SWDEV-203438 - [HIP] AllGather RCCL test issue
The test tries to launch a kernel on two devices at once and they need to communicate with each other.
For that, it uses a custom stream for each devices.
Problem is in getNullStream we used to call syncStreams all the time
and it was syncing all the streams even the ones on different devices.
So that made the second kernel launch (on 2n dev) to wait for the first kernel to finish which
would never occur since the first one was waiting for the second one.
The fix is to not call syncStreams from getNullStream because we sync already anyway prior in general.
Affected files ...
... //depot/stg/opencl/drivers/opencl/api/hip/hip_context.cpp#21 edit
... //depot/stg/opencl/drivers/opencl/api/hip/hip_event.cpp#16 edit
... //depot/stg/opencl/drivers/opencl/api/hip/hip_internal.hpp#40 edit
... //depot/stg/opencl/drivers/opencl/api/hip/hip_memory.cpp#70 edit
... //depot/stg/opencl/drivers/opencl/api/hip/hip_module.cpp#41 edit
... //depot/stg/opencl/drivers/opencl/api/hip/hip_stream.cpp#24 edit
[ROCm/hip commit: 184ee3f5a0]
+ Add a corresponding matcher cudaDeviceFuncCall to match only (__device__ or __global__) and not __host__ functions.
+ Add a corresponding device functions mapping:
only unsupported are listed, cause supported are exactly the same as of CUDA and do not need transformation;
make FindAndReplace for device functions separated from host API calls.
+ Add a test to distinguish device functions and user-defined.
[ROCm/hip commit: 6602fadc16]
The math library equivalents between CUDA-HIP are broken. This is a key feature for converting to AMD hardware. This fix corrects the broken link and moves the library equivalents to sit under the "Porting a New Cuda Project" header.
[ROCm/hip commit: 5a6eafcbf1]
The implementation for OCKL AS was recently removed from the device
library since that feature is now superseded by hostcall.
[ROCm/hip commit: 70023c9075]
Currently hipcc uses -O3 for hip-clang by default but uses -O0 if -g is used. This
causes surprise for users since -g should not affect default opt level.
[ROCm/hip commit: b046ec698b]
* occupancy.cpp with Makefile
* occupancy sample changes according tothe comments
* Changes according to the review comments
* Occupancy Sample Changes
* Changes according to review comments
[ROCm/hip commit: f807cc1a7b]
* first cut of the header implementation of cooperative group feature
* add diclarations for device library functions
* fixed various compile time issues in the CG headers
* enabled copy construction and copy assignment
* fixed a minor bug related to conditional compilation macro
* fixed few more CG constructor issues and added a unit testcase
* fixed typo
* extended unit testcase
* compute size of partitioned CG from mask
* bit of code refactoring
* removed boilerplate code
* fixed few of the review comments by Brian
* Changes to the sigantures of few grid and multi-grid related OCKL functions
* changes to declarations of OCKL functions related to CG feature
* removed all the block level support as it is not planned for 2.9
* Have taken care of review comments by Brian
* Have taken care of review comments by Brian
* removed unused functions which were initially intended to use in block level cg support
[ROCm/hip commit: d75dc4eb29]
+ Start to translate preprocessor's false conditional blocks too:
based on clang's https://reviews.llvm.org/D66597;
available only starting from LLVM 10.0 or trunk.
+ Option -skip-excluded-preprocessor-conditional-blocks for skipping excluded conditional blocks:
the default behavior for hipify-clang built with LLVM < 10.0;
false by default for hipify-clang built with LLVM 10 or trunk.
+ Add 4 preprocessor unit tests, 2 of which are LLVM 10.0 only
+ Update couple of existing tests by setting -skip-excluded-preprocessor-conditional-blocks option:
update lit testing accordingly
[ROCm/hip commit: 24be21495d]
1. Fix setting std c++ (11|14)
2. Get rid of WIN32, use MSVC instead
3. Use VERSION_GREATER_EQUAL and VERSION_MAJOR instead of logical expressions
[ROCm/hip commit: eeb4452b23]
Evgeny Mankov
foreman
Rahul Garg
Aditya Lad
Jeff Daily
mhbliao
Paul Fultz II
Yaxun (Sam) Liu
Rahul Garg
Nicholas Malaya
Sarbojit2019
Sameer Sahasrabuddhe
satyanveshd
mshivama