The following commit causes regressions in the ROCm/HIP support for Eigen
e55182ac09
I suspect the same breakages occur on the CUDA side too.
The above commit puts the EIGEN_CONSTEXPR attribute on `half_base` constructor. `half_base` is derived from `__half_raw`.
When compiling with GPU support, the definition of `__half_raw` gets picked up from the GPU Compiler specific header files (`hip_fp16.h`, `cuda_fp16.h`). Properly supporting the above commit would require adding the `constexpr` attribute to the `__half_raw` constructor (and other `*half*` routines) in those header files. While that is something we can explore in the future, for now we need to undo the above commit when compiling with GPU support, which is what this commit does.
This commit also reverts a small change in the `raw_uint16_to_half` routine made by the above commit. Similar to the case above, that change was leading to compile errors due to the fact that `__half_raw` has a different definition when compiling with DPU support.
If we have explicit conversion operators available (C++11) we define
explicit casts from bfloat16 to other types. If not (C++03), we don't
define conversion operators but rely on implicit conversion chains from
bfloat16 over float to other types.
Specialized `bfloat16_impl::float_to_bfloat16_rtne(float)` for normal floating point numbers, infinity and zero, in order to improve the performance of `bfloat16::bfloat16(const T&)` for integer argument types.
A reduction of more than 20% of the runtime duration of conversion from int to bfloat16 was observed, using Visual C++ 2019 on Windows 10.
Conversion from `bfloat16` to `float` and `double` is lossless. It seems natural to allow the conversion to be implicit, as the C++ language also support implicit conversion from a smaller to a larger floating point type.
Intel's OneDLL bfloat16 implementation also has an implicit `operator float()`: https://github.com/oneapi-src/oneDNN/blob/v1.5/src/common/bfloat16.hpp
- Guard fundamental types that are not available pre C++11
- Separate subsequent angle brackets >> by spaces
- Allow casting of Eigen::half and Eigen::bfloat16 to complex types
The original tensor casts were only defined for
`SrcCoeffRatio`:`TgtCoeffRatio` 1:1, 1:2, 2:1, 4:1. Here we add the
missing 1:N and 8:1.
We also add casting `Eigen::half` to/from `std::complex<T>`, which
was missing to make it consistent with `Eigen:bfloat16`, and
generalize the overload to work for any complex type.
Tests were added to `basicstuff`, `packetmath`, and
`cxx11_tensor_casts` to test all cast configurations.
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!)
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
This makes both the small and huge argument cases faster because:
- for small inputs this removes the last pselect
- for large inputs only the reduction part follows a scalar path,
the rest use the same SIMD path as the small-argument case.
- no FMA: 1ULP up to 3pi, 2ULP up to sin(25966) and cos(18838), fallback to std::sin/cos for larger inputs
- FMA: 1ULP up to sin(117435.992) and cos(71476.0625), fallback to std::sin/cos for larger inputs
