[libclc] f046f45 - libclc: Update tanh (#188215)
via cfe-commits
cfe-commits at lists.llvm.org
Tue Mar 24 04:01:36 PDT 2026
Author: Matt Arsenault
Date: 2026-03-24T12:01:30+01:00
New Revision: f046f4518dc7c707289b2fc8b748d3034cfa4c15
URL: https://github.com/llvm/llvm-project/commit/f046f4518dc7c707289b2fc8b748d3034cfa4c15
DIFF: https://github.com/llvm/llvm-project/commit/f046f4518dc7c707289b2fc8b748d3034cfa4c15.diff
LOG: libclc: Update tanh (#188215)
This was originally ported from rocm device libs in
f51df5ba8c4512dbeb7828ac0c34f89177b551d6. Merge in more
recent changes.
Added:
Modified:
libclc/clc/lib/generic/math/clc_tanh.cl
libclc/clc/lib/generic/math/clc_tanh.inc
Removed:
################################################################################
diff --git a/libclc/clc/lib/generic/math/clc_tanh.cl b/libclc/clc/lib/generic/math/clc_tanh.cl
index c0e6f60f6fe4c..cf9d5679c199d 100644
--- a/libclc/clc/lib/generic/math/clc_tanh.cl
+++ b/libclc/clc/lib/generic/math/clc_tanh.cl
@@ -7,15 +7,14 @@
//===----------------------------------------------------------------------===//
#include "clc/clc_convert.h"
-#include "clc/internal/clc.h"
+#include "clc/math/clc_copysign.h"
+#include "clc/math/clc_ep.h"
#include "clc/math/clc_exp.h"
-#include "clc/math/clc_fma.h"
+#include "clc/math/clc_exp2_fast.h"
+#include "clc/math/clc_fabs.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"
+#include "clc/math/clc_recip_fast.h"
+#include "clc/math/clc_tanh.h"
#define __CLC_BODY "clc_tanh.inc"
#include "clc/math/gentype.inc"
diff --git a/libclc/clc/lib/generic/math/clc_tanh.inc b/libclc/clc/lib/generic/math/clc_tanh.inc
index a25fd58fcbeaf..4daf879001b1d 100644
--- a/libclc/clc/lib/generic/math/clc_tanh.inc
+++ b/libclc/clc/lib/generic/math/clc_tanh.inc
@@ -8,130 +8,54 @@
#if __CLC_FPSIZE == 32
-_CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __clc_tanh(__CLC_GENTYPE x) {
- // The definition of tanh(x) is sinh(x)/cosh(x), which is also equivalent
- // to the following three formulae:
- // 1. (exp(x) - exp(-x))/(exp(x) + exp(-x))
- // 2. (1 - (2/(exp(2*x) + 1 )))
- // 3. (exp(2*x) - 1)/(exp(2*x) + 1)
- // but computationally, some formulae are better on some ranges.
-
- const __CLC_GENTYPE large_threshold = 0x1.0a2b24p+3f;
-
- __CLC_UINTN ux = __CLC_AS_UINTN(x);
- __CLC_UINTN aux = ux & EXSIGNBIT_SP32;
- __CLC_UINTN xs = ux ^ aux;
-
- __CLC_GENTYPE y = __CLC_AS_GENTYPE(aux);
- __CLC_GENTYPE y2 = y * y;
-
- __CLC_GENTYPE a1 = __clc_mad(
- y2, __clc_mad(y2, 0.4891631088530669873e-4F, -0.14628356048797849e-2F),
- -0.28192806108402678e0F);
- __CLC_GENTYPE b1 =
- __clc_mad(y2, 0.3427017942262751343e0F, 0.845784192581041099e0F);
-
- __CLC_GENTYPE a2 = __clc_mad(
- y2, __clc_mad(y2, 0.3827534993599483396e-4F, -0.12325644183611929e-2F),
- -0.24069858695196524e0F);
- __CLC_GENTYPE b2 =
- __clc_mad(y2, 0.292529068698052819e0F, 0.72209738473684982e0F);
-
- __CLC_INTN c = y < 0.9f;
- __CLC_GENTYPE a = c ? a1 : a2;
- __CLC_GENTYPE b = c ? b1 : b2;
- __CLC_GENTYPE zlo = __clc_mad(MATH_DIVIDE(a, b), y * y2, y);
-
- __CLC_GENTYPE p = __clc_exp(2.0f * y) + 1.0f;
- __CLC_GENTYPE zhi = 1.0F - MATH_DIVIDE(2.0F, p);
-
- __CLC_GENTYPE z = y <= 1.0f ? zlo : zhi;
- z = __CLC_AS_GENTYPE(xs | __CLC_AS_UINTN(z));
+static _CLC_OVERLOAD _CLC_CONST __CLC_FLOATN __clc_tanh_small(__CLC_FLOATN y) {
+ __CLC_FLOATN y2 = y * y;
+ __CLC_FLOATN p = __clc_mad(
+ y2,
+ __clc_mad(y2,
+ __clc_mad(y2, __clc_mad(y2, -0x1.758e7ap-8f, 0x1.521192p-6f),
+ -0x1.b8389cp-5f),
+ 0x1.110704p-3f),
+ -0x1.555532p-2f);
+ return __clc_mad(y2, y * p, y);
+}
- // Edge cases
- __CLC_GENTYPE sone = __CLC_AS_GENTYPE(0x3f800000U | xs);
- z = y > large_threshold ? sone : z;
- z = aux < 0x39000000 || aux > 0x7f800000 ? x : z;
+static _CLC_OVERLOAD _CLC_CONST __CLC_FLOATN __clc_tanh_large(__CLC_FLOATN y) {
+ __CLC_FLOATN t = __clc_exp(2.0f * y);
+ return __clc_mad(-2.0f, __clc_recip_fast(t + 1.0f), 1.0f);
+}
- return z;
+_CLC_DEF _CLC_OVERLOAD _CLC_CONST __CLC_FLOATN __clc_tanh(__CLC_FLOATN x) {
+ __CLC_FLOATN y = __clc_fabs(x);
+ __CLC_FLOATN z = y < 0.625f ? __clc_tanh_small(y) : __clc_tanh_large(y);
+ return __clc_copysign(z, x);
}
#elif __CLC_FPSIZE == 64
-_CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __clc_tanh(__CLC_GENTYPE x) {
- // The definition of tanh(x) is sinh(x)/cosh(x), which is also equivalent
- // to the following three formulae:
- // 1. (exp(x) - exp(-x))/(exp(x) + exp(-x))
- // 2. (1 - (2/(exp(2*x) + 1 )))
- // 3. (exp(2*x) - 1)/(exp(2*x) + 1)
- // but computationally, some formulae are better on some ranges.
-
- // The point at which e^-x is insignificant compared to e^x = ln(2^27)
- const __CLC_GENTYPE large_threshold = 0x1.2b708872320e2p+4;
-
- __CLC_ULONGN ux = __CLC_AS_ULONGN(x);
- __CLC_ULONGN ax = ux & ~SIGNBIT_DP64;
- __CLC_ULONGN sx = ux ^ ax;
- __CLC_GENTYPE y = __CLC_AS_GENTYPE(ax);
- __CLC_GENTYPE y2 = y * y;
-
- // y < 0.9
- __CLC_GENTYPE znl =
- __clc_fma(y2,
- __clc_fma(y2,
- __clc_fma(y2, -0.142077926378834722618091e-7,
- -0.200047621071909498730453e-3),
- -0.176016349003044679402273e-1),
- -0.274030424656179760118928e0);
-
- __CLC_GENTYPE zdl =
- __clc_fma(y2,
- __clc_fma(y2,
- __clc_fma(y2, 0.2091140262529164482568557e-3,
- 0.201562166026937652780575e-1),
- 0.381641414288328849317962e0),
- 0.822091273968539282568011e0);
-
- // 0.9 <= y <= 1
- __CLC_GENTYPE znm =
- __clc_fma(y2,
- __clc_fma(y2,
- __clc_fma(y2, -0.115475878996143396378318e-7,
- -0.165597043903549960486816e-3),
- -0.146173047288731678404066e-1),
- -0.227793870659088295252442e0);
-
- __CLC_GENTYPE zdm =
- __clc_fma(y2,
- __clc_fma(y2,
- __clc_fma(y2, 0.173076050126225961768710e-3,
- 0.167358775461896562588695e-1),
- 0.317204558977294374244770e0),
- 0.683381611977295894959554e0);
-
- __CLC_LONGN c = y < 0.9;
- __CLC_GENTYPE zn = c ? znl : znm;
- __CLC_GENTYPE zd = c ? zdl : zdm;
- __CLC_GENTYPE z = y + y * y2 * MATH_DIVIDE(zn, zd);
-
- // y > 1
- __CLC_GENTYPE p = __clc_exp(2.0 * y) + 1.0;
- __CLC_GENTYPE zg = 1.0 - 2.0 / p;
-
- z = y > 1.0 ? zg : z;
-
- // Other cases
- z = y < 0x1.0p-28 || ax > PINFBITPATT_DP64 ? x : z;
+_CLC_DEF _CLC_OVERLOAD _CLC_CONST __CLC_DOUBLEN __clc_tanh(__CLC_DOUBLEN x) {
+ __CLC_DOUBLEN y = __clc_fabs(x);
+ __CLC_EP_PAIR e = __clc_ep_exp_extended(__clc_ep_make_pair(y, 0.0));
+ __CLC_EP_PAIR ei = __clc_ep_recip(e);
+ __CLC_EP_PAIR t =
+ __clc_ep_fast_div(__clc_ep_fast_sub(e, ei), __clc_ep_fast_add(e, ei));
+ __CLC_DOUBLEN z = t.hi;
- z = y > large_threshold ? 1.0 : z;
+ z = y > 19.0625 ? 1.0 : z;
+ z = y < 0x1.0p-27 ? y : z;
- return __CLC_AS_GENTYPE(sx | __CLC_AS_ULONGN(z));
+ return __clc_copysign(z, x);
}
#elif __CLC_FPSIZE == 16
-_CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __clc_tanh(__CLC_GENTYPE x) {
- return __CLC_CONVERT_GENTYPE(__clc_tanh(__CLC_CONVERT_FLOATN(x)));
+_CLC_DEF _CLC_OVERLOAD _CLC_CONST __CLC_HALFN __clc_tanh(__CLC_HALFN hx) {
+ __CLC_FLOATN x = __CLC_CONVERT_FLOATN(hx) * (__CLC_FLOATN)M_LOG2E;
+ __CLC_FLOATN a = __clc_exp2_fast(x);
+ __CLC_FLOATN b = __clc_exp2_fast(-x);
+ __CLC_HALFN one = __clc_copysign(1.0h, hx);
+ __CLC_HALFN ret = __CLC_CONVERT_HALFN((a - b) * __clc_recip_fast(a + b));
+ return __clc_fabs(hx) > 4.5h ? one : ret;
}
#endif
More information about the cfe-commits
mailing list