See
<https://stackoverflow.com/questions/59709148/ensuring-that-eigen-uses-avx-vectorization-for-a-certain-operation>
for an explanation of the problem this solves.
In short, for some reason, before this commit the half-packet is
selected when the array / matrix size is not a multiple of
`unpacket_traits<PacketType>::size`, where `PacketType` starts out
being the full Packet.
For example, for some data of 100 `float`s, `Packet4f` will be
selected rather than `Packet8f`, because 100 is not a multiple of 8,
the size of `Packet8f`.
This commit switches to selecting the half-packet if the size is
less than the packet size, which seems to make more sense.
As I stated in the SO post I'm not sure that I'm understanding the
issue correctly, but this fix resolves the issue in my program. Moreover,
`make check` passes, with the exception of line 614 and 616 in
`test/packetmath.cpp`, which however also fail on master on my machine:
CHECK_CWISE1_IF(PacketTraits::HasBessel, numext::bessel_i0, internal::pbessel_i0);
...
CHECK_CWISE1_IF(PacketTraits::HasBessel, numext::bessel_i1, internal::pbessel_i1);
This provides a new op that matches std::rint and previous behavior of
pround. Also adds corresponding unsupported/../Tensor op.
Performance is the same as e. g. floor (tested SSE/AVX).
* Adding Missing operations for vector comparison in SYCL. This caused compiler error for vector comparison when compiling SYCL
* Fixing the compiler error for placement new in TensorForcedEval.h This caused compiler error when compiling SYCL backend
* Reducing the SYCL warning by removing the abort function inside the kernel
* Adding Strong inline to functions inside SYCL interop.
This fixes deprecated-copy warnings when compiling with GCC>=9
Also protect some additional Base-constructors from getting called by user code code (#1587)
This change re-instates the fast rational approximation of the logistic function for float32 in Eigen (removed in 66f07efeae), 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
This also adds pset1frombits helper to Packet[24]d.
Makes round ~45% slower for SSE: 1.65µs ± 1% before vs 2.45µs ± 2% after,
stil an order of magnitude faster than scalar version: 33.8µs ± 2%.
The following commit introduces compile errors when running eigen with hipcc
2918f85ba9
hipcc errors out because it requies the device attribute on the methods within the TensorBlockV2ResourceRequirements struct instroduced by the commit above. The fix is to add the device attribute to those methods
77b447c24e
While providing a 50% speedup on Haswell+ processors, the large relative error outside [-18, 18] in this approximation causes problems, e.g., when computing gradients of activation functions like softplus in neural networks.
Recent changes have introduced the following build error when compiling with HIPCC
---------
unsupported/test/../../Eigen/src/Core/GenericPacketMath.h:254:58: error: 'ldexp': no overloaded function has restriction specifiers that are compatible with the ambient context 'pldexp'
---------
The fix for the error is to pick the math function(s) from the global namespace (where they are declared as device functions in the HIP header files) when compiling with HIPCC.
* Unifying all loadLocalTile from lhs and rhs to an extract_block function.
* Adding get_tensor operation which was missing in TensorContractionMapper.
* Adding the -D method missing from cmake for Disable_Skinny Contraction operation.
* Wrapping all the indices in TensorScanSycl into Scan parameter struct.
* Fixing typo in Device SYCL
* Unifying load to private register for tall/skinny no shared
* Unifying load to vector tile for tensor-vector/vector-tensor operation
* Removing all the LHS/RHS class for extracting data from global
* Removing Outputfunction from TensorContractionSkinnyNoshared.
* Combining the local memory version of tall/skinny and normal tensor contraction into one kernel.
* Combining the no-local memory version of tall/skinny and normal tensor contraction into one kernel.
* Combining General Tensor-Vector and VectorTensor contraction into one kernel.
* Making double buffering optional for Tensor contraction when local memory is version is used.
* Modifying benchmark to accept custom Reduction Sizes
* Disabling AVX optimization for SYCL backend on the host to allow SSE optimization to the host
* Adding Test for SYCL
* Modifying SYCL CMake
Add a new EIGEN_HAS_INTRINSIC_INT128 macro, and use this instead of __SIZEOF_INT128__. This fixes related issues with TensorIntDiv.h when building with Clang for Windows, where support for 128-bit integer arithmetic is advertised but broken in practice.
2. Simplify handling of special cases by taking advantage of the fact that the
builtin vrsqrt approximation handles negative, zero and +inf arguments correctly.
This speeds up the SSE and AVX implementations by ~20%.
3. Make the Newton-Raphson formula used for rsqrt more numerically robust:
Before: y = y * (1.5 - x/2 * y^2)
After: y = y * (1.5 - y * (x/2) * y)
Forming y^2 can overflow for very large or very small (denormalized) values of x, while x*y ~= 1. For AVX512, this makes it possible to compute accurate results for denormal inputs down to ~1e-42 in single precision.
4. Add a faster double precision implementation for Knights Landing using the vrsqrt28 instruction and a single Newton-Raphson iteration.
Benchmark results: https://bitbucket.org/snippets/rmlarsen/5LBq9o
The errors were introduced by this commit : d38e6fbc27
After the above mentioned commit, some of the tests started failing with the following error
```
Building HIPCC object unsupported/test/CMakeFiles/cxx11_tensor_reduction_gpu_5.dir/cxx11_tensor_reduction_gpu_5_generated_cxx11_tensor_reduction_gpu.cu.o
In file included from /home/rocm-user/eigen/unsupported/test/cxx11_tensor_reduction_gpu.cu:16:
In file included from /home/rocm-user/eigen/unsupported/Eigen/CXX11/Tensor:29:
In file included from /home/rocm-user/eigen/unsupported/Eigen/CXX11/../SpecialFunctions:70:
/home/rocm-user/eigen/unsupported/Eigen/CXX11/../src/SpecialFunctions/SpecialFunctionsHalf.h:28:22: error: call to 'erf' is ambiguous
return Eigen::half(Eigen::numext::erf(static_cast<float>(a)));
^~~~~~~~~~~~~~~~~~
/home/rocm-user/eigen/unsupported/test/../../Eigen/src/Core/MathFunctions.h:1600:7: note: candidate function [with T = float]
float erf(const float &x) { return ::erff(x); }
^
/home/rocm-user/eigen/unsupported/Eigen/CXX11/../src/SpecialFunctions/SpecialFunctionsImpl.h:1897:5: note: candidate function [with Scalar = float]
erf(const Scalar& x) {
^
In file included from /home/rocm-user/eigen/unsupported/test/cxx11_tensor_reduction_gpu.cu:16:
In file included from /home/rocm-user/eigen/unsupported/Eigen/CXX11/Tensor:29:
In file included from /home/rocm-user/eigen/unsupported/Eigen/CXX11/../SpecialFunctions:75:
/home/rocm-user/eigen/unsupported/Eigen/CXX11/../src/SpecialFunctions/arch/GPU/GpuSpecialFunctions.h:87:23: error: call to 'erf' is ambiguous
return make_double2(erf(a.x), erf(a.y));
^~~
/home/rocm-user/eigen/unsupported/test/../../Eigen/src/Core/MathFunctions.h:1603:8: note: candidate function [with T = double]
double erf(const double &x) { return ::erf(x); }
^
/home/rocm-user/eigen/unsupported/Eigen/CXX11/../src/SpecialFunctions/SpecialFunctionsImpl.h:1897:5: note: candidate function [with Scalar = double]
erf(const Scalar& x) {
^
In file included from /home/rocm-user/eigen/unsupported/test/cxx11_tensor_reduction_gpu.cu:16:
In file included from /home/rocm-user/eigen/unsupported/Eigen/CXX11/Tensor:29:
In file included from /home/rocm-user/eigen/unsupported/Eigen/CXX11/../SpecialFunctions:75:
/home/rocm-user/eigen/unsupported/Eigen/CXX11/../src/SpecialFunctions/arch/GPU/GpuSpecialFunctions.h:87:33: error: call to 'erf' is ambiguous
return make_double2(erf(a.x), erf(a.y));
^~~
/home/rocm-user/eigen/unsupported/test/../../Eigen/src/Core/MathFunctions.h:1603:8: note: candidate function [with T = double]
double erf(const double &x) { return ::erf(x); }
^
/home/rocm-user/eigen/unsupported/Eigen/CXX11/../src/SpecialFunctions/SpecialFunctionsImpl.h:1897:5: note: candidate function [with Scalar = double]
erf(const Scalar& x) {
^
3 errors generated.
```
This PR fixes the compile error by removing the "old" implementation for "erf" (assuming that the "new" implementation is what we want going forward. from a GPU point-of-view both implementations are the same).
This PR also fixes what seems like a cut-n-paste error in the aforementioned commit
- Split SpecialFunctions files in to a separate BesselFunctions file.
In particular add:
- Modified bessel functions of the second kind k0, k1, k0e, k1e
- Bessel functions of the first kind j0, j1
- Bessel functions of the second kind y0, y1
The fixes needed are
* adding EIGEN_DEVICE_FUNC attribute to a couple of funcs (else HIPCC will error out when non-device funcs are called from global/device funcs)
* switching to using ::<math_func> instead std::<math_func> (only for HIPCC) in cases where the std::<math_func> is not recognized as a device func by HIPCC
* removing an errant "j" from a testcase (don't know how that made it in to begin with!)
- Move colamd implementation in its own namespace to avoid polluting the internal namespace with Ok, Status, etc.
- Fix signed/unsigned warning
- move some ugly free functions as member functions
This actually fixes an issue in unit-test packetmath_2 with pcmp_eq when it is compiled with clang. When pcmp_eq(Packet4f,Packet4f) is used instead of pcmp_eq(Packet2d,Packet2d), the unit-test does not pass due to NaN on ref vector.
Depending on instruction set, significant speedups are observed for the vectorized path:
log1p wall time is reduced 60-93% (2.5x - 15x speedup)
expm1 wall time is reduced 0-85% (1x - 7x speedup)
The scalar path is slower by 20-30% due to the extra branch needed to handle +infinity correctly.
Full benchmarks measured on Intel(R) Xeon(R) Gold 6154 here: https://bitbucket.org/snippets/rmlarsen/MXBkpM
The vec_vsx_ld/vec_vsx_st builtins were wrongly used for aligned load/store. In fact, they perform unaligned memory access and, even when the address is 16-byte aligned, they are much slower (at least 2x) than their aligned counterparts.
For double/Packet2d vec_xl/vec_xst should be prefered over vec_ld/vec_st, although the latter works when casted to float/Packet4f.
Silencing some weird warning with throw but some GCC versions. Such warning are not thrown by Clang.
If no offset is given, them it should be zero.
Also passes full address to vec_vsx_ld/st builtins.
Removes userless _EIGEN_ALIGNED_PTR & _EIGEN_MASK_ALIGNMENT.
Removes unnecessary casts.
* an interface for SYCL buffers to behave as a non-dereferenceable pointer
* an interface for placeholder accessor to behave like a pointer on both host and device
1. Fix buggy pcmp_eq and unit test for half types.
2. Add unit test for pselect and add specializations for SSE 4.1, AVX512, and half types.
3. Get rid of FIXME: Implement faster pnegate for half by XOR'ing with a sign bit mask.
That was hurting users with compilers that would object to proceed with
that:
"""
./Eigen/src/Core/products/GeneralMatrixVector.h:356:10: error: declaration shadows a static data member of 'general_matrix_vector_product<type-parameter-0-0, type-parameter-0-1, type-parameter-0-2, 1, ConjugateLhs, type-parameter-0-4, type-parameter-0-5, ConjugateRhs, Version>' [-Werror,-Wshadow]
LhsPacketSize = Traits::LhsPacketSize,
^
./Eigen/src/Core/products/GeneralMatrixVector.h:307:22: note: previous declaration is here
static const Index LhsPacketSize = Traits::LhsPacketSize;
"""
This fixes compilation issues with RealScalar types that are not implicitly castable from Index (e.g. ceres Jet types).
Reported by Peter Anderson-Sprecher via eMail
Make clipping outside [-18:18] consistent for vectorized and non-vectorized paths of scalar_logistic_op<float>.
Approved-by: Gael Guennebaud <g.gael@free.fr>
This helps avoids a conflict on certain Windows toolchains
(potentially due to some ADL name resolution bug) in the case
where aligned_free is defined in the global namespace. In any
case, tightening this up is harmless.
Change licensing of OrderingMethods/Amd.h and SparseCholesky/SimplicialCholesky_impl.h from LGPL to MPL2.
Approved-by: Gael Guennebaud <g.gael@free.fr>
This is also important to make sure that A.conjugate() * B.conjugate() does not evaluate
its arguments into temporaries (e.g., if A and B are fixed and small, or * fall back to lazyProduct)
The row-major matrix-vector multiplication code uses a threshold to
check if processing 8 rows at a time would thrash the cache.
This change introduces two modifications to this logic.
1. A smaller threshold for ARM and ARM64 devices.
The value of this threshold was determined empirically using a Pixel2
phone, by benchmarking a large number of matrix-vector products in the
range [1..4096]x[1..4096] and measuring performance separately on
small and little cores with frequency pinning.
On big (out-of-order) cores, this change has little to no impact. But
on the small (in-order) cores, the matrix-vector products are up to
700% faster. Especially on large matrices.
The motivation for this change was some internal code at Google which
was using hand-written NEON for implementing similar functionality,
processing the matrix one row at a time, which exhibited substantially
better performance than Eigen.
With the current change, Eigen handily beats that code.
2. Make the logic for choosing number of simultaneous rows apply
unifiormly to 8, 4 and 2 rows instead of just 8 rows.
Since the default threshold for non-ARM devices is essentially
unchanged (32000 -> 32 * 1024), this change has no impact on non-ARM
performance. This was verified by running the same set of benchmarks
on a Xeon desktop.
Prior to this change, a product with a LHS having 8 rows was faster with AVX-only than with AVX+FMA.
With AVX+FMA I measured a speed up of about x1.25 in such cases.
This changeset also includes:
* add HouseholderSequence::conjugateIf
* define int as the StorageIndex type for all dense solvers
* dedicated unit tests, including assertion checking
* _check_solve_assertion(): this method can be implemented in derived solver classes to implement custom checks
* CompleteOrthogonalDecompositions: add applyZOnTheLeftInPlace, fix scalar type in applyZAdjointOnTheLeftInPlace(), add missing assertions
* Cholesky: add missing assertions
* FullPivHouseholderQR: Corrected Scalar type in _solve_impl()
* BDCSVD: Unambiguous return type for ternary operator
* SVDBase: Corrected Scalar type in _solve_impl()
