The device side assertion calls printf to write out a message. In the
device compiler, printf is expanded into a series of hostcalls that
transmit the printf payload to the host. This expansion increases the
length of the kernel, resulting in sub-optimal compilation. The
solution is to ensure that the assert() implementation is not inlined
into the kernel.
Change-Id: Ia3a075461a755cf007218f262b0863e1926c76aa
[ROCm/clr commit: 160dfb5a1f]
With the PAL_ALWAYS_RESIDENT flag memory objects are resident at allocation time, no need to make them resident again before submit.
Also we should never evict anything with this setting, or we'll generate a VM fault.
Change-Id: Ieacc6af88ab4e09c20efd94100e148b2502e1d70
[ROCm/clr commit: fd09a7a23c]
hipStreamDestroy() by definition doesn't guarantee waiting for all stream operations.
Change-Id: Ibfd21c48d71741ebcbcb9898c8b5ac235bdc570b
[ROCm/clr commit: f2c2fd3d09]
Add extra API calls for managed memory: hipMemAdvise, hipMemPrefetchAsync,
hipMemRangeGetAttribute
Change-Id: Ifb09744acdf7582d278453707373c89c30e3ed03
[ROCm/clr commit: 4ee6d78763]
Positive and negative test cases for atomicAdd and
AtomicAddnoRet device functions
SWDEV-238517 for enhancing hip unit tests
Change-Id: Id20ba2550d20f224004f105cdcd087002cb80e56
[ROCm/clr commit: 00d46218b7]
Additional tests for following APIs
- hipOccupancyMaxActiveBlocksPerMultiprocessor
- hipOccupancyMaxPotentialBlockSize
SWDEV-238517 for enhancing hip unit tests
Change-Id: Ib9441c1366f46a082e10eb1a572bc7d8ebe1ee37
[ROCm/clr commit: 4eb65d58dd]
Currently all the logic checks if __HIP_ENABLE_PCH is set, so setting it
to OFF will still enable PCH.
Switching __HIP_ENABLE_PCH to be a CMake option will enable correct
build behaviour.
Change-Id: I77a663589ffb5f2595c6ad3d144eb9466adcf250
[ROCm/clr commit: 8034467ec3]
The change reuses HSA signals for dispatches as a wait signal.
Skipping the barrier requires to disable L2 cache for sysmem
allocations and extra tracking for HDP access with the large bar.
ROC_BARRIER_SYNC=0 activates the new logic. Barrier sync is
still used by default.
ROC_ACTIVE_WAIT=1 enables unconditional active wait in ROCr.
The change also consolidated ROCr wait logic under single function.
Change-Id: I6bd1be30aa88258da1b1f9de319ef5a45852afd8
[ROCm/clr commit: d9397590de]
Unit testing of CG threadRank api in multi-grid environment requires
some changes as grid id of particular device is not known apriori.
Change-Id: Ie7941a09337653f3ada5eedfb7c64a93a234498b
[ROCm/clr commit: 9c86cdb89d]
- Matching the output "hipcc --version" for Installed directory
to get the llvm compiler path, wasnt done properly.
- Adding fallback for HIP_CLANG_ROOT on ROCM_PATH if defined in env
- Changed the mode of get_filename_component to DIRECTORY
Signed-off-by: Pruthvi Madugundu <pruthvi.madugundu@amd.com>
Change-Id: Ia38e234f34cabb5b419906c7094b327083292b0e
[ROCm/clr commit: 5f667bf4e8]
[background] There is currently a compilation issue when both hip/hip_runtime.h and hip/hip_cooperative_groups.h included
in a file and hipcc used to compile it on NV platform. The issue is that an abort is defined in hip/nvcc_detail/hip_runtime.h
and it is also defined in the CUDA cooperative groups header (/cuda/include/cooperative_groups/details/helpers.h).
this is problematic and leads to a compilation issue in hipcc on NV platform.
Change-Id: I2ab6982ac4103822a1a4a0ced942cd604d6c19c1
[ROCm/clr commit: 35a3304acc]
The static members __HIP_Coordinates::x, ::y and ::z must be defined
outside the class. Otherwise, linker throws `undefined reference error`
when these definitions are needed in the HIP application.
Change-Id: Iabc09744b478c22e4b13cf9824877ec9cfdd4f7a
[ROCm/clr commit: 9b6d26f8b2]
Tao Sang
Chauncey Hui
Sameer Sahasrabuddhe
Yaxun (Sam) Liu
Vladislav Sytchenko
Jason Tang
Payam
German Andryeyev
sumanthtg
Sourabh Betigeri
Sarbojit Sarkar
kjayapra-amd
Rahul Garg
Jatin Chaudhary
mshivama
Pruthvi Madugundu
Srinivasan Subramanian
Aryan Salmanpour
Cole Nelson