[libclc] cba948b - libclc: Update atan (#188095)
via cfe-commits
cfe-commits at lists.llvm.org
Tue Mar 24 00:11:47 PDT 2026
Author: Matt Arsenault
Date: 2026-03-24T08:11:41+01:00
New Revision: cba948b2b93cc5d51f70f0fa0d66bc5cacef9d76
URL: https://github.com/llvm/llvm-project/commit/cba948b2b93cc5d51f70f0fa0d66bc5cacef9d76
DIFF: https://github.com/llvm/llvm-project/commit/cba948b2b93cc5d51f70f0fa0d66bc5cacef9d76.diff
LOG: libclc: Update atan (#188095)
This was originally ported from rocm device libs in
47882923c7b48c00d6c0ea9960b5457e957093c4. Merge in more
recent changes.
Added:
libclc/clc/include/clc/math/clc_atan_helpers.h
libclc/clc/include/clc/math/clc_atan_helpers_decl.inc
libclc/clc/lib/generic/math/clc_atan_helpers.cl
libclc/clc/lib/generic/math/clc_atan_helpers.inc
Modified:
libclc/clc/lib/generic/CMakeLists.txt
libclc/clc/lib/generic/math/clc_atan.cl
libclc/clc/lib/generic/math/clc_atan.inc
Removed:
################################################################################
diff --git a/libclc/clc/include/clc/math/clc_atan_helpers.h b/libclc/clc/include/clc/math/clc_atan_helpers.h
new file mode 100644
index 0000000000000..ad86cda15f0d7
--- /dev/null
+++ b/libclc/clc/include/clc/math/clc_atan_helpers.h
@@ -0,0 +1,17 @@
+//===----------------------------------------------------------------------===//
+//
+// 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_ATAN_HELPERS_H__
+#define __CLC_MATH_CLC_ATAN_HELPERS_H__
+
+#include "clc/internal/clc.h"
+
+#define __CLC_BODY "clc/math/clc_atan_helpers_decl.inc"
+#include "clc/math/gentype.inc"
+
+#endif // __CLC_MATH_CLC_ATAN_HELPERS_H__
diff --git a/libclc/clc/include/clc/math/clc_atan_helpers_decl.inc b/libclc/clc/include/clc/math/clc_atan_helpers_decl.inc
new file mode 100644
index 0000000000000..6b1aa27a9c60d
--- /dev/null
+++ b/libclc/clc/include/clc/math/clc_atan_helpers_decl.inc
@@ -0,0 +1,10 @@
+//===----------------------------------------------------------------------===//
+//
+// 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_DECL _CLC_OVERLOAD _CLC_CONST __CLC_GENTYPE
+__clc_atan_reduced(__CLC_GENTYPE v);
diff --git a/libclc/clc/lib/generic/CMakeLists.txt b/libclc/clc/lib/generic/CMakeLists.txt
index 8c5e73c11515d..8ad008a20bce3 100644
--- a/libclc/clc/lib/generic/CMakeLists.txt
+++ b/libclc/clc/lib/generic/CMakeLists.txt
@@ -62,6 +62,7 @@ libclc_configure_source_list(CLC_GENERIC_SOURCES
math/clc_asinh.cl
math/clc_asinpi.cl
math/clc_atan.cl
+ math/clc_atan_helpers.cl
math/clc_atan2.cl
math/clc_atan2pi.cl
math/clc_atanh.cl
diff --git a/libclc/clc/lib/generic/math/clc_atan.cl b/libclc/clc/lib/generic/math/clc_atan.cl
index 3219299dac2a2..768aa94d96333 100644
--- a/libclc/clc/lib/generic/math/clc_atan.cl
+++ b/libclc/clc/lib/generic/math/clc_atan.cl
@@ -9,11 +9,12 @@
#include "clc/clc_convert.h"
#include "clc/float/definitions.h"
#include "clc/internal/clc.h"
+#include "clc/math/clc_atan_helpers.h"
+#include "clc/math/clc_copysign.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/relational/clc_isnan.h"
+#include "clc/math/clc_recip_fast.h"
#define __CLC_BODY "clc_atan.inc"
#include "clc/math/gentype.inc"
diff --git a/libclc/clc/lib/generic/math/clc_atan.inc b/libclc/clc/lib/generic/math/clc_atan.inc
index 23136dbd74e02..83d849cc54590 100644
--- a/libclc/clc/lib/generic/math/clc_atan.inc
+++ b/libclc/clc/lib/generic/math/clc_atan.inc
@@ -8,161 +8,51 @@
#if __CLC_FPSIZE == 32
-_CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __clc_atan(__CLC_GENTYPE x) {
- const __CLC_GENTYPE piby2 = 1.5707963267948966f; // 0x3ff921fb54442d18
+_CLC_DEF _CLC_OVERLOAD _CLC_CONST __CLC_FLOATN __clc_atan(__CLC_FLOATN x) {
+ __CLC_FLOATN v = __clc_fabs(x);
+ __CLC_INTN g = v > 1.0f;
- __CLC_UINTN ux = __CLC_AS_UINTN(x);
- __CLC_UINTN aux = ux & EXSIGNBIT_SP32;
- __CLC_UINTN sx = ux ^ aux;
+ __CLC_FLOATN vi = __clc_recip_fast(v);
+ v = g ? vi : v;
- __CLC_GENTYPE spiby2 = __CLC_AS_GENTYPE(sx | __CLC_AS_UINTN(piby2));
+ __CLC_FLOATN a = __clc_atan_reduced(v);
- __CLC_GENTYPE v = __CLC_AS_GENTYPE(aux);
+ __CLC_FLOATN y = __clc_mad(0x1.ddcb02p-1f, 0x1.aee9d6p+0f, -a);
+ a = g ? y : a;
- // Return for NaN
- __CLC_GENTYPE ret = x;
-
- // 2^26 <= |x| <= Inf => atan(x) is close to piby2
- ret = aux <= PINFBITPATT_SP32 ? spiby2 : ret;
-
- // Reduce arguments 2^-19 <= |x| < 2^26
-
- // 39/16 <= x < 2^26
- x = -MATH_RECIP(v);
- __CLC_GENTYPE c = 1.57079632679489655800f; // atan(infinity)
-
- // 19/16 <= x < 39/16
- __CLC_INTN l = aux < 0x401c0000;
- __CLC_GENTYPE xx = MATH_DIVIDE(v - 1.5f, __clc_mad(v, 1.5f, 1.0f));
- x = l ? xx : x;
- c = l ? 9.82793723247329054082e-01f : c; // atan(1.5)
-
- // 11/16 <= x < 19/16
- l = aux < 0x3f980000U;
- xx = MATH_DIVIDE(v - 1.0f, 1.0f + v);
- x = l ? xx : x;
- c = l ? 7.85398163397448278999e-01f : c; // atan(1)
-
- // 7/16 <= x < 11/16
- l = aux < 0x3f300000;
- xx = MATH_DIVIDE(__clc_mad(v, 2.0f, -1.0f), 2.0f + v);
- x = l ? xx : x;
- c = l ? 4.63647609000806093515e-01f : c; // atan(0.5)
-
- // 2^-19 <= x < 7/16
- l = aux < 0x3ee00000;
- x = l ? v : x;
- c = l ? 0.0f : c;
+ return __clc_copysign(a, x);
+}
- // Core approximation: Remez(2,2) on [-7/16,7/16]
+#elif __CLC_FPSIZE == 64
- __CLC_GENTYPE s = x * x;
- __CLC_GENTYPE a = __clc_mad(s,
- __clc_mad(s, 0.470677934286149214138357545549e-2f,
- 0.192324546402108583211697690500f),
- 0.296528598819239217902158651186f);
+_CLC_DEF _CLC_OVERLOAD _CLC_CONST __CLC_DOUBLEN __clc_atan(__CLC_DOUBLEN x) {
+ __CLC_DOUBLEN v = __clc_fabs(x);
+ __CLC_LONGN g = v > 1.0;
- __CLC_GENTYPE b = __clc_mad(s,
- __clc_mad(s, 0.299309699959659728404442796915f,
- 0.111072499995399550138837673349e1f),
- 0.889585796862432286486651434570f);
+ v = g ? (1.0 / v) : v;
- __CLC_GENTYPE q = x * s * MATH_DIVIDE(a, b);
+ __CLC_DOUBLEN a = __clc_atan_reduced(v);
+ __CLC_DOUBLEN y = __clc_fma(0x1.dd9ad336a0500p-1, 0x1.af154eeb562d6p+0, -a);
+ a = g ? y : a;
- __CLC_GENTYPE z = c - (q - x);
- __CLC_GENTYPE zs = __CLC_AS_GENTYPE(sx | __CLC_AS_UINTN(z));
+ return __clc_copysign(a, x);
+}
- ret = aux < 0x4c800000 ? zs : ret;
+#elif __CLC_FPSIZE == 16
- // |x| < 2^-19
- ret = aux < 0x36000000 ? __CLC_AS_GENTYPE(ux) : ret;
- return ret;
-}
+_CLC_DEF _CLC_OVERLOAD _CLC_CONST __CLC_HALFN __clc_atan(__CLC_HALFN x) {
+ __CLC_HALFN v = __clc_fabs(x);
+ __CLC_SHORTN g = v > 1.0h;
-#elif __CLC_FPSIZE == 64
+ __CLC_HALFN vi = __clc_recip_fast(v);
+ v = g ? vi : v;
-_CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __clc_atan(__CLC_GENTYPE x) {
- const __CLC_GENTYPE piby2 = 1.5707963267948966e+00; // 0x3ff921fb54442d18
-
- __CLC_GENTYPE v = __clc_fabs(x);
-
- // 2^56 > v > 39/16
- __CLC_GENTYPE a = -1.0;
- __CLC_GENTYPE b = v;
- // (chi + clo) = arctan(infinity)
- __CLC_GENTYPE chi = 1.57079632679489655800e+00;
- __CLC_GENTYPE clo = 6.12323399573676480327e-17;
-
- __CLC_GENTYPE ta = v - 1.5;
- __CLC_GENTYPE tb = 1.0 + 1.5 * v;
- __CLC_LONGN l = v <= 0x1.38p+1; // 39/16 > v > 19/16
- a = l ? ta : a;
- b = l ? tb : b;
- // (chi + clo) = arctan(1.5)
- chi = l ? 9.82793723247329054082e-01 : chi;
- clo = l ? 1.39033110312309953701e-17 : clo;
-
- ta = v - 1.0;
- tb = 1.0 + v;
- l = v <= 0x1.3p+0; // 19/16 > v > 11/16
- a = l ? ta : a;
- b = l ? tb : b;
- // (chi + clo) = arctan(1.)
- chi = l ? 7.85398163397448278999e-01 : chi;
- clo = l ? 3.06161699786838240164e-17 : clo;
-
- ta = 2.0 * v - 1.0;
- tb = 2.0 + v;
- l = v <= 0x1.6p-1; // 11/16 > v > 7/16
- a = l ? ta : a;
- b = l ? tb : b;
- // (chi + clo) = arctan(0.5)
- chi = l ? 4.63647609000806093515e-01 : chi;
- clo = l ? 2.26987774529616809294e-17 : clo;
-
- l = v <= 0x1.cp-2; // v < 7/16
- a = l ? v : a;
- b = l ? 1.0 : b;
- ;
- chi = l ? 0.0 : chi;
- clo = l ? 0.0 : clo;
-
- // Core approximation: Remez(4,4) on [-7/16,7/16]
- __CLC_GENTYPE r = a / b;
- __CLC_GENTYPE s = r * r;
- __CLC_GENTYPE qn =
- __clc_fma(s,
- __clc_fma(s,
- __clc_fma(s,
- __clc_fma(s, 0.142316903342317766e-3,
- 0.304455919504853031e-1),
- 0.220638780716667420e0),
- 0.447677206805497472e0),
- 0.268297920532545909e0);
-
- __CLC_GENTYPE qd =
- __clc_fma(s,
- __clc_fma(s,
- __clc_fma(s,
- __clc_fma(s, 0.389525873944742195e-1,
- 0.424602594203847109e0),
- 0.141254259931958921e1),
- 0.182596787737507063e1),
- 0.804893761597637733e0);
-
- __CLC_GENTYPE q = r * s * qn / qd;
- r = chi - ((q - clo) - r);
-
- __CLC_GENTYPE z = __clc_isnan(x) ? x : piby2;
- z = v <= 0x1.0p+56 ? r : z;
- z = v < 0x1.0p-26 ? v : z;
- return x == v ? z : -z;
-}
+ __CLC_HALFN a = __clc_atan_reduced(v);
-#elif __CLC_FPSIZE == 16
+ __CLC_HALFN y = __clc_mad(0x1.ea8p-1h, 0x1.a3cp+0h, -a);
+ a = g ? y : a;
-_CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __clc_atan(__CLC_GENTYPE x) {
- return __CLC_CONVERT_GENTYPE(__clc_atan(__CLC_CONVERT_FLOATN(x)));
+ return __clc_copysign(a, x);
}
#endif
diff --git a/libclc/clc/lib/generic/math/clc_atan_helpers.cl b/libclc/clc/lib/generic/math/clc_atan_helpers.cl
new file mode 100644
index 0000000000000..aa623478c258b
--- /dev/null
+++ b/libclc/clc/lib/generic/math/clc_atan_helpers.cl
@@ -0,0 +1,14 @@
+//===----------------------------------------------------------------------===//
+//
+// 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/math/clc_atan_helpers.h"
+#include "clc/math/clc_mad.h"
+
+#define __CLC_BODY "clc_atan_helpers.inc"
+
+#include "clc/math/gentype.inc"
diff --git a/libclc/clc/lib/generic/math/clc_atan_helpers.inc b/libclc/clc/lib/generic/math/clc_atan_helpers.inc
new file mode 100644
index 0000000000000..ba3fb6d1904e3
--- /dev/null
+++ b/libclc/clc/lib/generic/math/clc_atan_helpers.inc
@@ -0,0 +1,53 @@
+//===----------------------------------------------------------------------===//
+//
+// 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
+
+_CLC_DEF _CLC_OVERLOAD _CLC_CONST __CLC_FLOATN
+__clc_atan_reduced(__CLC_FLOATN v) {
+ __CLC_FLOATN t = v * v;
+ __CLC_FLOATN z = __clc_mad(t, __clc_mad(t, __clc_mad(t, __clc_mad(t,
+ __clc_mad(t, __clc_mad(t, __clc_mad(t,
+ 0x1.5a54bp-9f, -0x1.f4b218p-7f), 0x1.53f67ep-5f), -0x1.2fa9aep-4f),
+ 0x1.b26364p-4f), -0x1.22c1ccp-3f), 0x1.99717ep-3f), -0x1.5554c4p-2f);
+
+ z = __clc_mad(v, t * z, v);
+ return z;
+}
+
+#elif __CLC_FPSIZE == 64
+
+_CLC_DEF _CLC_OVERLOAD _CLC_CONST __CLC_DOUBLEN
+__clc_atan_reduced(__CLC_DOUBLEN v) {
+ __CLC_DOUBLEN t = v * v;
+ __CLC_DOUBLEN z = __clc_mad(t, __clc_mad(t, __clc_mad(t, __clc_mad(t,
+ __clc_mad(t, __clc_mad(t, __clc_mad(t, __clc_mad(t,
+ __clc_mad(t, __clc_mad(t, __clc_mad(t, __clc_mad(t,
+ __clc_mad(t, __clc_mad(t, __clc_mad(t, __clc_mad(t,
+ __clc_mad(t, __clc_mad(t, __clc_mad(t,
+ 0x1.ba404b5e68a13p-17, -0x1.3e260bd3237f4p-13), 0x1.b2bb069efb384p-11), -0x1.7952daf56de9bp-9),
+ 0x1.d6d43a595c56fp-8), -0x1.c6ea4a57d9582p-7), 0x1.67e295f08b19fp-6), -0x1.e9ae6fc27006ap-6),
+ 0x1.2c15b5711927ap-5), -0x1.59976e82d3ff0p-5), 0x1.82d5d6ef28734p-5), -0x1.ae5ce6a214619p-5),
+ 0x1.e1bb48427b883p-5), -0x1.110e48b207f05p-4), 0x1.3b13657b87036p-4), -0x1.745d119378e4fp-4),
+ 0x1.c71c717e1913cp-4), -0x1.2492492376b7dp-3), 0x1.99999999952ccp-3), -0x1.5555555555523p-2);
+ z = __clc_mad(v, t*z, v);
+ return z;
+}
+
+#elif __CLC_FPSIZE == 16
+
+_CLC_DEF _CLC_OVERLOAD _CLC_CONST __CLC_HALFN
+__clc_atan_reduced(__CLC_HALFN v) {
+ __CLC_HALFN t = v * v;
+ __CLC_HALFN z = __clc_mad(
+ t, __clc_mad(t, __clc_mad(t, 0x1.938p-6h, -0x1.7f4p-4h), 0x1.7dcp-3h),
+ -0x1.54p-2h);
+ return __clc_mad(t, v * z, v);
+}
+
+#endif
More information about the cfe-commits
mailing list