This PR is a follow-up on PR# #1698 and it makes two more APIs (hipLaunchCooperativeKernel/hipLaunchCooperativeKernelMultiDevice) inline so that they can work correctly with lazy binding.
SWDEV-151670: Issue with 3D texture with 4 components
SWDEV-151671: Issue with 2D layered texture with 4 components
Fixed memcpy when memory is allocated with driver API's.
Github issues: #1755
Fixed 3D default case when array type is not set during memory allocation.
Reverting #1673, #1697 and #1707.
Support for hipMemcpyWithStream and memcpy optimizations, will be brought in again once issues seen with these are resolved independently.
[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.
* hipMemset et al can use HSA API directly for synchronous cases
* lock and flush stream in hipMemset, hold lock until complete
* move hipMemset async check to front of conditional
* use hsa_amd_memory_fill for additional sync memset cases
code cleanup/review for all memset calls
* Fix inversion of execution mutating value.
* ihipMemsetSync fall back to kernel if HSA memset fails
* Never fallback, never surrender.
* Allow NULL stream.
* Optimise memset kernel. Remove deadwood.
* Update hip_memory.cpp
* Clean up stream logic in sync memset
* Revert "Clean up stream logic in sync memset"
This reverts commit 6117dedf673367f44cc704192573a117a3d92477.
Texture 2D image mapping for pitched arrays:
github issue: Texture Object's Buffer seems to be Misaligned #886
JIRA ticket: SWDEV-199313
SWDEV-151670 : Fixed issue with 3D texture with 4 components
SWDEV-151671 : Issue with 2D layered texture with 4 components
The lifetime of the buffer given to
hsa_code_object_reader_create_from_memory must exceed that of the
code object reader. We need to create a copy of the code object
binary memory (file) that is kept allocated until the code object
reader is destroyed.
* [hip] add support for implicit kernel argument for multi-grid sync
* modified code for calculating the prev_sum
* change the impCoopArg type to size_t
* add memory clean up
* launch init_gws and main kernels into two separate loops
Added new memory API's hipMemAllocPitch, hipMemAllocHost, hipMemsetD16, hipMemsetD16Async, hipMemsetD8Async
Modified to support all scenarios hipMemcpyParam2DAsync, hipMemcpyParam2D.
satyanveshd
Rahul Garg
Evgeny Mankov
Aryan Salmanpour
ansurya
saleelk
Alex Voicu
Jeff Daily
Sarbojit2019
Rahul Garg
Laurent Morichetti
Anusha Godavarthy Surya
searlmc1
Nick Curtis
Vladislav Sytchenko
Evgeny
mhbliao
Siu Chi Chan