From e2f21465fea76a80966f12a20d0be36597f19b44 Mon Sep 17 00:00:00 2001 From: Antonio Sanchez Date: Wed, 2 Dec 2020 14:00:57 -0800 Subject: Special function implementations for half/bfloat16 packets. Current implementations fail to consider half-float packets, only half-float scalars. Added specializations for packets on AVX, AVX512 and NEON. Added tests to `special_packetmath`. The current `special_functions` tests would fail for half and bfloat16 due to lack of precision. The NEON tests also fail with precision issues and due to different handling of `sqrt(inf)`, so special functions bessel, ndtri have been disabled. Tested with AVX, AVX512. --- unsupported/test/special_packetmath.cpp | 57 +++++++++++++++++++-------------- 1 file changed, 33 insertions(+), 24 deletions(-) (limited to 'unsupported/test/special_packetmath.cpp') diff --git a/unsupported/test/special_packetmath.cpp b/unsupported/test/special_packetmath.cpp index 87b87351b..4f426eb8a 100644 --- a/unsupported/test/special_packetmath.cpp +++ b/unsupported/test/special_packetmath.cpp @@ -8,6 +8,7 @@ // Public License v. 2.0. If a copy of the MPL was not distributed // with this file, You can obtain one at http://mozilla.org/MPL/2.0/. +#include #include "packetmath_test_shared.h" #include "../Eigen/SpecialFunctions" @@ -43,42 +44,48 @@ template void packetmath_real() } { for (int i=0; i(0,1); + data1[i] = internal::random(Scalar(0),Scalar(1)); } CHECK_CWISE1_IF(internal::packet_traits::HasNdtri, numext::ndtri, internal::pndtri); } #endif // EIGEN_HAS_C99_MATH // For bessel_i*e and bessel_j*, the valid range is negative reals. - for (int i=0; i(-1,1) * std::pow(Scalar(10), internal::random(-6,6)); - data2[i] = internal::random(-1,1) * std::pow(Scalar(10), internal::random(-6,6)); - } + const int max_exponent = numext::mini(std::numeric_limits::max_exponent10-1, 6); + for (int i=0; i(Scalar(-1),Scalar(1)) * Scalar(std::pow(Scalar(10), internal::random(Scalar(-max_exponent),Scalar(max_exponent)))); + data2[i] = internal::random(Scalar(-1),Scalar(1)) * Scalar(std::pow(Scalar(10), internal::random(Scalar(-max_exponent),Scalar(max_exponent)))); + } - CHECK_CWISE1_IF(PacketTraits::HasBessel, numext::bessel_i0e, internal::pbessel_i0e); - CHECK_CWISE1_IF(PacketTraits::HasBessel, numext::bessel_i1e, internal::pbessel_i1e); - CHECK_CWISE1_IF(PacketTraits::HasBessel, numext::bessel_j0, internal::pbessel_j0); - CHECK_CWISE1_IF(PacketTraits::HasBessel, numext::bessel_j1, internal::pbessel_j1); + CHECK_CWISE1_IF(PacketTraits::HasBessel, numext::bessel_i0e, internal::pbessel_i0e); + CHECK_CWISE1_IF(PacketTraits::HasBessel, numext::bessel_i1e, internal::pbessel_i1e); + CHECK_CWISE1_IF(PacketTraits::HasBessel, numext::bessel_j0, internal::pbessel_j0); + CHECK_CWISE1_IF(PacketTraits::HasBessel, numext::bessel_j1, internal::pbessel_j1); + } // Use a smaller data range for the bessel_i* as these can become very large. // Following #1693, we also restrict this range further to avoid inf's due to // differences in pexp and exp. for (int i=0; i(0.01,1) * std::pow( - Scalar(9), internal::random(-1,2)); - data2[i] = internal::random(0.01,1) * std::pow( - Scalar(9), internal::random(-1,2)); + data1[i] = internal::random(Scalar(0.01),Scalar(1)) * + Scalar(std::pow(Scalar(9), internal::random(Scalar(-1),Scalar(2)))); + data2[i] = internal::random(Scalar(0.01),Scalar(1)) * + Scalar(std::pow(Scalar(9), internal::random(Scalar(-1),Scalar(2)))); } CHECK_CWISE1_IF(PacketTraits::HasBessel, numext::bessel_i0, internal::pbessel_i0); CHECK_CWISE1_IF(PacketTraits::HasBessel, numext::bessel_i1, internal::pbessel_i1); // y_i, and k_i are valid for x > 0. - for (int i=0; i(0.01,1) * std::pow(Scalar(10), internal::random(-2,5)); - data2[i] = internal::random(0.01,1) * std::pow(Scalar(10), internal::random(-2,5)); + const int max_exponent = numext::mini(std::numeric_limits::max_exponent10-1, 5); + for (int i=0; i(Scalar(0.01),Scalar(1)) * Scalar(std::pow(Scalar(10), internal::random(Scalar(-2),Scalar(max_exponent)))); + data2[i] = internal::random(Scalar(0.01),Scalar(1)) * Scalar(std::pow(Scalar(10), internal::random(Scalar(-2),Scalar(max_exponent)))); + } } // TODO(srvasude): Re-enable this test once properly investigated why the @@ -91,20 +98,20 @@ template void packetmath_real() // Following #1693, we restrict the range for exp to avoid zeroing out too // fast. for (int i=0; i(0.01,1) * std::pow( - Scalar(9), internal::random(-1,2)); - data2[i] = internal::random(0.01,1) * std::pow( - Scalar(9), internal::random(-1,2)); + data1[i] = internal::random(Scalar(0.01),Scalar(1)) * + Scalar(std::pow(Scalar(9), internal::random(Scalar(-1),Scalar(2)))); + data2[i] = internal::random(Scalar(0.01),Scalar(1)) * + Scalar(std::pow(Scalar(9), internal::random(Scalar(-1),Scalar(2)))); } CHECK_CWISE1_IF(PacketTraits::HasBessel, numext::bessel_k0, internal::pbessel_k0); CHECK_CWISE1_IF(PacketTraits::HasBessel, numext::bessel_k1, internal::pbessel_k1); for (int i=0; i(0.01,1) * std::pow( - Scalar(10), internal::random(-1,2)); - data2[i] = internal::random(0.01,1) * std::pow( - Scalar(10), internal::random(-1,2)); + data1[i] = internal::random(Scalar(0.01),Scalar(1)) * + Scalar(std::pow(Scalar(10), internal::random(Scalar(-1),Scalar(2)))); + data2[i] = internal::random(Scalar(0.01),Scalar(1)) * + Scalar(std::pow(Scalar(10), internal::random(Scalar(-1),Scalar(2)))); } #if EIGEN_HAS_C99_MATH && (__cplusplus > 199711L) @@ -135,6 +142,8 @@ EIGEN_DECLARE_TEST(special_packetmath) CALL_SUBTEST_1( test::runner::run() ); CALL_SUBTEST_2( test::runner::run() ); + CALL_SUBTEST_3( test::runner::run() ); + CALL_SUBTEST_4( test::runner::run() ); g_first_pass = false; } } -- cgit v1.2.3