[libc][C23][math] Implement cospif function correctly rounded for all…

… rounding modes (#97464) I also fixed a comment in sinpif.cpp in the first commit. Should this be included in this PR? All tests were passed, including the exhaustive test. CC: @lntue
llvm · Jul 6, 2024 · f8834ed · f8834ed
1 parent de88b2c
commit f8834ed
Show file tree

Hide file tree

Showing 18 changed files with 406 additions and 4 deletions.
diff --git a/libc/config/darwin/arm/entrypoints.txt b/libc/config/darwin/arm/entrypoints.txt
@@ -132,6 +132,7 @@ set(TARGET_LIBM_ENTRYPOINTS
     libc.src.math.coshf
     libc.src.math.cos
     libc.src.math.cosf
+    libc.src.math.cospif
     libc.src.math.erff
     libc.src.math.exp
     libc.src.math.expf

diff --git a/libc/config/linux/aarch64/entrypoints.txt b/libc/config/linux/aarch64/entrypoints.txt
@@ -346,6 +346,7 @@ set(TARGET_LIBM_ENTRYPOINTS
     libc.src.math.cos
     libc.src.math.cosf
     libc.src.math.coshf
+    libc.src.math.cospif
     libc.src.math.erff
     libc.src.math.exp
     libc.src.math.exp10

diff --git a/libc/config/linux/riscv/entrypoints.txt b/libc/config/linux/riscv/entrypoints.txt
@@ -354,6 +354,7 @@ set(TARGET_LIBM_ENTRYPOINTS
     libc.src.math.cos
     libc.src.math.cosf
     libc.src.math.coshf
+    libc.src.math.cospif
     libc.src.math.erff
     libc.src.math.exp
     libc.src.math.exp10

diff --git a/libc/config/linux/x86_64/entrypoints.txt b/libc/config/linux/x86_64/entrypoints.txt
@@ -371,6 +371,7 @@ set(TARGET_LIBM_ENTRYPOINTS
     libc.src.math.cos
     libc.src.math.cosf
     libc.src.math.coshf
+    libc.src.math.cospif
     libc.src.math.erff
     libc.src.math.exp
     libc.src.math.exp10

diff --git a/libc/docs/math/index.rst b/libc/docs/math/index.rst
@@ -274,7 +274,7 @@ Higher Math Functions
 +-----------+------------------+-----------------+------------------------+----------------------+------------------------+------------------------+----------------------------+
 | cosh      | |check|          |                 |                        |                      |                        | 7.12.5.4               | F.10.2.4                   |
 +-----------+------------------+-----------------+------------------------+----------------------+------------------------+------------------------+----------------------------+
-| cospi     |                  |                 |                        |                      |                        | 7.12.4.12              | F.10.1.12                  |
+| cospi     | |check|          |                 |                        |                      |                        | 7.12.4.12              | F.10.1.12                  |
 +-----------+------------------+-----------------+------------------------+----------------------+------------------------+------------------------+----------------------------+
 | dsqrt     | N/A              | N/A             |                        | N/A                  |                        | 7.12.14.6              | F.10.11                    |
 +-----------+------------------+-----------------+------------------------+----------------------+------------------------+------------------------+----------------------------+

diff --git a/libc/src/math/CMakeLists.txt b/libc/src/math/CMakeLists.txt
@@ -81,6 +81,7 @@ add_math_entrypoint_object(cos)
 add_math_entrypoint_object(cosf)
 add_math_entrypoint_object(cosh)
 add_math_entrypoint_object(coshf)
+add_math_entrypoint_object(cospif)
 
 add_math_entrypoint_object(erf)
 add_math_entrypoint_object(erff)

diff --git a/libc/src/math/cospif.h b/libc/src/math/cospif.h
@@ -0,0 +1,18 @@
+//===-- Implementation header for cospif ------------------------*- C++ -*-===//
+//
+// 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 LLVM_LIBC_SRC_MATH_COSPIF_H
+#define LLVM_LIBC_SRC_MATH_COSPIF_H
+
+namespace LIBC_NAMESPACE {
+
+float cospif(float x);
+
+} // namespace LIBC_NAMESPACE
+
+#endif // LLVM_LIBC_SRC_MATH_COSPIF_H
diff --git a/libc/src/math/generic/CMakeLists.txt b/libc/src/math/generic/CMakeLists.txt
@@ -217,6 +217,23 @@ add_entrypoint_object(
     -O3
 )
 
+add_entrypoint_object(
+  cospif
+  SRCS
+    cospif.cpp
+  HDRS
+    ../cospif.h
+  DEPENDS
+    .sincosf_utils
+    libc.src.__support.FPUtil.fenv_impl
+    libc.src.__support.FPUtil.fp_bits
+    libc.src.__support.FPUtil.fma
+    libc.src.__support.FPUtil.multiply_add
+    libc.src.__support.macros.optimization
+  COMPILE_OPTIONS
+    -O3
+)
+
 add_entrypoint_object(
   sin
   SRCS

diff --git a/libc/src/math/generic/cospif.cpp b/libc/src/math/generic/cospif.cpp
@@ -0,0 +1,96 @@
+//===-- Single-precision cospi function -----------------------------------===//
+//
+// 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 "src/math/cospif.h"
+#include "sincosf_utils.h"
+#include "src/__support/FPUtil/FEnvImpl.h"
+#include "src/__support/FPUtil/FPBits.h"
+#include "src/__support/FPUtil/multiply_add.h"
+#include "src/__support/common.h"
+#include "src/__support/macros/optimization.h"            // LIBC_UNLIKELY
+#include "src/__support/macros/properties/cpu_features.h" // LIBC_TARGET_CPU_HAS_FMA
+
+namespace LIBC_NAMESPACE {
+
+LLVM_LIBC_FUNCTION(float, cospif, (float x)) {
+  using FPBits = typename fputil::FPBits<float>;
+
+  FPBits xbits(x);
+  Sign xsign = xbits.sign();
+  xbits.set_sign(Sign::POS);
+
+  uint32_t x_abs = xbits.uintval();
+  double xd = static_cast<double>(xbits.get_val());
+
+  // Range reduction:
+  // For |x| > 1/32, we perform range reduction as follows:
+  // Find k and y such that:
+  //   x = (k + y) * 1/32
+  //   k is an integer
+  //   |y| < 0.5
+  //
+  // This is done by performing:
+  //   k = round(x * 32)
+  //   y = x * 32 - k
+  //
+  // Once k and y are computed, we then deduce the answer by the cosine of sum
+  // formula:
+  //   cospi(x) = cos((k + y)*pi/32)
+  //          = cos(y*pi/32) * cos(k*pi/32) - sin(y*pi/32) * sin(k*pi/32)
+  // The values of sin(k*pi/32) and cos(k*pi/32) for k = 0..63 are precomputed
+  // and stored using a vector of 32 doubles. Sin(y*pi/32) and cos(y*pi/32) are
+  // computed using degree-7 and degree-6 minimax polynomials generated by
+  // Sollya respectively.
+
+  // The exhautive test passes for smaller values
+  if (LIBC_UNLIKELY(x_abs < 0x38A2'F984U)) {
+
+#if defined(LIBC_TARGET_CPU_HAS_FMA)
+    return fputil::multiply_add(xbits.get_val(), -0x1.0p-25f, 1.0f);
+#else
+    return static_cast<float>(fputil::multiply_add(xd, -0x1.0p-25, 1.0));
+#endif // LIBC_TARGET_CPU_HAS_FMA
+  }
+
+  // Numbers greater or equal to 2^23 are always integers or NaN
+  if (LIBC_UNLIKELY(x_abs >= 0x4B00'0000)) {
+
+    if (LIBC_UNLIKELY(x_abs < 0x4B80'0000)) {
+      return (x_abs & 0x1) ? -1.0f : 1.0f;
+    }
+
+    // x is inf or nan.
+    if (LIBC_UNLIKELY(x_abs >= 0x7f80'0000U)) {
+      if (x_abs == 0x7f80'0000U) {
+        fputil::set_errno_if_required(EDOM);
+        fputil::raise_except_if_required(FE_INVALID);
+      }
+      return x + FPBits::quiet_nan().get_val();
+    }
+
+    return 1.0f;
+  }
+
+  // Combine the results with the sine of sum formula:
+  //   cos(pi * x) = cos((k + y)*pi/32)
+  //          = cos(y*pi/32) * cos(k*pi/32) - sin(y*pi/32) * sin(k*pi/32)
+  //          = (cosm1_y + 1) * cos_k - sin_y * sin_k
+  //          = (cosm1_y * cos_k + cos_k) - sin_y * sin_k
+  double sin_k, cos_k, sin_y, cosm1_y;
+
+  sincospif_eval(xd, sin_k, cos_k, sin_y, cosm1_y);
+
+  if (LIBC_UNLIKELY(sin_y == 0 && cos_k == 0)) {
+    return FPBits::zero(xsign).get_val();
+  }
+
+  return static_cast<float>(fputil::multiply_add(
+      sin_y, -sin_k, fputil::multiply_add(cosm1_y, cos_k, cos_k)));
+}
+
+} // namespace LIBC_NAMESPACE
diff --git a/libc/src/math/generic/sinpif.cpp b/libc/src/math/generic/sinpif.cpp
@@ -26,13 +26,13 @@ LLVM_LIBC_FUNCTION(float, sinpif, (float x)) {
   double xd = static_cast<double>(x);
 
   // Range reduction:
-  // For |x| > pi/32, we perform range reduction as follows:
+  // For |x| > 1/32, we perform range reduction as follows:
   // Find k and y such that:
   //   x = (k + y) * 1/32
   //   k is an integer
   //   |y| < 0.5
-  // For small range (|x| < 2^45 when FMA instructions are available, 2^22
-  // otherwise), this is done by performing:
+  //
+  // This is done by performing:
   //   k = round(x * 32)
   //   y = x * 32 - k
   //

diff --git a/libc/test/src/math/CMakeLists.txt b/libc/test/src/math/CMakeLists.txt
@@ -28,6 +28,22 @@ add_fp_unittest(
     libc.src.__support.FPUtil.fp_bits
 )
 
+add_fp_unittest(
+  cospif_test
+  NEED_MPFR
+  SUITE
+    libc-math-unittests
+  SRCS
+    cospif_test.cpp
+  HDRS
+    sdcomp26094.h
+  DEPENDS
+    libc.src.errno.errno
+    libc.src.math.cospif
+    libc.src.__support.CPP.array
+    libc.src.__support.FPUtil.fp_bits
+)
+
 add_fp_unittest(
   sinf_test
   NEED_MPFR

diff --git a/libc/test/src/math/cospif_test.cpp b/libc/test/src/math/cospif_test.cpp
@@ -0,0 +1,120 @@
+//===-- Unittests for cospif ----------------------------------------------===//
+//
+// 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 "src/errno/libc_errno.h"
+#include "src/math/cospif.h"
+#include "test/UnitTest/FPMatcher.h"
+#include "test/src/math/sdcomp26094.h"
+#include "utils/MPFRWrapper/MPFRUtils.h"
+
+using LlvmLibcCospifTest = LIBC_NAMESPACE::testing::FPTest<float>;
+
+using LIBC_NAMESPACE::testing::SDCOMP26094_VALUES;
+
+namespace mpfr = LIBC_NAMESPACE::testing::mpfr;
+
+TEST_F(LlvmLibcCospifTest, SpecialNumbers) {
+  LIBC_NAMESPACE::libc_errno = 0;
+
+  EXPECT_FP_EQ(aNaN, LIBC_NAMESPACE::cospif(aNaN));
+  EXPECT_MATH_ERRNO(0);
+
+  EXPECT_FP_EQ(1.0f, LIBC_NAMESPACE::cospif(0.0f));
+  EXPECT_MATH_ERRNO(0);
+
+  EXPECT_FP_EQ(1.0f, LIBC_NAMESPACE::cospif(-0.0f));
+  EXPECT_MATH_ERRNO(0);
+
+  EXPECT_FP_EQ(aNaN, LIBC_NAMESPACE::cospif(inf));
+  EXPECT_MATH_ERRNO(EDOM);
+
+  EXPECT_FP_EQ(aNaN, LIBC_NAMESPACE::cospif(neg_inf));
+  EXPECT_MATH_ERRNO(EDOM);
+}
+
+TEST_F(LlvmLibcCospifTest, SpecificBitPatterns) {
+  constexpr int N = 36;
+  constexpr uint32_t INPUTS[N] = {
+      0x3f06'0a92U, // x = pi/6
+      0x3f3a'dc51U, // x = 0x1.75b8a2p-1f
+      0x3f49'0fdbU, // x = pi/4
+      0x3f86'0a92U, // x = pi/3
+      0x3fa7'832aU, // x = 0x1.4f0654p+0f
+      0x3fc9'0fdbU, // x = pi/2
+      0x4017'1973U, // x = 0x1.2e32e6p+1f
+      0x4049'0fdbU, // x = pi
+      0x4096'cbe4U, // x = 0x1.2d97c8p+2f
+      0x40c9'0fdbU, // x = 2*pi
+      0x433b'7490U, // x = 0x1.76e92p+7f
+      0x437c'e5f1U, // x = 0x1.f9cbe2p+7f
+      0x4619'9998U, // x = 0x1.33333p+13f
+      0x474d'246fU, // x = 0x1.9a48dep+15f
+      0x4afd'ece4U, // x = 0x1.fbd9c8p+22f
+      0x4c23'32e9U, // x = 0x1.4665d2p+25f
+      0x50a3'e87fU, // x = 0x1.47d0fep+34f
+      0x5239'47f6U, // x = 0x1.728fecp+37f
+      0x53b1'46a6U, // x = 0x1.628d4cp+40f
+      0x55ca'fb2aU, // x = 0x1.95f654p+44f
+      0x588e'f060U, // x = 0x1.1de0cp+50f
+      0x5c07'bcd0U, // x = 0x1.0f79ap+57f
+      0x5ebc'fddeU, // x = 0x1.79fbbcp+62f
+      0x5fa6'eba7U, // x = 0x1.4dd74ep+64f
+      0x61a4'0b40U, // x = 0x1.48168p+68f
+      0x6386'134eU, // x = 0x1.0c269cp+72f
+      0x6589'8498U, // x = 0x1.13093p+76f
+      0x6600'0001U, // x = 0x1.000002p+77f
+      0x664e'46e4U, // x = 0x1.9c8dc8p+77f
+      0x66b0'14aaU, // x = 0x1.602954p+78f
+      0x67a9'242bU, // x = 0x1.524856p+80f
+      0x6a19'76f1U, // x = 0x1.32ede2p+85f
+      0x6c55'da58U, // x = 0x1.abb4bp+89f
+      0x6f79'be45U, // x = 0x1.f37c8ap+95f
+      0x7276'69d4U, // x = 0x1.ecd3a8p+101f
+      0x7758'4625U, // x = 0x1.b08c4ap+111f
+  };
+
+  for (int i = 0; i < N; ++i) {
+    float x = FPBits(INPUTS[i]).get_val();
+    EXPECT_MPFR_MATCH_ALL_ROUNDING(mpfr::Operation::Cospi, x,
+                                   LIBC_NAMESPACE::cospif(x), 0.5);
+    EXPECT_MPFR_MATCH_ALL_ROUNDING(mpfr::Operation::Cospi, -x,
+                                   LIBC_NAMESPACE::cospif(-x), 0.5);
+  }
+}
+
+// For small values, sinpi(x) is pi * x.
+TEST_F(LlvmLibcCospifTest, SmallValues) {
+  float x = FPBits(0x1780'0000U).get_val();
+  EXPECT_MPFR_MATCH_ALL_ROUNDING(mpfr::Operation::Cospi, x,
+                                 LIBC_NAMESPACE::cospif(x), 0.5);
+
+  x = FPBits(0x0040'0000U).get_val();
+  EXPECT_MPFR_MATCH_ALL_ROUNDING(mpfr::Operation::Cospi, x,
+                                 LIBC_NAMESPACE::cospif(x), 0.5);
+}
+
+// SDCOMP-26094: check sinfpi in the cases for which the range reducer
+// returns values furthest beyond its nominal upper bound of pi/4.
+TEST_F(LlvmLibcCospifTest, SDCOMP_26094) {
+  for (uint32_t v : SDCOMP26094_VALUES) {
+    float x = FPBits((v)).get_val();
+    EXPECT_MPFR_MATCH_ALL_ROUNDING(mpfr::Operation::Cospi, x,
+                                   LIBC_NAMESPACE::cospif(x), 0.5);
+  }
+}
+
+// sinpi(-n) = -0.0
+// sinpi(+n) = +0.0
+TEST_F(LlvmLibcCospifTest, SignedZeros) {
+  EXPECT_FP_EQ(0.0, LIBC_NAMESPACE::cospif(100.5f));
+  EXPECT_FP_EQ(-0.0, LIBC_NAMESPACE::cospif(-100.5f));
+  EXPECT_FP_EQ(0.0, LIBC_NAMESPACE::cospif(45678.5f));
+  EXPECT_FP_EQ(-0.0, LIBC_NAMESPACE::cospif(-45678.5f));
+  EXPECT_FP_EQ(0.0, LIBC_NAMESPACE::cospif(8000000.5f));
+  EXPECT_FP_EQ(-0.0, LIBC_NAMESPACE::cospif(-8000000.5f));
+}
diff --git a/libc/test/src/math/exhaustive/CMakeLists.txt b/libc/test/src/math/exhaustive/CMakeLists.txt
@@ -74,6 +74,22 @@ add_fp_unittest(
     -lpthread
 )
 
+add_fp_unittest(
+  cospif_test
+  NO_RUN_POSTBUILD
+  NEED_MPFR
+  SUITE
+    libc_math_exhaustive_tests
+  SRCS
+    cospif_test.cpp
+  DEPENDS
+    .exhaustive_test
+    libc.src.math.cospif
+    libc.src.__support.FPUtil.fp_bits
+  LINK_LIBRARIES
+    -lpthread
+)
+
 add_fp_unittest(
   sincosf_test
   NO_RUN_POSTBUILD