - // splitexp and scaleDouble functions as for exp_amd(). - // abs(x) < small_threshold: - // compute p = exp(y) - 1 and then z = 0.5*(p+(p/(p+1.0))) - // cosh(x) is then z. - - const __CLC_GENTYPE max_cosh_arg = 0x1.65a9fap+6f; - const __CLC_GENTYPE small_threshold = 0x1.0a2b24p+3f; - - __CLC_GENTYPE y = __clc_fabs(x); - __CLC_UINTN aux = __CLC_AS_UINTN(y); - - // Find the integer part y0 of y and the increment dy = y - y0. We then - // compute z = sinh(y) = sinh(y0)cosh(dy) + cosh(y0)sinh(dy) z = cosh(y) = - // cosh(y0)cosh(dy) + sinh(y0)sinh(dy) where sinh(y0) and cosh(y0) are - // tabulated above. - - __CLC_INTN ind = __CLC_CONVERT_INTN(y); - ind = __CLC_CONVERT_UINTN(ind) > 36U ? 0 : ind; - - __CLC_GENTYPE dy = y - __CLC_CONVERT_GENTYPE(ind); - __CLC_GENTYPE dy2 = dy * dy; - - __CLC_GENTYPE sdy = __clc_mad( - dy2, - __clc_mad( - dy2, - __clc_mad( - dy2, - __clc_mad( - dy2, - __clc_mad(dy2, - __clc_mad(dy2, 0.7746188980094184251527126e-12f, - 0.160576793121939886190847e-9f), - 0.250521176994133472333666e-7f), - 0.275573191913636406057211e-5f), - 0.198412698413242405162014e-3f), - 0.833333333333329931873097e-2f), - 0.166666666666666667013899e0f); - sdy = __clc_mad(sdy, dy * dy2, dy); - - __CLC_GENTYPE cdy = __clc_mad( - dy2, - __clc_mad( - dy2, - __clc_mad( - dy2, - __clc_mad( - dy2, - __clc_mad(dy2, - __clc_mad(dy2, 0.1163921388172173692062032e-10f, - 0.208744349831471353536305e-8f), - 0.275573350756016588011357e-6f), - 0.248015872460622433115785e-4f), - 0.138888888889814854814536e-2f), - 0.416666666666660876512776e-1f), - 0.500000000000000005911074e0f); - cdy = __clc_mad(cdy, dy2, 1.0f); - - __CLC_GENTYPE sinhcoshh = __CLC_USE_TABLE(sinhcosh_tbl_head, ind); - __CLC_GENTYPE sinhcosht = __CLC_USE_TABLE(sinhcosh_tbl_tail, ind); - __CLC_GENTYPE z = __clc_mad(sinhcoshh, sdy, sinhcosht * cdy); - - // When exp(-x) is insignificant compared to exp(x), return exp(x)/2 - __CLC_GENTYPE t = __clc_exp(y - 0x1.62e500p-1f); - __CLC_GENTYPE zsmall = __clc_mad(0x1.a0210ep-18f, t, t); - z = y >= small_threshold ? zsmall : z; - - // Corner cases - z = y >= max_cosh_arg ? __CLC_AS_GENTYPE((__CLC_UINTN)PINFBITPATT_SP32) : z; - z = aux > PINFBITPATT_SP32 ? __CLC_GENTYPE_NAN : z; - z = aux < 0x38800000 ? 1.0f : z; +_CLC_DEF _CLC_OVERLOAD _CLC_CONST __CLC_FLOATN __clc_cosh(__CLC_FLOATN x) { + const __CLC_EP_PAIR ln2 = __clc_ep_make_pair(__CLC_FP_LIT(0x1.62e430p-1), + __CLC_FP_LIT(-0x1.05c610p-29)); + __CLC_FLOATN absx = __clc_fabs(x); + __CLC_EP_PAIR e = __clc_ep_exp_extended(__clc_ep_sub(absx, ln2)); + __CLC_EP_PAIR c = __clc_ep_fast_add(e, __clc_ep_ldexp(__clc_ep_recip(e), -2)); + __CLC_FLOATN z = absx > 0x1.65a9f8p+6f ? __CLC_GENTYPE_INF : c.hi; return z; } #elif __CLC_FPSIZE == 64 -_CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __clc_cosh(__CLC_GENTYPE x) { - // After dealing with special cases the computation is split into - // regions as follows: - // - // abs(x) >= max_cosh_arg: - // cosh(x) = sign(x)*Inf - // - // abs(x) >= small_threshold: - // cosh(x) = sign(x)*exp(abs(x))/2 computed using the - // splitexp and scaleDouble functions as for exp_amd(). - // - // abs(x) < small_threshold: - // compute p = exp(y) - 1 and then z = 0.5*(p+(p/(p+1.0))) - // cosh(x) is then sign(x)*z. - - // This is ln(2^1025) = 0x408633ce8fb9f87e - const __CLC_GENTYPE max_cosh_arg = 7.10475860073943977113e+02; - - // This is where exp(-x) is insignificant compared to exp(x) = ln(2^27) - const __CLC_GENTYPE small_threshold = 0x1.2b708872320e2p+4; - - __CLC_GENTYPE y = __clc_fabs(x); - - // In this range we find the integer part y0 of y - // and the increment dy = y - y0. We then compute - // z = cosh(y) = cosh(y0)cosh(dy) + sinh(y0)sinh(dy) - // where sinh(y0) and cosh(y0) are tabulated above. - - __CLC_INTN ind = __clc_min(__CLC_CONVERT_INTN(y), 36); - __CLC_GENTYPE dy = y - __CLC_CONVERT_GENTYPE(ind); - __CLC_GENTYPE dy2 = dy * dy; - - __CLC_GENTYPE sdy = - dy * dy2 * - __clc_fma( - dy2, - __clc_fma( - dy2, - __clc_fma( - dy2, - __clc_fma( - dy2, - __clc_fma(dy2, - __clc_fma(dy2, 0.7746188980094184251527126e-12, - 0.160576793121939886190847e-9), - 0.250521176994133472333666e-7), - 0.275573191913636406057211e-5), - 0.198412698413242405162014e-3), - 0.833333333333329931873097e-2), - 0.166666666666666667013899e0); - - __CLC_GENTYPE cdy = - dy2 * - __clc_fma( - dy2, - __clc_fma( - dy2, - __clc_fma( - dy2, - __clc_fma( - dy2, - __clc_fma(dy2, - __clc_fma(dy2, 0.1163921388172173692062032e-10, - 0.208744349831471353536305e-8), - 0.275573350756016588011357e-6), - 0.248015872460622433115785e-4), - 0.138888888889814854814536e-2), - 0.416666666666660876512776e-1), - 0.500000000000000005911074e0); - - // At this point sinh(dy) is approximated by dy + sdy, - // and cosh(dy) is approximated by 1 + cdy. - __CLC_GENTYPE cl = __CLC_USE_TABLE(cosh_tbl_head, ind); - __CLC_GENTYPE ct = __CLC_USE_TABLE(cosh_tbl_tail, ind); - __CLC_GENTYPE sl = __CLC_USE_TABLE(sinh_tbl_head, ind); - __CLC_GENTYPE st = __CLC_USE_TABLE(sinh_tbl_tail, ind); - - __CLC_GENTYPE z = - __clc_fma( - sl, dy, - __clc_fma(sl, sdy, - __clc_fma(cl, cdy, - __clc_fma(st, dy, __clc_fma(st, sdy, ct * cdy)) + - ct))) + - cl; - - // Other cases - z = y < 0x1.0p-28 ? 1.0 : z; - - __CLC_GENTYPE t = __clc_exp(y - 0x1.62e42fefa3800p-1); - t = __clc_fma(t, -0x1.ef35793c76641p-45, t); - z = y >= small_threshold ? t : z; - - z = y >= max_cosh_arg ? __CLC_AS_GENTYPE((__CLC_ULONGN)PINFBITPATT_DP64) : z; +_CLC_DEF _CLC_OVERLOAD _CLC_CONST __CLC_DOUBLEN __clc_cosh(__CLC_DOUBLEN x) { + const __CLC_EP_PAIR ln2 = __clc_ep_make_pair( + __CLC_FP_LIT(0x1.62e42fefa39efp-1), __CLC_FP_LIT(0x1.abc9e3b39803fp-56)); - z = __clc_isinf(x) || __clc_isnan(x) ? y : z; + x = __clc_fabs(x); + __CLC_EP_PAIR e = __clc_ep_exp_extended(__clc_ep_sub(x, ln2)); + __CLC_EP_PAIR c = __clc_ep_fast_add(e, __clc_ep_ldexp(__clc_ep_recip(e), -2)); + __CLC_DOUBLEN z = x >= 0x1.633ce8fb9f87ep+9 ? __CLC_GENTYPE_INF : c.hi; return z; } #elif __CLC_FPSIZE == 16 -_CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __clc_cosh(__CLC_GENTYPE x) { - return __CLC_CONVERT_GENTYPE(__clc_cosh(__CLC_CONVERT_FLOATN(x))); +_CLC_DEF _CLC_OVERLOAD _CLC_CONST __CLC_HALFN __clc_cosh(__CLC_HALFN x) { + __CLC_FLOATN x_float = __CLC_CONVERT_FLOATN(x) * (__CLC_FLOATN)M_LOG2E; + __CLC_FLOATN result = + 0.5f * (__clc_exp2_fast(x_float) + __clc_exp2_fast(-x_float)); + return __CLC_CONVERT_HALFN(result); } #endif _______________________________________________ cfe-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
