The maxSharedMemoryPerMultiProcessor attribute is meant to describe
the number of bytes of shared memory (LDS space in AMD terminology)
in each SM (CU in AMD terminology). For instance, on AMD GPUs this
is often 64KB per CU, and some Nvidia GPUs it's 96KB per SM.
This shared memory is a different address space from the normal
global memory. However, the current HIP-HCC properties fill this
in with a size that matches the totalGlboalMem property. This gives
a drastically too-high calculation for the amount of LDS space that
each CU has -- tens of GBs vs. 10s of KBs.
This patch fixes this by pulling the maxSharedMemoryPerMultiProcessor
property from the HSA pool that describes how much workgroup-local
space is available on each CU. The HSA runtime eventually pulls
this from the topology information about LDSSizeInKB, defined as
"Size of Local Data Store in Kilobytes per SIMD".
Previously, this HSA query was used to fill in the value of the
sharedMemPerBlock property. On today's AMD GPUs, we know that
the amount of LDS avaialble to the workgroup is identical to the
amount of LDS space in the CU. However, in the future this may
differ. As such, this patch changes around the order and fills
in the "PerMultiProcessor" property from the HSA query (since
what's what the query is defined to return), and then separately
fills in the "PerBlock" property as we know it.
Fix two errors in hipOccupancyMaxActiveBlocksPerMultiprocessor.
1) Fix a possible segfault if the user passed in a null pointer for
the numBlocks value.
2) Handle the situation when the user is asking for a block size
that is larger than what the target device can hold within a
single block.
Currently there is a clang bug on Windows causing duplicate -mllvm options in clang -cc1.
Tempoarily disable -mllvm options for HIP-Clang on Windows until the bug is fixed.
Change-Id: I3a4393ba7745989398dc6c6001722837dad18704
- Need to check the availability of `__has_attribute` builtin macro
instead of compiler versions. That's more reliable and portable among
various compilers.
- Provides a very basic support of vectors for unknown compilers.
There are now two implementations of printf in HIP:
1. The implemenation for HCC is controlled by the HC_FEATURE_PRINTF
macro, and it works only with the HCC compiler used in combination
with the HCC runtime.
2. The implementation for hip-clang requires the VDI runtime, and is
always enabled with that combination.
There were several error messages that appeared even if the hipEnvVarDriver.exe test passes and executes successfully. Now it is cleaned up. The following are those instances:
* When popen searches for directed_test directory but does not find it, it outputs an error, then finds the hipEnvVar at the same level. Currently the fix will prompt the test to only output an error if both searches for hipEnvVar fails.
* When assertion is used towards the later half of the test, conditions were set to specifically hide the devices, resulting in No Hip Device detected in the latter half of the test. The fix will make these errors not appear as they are intended to not find any devices. Assertions themselves are untouched.
HipEnvVarDriver.cpp has also been refactored. Reading HipEnvVar will now happen in a helper function for getDeviceNumber and getDevicePCIBusNumRemote, as the code to read HipEnvVar were really similar in them.
The existing one can have issues on certain systems, therefore this limits use of direct memcpy via largeBAR to sizes where it is unequivocally better.
Also addresses SWDEV-220030 and SWDEV-222237.
This is a quick workaround to match HCC behavior for performance since inlining usually
results in more optimization opportunities therefore better performance.
We will fine tuning inline threashold later.
* Use deque instead of vector for code readers so that the iterators and references will be stable
* Fix compile error
* Assign the iterator
* Add multithreaded test
* Make threads a multiple of hardware concurrency
* Output on failure
* Add setDevice to try and initialize the context on cuda
* Create context for cuda
* Set context on each thread
* Reduce threads on cuda
* Skip test on cuda
* Try to initialize the primary context on cuda
* Push ctx to the stack as current
* Revert "Push ctx to the stack as current"
This reverts commit bff8cbe950.
* Revert "Try to initialize the primary context on cuda"
This reverts commit fd98514113.
* updated test for nvidia path
* Add c++11 option for nvcc
Co-authored-by: satyanveshd <53337087+satyanveshd@users.noreply.github.com>
Some PyTorch unit tests have regression. Disabling cov3 to allow more
time to debug and unblock PyTorch
Change-Id: Iba7f425ef3499c20c42ec45d9152b5d27ce97d03
Aryan Salmanpour
Joseph Greathouse
Yaxun (Sam) Liu
Maneesh Gupta
mhbliao
Evgeny Mankov
Alex Voicu
srinivamd
Nick Curtis
Sameer Sahasrabuddhe
Jatin Chaudhary
Siu Chi Chan
saleelk
Rahul Garg
jiabaxie
Sarbojit2019
eshcherb
vsytch