[Background] it was found that if lazy linking used for a library that calls hipExtLaunchMultiKernelMultiDevice API then this API can get the wrong program_state object for looking up device kernels leading to a "No device code available" error in this API.
To fix this issue, the API was refactored to be inline and get and pass the correct program_state to an internal hip API to request a multi-device kernel launch.
[ROCm/hip commit: 68cc787781]
SWDEV-212749:
o Recent changes to “add support for extended launch” require hip_runtime.h to be include in hip_ext.h
o Order in which external applications include hip_hcc.h/hip_runtime.h causes compilation failure
[ROCm/hip commit: e60dec51da]
Handled the HCC version check appropriately as few of the directed tests (SWDEV-212161) were failing when hcc was bumped to 3.0.
[ROCm/hip commit: 6b06911ef1]
This will fix issue #1621. It also adds tests for is_callable with c++11, c++14, and c++17.
The fallback implementation was completely broken so I rewrote it so it pass the tests as well. This should be used instead of PR #1631.
[ROCm/hip commit: 8519a1411c]
- Fix 2 runtime API prototypes
`hipOccupancyMaxActiveBlocksPerMultiprocessor` and
`hipOccupancyMaxActiveBlocksPerMultiprocessorWithFlags`
- Add missing function templates of them in hip-clang.
[ROCm/hip commit: 5c8a7521f4]
This fixes a deadlock introduced by the switch to TTAS loops, and is therefore mildly urgent (to prevent the CI from hoovering in the broken code).
[ROCm/hip commit: a855a13c22]
Adds hipMemcpy2DFromArray and hipMemcpy2DFromArrayAsync equivalent to cudaMemcpy2DFromArray and cudaMemcpy2DFromArrayAsync.
[ROCm/hip commit: 356765a223]
* Make CAS loops use the TTAS idiom.
* More efficient re-formulation of TTAS.
* Fix typo.
* The typo was not quite a typo
[ROCm/hip commit: 9ba25b42c8]
* [HIP] Introduce library_types.h as a common header for libs
[Reason]
Currently, hipFFT, hipBLAS and other HIP libs use their own data types, prefixed with HIPFFT or HIPBLAS, whereas in CUDA those types are common and declared in library_types.h
[TODO]
Switch hipFFT, hipBLAS and other HIP libs to use common library_types.h.
* [HIP] Move include for library_types.h to hip_runtime.h
[Reason]
Repeat CUDA's behaviour, where library_types.h is included in cuda_runtime.h
[ROCm/hip commit: 94eb4155dd]
This fixes the usage of an uninitialized cdattr variable in hipDeviceGetAttribute for the CUDA backend when taking the switch default, as detailed in #1317.
Note that the directed_tests/runtimeApi/device/hipGetDeviceAttribute.tst test fails for me, but it already did before applying this patch. Let's see what CI says!
[ROCm/hip commit: 9ababa4276]
Added new memory API's hipMemAllocPitch, hipMemAllocHost, hipMemsetD16, hipMemsetD16Async, hipMemsetD8Async
Modified to support all scenarios hipMemcpyParam2DAsync, hipMemcpyParam2D.
[ROCm/hip commit: ba9c6e13e4]
Aryan Salmanpour
Maneesh Gupta
Alex Voicu
ansurya
satyanveshd
Rahul Garg
Paul Fultz II
mhbliao
Nick Curtis
Sarbojit2019
Michael LIAO
Anusha Godavarthy Surya
gandryey
searlmc1
Vladislav Sytchenko
Evgeny Mankov
Philip Salzmann
Yaxun (Sam) Liu