eigen/test/packetmath.cpp

797 lines
30 KiB
C++
Raw Normal View History

// This file is part of Eigen, a lightweight C++ template library
// for linear algebra.
//
2010-06-25 05:21:58 +08:00
// Copyright (C) 2008-2009 Gael Guennebaud <gael.guennebaud@inria.fr>
2008-11-24 21:40:43 +08:00
// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
//
// This Source Code Form is subject to the terms of the Mozilla
// 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 "packetmath_test_shared.h"
Add missing packet ops for bool, and make it pass the same packet op unit tests as other arithmetic types. This change also contains a few minor cleanups: 1. Remove packet op pnot, which is not needed for anything other than pcmp_le_or_nan, which can be done in other ways. 2. Remove the "HasInsert" enum, which is no longer needed since we removed the corresponding packet ops. 3. Add faster pselect op for Packet4i when SSE4.1 is supported. Among other things, this makes the fast transposeInPlace() method available for Matrix<bool>. Run on ************** (72 X 2994 MHz CPUs); 2020-05-09T10:51:02.372347913-07:00 CPU: Intel Skylake Xeon with HyperThreading (36 cores) dL1:32KB dL2:1024KB dL3:24MB Benchmark Time(ns) CPU(ns) Iterations ----------------------------------------------------------------------- BM_TransposeInPlace<float>/4 9.77 9.77 71670320 BM_TransposeInPlace<float>/8 21.9 21.9 31929525 BM_TransposeInPlace<float>/16 66.6 66.6 10000000 BM_TransposeInPlace<float>/32 243 243 2879561 BM_TransposeInPlace<float>/59 844 844 829767 BM_TransposeInPlace<float>/64 933 933 750567 BM_TransposeInPlace<float>/128 3944 3945 177405 BM_TransposeInPlace<float>/256 16853 16853 41457 BM_TransposeInPlace<float>/512 204952 204968 3448 BM_TransposeInPlace<float>/1k 1053889 1053861 664 BM_TransposeInPlace<bool>/4 14.4 14.4 48637301 BM_TransposeInPlace<bool>/8 36.0 36.0 19370222 BM_TransposeInPlace<bool>/16 31.5 31.5 22178902 BM_TransposeInPlace<bool>/32 111 111 6272048 BM_TransposeInPlace<bool>/59 626 626 1000000 BM_TransposeInPlace<bool>/64 428 428 1632689 BM_TransposeInPlace<bool>/128 1677 1677 417377 BM_TransposeInPlace<bool>/256 7126 7126 96264 BM_TransposeInPlace<bool>/512 29021 29024 24165 BM_TransposeInPlace<bool>/1k 116321 116330 6068
2020-05-12 04:23:31 +08:00
template <typename T>
inline T REF_ADD(const T& a, const T& b) { return a + b;}
template <typename T>
inline T REF_SUB(const T& a, const T& b) { return a - b;}
template <typename T>
inline T REF_MUL(const T& a, const T& b) { return a * b;}
template <typename T>
inline T REF_DIV(const T& a, const T& b) { return a / b;}
template <typename T>
inline T REF_ABS_DIFF(const T& a, const T& b) { return a>b ? a - b : b-a;}
// Specializations for bool.
Add missing packet ops for bool, and make it pass the same packet op unit tests as other arithmetic types. This change also contains a few minor cleanups: 1. Remove packet op pnot, which is not needed for anything other than pcmp_le_or_nan, which can be done in other ways. 2. Remove the "HasInsert" enum, which is no longer needed since we removed the corresponding packet ops. 3. Add faster pselect op for Packet4i when SSE4.1 is supported. Among other things, this makes the fast transposeInPlace() method available for Matrix<bool>. Run on ************** (72 X 2994 MHz CPUs); 2020-05-09T10:51:02.372347913-07:00 CPU: Intel Skylake Xeon with HyperThreading (36 cores) dL1:32KB dL2:1024KB dL3:24MB Benchmark Time(ns) CPU(ns) Iterations ----------------------------------------------------------------------- BM_TransposeInPlace<float>/4 9.77 9.77 71670320 BM_TransposeInPlace<float>/8 21.9 21.9 31929525 BM_TransposeInPlace<float>/16 66.6 66.6 10000000 BM_TransposeInPlace<float>/32 243 243 2879561 BM_TransposeInPlace<float>/59 844 844 829767 BM_TransposeInPlace<float>/64 933 933 750567 BM_TransposeInPlace<float>/128 3944 3945 177405 BM_TransposeInPlace<float>/256 16853 16853 41457 BM_TransposeInPlace<float>/512 204952 204968 3448 BM_TransposeInPlace<float>/1k 1053889 1053861 664 BM_TransposeInPlace<bool>/4 14.4 14.4 48637301 BM_TransposeInPlace<bool>/8 36.0 36.0 19370222 BM_TransposeInPlace<bool>/16 31.5 31.5 22178902 BM_TransposeInPlace<bool>/32 111 111 6272048 BM_TransposeInPlace<bool>/59 626 626 1000000 BM_TransposeInPlace<bool>/64 428 428 1632689 BM_TransposeInPlace<bool>/128 1677 1677 417377 BM_TransposeInPlace<bool>/256 7126 7126 96264 BM_TransposeInPlace<bool>/512 29021 29024 24165 BM_TransposeInPlace<bool>/1k 116321 116330 6068
2020-05-12 04:23:31 +08:00
template <>
inline bool REF_ADD(const bool& a, const bool& b) { return a || b;}
template <>
inline bool REF_SUB(const bool& a, const bool& b) { return a ^ b;}
template <>
inline bool REF_MUL(const bool& a, const bool& b) { return a && b;}
2020-03-27 04:18:19 +08:00
template<typename FromScalar, typename FromPacket, typename ToScalar, typename ToPacket, bool CanCast = false>
struct test_cast_helper;
template<typename FromScalar, typename FromPacket, typename ToScalar, typename ToPacket>
struct test_cast_helper<FromScalar, FromPacket, ToScalar, ToPacket, false> {
static void run() {}
};
template<typename FromScalar, typename FromPacket, typename ToScalar, typename ToPacket>
struct test_cast_helper<FromScalar, FromPacket, ToScalar, ToPacket, true> {
static void run() {
static const int PacketSize = internal::unpacket_traits<FromPacket>::size;
EIGEN_ALIGN_MAX FromScalar data1[PacketSize];
EIGEN_ALIGN_MAX ToScalar data2[PacketSize];
EIGEN_ALIGN_MAX ToScalar ref[PacketSize];
// Construct a packet of scalars that will not overflow when casting
for (int i=0; i<PacketSize; ++i) {
const FromScalar from_scalar = Array<FromScalar,1,1>::Random().value();
const ToScalar to_scalar = Array<ToScalar,1,1>::Random().value();
const FromScalar c = sizeof(ToScalar) > sizeof(FromScalar) ? static_cast<FromScalar>(to_scalar) : from_scalar;
data1[i] = (NumTraits<FromScalar>::IsSigned && !NumTraits<ToScalar>::IsSigned) ? numext::abs(c) : c;
}
for (int i=0; i<PacketSize; ++i)
ref[i] = static_cast<const ToScalar>(data1[i]);
internal::pstore(data2, internal::pcast<FromPacket, ToPacket>(internal::pload<FromPacket>(data1)));
VERIFY(test::areApprox(ref, data2, PacketSize) && "internal::pcast<>");
2020-03-27 04:18:19 +08:00
}
};
template<typename FromPacket, typename ToScalar>
void test_cast() {
2020-03-28 01:05:39 +08:00
typedef typename internal::unpacket_traits<FromPacket>::type FromScalar;
typedef typename internal::packet_traits<FromScalar> FromPacketTraits;
2020-03-27 04:18:19 +08:00
typedef typename internal::packet_traits<ToScalar>::type Full;
typedef typename internal::unpacket_traits<Full>::half Half;
typedef typename internal::unpacket_traits<typename internal::unpacket_traits<Full>::half>::half Quarter;
static const int PacketSize = internal::unpacket_traits<FromPacket>::size;
static const bool CanCast =
2020-03-28 01:05:39 +08:00
FromPacketTraits::HasCast &&
(PacketSize == internal::unpacket_traits<Full>::size ||
2020-03-27 04:18:19 +08:00
PacketSize == internal::unpacket_traits<Half>::size ||
2020-03-28 01:05:39 +08:00
PacketSize == internal::unpacket_traits<Quarter>::size);
2020-03-27 04:18:19 +08:00
typedef typename internal::conditional<internal::unpacket_traits<Quarter>::size == PacketSize, Quarter,
typename internal::conditional<internal::unpacket_traits<Half>::size == PacketSize, Half, Full>::type>::type
ToPacket;
test_cast_helper<FromScalar, FromPacket, ToScalar, ToPacket, CanCast>::run();
}
Add missing packet ops for bool, and make it pass the same packet op unit tests as other arithmetic types. This change also contains a few minor cleanups: 1. Remove packet op pnot, which is not needed for anything other than pcmp_le_or_nan, which can be done in other ways. 2. Remove the "HasInsert" enum, which is no longer needed since we removed the corresponding packet ops. 3. Add faster pselect op for Packet4i when SSE4.1 is supported. Among other things, this makes the fast transposeInPlace() method available for Matrix<bool>. Run on ************** (72 X 2994 MHz CPUs); 2020-05-09T10:51:02.372347913-07:00 CPU: Intel Skylake Xeon with HyperThreading (36 cores) dL1:32KB dL2:1024KB dL3:24MB Benchmark Time(ns) CPU(ns) Iterations ----------------------------------------------------------------------- BM_TransposeInPlace<float>/4 9.77 9.77 71670320 BM_TransposeInPlace<float>/8 21.9 21.9 31929525 BM_TransposeInPlace<float>/16 66.6 66.6 10000000 BM_TransposeInPlace<float>/32 243 243 2879561 BM_TransposeInPlace<float>/59 844 844 829767 BM_TransposeInPlace<float>/64 933 933 750567 BM_TransposeInPlace<float>/128 3944 3945 177405 BM_TransposeInPlace<float>/256 16853 16853 41457 BM_TransposeInPlace<float>/512 204952 204968 3448 BM_TransposeInPlace<float>/1k 1053889 1053861 664 BM_TransposeInPlace<bool>/4 14.4 14.4 48637301 BM_TransposeInPlace<bool>/8 36.0 36.0 19370222 BM_TransposeInPlace<bool>/16 31.5 31.5 22178902 BM_TransposeInPlace<bool>/32 111 111 6272048 BM_TransposeInPlace<bool>/59 626 626 1000000 BM_TransposeInPlace<bool>/64 428 428 1632689 BM_TransposeInPlace<bool>/128 1677 1677 417377 BM_TransposeInPlace<bool>/256 7126 7126 96264 BM_TransposeInPlace<bool>/512 29021 29024 24165 BM_TransposeInPlace<bool>/1k 116321 116330 6068
2020-05-12 04:23:31 +08:00
template<typename Scalar,typename Packet>
void packetmath_boolean_mask_ops()
{
const int PacketSize = internal::unpacket_traits<Packet>::size;
const int size = 2*PacketSize;
EIGEN_ALIGN_MAX Scalar data1[size];
EIGEN_ALIGN_MAX Scalar data2[size];
EIGEN_ALIGN_MAX Scalar ref[size];
for (int i=0; i<size; ++i)
{
data1[i] = internal::random<Scalar>();
}
Add missing packet ops for bool, and make it pass the same packet op unit tests as other arithmetic types. This change also contains a few minor cleanups: 1. Remove packet op pnot, which is not needed for anything other than pcmp_le_or_nan, which can be done in other ways. 2. Remove the "HasInsert" enum, which is no longer needed since we removed the corresponding packet ops. 3. Add faster pselect op for Packet4i when SSE4.1 is supported. Among other things, this makes the fast transposeInPlace() method available for Matrix<bool>. Run on ************** (72 X 2994 MHz CPUs); 2020-05-09T10:51:02.372347913-07:00 CPU: Intel Skylake Xeon with HyperThreading (36 cores) dL1:32KB dL2:1024KB dL3:24MB Benchmark Time(ns) CPU(ns) Iterations ----------------------------------------------------------------------- BM_TransposeInPlace<float>/4 9.77 9.77 71670320 BM_TransposeInPlace<float>/8 21.9 21.9 31929525 BM_TransposeInPlace<float>/16 66.6 66.6 10000000 BM_TransposeInPlace<float>/32 243 243 2879561 BM_TransposeInPlace<float>/59 844 844 829767 BM_TransposeInPlace<float>/64 933 933 750567 BM_TransposeInPlace<float>/128 3944 3945 177405 BM_TransposeInPlace<float>/256 16853 16853 41457 BM_TransposeInPlace<float>/512 204952 204968 3448 BM_TransposeInPlace<float>/1k 1053889 1053861 664 BM_TransposeInPlace<bool>/4 14.4 14.4 48637301 BM_TransposeInPlace<bool>/8 36.0 36.0 19370222 BM_TransposeInPlace<bool>/16 31.5 31.5 22178902 BM_TransposeInPlace<bool>/32 111 111 6272048 BM_TransposeInPlace<bool>/59 626 626 1000000 BM_TransposeInPlace<bool>/64 428 428 1632689 BM_TransposeInPlace<bool>/128 1677 1677 417377 BM_TransposeInPlace<bool>/256 7126 7126 96264 BM_TransposeInPlace<bool>/512 29021 29024 24165 BM_TransposeInPlace<bool>/1k 116321 116330 6068
2020-05-12 04:23:31 +08:00
CHECK_CWISE1(internal::ptrue, internal::ptrue);
CHECK_CWISE2_IF(true, internal::pandnot, internal::pandnot);
for (int i = 0; i < PacketSize; ++i) {
data1[i] = Scalar(i);
data1[i + PacketSize] = internal::random<bool>() ? data1[i] : Scalar(0);
}
CHECK_CWISE2_IF(true, internal::pcmp_eq, internal::pcmp_eq);
}
// Packet16b representing bool does not support ptrue, pandnot or pcmp_eq, since the scalar path
// (for some compilers) compute the bitwise and with 0x1 of the results to keep the value in [0,1].
#ifdef EIGEN_PACKET_MATH_SSE_H
Add missing packet ops for bool, and make it pass the same packet op unit tests as other arithmetic types. This change also contains a few minor cleanups: 1. Remove packet op pnot, which is not needed for anything other than pcmp_le_or_nan, which can be done in other ways. 2. Remove the "HasInsert" enum, which is no longer needed since we removed the corresponding packet ops. 3. Add faster pselect op for Packet4i when SSE4.1 is supported. Among other things, this makes the fast transposeInPlace() method available for Matrix<bool>. Run on ************** (72 X 2994 MHz CPUs); 2020-05-09T10:51:02.372347913-07:00 CPU: Intel Skylake Xeon with HyperThreading (36 cores) dL1:32KB dL2:1024KB dL3:24MB Benchmark Time(ns) CPU(ns) Iterations ----------------------------------------------------------------------- BM_TransposeInPlace<float>/4 9.77 9.77 71670320 BM_TransposeInPlace<float>/8 21.9 21.9 31929525 BM_TransposeInPlace<float>/16 66.6 66.6 10000000 BM_TransposeInPlace<float>/32 243 243 2879561 BM_TransposeInPlace<float>/59 844 844 829767 BM_TransposeInPlace<float>/64 933 933 750567 BM_TransposeInPlace<float>/128 3944 3945 177405 BM_TransposeInPlace<float>/256 16853 16853 41457 BM_TransposeInPlace<float>/512 204952 204968 3448 BM_TransposeInPlace<float>/1k 1053889 1053861 664 BM_TransposeInPlace<bool>/4 14.4 14.4 48637301 BM_TransposeInPlace<bool>/8 36.0 36.0 19370222 BM_TransposeInPlace<bool>/16 31.5 31.5 22178902 BM_TransposeInPlace<bool>/32 111 111 6272048 BM_TransposeInPlace<bool>/59 626 626 1000000 BM_TransposeInPlace<bool>/64 428 428 1632689 BM_TransposeInPlace<bool>/128 1677 1677 417377 BM_TransposeInPlace<bool>/256 7126 7126 96264 BM_TransposeInPlace<bool>/512 29021 29024 24165 BM_TransposeInPlace<bool>/1k 116321 116330 6068
2020-05-12 04:23:31 +08:00
template<>
void packetmath_boolean_mask_ops<bool, internal::Packet16b>()
{
}
#endif
template<typename Scalar,typename Packet> void packetmath()
{
typedef internal::packet_traits<Scalar> PacketTraits;
const int PacketSize = internal::unpacket_traits<Packet>::size;
typedef typename NumTraits<Scalar>::Real RealScalar;
if (g_first_pass)
std::cerr << "=== Testing packet of type '" << typeid(Packet).name()
<< "' and scalar type '" << typeid(Scalar).name()
<< "' and size '" << PacketSize << "' ===\n" ;
2014-01-30 03:43:05 +08:00
const int max_size = PacketSize > 4 ? PacketSize : 4;
const int size = PacketSize*max_size;
EIGEN_ALIGN_MAX Scalar data1[size];
EIGEN_ALIGN_MAX Scalar data2[size];
Adding lowlevel APIs for optimized RHS packet load in TensorFlow SpatialConvolution Low-level APIs are added in order to optimized packet load in gemm_pack_rhs in TensorFlow SpatialConvolution. The optimization is for scenario when a packet is split across 2 adjacent columns. In this case we read it as two 'partial' packets and then merge these into 1. Currently this only works for Packet16f (AVX512) and Packet8f (AVX2). We plan to add this for other packet types (such as Packet8d) also. This optimization shows significant speedup in SpatialConvolution with certain parameters. Some examples are below. Benchmark parameters are specified as: Batch size, Input dim, Depth, Num of filters, Filter dim Speedup numbers are specified for number of threads 1, 2, 4, 8, 16. AVX512: Parameters | Speedup (Num of threads: 1, 2, 4, 8, 16) ----------------------------|------------------------------------------ 128, 24x24, 3, 64, 5x5 |2.18X, 2.13X, 1.73X, 1.64X, 1.66X 128, 24x24, 1, 64, 8x8 |2.00X, 1.98X, 1.93X, 1.91X, 1.91X 32, 24x24, 3, 64, 5x5 |2.26X, 2.14X, 2.17X, 2.22X, 2.33X 128, 24x24, 3, 64, 3x3 |1.51X, 1.45X, 1.45X, 1.67X, 1.57X 32, 14x14, 24, 64, 5x5 |1.21X, 1.19X, 1.16X, 1.70X, 1.17X 128, 128x128, 3, 96, 11x11 |2.17X, 2.18X, 2.19X, 2.20X, 2.18X AVX2: Parameters | Speedup (Num of threads: 1, 2, 4, 8, 16) ----------------------------|------------------------------------------ 128, 24x24, 3, 64, 5x5 | 1.66X, 1.65X, 1.61X, 1.56X, 1.49X 32, 24x24, 3, 64, 5x5 | 1.71X, 1.63X, 1.77X, 1.58X, 1.68X 128, 24x24, 1, 64, 5x5 | 1.44X, 1.40X, 1.38X, 1.37X, 1.33X 128, 24x24, 3, 64, 3x3 | 1.68X, 1.63X, 1.58X, 1.56X, 1.62X 128, 128x128, 3, 96, 11x11 | 1.36X, 1.36X, 1.37X, 1.37X, 1.37X In the higher level benchmark cifar10, we observe a runtime improvement of around 6% for AVX512 on Intel Skylake server (8 cores). On lower level PackRhs micro-benchmarks specified in TensorFlow tensorflow/core/kernels/eigen_spatial_convolutions_test.cc, we observe the following runtime numbers: AVX512: Parameters | Runtime without patch (ns) | Runtime with patch (ns) | Speedup ---------------------------------------------------------------|----------------------------|-------------------------|--------- BM_RHS_NAME(PackRhs, 128, 24, 24, 3, 64, 5, 5, 1, 1, 256, 56) | 41350 | 15073 | 2.74X BM_RHS_NAME(PackRhs, 32, 64, 64, 32, 64, 5, 5, 1, 1, 256, 56) | 7277 | 7341 | 0.99X BM_RHS_NAME(PackRhs, 32, 64, 64, 32, 64, 5, 5, 2, 2, 256, 56) | 8675 | 8681 | 1.00X BM_RHS_NAME(PackRhs, 32, 64, 64, 30, 64, 5, 5, 1, 1, 256, 56) | 24155 | 16079 | 1.50X BM_RHS_NAME(PackRhs, 32, 64, 64, 30, 64, 5, 5, 2, 2, 256, 56) | 25052 | 17152 | 1.46X BM_RHS_NAME(PackRhs, 32, 256, 256, 4, 16, 8, 8, 1, 1, 256, 56) | 18269 | 18345 | 1.00X BM_RHS_NAME(PackRhs, 32, 256, 256, 4, 16, 8, 8, 2, 4, 256, 56) | 19468 | 19872 | 0.98X BM_RHS_NAME(PackRhs, 32, 64, 64, 4, 16, 3, 3, 1, 1, 36, 432) | 156060 | 42432 | 3.68X BM_RHS_NAME(PackRhs, 32, 64, 64, 4, 16, 3, 3, 2, 2, 36, 432) | 132701 | 36944 | 3.59X AVX2: Parameters | Runtime without patch (ns) | Runtime with patch (ns) | Speedup ---------------------------------------------------------------|----------------------------|-------------------------|--------- BM_RHS_NAME(PackRhs, 128, 24, 24, 3, 64, 5, 5, 1, 1, 256, 56) | 26233 | 12393 | 2.12X BM_RHS_NAME(PackRhs, 32, 64, 64, 32, 64, 5, 5, 1, 1, 256, 56) | 6091 | 6062 | 1.00X BM_RHS_NAME(PackRhs, 32, 64, 64, 32, 64, 5, 5, 2, 2, 256, 56) | 7427 | 7408 | 1.00X BM_RHS_NAME(PackRhs, 32, 64, 64, 30, 64, 5, 5, 1, 1, 256, 56) | 23453 | 20826 | 1.13X BM_RHS_NAME(PackRhs, 32, 64, 64, 30, 64, 5, 5, 2, 2, 256, 56) | 23167 | 22091 | 1.09X BM_RHS_NAME(PackRhs, 32, 256, 256, 4, 16, 8, 8, 1, 1, 256, 56) | 23422 | 23682 | 0.99X BM_RHS_NAME(PackRhs, 32, 256, 256, 4, 16, 8, 8, 2, 4, 256, 56) | 23165 | 23663 | 0.98X BM_RHS_NAME(PackRhs, 32, 64, 64, 4, 16, 3, 3, 1, 1, 36, 432) | 72689 | 44969 | 1.62X BM_RHS_NAME(PackRhs, 32, 64, 64, 4, 16, 3, 3, 2, 2, 36, 432) | 61732 | 39779 | 1.55X All benchmarks on Intel Skylake server with 8 cores.
2019-04-20 14:46:43 +08:00
EIGEN_ALIGN_MAX Scalar data3[size];
EIGEN_ALIGN_MAX Scalar ref[size];
RealScalar refvalue = RealScalar(0);
for (int i=0; i<size; ++i)
{
data1[i] = internal::random<Scalar>()/RealScalar(PacketSize);
data2[i] = internal::random<Scalar>()/RealScalar(PacketSize);
refvalue = (std::max)(refvalue, numext::abs(data1[i]));
}
internal::pstore(data2, internal::pload<Packet>(data1));
VERIFY(test::areApprox(data1, data2, PacketSize) && "aligned load/store");
for (int offset=0; offset<PacketSize; ++offset)
{
internal::pstore(data2, internal::ploadu<Packet>(data1+offset));
VERIFY(test::areApprox(data1+offset, data2, PacketSize) && "internal::ploadu");
}
for (int offset=0; offset<PacketSize; ++offset)
{
internal::pstoreu(data2+offset, internal::pload<Packet>(data1));
VERIFY(test::areApprox(data1, data2+offset, PacketSize) && "internal::pstoreu");
}
Adding lowlevel APIs for optimized RHS packet load in TensorFlow SpatialConvolution Low-level APIs are added in order to optimized packet load in gemm_pack_rhs in TensorFlow SpatialConvolution. The optimization is for scenario when a packet is split across 2 adjacent columns. In this case we read it as two 'partial' packets and then merge these into 1. Currently this only works for Packet16f (AVX512) and Packet8f (AVX2). We plan to add this for other packet types (such as Packet8d) also. This optimization shows significant speedup in SpatialConvolution with certain parameters. Some examples are below. Benchmark parameters are specified as: Batch size, Input dim, Depth, Num of filters, Filter dim Speedup numbers are specified for number of threads 1, 2, 4, 8, 16. AVX512: Parameters | Speedup (Num of threads: 1, 2, 4, 8, 16) ----------------------------|------------------------------------------ 128, 24x24, 3, 64, 5x5 |2.18X, 2.13X, 1.73X, 1.64X, 1.66X 128, 24x24, 1, 64, 8x8 |2.00X, 1.98X, 1.93X, 1.91X, 1.91X 32, 24x24, 3, 64, 5x5 |2.26X, 2.14X, 2.17X, 2.22X, 2.33X 128, 24x24, 3, 64, 3x3 |1.51X, 1.45X, 1.45X, 1.67X, 1.57X 32, 14x14, 24, 64, 5x5 |1.21X, 1.19X, 1.16X, 1.70X, 1.17X 128, 128x128, 3, 96, 11x11 |2.17X, 2.18X, 2.19X, 2.20X, 2.18X AVX2: Parameters | Speedup (Num of threads: 1, 2, 4, 8, 16) ----------------------------|------------------------------------------ 128, 24x24, 3, 64, 5x5 | 1.66X, 1.65X, 1.61X, 1.56X, 1.49X 32, 24x24, 3, 64, 5x5 | 1.71X, 1.63X, 1.77X, 1.58X, 1.68X 128, 24x24, 1, 64, 5x5 | 1.44X, 1.40X, 1.38X, 1.37X, 1.33X 128, 24x24, 3, 64, 3x3 | 1.68X, 1.63X, 1.58X, 1.56X, 1.62X 128, 128x128, 3, 96, 11x11 | 1.36X, 1.36X, 1.37X, 1.37X, 1.37X In the higher level benchmark cifar10, we observe a runtime improvement of around 6% for AVX512 on Intel Skylake server (8 cores). On lower level PackRhs micro-benchmarks specified in TensorFlow tensorflow/core/kernels/eigen_spatial_convolutions_test.cc, we observe the following runtime numbers: AVX512: Parameters | Runtime without patch (ns) | Runtime with patch (ns) | Speedup ---------------------------------------------------------------|----------------------------|-------------------------|--------- BM_RHS_NAME(PackRhs, 128, 24, 24, 3, 64, 5, 5, 1, 1, 256, 56) | 41350 | 15073 | 2.74X BM_RHS_NAME(PackRhs, 32, 64, 64, 32, 64, 5, 5, 1, 1, 256, 56) | 7277 | 7341 | 0.99X BM_RHS_NAME(PackRhs, 32, 64, 64, 32, 64, 5, 5, 2, 2, 256, 56) | 8675 | 8681 | 1.00X BM_RHS_NAME(PackRhs, 32, 64, 64, 30, 64, 5, 5, 1, 1, 256, 56) | 24155 | 16079 | 1.50X BM_RHS_NAME(PackRhs, 32, 64, 64, 30, 64, 5, 5, 2, 2, 256, 56) | 25052 | 17152 | 1.46X BM_RHS_NAME(PackRhs, 32, 256, 256, 4, 16, 8, 8, 1, 1, 256, 56) | 18269 | 18345 | 1.00X BM_RHS_NAME(PackRhs, 32, 256, 256, 4, 16, 8, 8, 2, 4, 256, 56) | 19468 | 19872 | 0.98X BM_RHS_NAME(PackRhs, 32, 64, 64, 4, 16, 3, 3, 1, 1, 36, 432) | 156060 | 42432 | 3.68X BM_RHS_NAME(PackRhs, 32, 64, 64, 4, 16, 3, 3, 2, 2, 36, 432) | 132701 | 36944 | 3.59X AVX2: Parameters | Runtime without patch (ns) | Runtime with patch (ns) | Speedup ---------------------------------------------------------------|----------------------------|-------------------------|--------- BM_RHS_NAME(PackRhs, 128, 24, 24, 3, 64, 5, 5, 1, 1, 256, 56) | 26233 | 12393 | 2.12X BM_RHS_NAME(PackRhs, 32, 64, 64, 32, 64, 5, 5, 1, 1, 256, 56) | 6091 | 6062 | 1.00X BM_RHS_NAME(PackRhs, 32, 64, 64, 32, 64, 5, 5, 2, 2, 256, 56) | 7427 | 7408 | 1.00X BM_RHS_NAME(PackRhs, 32, 64, 64, 30, 64, 5, 5, 1, 1, 256, 56) | 23453 | 20826 | 1.13X BM_RHS_NAME(PackRhs, 32, 64, 64, 30, 64, 5, 5, 2, 2, 256, 56) | 23167 | 22091 | 1.09X BM_RHS_NAME(PackRhs, 32, 256, 256, 4, 16, 8, 8, 1, 1, 256, 56) | 23422 | 23682 | 0.99X BM_RHS_NAME(PackRhs, 32, 256, 256, 4, 16, 8, 8, 2, 4, 256, 56) | 23165 | 23663 | 0.98X BM_RHS_NAME(PackRhs, 32, 64, 64, 4, 16, 3, 3, 1, 1, 36, 432) | 72689 | 44969 | 1.62X BM_RHS_NAME(PackRhs, 32, 64, 64, 4, 16, 3, 3, 2, 2, 36, 432) | 61732 | 39779 | 1.55X All benchmarks on Intel Skylake server with 8 cores.
2019-04-20 14:46:43 +08:00
if (internal::unpacket_traits<Packet>::masked_load_available)
{
test::packet_helper<internal::unpacket_traits<Packet>::masked_load_available, Packet> h;
Adding lowlevel APIs for optimized RHS packet load in TensorFlow SpatialConvolution Low-level APIs are added in order to optimized packet load in gemm_pack_rhs in TensorFlow SpatialConvolution. The optimization is for scenario when a packet is split across 2 adjacent columns. In this case we read it as two 'partial' packets and then merge these into 1. Currently this only works for Packet16f (AVX512) and Packet8f (AVX2). We plan to add this for other packet types (such as Packet8d) also. This optimization shows significant speedup in SpatialConvolution with certain parameters. Some examples are below. Benchmark parameters are specified as: Batch size, Input dim, Depth, Num of filters, Filter dim Speedup numbers are specified for number of threads 1, 2, 4, 8, 16. AVX512: Parameters | Speedup (Num of threads: 1, 2, 4, 8, 16) ----------------------------|------------------------------------------ 128, 24x24, 3, 64, 5x5 |2.18X, 2.13X, 1.73X, 1.64X, 1.66X 128, 24x24, 1, 64, 8x8 |2.00X, 1.98X, 1.93X, 1.91X, 1.91X 32, 24x24, 3, 64, 5x5 |2.26X, 2.14X, 2.17X, 2.22X, 2.33X 128, 24x24, 3, 64, 3x3 |1.51X, 1.45X, 1.45X, 1.67X, 1.57X 32, 14x14, 24, 64, 5x5 |1.21X, 1.19X, 1.16X, 1.70X, 1.17X 128, 128x128, 3, 96, 11x11 |2.17X, 2.18X, 2.19X, 2.20X, 2.18X AVX2: Parameters | Speedup (Num of threads: 1, 2, 4, 8, 16) ----------------------------|------------------------------------------ 128, 24x24, 3, 64, 5x5 | 1.66X, 1.65X, 1.61X, 1.56X, 1.49X 32, 24x24, 3, 64, 5x5 | 1.71X, 1.63X, 1.77X, 1.58X, 1.68X 128, 24x24, 1, 64, 5x5 | 1.44X, 1.40X, 1.38X, 1.37X, 1.33X 128, 24x24, 3, 64, 3x3 | 1.68X, 1.63X, 1.58X, 1.56X, 1.62X 128, 128x128, 3, 96, 11x11 | 1.36X, 1.36X, 1.37X, 1.37X, 1.37X In the higher level benchmark cifar10, we observe a runtime improvement of around 6% for AVX512 on Intel Skylake server (8 cores). On lower level PackRhs micro-benchmarks specified in TensorFlow tensorflow/core/kernels/eigen_spatial_convolutions_test.cc, we observe the following runtime numbers: AVX512: Parameters | Runtime without patch (ns) | Runtime with patch (ns) | Speedup ---------------------------------------------------------------|----------------------------|-------------------------|--------- BM_RHS_NAME(PackRhs, 128, 24, 24, 3, 64, 5, 5, 1, 1, 256, 56) | 41350 | 15073 | 2.74X BM_RHS_NAME(PackRhs, 32, 64, 64, 32, 64, 5, 5, 1, 1, 256, 56) | 7277 | 7341 | 0.99X BM_RHS_NAME(PackRhs, 32, 64, 64, 32, 64, 5, 5, 2, 2, 256, 56) | 8675 | 8681 | 1.00X BM_RHS_NAME(PackRhs, 32, 64, 64, 30, 64, 5, 5, 1, 1, 256, 56) | 24155 | 16079 | 1.50X BM_RHS_NAME(PackRhs, 32, 64, 64, 30, 64, 5, 5, 2, 2, 256, 56) | 25052 | 17152 | 1.46X BM_RHS_NAME(PackRhs, 32, 256, 256, 4, 16, 8, 8, 1, 1, 256, 56) | 18269 | 18345 | 1.00X BM_RHS_NAME(PackRhs, 32, 256, 256, 4, 16, 8, 8, 2, 4, 256, 56) | 19468 | 19872 | 0.98X BM_RHS_NAME(PackRhs, 32, 64, 64, 4, 16, 3, 3, 1, 1, 36, 432) | 156060 | 42432 | 3.68X BM_RHS_NAME(PackRhs, 32, 64, 64, 4, 16, 3, 3, 2, 2, 36, 432) | 132701 | 36944 | 3.59X AVX2: Parameters | Runtime without patch (ns) | Runtime with patch (ns) | Speedup ---------------------------------------------------------------|----------------------------|-------------------------|--------- BM_RHS_NAME(PackRhs, 128, 24, 24, 3, 64, 5, 5, 1, 1, 256, 56) | 26233 | 12393 | 2.12X BM_RHS_NAME(PackRhs, 32, 64, 64, 32, 64, 5, 5, 1, 1, 256, 56) | 6091 | 6062 | 1.00X BM_RHS_NAME(PackRhs, 32, 64, 64, 32, 64, 5, 5, 2, 2, 256, 56) | 7427 | 7408 | 1.00X BM_RHS_NAME(PackRhs, 32, 64, 64, 30, 64, 5, 5, 1, 1, 256, 56) | 23453 | 20826 | 1.13X BM_RHS_NAME(PackRhs, 32, 64, 64, 30, 64, 5, 5, 2, 2, 256, 56) | 23167 | 22091 | 1.09X BM_RHS_NAME(PackRhs, 32, 256, 256, 4, 16, 8, 8, 1, 1, 256, 56) | 23422 | 23682 | 0.99X BM_RHS_NAME(PackRhs, 32, 256, 256, 4, 16, 8, 8, 2, 4, 256, 56) | 23165 | 23663 | 0.98X BM_RHS_NAME(PackRhs, 32, 64, 64, 4, 16, 3, 3, 1, 1, 36, 432) | 72689 | 44969 | 1.62X BM_RHS_NAME(PackRhs, 32, 64, 64, 4, 16, 3, 3, 2, 2, 36, 432) | 61732 | 39779 | 1.55X All benchmarks on Intel Skylake server with 8 cores.
2019-04-20 14:46:43 +08:00
unsigned long long max_umask = (0x1ull << PacketSize);
Adding lowlevel APIs for optimized RHS packet load in TensorFlow SpatialConvolution Low-level APIs are added in order to optimized packet load in gemm_pack_rhs in TensorFlow SpatialConvolution. The optimization is for scenario when a packet is split across 2 adjacent columns. In this case we read it as two 'partial' packets and then merge these into 1. Currently this only works for Packet16f (AVX512) and Packet8f (AVX2). We plan to add this for other packet types (such as Packet8d) also. This optimization shows significant speedup in SpatialConvolution with certain parameters. Some examples are below. Benchmark parameters are specified as: Batch size, Input dim, Depth, Num of filters, Filter dim Speedup numbers are specified for number of threads 1, 2, 4, 8, 16. AVX512: Parameters | Speedup (Num of threads: 1, 2, 4, 8, 16) ----------------------------|------------------------------------------ 128, 24x24, 3, 64, 5x5 |2.18X, 2.13X, 1.73X, 1.64X, 1.66X 128, 24x24, 1, 64, 8x8 |2.00X, 1.98X, 1.93X, 1.91X, 1.91X 32, 24x24, 3, 64, 5x5 |2.26X, 2.14X, 2.17X, 2.22X, 2.33X 128, 24x24, 3, 64, 3x3 |1.51X, 1.45X, 1.45X, 1.67X, 1.57X 32, 14x14, 24, 64, 5x5 |1.21X, 1.19X, 1.16X, 1.70X, 1.17X 128, 128x128, 3, 96, 11x11 |2.17X, 2.18X, 2.19X, 2.20X, 2.18X AVX2: Parameters | Speedup (Num of threads: 1, 2, 4, 8, 16) ----------------------------|------------------------------------------ 128, 24x24, 3, 64, 5x5 | 1.66X, 1.65X, 1.61X, 1.56X, 1.49X 32, 24x24, 3, 64, 5x5 | 1.71X, 1.63X, 1.77X, 1.58X, 1.68X 128, 24x24, 1, 64, 5x5 | 1.44X, 1.40X, 1.38X, 1.37X, 1.33X 128, 24x24, 3, 64, 3x3 | 1.68X, 1.63X, 1.58X, 1.56X, 1.62X 128, 128x128, 3, 96, 11x11 | 1.36X, 1.36X, 1.37X, 1.37X, 1.37X In the higher level benchmark cifar10, we observe a runtime improvement of around 6% for AVX512 on Intel Skylake server (8 cores). On lower level PackRhs micro-benchmarks specified in TensorFlow tensorflow/core/kernels/eigen_spatial_convolutions_test.cc, we observe the following runtime numbers: AVX512: Parameters | Runtime without patch (ns) | Runtime with patch (ns) | Speedup ---------------------------------------------------------------|----------------------------|-------------------------|--------- BM_RHS_NAME(PackRhs, 128, 24, 24, 3, 64, 5, 5, 1, 1, 256, 56) | 41350 | 15073 | 2.74X BM_RHS_NAME(PackRhs, 32, 64, 64, 32, 64, 5, 5, 1, 1, 256, 56) | 7277 | 7341 | 0.99X BM_RHS_NAME(PackRhs, 32, 64, 64, 32, 64, 5, 5, 2, 2, 256, 56) | 8675 | 8681 | 1.00X BM_RHS_NAME(PackRhs, 32, 64, 64, 30, 64, 5, 5, 1, 1, 256, 56) | 24155 | 16079 | 1.50X BM_RHS_NAME(PackRhs, 32, 64, 64, 30, 64, 5, 5, 2, 2, 256, 56) | 25052 | 17152 | 1.46X BM_RHS_NAME(PackRhs, 32, 256, 256, 4, 16, 8, 8, 1, 1, 256, 56) | 18269 | 18345 | 1.00X BM_RHS_NAME(PackRhs, 32, 256, 256, 4, 16, 8, 8, 2, 4, 256, 56) | 19468 | 19872 | 0.98X BM_RHS_NAME(PackRhs, 32, 64, 64, 4, 16, 3, 3, 1, 1, 36, 432) | 156060 | 42432 | 3.68X BM_RHS_NAME(PackRhs, 32, 64, 64, 4, 16, 3, 3, 2, 2, 36, 432) | 132701 | 36944 | 3.59X AVX2: Parameters | Runtime without patch (ns) | Runtime with patch (ns) | Speedup ---------------------------------------------------------------|----------------------------|-------------------------|--------- BM_RHS_NAME(PackRhs, 128, 24, 24, 3, 64, 5, 5, 1, 1, 256, 56) | 26233 | 12393 | 2.12X BM_RHS_NAME(PackRhs, 32, 64, 64, 32, 64, 5, 5, 1, 1, 256, 56) | 6091 | 6062 | 1.00X BM_RHS_NAME(PackRhs, 32, 64, 64, 32, 64, 5, 5, 2, 2, 256, 56) | 7427 | 7408 | 1.00X BM_RHS_NAME(PackRhs, 32, 64, 64, 30, 64, 5, 5, 1, 1, 256, 56) | 23453 | 20826 | 1.13X BM_RHS_NAME(PackRhs, 32, 64, 64, 30, 64, 5, 5, 2, 2, 256, 56) | 23167 | 22091 | 1.09X BM_RHS_NAME(PackRhs, 32, 256, 256, 4, 16, 8, 8, 1, 1, 256, 56) | 23422 | 23682 | 0.99X BM_RHS_NAME(PackRhs, 32, 256, 256, 4, 16, 8, 8, 2, 4, 256, 56) | 23165 | 23663 | 0.98X BM_RHS_NAME(PackRhs, 32, 64, 64, 4, 16, 3, 3, 1, 1, 36, 432) | 72689 | 44969 | 1.62X BM_RHS_NAME(PackRhs, 32, 64, 64, 4, 16, 3, 3, 2, 2, 36, 432) | 61732 | 39779 | 1.55X All benchmarks on Intel Skylake server with 8 cores.
2019-04-20 14:46:43 +08:00
for (int offset=0; offset<PacketSize; ++offset)
{
for (unsigned long long umask=0; umask<max_umask; ++umask)
{
h.store(data2, h.load(data1+offset, umask));
for (int k=0; k<PacketSize; ++k)
data3[k] = ((umask & ( 0x1ull << k )) >> k) ? data1[k+offset] : Scalar(0);
VERIFY(test::areApprox(data3, data2, PacketSize) && "internal::ploadu masked");
Adding lowlevel APIs for optimized RHS packet load in TensorFlow SpatialConvolution Low-level APIs are added in order to optimized packet load in gemm_pack_rhs in TensorFlow SpatialConvolution. The optimization is for scenario when a packet is split across 2 adjacent columns. In this case we read it as two 'partial' packets and then merge these into 1. Currently this only works for Packet16f (AVX512) and Packet8f (AVX2). We plan to add this for other packet types (such as Packet8d) also. This optimization shows significant speedup in SpatialConvolution with certain parameters. Some examples are below. Benchmark parameters are specified as: Batch size, Input dim, Depth, Num of filters, Filter dim Speedup numbers are specified for number of threads 1, 2, 4, 8, 16. AVX512: Parameters | Speedup (Num of threads: 1, 2, 4, 8, 16) ----------------------------|------------------------------------------ 128, 24x24, 3, 64, 5x5 |2.18X, 2.13X, 1.73X, 1.64X, 1.66X 128, 24x24, 1, 64, 8x8 |2.00X, 1.98X, 1.93X, 1.91X, 1.91X 32, 24x24, 3, 64, 5x5 |2.26X, 2.14X, 2.17X, 2.22X, 2.33X 128, 24x24, 3, 64, 3x3 |1.51X, 1.45X, 1.45X, 1.67X, 1.57X 32, 14x14, 24, 64, 5x5 |1.21X, 1.19X, 1.16X, 1.70X, 1.17X 128, 128x128, 3, 96, 11x11 |2.17X, 2.18X, 2.19X, 2.20X, 2.18X AVX2: Parameters | Speedup (Num of threads: 1, 2, 4, 8, 16) ----------------------------|------------------------------------------ 128, 24x24, 3, 64, 5x5 | 1.66X, 1.65X, 1.61X, 1.56X, 1.49X 32, 24x24, 3, 64, 5x5 | 1.71X, 1.63X, 1.77X, 1.58X, 1.68X 128, 24x24, 1, 64, 5x5 | 1.44X, 1.40X, 1.38X, 1.37X, 1.33X 128, 24x24, 3, 64, 3x3 | 1.68X, 1.63X, 1.58X, 1.56X, 1.62X 128, 128x128, 3, 96, 11x11 | 1.36X, 1.36X, 1.37X, 1.37X, 1.37X In the higher level benchmark cifar10, we observe a runtime improvement of around 6% for AVX512 on Intel Skylake server (8 cores). On lower level PackRhs micro-benchmarks specified in TensorFlow tensorflow/core/kernels/eigen_spatial_convolutions_test.cc, we observe the following runtime numbers: AVX512: Parameters | Runtime without patch (ns) | Runtime with patch (ns) | Speedup ---------------------------------------------------------------|----------------------------|-------------------------|--------- BM_RHS_NAME(PackRhs, 128, 24, 24, 3, 64, 5, 5, 1, 1, 256, 56) | 41350 | 15073 | 2.74X BM_RHS_NAME(PackRhs, 32, 64, 64, 32, 64, 5, 5, 1, 1, 256, 56) | 7277 | 7341 | 0.99X BM_RHS_NAME(PackRhs, 32, 64, 64, 32, 64, 5, 5, 2, 2, 256, 56) | 8675 | 8681 | 1.00X BM_RHS_NAME(PackRhs, 32, 64, 64, 30, 64, 5, 5, 1, 1, 256, 56) | 24155 | 16079 | 1.50X BM_RHS_NAME(PackRhs, 32, 64, 64, 30, 64, 5, 5, 2, 2, 256, 56) | 25052 | 17152 | 1.46X BM_RHS_NAME(PackRhs, 32, 256, 256, 4, 16, 8, 8, 1, 1, 256, 56) | 18269 | 18345 | 1.00X BM_RHS_NAME(PackRhs, 32, 256, 256, 4, 16, 8, 8, 2, 4, 256, 56) | 19468 | 19872 | 0.98X BM_RHS_NAME(PackRhs, 32, 64, 64, 4, 16, 3, 3, 1, 1, 36, 432) | 156060 | 42432 | 3.68X BM_RHS_NAME(PackRhs, 32, 64, 64, 4, 16, 3, 3, 2, 2, 36, 432) | 132701 | 36944 | 3.59X AVX2: Parameters | Runtime without patch (ns) | Runtime with patch (ns) | Speedup ---------------------------------------------------------------|----------------------------|-------------------------|--------- BM_RHS_NAME(PackRhs, 128, 24, 24, 3, 64, 5, 5, 1, 1, 256, 56) | 26233 | 12393 | 2.12X BM_RHS_NAME(PackRhs, 32, 64, 64, 32, 64, 5, 5, 1, 1, 256, 56) | 6091 | 6062 | 1.00X BM_RHS_NAME(PackRhs, 32, 64, 64, 32, 64, 5, 5, 2, 2, 256, 56) | 7427 | 7408 | 1.00X BM_RHS_NAME(PackRhs, 32, 64, 64, 30, 64, 5, 5, 1, 1, 256, 56) | 23453 | 20826 | 1.13X BM_RHS_NAME(PackRhs, 32, 64, 64, 30, 64, 5, 5, 2, 2, 256, 56) | 23167 | 22091 | 1.09X BM_RHS_NAME(PackRhs, 32, 256, 256, 4, 16, 8, 8, 1, 1, 256, 56) | 23422 | 23682 | 0.99X BM_RHS_NAME(PackRhs, 32, 256, 256, 4, 16, 8, 8, 2, 4, 256, 56) | 23165 | 23663 | 0.98X BM_RHS_NAME(PackRhs, 32, 64, 64, 4, 16, 3, 3, 1, 1, 36, 432) | 72689 | 44969 | 1.62X BM_RHS_NAME(PackRhs, 32, 64, 64, 4, 16, 3, 3, 2, 2, 36, 432) | 61732 | 39779 | 1.55X All benchmarks on Intel Skylake server with 8 cores.
2019-04-20 14:46:43 +08:00
}
}
}
if (internal::unpacket_traits<Packet>::masked_store_available)
{
test::packet_helper<internal::unpacket_traits<Packet>::masked_store_available, Packet> h;
unsigned long long max_umask = (0x1ull << PacketSize);
for (int offset=0; offset<PacketSize; ++offset)
{
for (unsigned long long umask=0; umask<max_umask; ++umask)
{
internal::pstore(data2, internal::pset1<Packet>(Scalar(0)));
h.store(data2, h.loadu(data1+offset), umask);
for (int k=0; k<PacketSize; ++k)
data3[k] = ((umask & ( 0x1ull << k )) >> k) ? data1[k+offset] : Scalar(0);
VERIFY(test::areApprox(data3, data2, PacketSize) && "internal::pstoreu masked");
}
}
Adding lowlevel APIs for optimized RHS packet load in TensorFlow SpatialConvolution Low-level APIs are added in order to optimized packet load in gemm_pack_rhs in TensorFlow SpatialConvolution. The optimization is for scenario when a packet is split across 2 adjacent columns. In this case we read it as two 'partial' packets and then merge these into 1. Currently this only works for Packet16f (AVX512) and Packet8f (AVX2). We plan to add this for other packet types (such as Packet8d) also. This optimization shows significant speedup in SpatialConvolution with certain parameters. Some examples are below. Benchmark parameters are specified as: Batch size, Input dim, Depth, Num of filters, Filter dim Speedup numbers are specified for number of threads 1, 2, 4, 8, 16. AVX512: Parameters | Speedup (Num of threads: 1, 2, 4, 8, 16) ----------------------------|------------------------------------------ 128, 24x24, 3, 64, 5x5 |2.18X, 2.13X, 1.73X, 1.64X, 1.66X 128, 24x24, 1, 64, 8x8 |2.00X, 1.98X, 1.93X, 1.91X, 1.91X 32, 24x24, 3, 64, 5x5 |2.26X, 2.14X, 2.17X, 2.22X, 2.33X 128, 24x24, 3, 64, 3x3 |1.51X, 1.45X, 1.45X, 1.67X, 1.57X 32, 14x14, 24, 64, 5x5 |1.21X, 1.19X, 1.16X, 1.70X, 1.17X 128, 128x128, 3, 96, 11x11 |2.17X, 2.18X, 2.19X, 2.20X, 2.18X AVX2: Parameters | Speedup (Num of threads: 1, 2, 4, 8, 16) ----------------------------|------------------------------------------ 128, 24x24, 3, 64, 5x5 | 1.66X, 1.65X, 1.61X, 1.56X, 1.49X 32, 24x24, 3, 64, 5x5 | 1.71X, 1.63X, 1.77X, 1.58X, 1.68X 128, 24x24, 1, 64, 5x5 | 1.44X, 1.40X, 1.38X, 1.37X, 1.33X 128, 24x24, 3, 64, 3x3 | 1.68X, 1.63X, 1.58X, 1.56X, 1.62X 128, 128x128, 3, 96, 11x11 | 1.36X, 1.36X, 1.37X, 1.37X, 1.37X In the higher level benchmark cifar10, we observe a runtime improvement of around 6% for AVX512 on Intel Skylake server (8 cores). On lower level PackRhs micro-benchmarks specified in TensorFlow tensorflow/core/kernels/eigen_spatial_convolutions_test.cc, we observe the following runtime numbers: AVX512: Parameters | Runtime without patch (ns) | Runtime with patch (ns) | Speedup ---------------------------------------------------------------|----------------------------|-------------------------|--------- BM_RHS_NAME(PackRhs, 128, 24, 24, 3, 64, 5, 5, 1, 1, 256, 56) | 41350 | 15073 | 2.74X BM_RHS_NAME(PackRhs, 32, 64, 64, 32, 64, 5, 5, 1, 1, 256, 56) | 7277 | 7341 | 0.99X BM_RHS_NAME(PackRhs, 32, 64, 64, 32, 64, 5, 5, 2, 2, 256, 56) | 8675 | 8681 | 1.00X BM_RHS_NAME(PackRhs, 32, 64, 64, 30, 64, 5, 5, 1, 1, 256, 56) | 24155 | 16079 | 1.50X BM_RHS_NAME(PackRhs, 32, 64, 64, 30, 64, 5, 5, 2, 2, 256, 56) | 25052 | 17152 | 1.46X BM_RHS_NAME(PackRhs, 32, 256, 256, 4, 16, 8, 8, 1, 1, 256, 56) | 18269 | 18345 | 1.00X BM_RHS_NAME(PackRhs, 32, 256, 256, 4, 16, 8, 8, 2, 4, 256, 56) | 19468 | 19872 | 0.98X BM_RHS_NAME(PackRhs, 32, 64, 64, 4, 16, 3, 3, 1, 1, 36, 432) | 156060 | 42432 | 3.68X BM_RHS_NAME(PackRhs, 32, 64, 64, 4, 16, 3, 3, 2, 2, 36, 432) | 132701 | 36944 | 3.59X AVX2: Parameters | Runtime without patch (ns) | Runtime with patch (ns) | Speedup ---------------------------------------------------------------|----------------------------|-------------------------|--------- BM_RHS_NAME(PackRhs, 128, 24, 24, 3, 64, 5, 5, 1, 1, 256, 56) | 26233 | 12393 | 2.12X BM_RHS_NAME(PackRhs, 32, 64, 64, 32, 64, 5, 5, 1, 1, 256, 56) | 6091 | 6062 | 1.00X BM_RHS_NAME(PackRhs, 32, 64, 64, 32, 64, 5, 5, 2, 2, 256, 56) | 7427 | 7408 | 1.00X BM_RHS_NAME(PackRhs, 32, 64, 64, 30, 64, 5, 5, 1, 1, 256, 56) | 23453 | 20826 | 1.13X BM_RHS_NAME(PackRhs, 32, 64, 64, 30, 64, 5, 5, 2, 2, 256, 56) | 23167 | 22091 | 1.09X BM_RHS_NAME(PackRhs, 32, 256, 256, 4, 16, 8, 8, 1, 1, 256, 56) | 23422 | 23682 | 0.99X BM_RHS_NAME(PackRhs, 32, 256, 256, 4, 16, 8, 8, 2, 4, 256, 56) | 23165 | 23663 | 0.98X BM_RHS_NAME(PackRhs, 32, 64, 64, 4, 16, 3, 3, 1, 1, 36, 432) | 72689 | 44969 | 1.62X BM_RHS_NAME(PackRhs, 32, 64, 64, 4, 16, 3, 3, 2, 2, 36, 432) | 61732 | 39779 | 1.55X All benchmarks on Intel Skylake server with 8 cores.
2019-04-20 14:46:43 +08:00
}
VERIFY((!PacketTraits::Vectorizable) || PacketTraits::HasAdd);
VERIFY((!PacketTraits::Vectorizable) || PacketTraits::HasSub);
VERIFY((!PacketTraits::Vectorizable) || PacketTraits::HasMul);
CHECK_CWISE2_IF(PacketTraits::HasAdd, REF_ADD, internal::padd);
CHECK_CWISE2_IF(PacketTraits::HasSub, REF_SUB, internal::psub);
CHECK_CWISE2_IF(PacketTraits::HasMul, REF_MUL, internal::pmul);
CHECK_CWISE2_IF(PacketTraits::HasDiv, REF_DIV, internal::pdiv);
if (PacketTraits::HasNegate)
CHECK_CWISE1(internal::negate, internal::pnegate);
CHECK_CWISE1(numext::conj, internal::pconj);
2011-02-24 05:22:10 +08:00
for(int offset=0;offset<3;++offset)
2011-02-24 02:24:26 +08:00
{
for (int i=0; i<PacketSize; ++i)
ref[i] = data1[offset];
internal::pstore(data2, internal::pset1<Packet>(data1[offset]));
VERIFY(test::areApprox(ref, data2, PacketSize) && "internal::pset1");
2011-02-24 02:24:26 +08:00
}
2014-04-25 17:21:18 +08:00
{
for (int i=0; i<PacketSize*4; ++i)
2014-04-25 17:46:22 +08:00
ref[i] = data1[i/PacketSize];
2014-04-25 17:21:18 +08:00
Packet A0, A1, A2, A3;
2014-04-25 17:46:22 +08:00
internal::pbroadcast4<Packet>(data1, A0, A1, A2, A3);
2014-04-25 17:21:18 +08:00
internal::pstore(data2+0*PacketSize, A0);
internal::pstore(data2+1*PacketSize, A1);
internal::pstore(data2+2*PacketSize, A2);
internal::pstore(data2+3*PacketSize, A3);
VERIFY(test::areApprox(ref, data2, 4*PacketSize) && "internal::pbroadcast4");
2014-04-25 17:21:18 +08:00
}
2014-04-25 17:21:18 +08:00
{
for (int i=0; i<PacketSize*2; ++i)
2014-04-25 17:46:22 +08:00
ref[i] = data1[i/PacketSize];
2014-05-05 21:03:29 +08:00
Packet A0, A1;
2014-04-25 17:46:22 +08:00
internal::pbroadcast2<Packet>(data1, A0, A1);
2014-04-25 17:21:18 +08:00
internal::pstore(data2+0*PacketSize, A0);
internal::pstore(data2+1*PacketSize, A1);
VERIFY(test::areApprox(ref, data2, 2*PacketSize) && "internal::pbroadcast2");
2014-04-25 17:21:18 +08:00
}
VERIFY(internal::isApprox(data1[0], internal::pfirst(internal::pload<Packet>(data1))) && "internal::pfirst");
if(PacketSize>1)
{
// apply different offsets to check that ploaddup is robust to unaligned inputs
2011-02-24 05:22:10 +08:00
for(int offset=0;offset<4;++offset)
{
for(int i=0;i<PacketSize/2;++i)
ref[2*i+0] = ref[2*i+1] = data1[offset+i];
internal::pstore(data2,internal::ploaddup<Packet>(data1+offset));
VERIFY(test::areApprox(ref, data2, PacketSize) && "ploaddup");
2011-02-24 05:22:10 +08:00
}
}
if(PacketSize>2)
{
// apply different offsets to check that ploadquad is robust to unaligned inputs
for(int offset=0;offset<4;++offset)
{
for(int i=0;i<PacketSize/4;++i)
ref[4*i+0] = ref[4*i+1] = ref[4*i+2] = ref[4*i+3] = data1[offset+i];
internal::pstore(data2,internal::ploadquad<Packet>(data1+offset));
VERIFY(test::areApprox(ref, data2, PacketSize) && "ploadquad");
}
}
ref[0] = Scalar(0);
for (int i=0; i<PacketSize; ++i)
ref[0] += data1[i];
VERIFY(test::isApproxAbs(ref[0], internal::predux(internal::pload<Packet>(data1)), refvalue) && "internal::predux");
if(PacketSize==8 && internal::unpacket_traits<typename internal::unpacket_traits<Packet>::half>::size ==4) // so far, predux_half_downto4 is only required in such a case
{
int HalfPacketSize = PacketSize>4 ? PacketSize/2 : PacketSize;
for (int i=0; i<HalfPacketSize; ++i)
ref[i] = Scalar(0);
for (int i=0; i<PacketSize; ++i)
ref[i%HalfPacketSize] += data1[i];
internal::pstore(data2, internal::predux_half_dowto4(internal::pload<Packet>(data1)));
VERIFY(test::areApprox(ref, data2, HalfPacketSize) && "internal::predux_half_dowto4");
}
2009-03-10 02:40:09 +08:00
ref[0] = Scalar(1);
for (int i=0; i<PacketSize; ++i)
Add missing packet ops for bool, and make it pass the same packet op unit tests as other arithmetic types. This change also contains a few minor cleanups: 1. Remove packet op pnot, which is not needed for anything other than pcmp_le_or_nan, which can be done in other ways. 2. Remove the "HasInsert" enum, which is no longer needed since we removed the corresponding packet ops. 3. Add faster pselect op for Packet4i when SSE4.1 is supported. Among other things, this makes the fast transposeInPlace() method available for Matrix<bool>. Run on ************** (72 X 2994 MHz CPUs); 2020-05-09T10:51:02.372347913-07:00 CPU: Intel Skylake Xeon with HyperThreading (36 cores) dL1:32KB dL2:1024KB dL3:24MB Benchmark Time(ns) CPU(ns) Iterations ----------------------------------------------------------------------- BM_TransposeInPlace<float>/4 9.77 9.77 71670320 BM_TransposeInPlace<float>/8 21.9 21.9 31929525 BM_TransposeInPlace<float>/16 66.6 66.6 10000000 BM_TransposeInPlace<float>/32 243 243 2879561 BM_TransposeInPlace<float>/59 844 844 829767 BM_TransposeInPlace<float>/64 933 933 750567 BM_TransposeInPlace<float>/128 3944 3945 177405 BM_TransposeInPlace<float>/256 16853 16853 41457 BM_TransposeInPlace<float>/512 204952 204968 3448 BM_TransposeInPlace<float>/1k 1053889 1053861 664 BM_TransposeInPlace<bool>/4 14.4 14.4 48637301 BM_TransposeInPlace<bool>/8 36.0 36.0 19370222 BM_TransposeInPlace<bool>/16 31.5 31.5 22178902 BM_TransposeInPlace<bool>/32 111 111 6272048 BM_TransposeInPlace<bool>/59 626 626 1000000 BM_TransposeInPlace<bool>/64 428 428 1632689 BM_TransposeInPlace<bool>/128 1677 1677 417377 BM_TransposeInPlace<bool>/256 7126 7126 96264 BM_TransposeInPlace<bool>/512 29021 29024 24165 BM_TransposeInPlace<bool>/1k 116321 116330 6068
2020-05-12 04:23:31 +08:00
ref[0] = REF_MUL(ref[0], data1[i]);
VERIFY(internal::isApprox(ref[0], internal::predux_mul(internal::pload<Packet>(data1))) && "internal::predux_mul");
2009-03-10 02:40:09 +08:00
for (int i=0; i<PacketSize; ++i)
ref[i] = data1[PacketSize-i-1];
internal::pstore(data2, internal::preverse(internal::pload<Packet>(data1)));
VERIFY(test::areApprox(ref, data2, PacketSize) && "internal::preverse");
internal::PacketBlock<Packet> kernel;
for (int i=0; i<PacketSize; ++i) {
kernel.packet[i] = internal::pload<Packet>(data1+i*PacketSize);
}
ptranspose(kernel);
for (int i=0; i<PacketSize; ++i) {
internal::pstore(data2, kernel.packet[i]);
for (int j = 0; j < PacketSize; ++j) {
VERIFY(test::isApproxAbs(data2[j], data1[i+j*PacketSize], refvalue) && "ptranspose");
}
}
Add missing packet ops for bool, and make it pass the same packet op unit tests as other arithmetic types. This change also contains a few minor cleanups: 1. Remove packet op pnot, which is not needed for anything other than pcmp_le_or_nan, which can be done in other ways. 2. Remove the "HasInsert" enum, which is no longer needed since we removed the corresponding packet ops. 3. Add faster pselect op for Packet4i when SSE4.1 is supported. Among other things, this makes the fast transposeInPlace() method available for Matrix<bool>. Run on ************** (72 X 2994 MHz CPUs); 2020-05-09T10:51:02.372347913-07:00 CPU: Intel Skylake Xeon with HyperThreading (36 cores) dL1:32KB dL2:1024KB dL3:24MB Benchmark Time(ns) CPU(ns) Iterations ----------------------------------------------------------------------- BM_TransposeInPlace<float>/4 9.77 9.77 71670320 BM_TransposeInPlace<float>/8 21.9 21.9 31929525 BM_TransposeInPlace<float>/16 66.6 66.6 10000000 BM_TransposeInPlace<float>/32 243 243 2879561 BM_TransposeInPlace<float>/59 844 844 829767 BM_TransposeInPlace<float>/64 933 933 750567 BM_TransposeInPlace<float>/128 3944 3945 177405 BM_TransposeInPlace<float>/256 16853 16853 41457 BM_TransposeInPlace<float>/512 204952 204968 3448 BM_TransposeInPlace<float>/1k 1053889 1053861 664 BM_TransposeInPlace<bool>/4 14.4 14.4 48637301 BM_TransposeInPlace<bool>/8 36.0 36.0 19370222 BM_TransposeInPlace<bool>/16 31.5 31.5 22178902 BM_TransposeInPlace<bool>/32 111 111 6272048 BM_TransposeInPlace<bool>/59 626 626 1000000 BM_TransposeInPlace<bool>/64 428 428 1632689 BM_TransposeInPlace<bool>/128 1677 1677 417377 BM_TransposeInPlace<bool>/256 7126 7126 96264 BM_TransposeInPlace<bool>/512 29021 29024 24165 BM_TransposeInPlace<bool>/1k 116321 116330 6068
2020-05-12 04:23:31 +08:00
if (PacketTraits::HasBlend) {
Packet thenPacket = internal::pload<Packet>(data1);
Packet elsePacket = internal::pload<Packet>(data2);
EIGEN_ALIGN_MAX internal::Selector<PacketSize> selector;
for (int i = 0; i < PacketSize; ++i) {
selector.select[i] = i;
}
Packet blend = internal::pblend(selector, thenPacket, elsePacket);
EIGEN_ALIGN_MAX Scalar result[size];
internal::pstore(result, blend);
for (int i = 0; i < PacketSize; ++i) {
VERIFY(test::isApproxAbs(result[i], (selector.select[i] ? data1[i] : data2[i]), refvalue));
}
}
{
for (int i = 0; i < PacketSize; ++i) {
// "if" mask
unsigned char v = internal::random<bool>() ? 0xff : 0;
char* bytes = (char*)(data1+i);
for(int k=0; k<int(sizeof(Scalar)); ++k) {
bytes[k] = v;
}
// "then" packet
data1[i+PacketSize] = internal::random<Scalar>();
// "else" packet
data1[i+2*PacketSize] = internal::random<Scalar>();
}
CHECK_CWISE3_IF(true, internal::pselect, internal::pselect);
}
2020-03-20 01:05:13 +08:00
CHECK_CWISE1_IF(PacketTraits::HasSqrt, numext::sqrt, internal::psqrt);
for (int i=0; i<size; ++i)
{
data1[i] = internal::random<Scalar>();
}
Add missing packet ops for bool, and make it pass the same packet op unit tests as other arithmetic types. This change also contains a few minor cleanups: 1. Remove packet op pnot, which is not needed for anything other than pcmp_le_or_nan, which can be done in other ways. 2. Remove the "HasInsert" enum, which is no longer needed since we removed the corresponding packet ops. 3. Add faster pselect op for Packet4i when SSE4.1 is supported. Among other things, this makes the fast transposeInPlace() method available for Matrix<bool>. Run on ************** (72 X 2994 MHz CPUs); 2020-05-09T10:51:02.372347913-07:00 CPU: Intel Skylake Xeon with HyperThreading (36 cores) dL1:32KB dL2:1024KB dL3:24MB Benchmark Time(ns) CPU(ns) Iterations ----------------------------------------------------------------------- BM_TransposeInPlace<float>/4 9.77 9.77 71670320 BM_TransposeInPlace<float>/8 21.9 21.9 31929525 BM_TransposeInPlace<float>/16 66.6 66.6 10000000 BM_TransposeInPlace<float>/32 243 243 2879561 BM_TransposeInPlace<float>/59 844 844 829767 BM_TransposeInPlace<float>/64 933 933 750567 BM_TransposeInPlace<float>/128 3944 3945 177405 BM_TransposeInPlace<float>/256 16853 16853 41457 BM_TransposeInPlace<float>/512 204952 204968 3448 BM_TransposeInPlace<float>/1k 1053889 1053861 664 BM_TransposeInPlace<bool>/4 14.4 14.4 48637301 BM_TransposeInPlace<bool>/8 36.0 36.0 19370222 BM_TransposeInPlace<bool>/16 31.5 31.5 22178902 BM_TransposeInPlace<bool>/32 111 111 6272048 BM_TransposeInPlace<bool>/59 626 626 1000000 BM_TransposeInPlace<bool>/64 428 428 1632689 BM_TransposeInPlace<bool>/128 1677 1677 417377 BM_TransposeInPlace<bool>/256 7126 7126 96264 BM_TransposeInPlace<bool>/512 29021 29024 24165 BM_TransposeInPlace<bool>/1k 116321 116330 6068
2020-05-12 04:23:31 +08:00
CHECK_CWISE1(internal::pzero, internal::pzero);
CHECK_CWISE2_IF(true, internal::por, internal::por);
CHECK_CWISE2_IF(true, internal::pxor, internal::pxor);
CHECK_CWISE2_IF(true, internal::pand, internal::pand);
Add missing packet ops for bool, and make it pass the same packet op unit tests as other arithmetic types. This change also contains a few minor cleanups: 1. Remove packet op pnot, which is not needed for anything other than pcmp_le_or_nan, which can be done in other ways. 2. Remove the "HasInsert" enum, which is no longer needed since we removed the corresponding packet ops. 3. Add faster pselect op for Packet4i when SSE4.1 is supported. Among other things, this makes the fast transposeInPlace() method available for Matrix<bool>. Run on ************** (72 X 2994 MHz CPUs); 2020-05-09T10:51:02.372347913-07:00 CPU: Intel Skylake Xeon with HyperThreading (36 cores) dL1:32KB dL2:1024KB dL3:24MB Benchmark Time(ns) CPU(ns) Iterations ----------------------------------------------------------------------- BM_TransposeInPlace<float>/4 9.77 9.77 71670320 BM_TransposeInPlace<float>/8 21.9 21.9 31929525 BM_TransposeInPlace<float>/16 66.6 66.6 10000000 BM_TransposeInPlace<float>/32 243 243 2879561 BM_TransposeInPlace<float>/59 844 844 829767 BM_TransposeInPlace<float>/64 933 933 750567 BM_TransposeInPlace<float>/128 3944 3945 177405 BM_TransposeInPlace<float>/256 16853 16853 41457 BM_TransposeInPlace<float>/512 204952 204968 3448 BM_TransposeInPlace<float>/1k 1053889 1053861 664 BM_TransposeInPlace<bool>/4 14.4 14.4 48637301 BM_TransposeInPlace<bool>/8 36.0 36.0 19370222 BM_TransposeInPlace<bool>/16 31.5 31.5 22178902 BM_TransposeInPlace<bool>/32 111 111 6272048 BM_TransposeInPlace<bool>/59 626 626 1000000 BM_TransposeInPlace<bool>/64 428 428 1632689 BM_TransposeInPlace<bool>/128 1677 1677 417377 BM_TransposeInPlace<bool>/256 7126 7126 96264 BM_TransposeInPlace<bool>/512 29021 29024 24165 BM_TransposeInPlace<bool>/1k 116321 116330 6068
2020-05-12 04:23:31 +08:00
packetmath_boolean_mask_ops<Scalar, Packet>();
}
Add missing packet ops for bool, and make it pass the same packet op unit tests as other arithmetic types. This change also contains a few minor cleanups: 1. Remove packet op pnot, which is not needed for anything other than pcmp_le_or_nan, which can be done in other ways. 2. Remove the "HasInsert" enum, which is no longer needed since we removed the corresponding packet ops. 3. Add faster pselect op for Packet4i when SSE4.1 is supported. Among other things, this makes the fast transposeInPlace() method available for Matrix<bool>. Run on ************** (72 X 2994 MHz CPUs); 2020-05-09T10:51:02.372347913-07:00 CPU: Intel Skylake Xeon with HyperThreading (36 cores) dL1:32KB dL2:1024KB dL3:24MB Benchmark Time(ns) CPU(ns) Iterations ----------------------------------------------------------------------- BM_TransposeInPlace<float>/4 9.77 9.77 71670320 BM_TransposeInPlace<float>/8 21.9 21.9 31929525 BM_TransposeInPlace<float>/16 66.6 66.6 10000000 BM_TransposeInPlace<float>/32 243 243 2879561 BM_TransposeInPlace<float>/59 844 844 829767 BM_TransposeInPlace<float>/64 933 933 750567 BM_TransposeInPlace<float>/128 3944 3945 177405 BM_TransposeInPlace<float>/256 16853 16853 41457 BM_TransposeInPlace<float>/512 204952 204968 3448 BM_TransposeInPlace<float>/1k 1053889 1053861 664 BM_TransposeInPlace<bool>/4 14.4 14.4 48637301 BM_TransposeInPlace<bool>/8 36.0 36.0 19370222 BM_TransposeInPlace<bool>/16 31.5 31.5 22178902 BM_TransposeInPlace<bool>/32 111 111 6272048 BM_TransposeInPlace<bool>/59 626 626 1000000 BM_TransposeInPlace<bool>/64 428 428 1632689 BM_TransposeInPlace<bool>/128 1677 1677 417377 BM_TransposeInPlace<bool>/256 7126 7126 96264 BM_TransposeInPlace<bool>/512 29021 29024 24165 BM_TransposeInPlace<bool>/1k 116321 116330 6068
2020-05-12 04:23:31 +08:00
template<typename Scalar,typename Packet> void packetmath_real()
{
typedef internal::packet_traits<Scalar> PacketTraits;
const int PacketSize = internal::unpacket_traits<Packet>::size;
const int size = PacketSize*4;
EIGEN_ALIGN_MAX Scalar data1[PacketSize*4];
EIGEN_ALIGN_MAX Scalar data2[PacketSize*4];
EIGEN_ALIGN_MAX Scalar ref[PacketSize*4];
for (int i=0; i<size; ++i)
{
data1[i] = internal::random<Scalar>(0,1) * std::pow(Scalar(10), internal::random<Scalar>(-6,6));
data2[i] = internal::random<Scalar>(0,1) * std::pow(Scalar(10), internal::random<Scalar>(-6,6));
}
if(internal::random<float>(0,1)<0.1f)
data1[internal::random<int>(0, PacketSize)] = 0;
CHECK_CWISE1_IF(PacketTraits::HasLog, std::log, internal::plog);
CHECK_CWISE1_IF(PacketTraits::HasRsqrt, Scalar(1)/std::sqrt, internal::prsqrt);
for (int i=0; i<size; ++i)
{
data1[i] = internal::random<Scalar>(-1,1) * std::pow(Scalar(10), internal::random<Scalar>(-3,3));
data2[i] = internal::random<Scalar>(-1,1) * std::pow(Scalar(10), internal::random<Scalar>(-3,3));
}
CHECK_CWISE1_IF(PacketTraits::HasSin, std::sin, internal::psin);
CHECK_CWISE1_IF(PacketTraits::HasCos, std::cos, internal::pcos);
CHECK_CWISE1_IF(PacketTraits::HasTan, std::tan, internal::ptan);
CHECK_CWISE1_IF(PacketTraits::HasRound, numext::round, internal::pround);
CHECK_CWISE1_IF(PacketTraits::HasCeil, numext::ceil, internal::pceil);
CHECK_CWISE1_IF(PacketTraits::HasFloor, numext::floor, internal::pfloor);
CHECK_CWISE1_IF(PacketTraits::HasRint, numext::rint, internal::print);
// See bug 1785.
for (int i=0; i<size; ++i)
{
data1[i] = -1.5 + i;
data2[i] = -1.5 + i;
}
CHECK_CWISE1_IF(PacketTraits::HasRound, numext::round, internal::pround);
CHECK_CWISE1_IF(PacketTraits::HasRint, numext::rint, internal::print);
for (int i=0; i<size; ++i)
{
data1[i] = internal::random<Scalar>(-1,1);
data2[i] = internal::random<Scalar>(-1,1);
}
CHECK_CWISE1_IF(PacketTraits::HasASin, std::asin, internal::pasin);
CHECK_CWISE1_IF(PacketTraits::HasACos, std::acos, internal::pacos);
for (int i=0; i<size; ++i)
{
data1[i] = internal::random<Scalar>(-87,88);
data2[i] = internal::random<Scalar>(-87,88);
}
CHECK_CWISE1_IF(PacketTraits::HasExp, std::exp, internal::pexp);
2016-02-11 09:41:47 +08:00
for (int i=0; i<size; ++i)
{
data1[i] = internal::random<Scalar>(-1,1) * std::pow(Scalar(10), internal::random<Scalar>(-6,6));
data2[i] = internal::random<Scalar>(-1,1) * std::pow(Scalar(10), internal::random<Scalar>(-6,6));
}
Improve accuracy of fast approximate tanh and the logistic functions in Eigen, such that they preserve relative accuracy to within a few ULPs where their function values tend to zero (around x=0 for tanh, and for large negative x for the logistic function). This change re-instates the fast rational approximation of the logistic function for float32 in Eigen (removed in https://gitlab.com/libeigen/eigen/commit/66f07efeaed39d6a67005343d7e0caf7d9eeacdb), but uses the more accurate approximation 1/(1+exp(-1)) ~= exp(x) below -9. The exponential is only calculated on the vectorized path if at least one element in the SIMD input vector is less than -9. This change also contains a few improvements to speed up the original float specialization of logistic: - Introduce EIGEN_PREDICT_{FALSE,TRUE} for __builtin_predict and use it to predict that the logistic-only path is most likely (~2-3% speedup for the common case). - Carefully set the upper clipping point to the smallest x where the approximation evaluates to exactly 1. This saves the explicit clamping of the output (~7% speedup). The increased accuracy for tanh comes at a cost of 10-20% depending on instruction set. The benchmarks below repeated calls u = v.logistic() (u = v.tanh(), respectively) where u and v are of type Eigen::ArrayXf, have length 8k, and v contains random numbers in [-1,1]. Benchmark numbers for logistic: Before: Benchmark Time(ns) CPU(ns) Iterations ----------------------------------------------------------------- SSE BM_eigen_logistic_float 4467 4468 155835 model_time: 4827 AVX BM_eigen_logistic_float 2347 2347 299135 model_time: 2926 AVX+FMA BM_eigen_logistic_float 1467 1467 476143 model_time: 2926 AVX512 BM_eigen_logistic_float 805 805 858696 model_time: 1463 After: Benchmark Time(ns) CPU(ns) Iterations ----------------------------------------------------------------- SSE BM_eigen_logistic_float 2589 2590 270264 model_time: 4827 AVX BM_eigen_logistic_float 1428 1428 489265 model_time: 2926 AVX+FMA BM_eigen_logistic_float 1059 1059 662255 model_time: 2926 AVX512 BM_eigen_logistic_float 673 673 1000000 model_time: 1463 Benchmark numbers for tanh: Before: Benchmark Time(ns) CPU(ns) Iterations ----------------------------------------------------------------- SSE BM_eigen_tanh_float 2391 2391 292624 model_time: 4242 AVX BM_eigen_tanh_float 1256 1256 554662 model_time: 2633 AVX+FMA BM_eigen_tanh_float 823 823 866267 model_time: 1609 AVX512 BM_eigen_tanh_float 443 443 1578999 model_time: 805 After: Benchmark Time(ns) CPU(ns) Iterations ----------------------------------------------------------------- SSE BM_eigen_tanh_float 2588 2588 273531 model_time: 4242 AVX BM_eigen_tanh_float 1536 1536 452321 model_time: 2633 AVX+FMA BM_eigen_tanh_float 1007 1007 694681 model_time: 1609 AVX512 BM_eigen_tanh_float 471 471 1472178 model_time: 805
2019-12-17 05:33:42 +08:00
data1[0] = 1e-20;
2016-02-11 09:41:47 +08:00
CHECK_CWISE1_IF(PacketTraits::HasTanh, std::tanh, internal::ptanh);
if(PacketTraits::HasExp && PacketSize>=2)
{
data1[0] = std::numeric_limits<Scalar>::quiet_NaN();
data1[1] = std::numeric_limits<Scalar>::epsilon();
test::packet_helper<PacketTraits::HasExp,Packet> h;
h.store(data2, internal::pexp(h.load(data1)));
VERIFY((numext::isnan)(data2[0]));
VERIFY_IS_EQUAL(std::exp(std::numeric_limits<Scalar>::epsilon()), data2[1]);
data1[0] = -std::numeric_limits<Scalar>::epsilon();
data1[1] = 0;
h.store(data2, internal::pexp(h.load(data1)));
VERIFY_IS_EQUAL(std::exp(-std::numeric_limits<Scalar>::epsilon()), data2[0]);
VERIFY_IS_EQUAL(std::exp(Scalar(0)), data2[1]);
data1[0] = (std::numeric_limits<Scalar>::min)();
data1[1] = -(std::numeric_limits<Scalar>::min)();
h.store(data2, internal::pexp(h.load(data1)));
VERIFY_IS_EQUAL(std::exp((std::numeric_limits<Scalar>::min)()), data2[0]);
VERIFY_IS_EQUAL(std::exp(-(std::numeric_limits<Scalar>::min)()), data2[1]);
data1[0] = std::numeric_limits<Scalar>::denorm_min();
data1[1] = -std::numeric_limits<Scalar>::denorm_min();
h.store(data2, internal::pexp(h.load(data1)));
VERIFY_IS_EQUAL(std::exp(std::numeric_limits<Scalar>::denorm_min()), data2[0]);
VERIFY_IS_EQUAL(std::exp(-std::numeric_limits<Scalar>::denorm_min()), data2[1]);
}
2016-05-11 07:21:43 +08:00
if (PacketTraits::HasTanh) {
2016-09-22 17:18:52 +08:00
// NOTE this test migh fail with GCC prior to 6.3, see MathFunctionsImpl.h for details.
2016-05-11 07:21:43 +08:00
data1[0] = std::numeric_limits<Scalar>::quiet_NaN();
test::packet_helper<internal::packet_traits<Scalar>::HasTanh,Packet> h;
2016-05-11 07:21:43 +08:00
h.store(data2, internal::ptanh(h.load(data1)));
VERIFY((numext::isnan)(data2[0]));
}
{
internal::scalar_logistic_op<Scalar> logistic;
for (int i=0; i<size; ++i)
{
data1[i] = internal::random<Scalar>(-20,20);
}
internal::pstore(data2, logistic.packetOp(internal::pload<Packet>(data1)));
for (int i=0; i<PacketSize; ++i) {
VERIFY_IS_APPROX(data2[i],logistic(data1[i]));
#ifdef EIGEN_VECTORIZE // don't check for exactness when using the i387 FPU
VERIFY_IS_EQUAL(data2[i],logistic(data1[i]));
#endif
}
}
#if EIGEN_HAS_C99_MATH && (__cplusplus > 199711L)
data1[0] = std::numeric_limits<Scalar>::infinity();
data1[1] = Scalar(-1);
CHECK_CWISE1_IF(PacketTraits::HasLog1p, std::log1p, internal::plog1p);
data1[0] = std::numeric_limits<Scalar>::infinity();
data1[1] = -std::numeric_limits<Scalar>::infinity();
CHECK_CWISE1_IF(PacketTraits::HasExpm1, std::expm1, internal::pexpm1);
#endif
if(PacketSize>=2)
{
data1[0] = std::numeric_limits<Scalar>::quiet_NaN();
data1[1] = std::numeric_limits<Scalar>::epsilon();
if(PacketTraits::HasLog)
{
test::packet_helper<PacketTraits::HasLog,Packet> h;
h.store(data2, internal::plog(h.load(data1)));
VERIFY((numext::isnan)(data2[0]));
VERIFY_IS_EQUAL(std::log(std::numeric_limits<Scalar>::epsilon()), data2[1]);
data1[0] = -std::numeric_limits<Scalar>::epsilon();
data1[1] = 0;
h.store(data2, internal::plog(h.load(data1)));
VERIFY((numext::isnan)(data2[0]));
VERIFY_IS_EQUAL(std::log(Scalar(0)), data2[1]);
data1[0] = (std::numeric_limits<Scalar>::min)();
data1[1] = -(std::numeric_limits<Scalar>::min)();
h.store(data2, internal::plog(h.load(data1)));
VERIFY_IS_EQUAL(std::log((std::numeric_limits<Scalar>::min)()), data2[0]);
VERIFY((numext::isnan)(data2[1]));
data1[0] = std::numeric_limits<Scalar>::denorm_min();
data1[1] = -std::numeric_limits<Scalar>::denorm_min();
h.store(data2, internal::plog(h.load(data1)));
// VERIFY_IS_EQUAL(std::log(std::numeric_limits<Scalar>::denorm_min()), data2[0]);
VERIFY((numext::isnan)(data2[1]));
data1[0] = Scalar(-1.0f);
h.store(data2, internal::plog(h.load(data1)));
VERIFY((numext::isnan)(data2[0]));
data1[0] = std::numeric_limits<Scalar>::infinity();
h.store(data2, internal::plog(h.load(data1)));
VERIFY((numext::isinf)(data2[0]));
}
if(PacketTraits::HasLog1p) {
test::packet_helper<PacketTraits::HasLog1p,Packet> h;
data1[0] = Scalar(-2);
data1[1] = -std::numeric_limits<Scalar>::infinity();
h.store(data2, internal::plog1p(h.load(data1)));
VERIFY((numext::isnan)(data2[0]));
VERIFY((numext::isnan)(data2[1]));
}
if(PacketTraits::HasSqrt)
{
test::packet_helper<PacketTraits::HasSqrt,Packet> h;
data1[0] = Scalar(-1.0f);
2018-12-27 18:20:47 +08:00
data1[1] = -std::numeric_limits<Scalar>::denorm_min();
h.store(data2, internal::psqrt(h.load(data1)));
VERIFY((numext::isnan)(data2[0]));
VERIFY((numext::isnan)(data2[1]));
}
if(PacketTraits::HasCos)
{
test::packet_helper<PacketTraits::HasCos,Packet> h;
for(Scalar k = 1; k<Scalar(10000)/std::numeric_limits<Scalar>::epsilon(); k*=2)
{
for(int k1=0;k1<=1; ++k1)
{
data1[0] = (2*k+k1 )*Scalar(EIGEN_PI)/2 * internal::random<Scalar>(0.8,1.2);
data1[1] = (2*k+2+k1)*Scalar(EIGEN_PI)/2 * internal::random<Scalar>(0.8,1.2);
h.store(data2, internal::pcos(h.load(data1)));
h.store(data2+PacketSize, internal::psin(h.load(data1)));
VERIFY(data2[0]<=Scalar(1.) && data2[0]>=Scalar(-1.));
VERIFY(data2[1]<=Scalar(1.) && data2[1]>=Scalar(-1.));
VERIFY(data2[PacketSize+0]<=Scalar(1.) && data2[PacketSize+0]>=Scalar(-1.));
VERIFY(data2[PacketSize+1]<=Scalar(1.) && data2[PacketSize+1]>=Scalar(-1.));
VERIFY_IS_APPROX(numext::abs2(data2[0])+numext::abs2(data2[PacketSize+0]), Scalar(1));
VERIFY_IS_APPROX(numext::abs2(data2[1])+numext::abs2(data2[PacketSize+1]), Scalar(1));
}
}
data1[0] = std::numeric_limits<Scalar>::infinity();
data1[1] = -std::numeric_limits<Scalar>::infinity();
h.store(data2, internal::psin(h.load(data1)));
VERIFY((numext::isnan)(data2[0]));
VERIFY((numext::isnan)(data2[1]));
h.store(data2, internal::pcos(h.load(data1)));
VERIFY((numext::isnan)(data2[0]));
VERIFY((numext::isnan)(data2[1]));
data1[0] = std::numeric_limits<Scalar>::quiet_NaN();
h.store(data2, internal::psin(h.load(data1)));
VERIFY((numext::isnan)(data2[0]));
h.store(data2, internal::pcos(h.load(data1)));
VERIFY((numext::isnan)(data2[0]));
data1[0] = -Scalar(0.);
h.store(data2, internal::psin(h.load(data1)));
VERIFY( internal::biteq(data2[0], data1[0]) );
h.store(data2, internal::pcos(h.load(data1)));
VERIFY_IS_EQUAL(data2[0], Scalar(1));
}
}
2013-03-21 01:28:40 +08:00
}
template<typename Scalar,typename Packet> void packetmath_notcomplex()
2013-03-21 01:28:40 +08:00
{
typedef internal::packet_traits<Scalar> PacketTraits;
const int PacketSize = internal::unpacket_traits<Packet>::size;
2013-03-21 01:28:40 +08:00
EIGEN_ALIGN_MAX Scalar data1[PacketSize*4];
EIGEN_ALIGN_MAX Scalar data2[PacketSize*4];
EIGEN_ALIGN_MAX Scalar ref[PacketSize*4];
Array<Scalar,Dynamic,1>::Map(data1, PacketSize*4).setRandom();
2020-03-27 04:18:19 +08:00
if (PacketTraits::HasCast) {
test_cast<Packet, float>();
test_cast<Packet, double>();
test_cast<Packet, int8_t>();
test_cast<Packet, uint8_t>();
test_cast<Packet, int16_t>();
test_cast<Packet, uint16_t>();
test_cast<Packet, int32_t>();
test_cast<Packet, uint32_t>();
test_cast<Packet, int64_t>();
test_cast<Packet, uint64_t>();
}
ref[0] = data1[0];
for (int i=0; i<PacketSize; ++i)
ref[0] = (std::min)(ref[0],data1[i]);
VERIFY(internal::isApprox(ref[0], internal::predux_min(internal::pload<Packet>(data1))) && "internal::predux_min");
VERIFY((!PacketTraits::Vectorizable) || PacketTraits::HasMin);
VERIFY((!PacketTraits::Vectorizable) || PacketTraits::HasMax);
CHECK_CWISE2_IF(PacketTraits::HasMin, (std::min), internal::pmin);
CHECK_CWISE2_IF(PacketTraits::HasMax, (std::max), internal::pmax);
CHECK_CWISE1(numext::abs, internal::pabs);
CHECK_CWISE2_IF(PacketTraits::HasAbsDiff, REF_ABS_DIFF, internal::pabsdiff);
ref[0] = data1[0];
for (int i=0; i<PacketSize; ++i)
ref[0] = (std::max)(ref[0],data1[i]);
VERIFY(internal::isApprox(ref[0], internal::predux_max(internal::pload<Packet>(data1))) && "internal::predux_max");
2011-05-19 03:11:03 +08:00
for (int i=0; i<PacketSize; ++i)
ref[i] = data1[0]+Scalar(i);
internal::pstore(data2, internal::plset<Packet>(data1[0]));
VERIFY(test::areApprox(ref, data2, PacketSize) && "internal::plset");
{
unsigned char* data1_bits = reinterpret_cast<unsigned char*>(data1);
// predux_all - not needed yet
// for (unsigned int i=0; i<PacketSize*sizeof(Scalar); ++i) data1_bits[i] = 0xff;
// VERIFY(internal::predux_all(internal::pload<Packet>(data1)) && "internal::predux_all(1111)");
// for(int k=0; k<PacketSize; ++k)
// {
// for (unsigned int i=0; i<sizeof(Scalar); ++i) data1_bits[k*sizeof(Scalar)+i] = 0x0;
// VERIFY( (!internal::predux_all(internal::pload<Packet>(data1))) && "internal::predux_all(0101)");
// for (unsigned int i=0; i<sizeof(Scalar); ++i) data1_bits[k*sizeof(Scalar)+i] = 0xff;
// }
// predux_any
for (unsigned int i=0; i<PacketSize*sizeof(Scalar); ++i) data1_bits[i] = 0x0;
VERIFY( (!internal::predux_any(internal::pload<Packet>(data1))) && "internal::predux_any(0000)");
for(int k=0; k<PacketSize; ++k)
{
for (unsigned int i=0; i<sizeof(Scalar); ++i) data1_bits[k*sizeof(Scalar)+i] = 0xff;
VERIFY( internal::predux_any(internal::pload<Packet>(data1)) && "internal::predux_any(0101)");
for (unsigned int i=0; i<sizeof(Scalar); ++i) data1_bits[k*sizeof(Scalar)+i] = 0x00;
}
}
}
template<typename Scalar,typename Packet,bool ConjLhs,bool ConjRhs> void test_conj_helper(Scalar* data1, Scalar* data2, Scalar* ref, Scalar* pval)
2011-02-24 02:24:26 +08:00
{
const int PacketSize = internal::unpacket_traits<Packet>::size;
2011-02-24 02:24:26 +08:00
internal::conj_if<ConjLhs> cj0;
internal::conj_if<ConjRhs> cj1;
internal::conj_helper<Scalar,Scalar,ConjLhs,ConjRhs> cj;
internal::conj_helper<Packet,Packet,ConjLhs,ConjRhs> pcj;
2011-02-24 02:24:26 +08:00
for(int i=0;i<PacketSize;++i)
{
ref[i] = cj0(data1[i]) * cj1(data2[i]);
VERIFY(internal::isApprox(ref[i], cj.pmul(data1[i],data2[i])) && "conj_helper pmul");
}
internal::pstore(pval,pcj.pmul(internal::pload<Packet>(data1),internal::pload<Packet>(data2)));
VERIFY(test::areApprox(ref, pval, PacketSize) && "conj_helper pmul");
2011-02-24 02:24:26 +08:00
for(int i=0;i<PacketSize;++i)
{
Scalar tmp = ref[i];
ref[i] += cj0(data1[i]) * cj1(data2[i]);
VERIFY(internal::isApprox(ref[i], cj.pmadd(data1[i],data2[i],tmp)) && "conj_helper pmadd");
}
internal::pstore(pval,pcj.pmadd(internal::pload<Packet>(data1),internal::pload<Packet>(data2),internal::pload<Packet>(pval)));
VERIFY(test::areApprox(ref, pval, PacketSize) && "conj_helper pmadd");
2011-02-24 02:24:26 +08:00
}
template<typename Scalar,typename Packet> void packetmath_complex()
{
const int PacketSize = internal::unpacket_traits<Packet>::size;
const int size = PacketSize*4;
EIGEN_ALIGN_MAX Scalar data1[PacketSize*4];
EIGEN_ALIGN_MAX Scalar data2[PacketSize*4];
EIGEN_ALIGN_MAX Scalar ref[PacketSize*4];
EIGEN_ALIGN_MAX Scalar pval[PacketSize*4];
for (int i=0; i<size; ++i)
{
data1[i] = internal::random<Scalar>() * Scalar(1e2);
data2[i] = internal::random<Scalar>() * Scalar(1e2);
}
test_conj_helper<Scalar,Packet,false,false> (data1,data2,ref,pval);
test_conj_helper<Scalar,Packet,false,true> (data1,data2,ref,pval);
test_conj_helper<Scalar,Packet,true,false> (data1,data2,ref,pval);
test_conj_helper<Scalar,Packet,true,true> (data1,data2,ref,pval);
2011-02-23 21:20:33 +08:00
{
for(int i=0;i<PacketSize;++i)
ref[i] = Scalar(std::imag(data1[i]),std::real(data1[i]));
internal::pstore(pval,internal::pcplxflip(internal::pload<Packet>(data1)));
VERIFY(test::areApprox(ref, pval, PacketSize) && "pcplxflip");
2011-02-23 21:20:33 +08:00
}
}
template<typename Scalar,typename Packet> void packetmath_scatter_gather()
{
typedef typename NumTraits<Scalar>::Real RealScalar;
const int PacketSize = internal::unpacket_traits<Packet>::size;
EIGEN_ALIGN_MAX Scalar data1[PacketSize];
RealScalar refvalue = 0;
for (int i=0; i<PacketSize; ++i) {
data1[i] = internal::random<Scalar>()/RealScalar(PacketSize);
}
2014-07-09 22:01:24 +08:00
int stride = internal::random<int>(1,20);
EIGEN_ALIGN_MAX Scalar buffer[PacketSize*20];
memset(buffer, 0, 20*PacketSize*sizeof(Scalar));
Packet packet = internal::pload<Packet>(data1);
2014-07-09 22:01:24 +08:00
internal::pscatter<Scalar, Packet>(buffer, packet, stride);
2014-07-09 22:01:24 +08:00
for (int i = 0; i < PacketSize*20; ++i) {
if ((i%stride) == 0 && i<stride*PacketSize) {
VERIFY(
test::isApproxAbs(buffer[i], data1[i/stride], refvalue) && "pscatter");
} else {
VERIFY(
test::isApproxAbs(buffer[i], Scalar(0), refvalue) && "pscatter");
}
}
for (int i=0; i<PacketSize*7; ++i) {
buffer[i] = internal::random<Scalar>()/RealScalar(PacketSize);
}
packet = internal::pgather<Scalar, Packet>(buffer, 7);
internal::pstore(data1, packet);
for (int i = 0; i < PacketSize; ++i) {
VERIFY(test::isApproxAbs(data1[i], buffer[i*7], refvalue) && "pgather");
}
}
namespace Eigen {
namespace test {
template<typename Scalar,typename PacketType>
struct runall<Scalar,PacketType,false,false> { // i.e. float or double
static void run() {
packetmath<Scalar,PacketType>();
packetmath_scatter_gather<Scalar,PacketType>();
packetmath_notcomplex<Scalar,PacketType>();
packetmath_real<Scalar,PacketType>();
}
};
template<typename Scalar,typename PacketType>
struct runall<Scalar,PacketType,false,true> { // i.e. int
static void run() {
packetmath<Scalar,PacketType>();
packetmath_scatter_gather<Scalar,PacketType>();
packetmath_notcomplex<Scalar,PacketType>();
}
};
template<typename Scalar,typename PacketType>
struct runall<Scalar,PacketType,true,false> { // i.e. complex
static void run() {
packetmath<Scalar,PacketType>();
packetmath_scatter_gather<Scalar,PacketType>();
packetmath_complex<Scalar,PacketType>();
}
};
}
}
EIGEN_DECLARE_TEST(packetmath)
{
g_first_pass = true;
for(int i = 0; i < g_repeat; i++) {
CALL_SUBTEST_1( test::runner<float>::run() );
CALL_SUBTEST_2( test::runner<double>::run() );
CALL_SUBTEST_3( test::runner<int8_t>::run() );
CALL_SUBTEST_4( test::runner<uint8_t>::run() );
CALL_SUBTEST_5( test::runner<int16_t>::run() );
CALL_SUBTEST_6( test::runner<uint16_t>::run() );
CALL_SUBTEST_7( test::runner<int32_t>::run() );
CALL_SUBTEST_8( test::runner<uint32_t>::run() );
CALL_SUBTEST_9( test::runner<int64_t>::run() );
CALL_SUBTEST_10( test::runner<uint64_t>::run() );
CALL_SUBTEST_11( test::runner<std::complex<float> >::run() );
CALL_SUBTEST_12( test::runner<std::complex<double> >::run() );
CALL_SUBTEST_13(( packetmath<half,internal::packet_traits<half>::type>() ));
#ifdef EIGEN_PACKET_MATH_SSE_H
Add missing packet ops for bool, and make it pass the same packet op unit tests as other arithmetic types. This change also contains a few minor cleanups: 1. Remove packet op pnot, which is not needed for anything other than pcmp_le_or_nan, which can be done in other ways. 2. Remove the "HasInsert" enum, which is no longer needed since we removed the corresponding packet ops. 3. Add faster pselect op for Packet4i when SSE4.1 is supported. Among other things, this makes the fast transposeInPlace() method available for Matrix<bool>. Run on ************** (72 X 2994 MHz CPUs); 2020-05-09T10:51:02.372347913-07:00 CPU: Intel Skylake Xeon with HyperThreading (36 cores) dL1:32KB dL2:1024KB dL3:24MB Benchmark Time(ns) CPU(ns) Iterations ----------------------------------------------------------------------- BM_TransposeInPlace<float>/4 9.77 9.77 71670320 BM_TransposeInPlace<float>/8 21.9 21.9 31929525 BM_TransposeInPlace<float>/16 66.6 66.6 10000000 BM_TransposeInPlace<float>/32 243 243 2879561 BM_TransposeInPlace<float>/59 844 844 829767 BM_TransposeInPlace<float>/64 933 933 750567 BM_TransposeInPlace<float>/128 3944 3945 177405 BM_TransposeInPlace<float>/256 16853 16853 41457 BM_TransposeInPlace<float>/512 204952 204968 3448 BM_TransposeInPlace<float>/1k 1053889 1053861 664 BM_TransposeInPlace<bool>/4 14.4 14.4 48637301 BM_TransposeInPlace<bool>/8 36.0 36.0 19370222 BM_TransposeInPlace<bool>/16 31.5 31.5 22178902 BM_TransposeInPlace<bool>/32 111 111 6272048 BM_TransposeInPlace<bool>/59 626 626 1000000 BM_TransposeInPlace<bool>/64 428 428 1632689 BM_TransposeInPlace<bool>/128 1677 1677 417377 BM_TransposeInPlace<bool>/256 7126 7126 96264 BM_TransposeInPlace<bool>/512 29021 29024 24165 BM_TransposeInPlace<bool>/1k 116321 116330 6068
2020-05-12 04:23:31 +08:00
CALL_SUBTEST_14(( packetmath<bool,internal::packet_traits<bool>::type>() ));
#endif
g_first_pass = false;
}
}