gtbercea (Gheorghe-Teodor Bercea)
User

Projects

User does not belong to any projects.

User Details

User Since
Dec 29 2016, 12:44 AM (49 w, 4 d)

Recent Activity

Mon, Dec 4

gtbercea added a comment to D38978: [OpenMP] Enable the lowering of implicitly shared variables in OpenMP GPU-offloaded target regions to the GPU shared memory.

ping

Mon, Dec 4, 10:43 AM

Tue, Nov 28

gtbercea added a comment to D38978: [OpenMP] Enable the lowering of implicitly shared variables in OpenMP GPU-offloaded target regions to the GPU shared memory.

@tra @hfinkel

Tue, Nov 28, 7:17 AM

Mon, Nov 27

gtbercea added inline comments to D14254: [OpenMP] Initial implementation of OpenMP offloading library - libomptarget device RTLs..
Mon, Nov 27, 5:42 PM · Restricted Project
gtbercea added a comment to D38978: [OpenMP] Enable the lowering of implicitly shared variables in OpenMP GPU-offloaded target regions to the GPU shared memory.

ping

Mon, Nov 27, 7:56 AM

Fri, Nov 24

gtbercea updated the diff for D40451: [OpenMP] Add function attribute for triggering shared memory lowering in the LLVM backend.
Fri, Nov 24, 4:01 PM
gtbercea created D40451: [OpenMP] Add function attribute for triggering shared memory lowering in the LLVM backend.
Fri, Nov 24, 3:58 PM
gtbercea updated the diff for D38978: [OpenMP] Enable the lowering of implicitly shared variables in OpenMP GPU-offloaded target regions to the GPU shared memory.

Add regression tests and allow for shared memory lowering to be disabled at function level.

Fri, Nov 24, 3:22 PM

Tue, Nov 21

gtbercea closed D38976: [OpenMP] Add implicit data sharing support when offloading to NVIDIA GPUs using OpenMP device offloading.
Tue, Nov 21, 7:55 AM

Mon, Nov 20

gtbercea accepted D40250: [OpenMP] Consistently use cubin extension for nvlink.

LG

Mon, Nov 20, 4:47 PM
gtbercea added inline comments to D40250: [OpenMP] Consistently use cubin extension for nvlink.
Mon, Nov 20, 10:33 AM
gtbercea added inline comments to D40250: [OpenMP] Consistently use cubin extension for nvlink.
Mon, Nov 20, 9:01 AM
gtbercea added inline comments to D40250: [OpenMP] Consistently use cubin extension for nvlink.
Mon, Nov 20, 8:46 AM

Nov 3 2017

gtbercea updated the diff for D38976: [OpenMP] Add implicit data sharing support when offloading to NVIDIA GPUs using OpenMP device offloading.

Remove blocks.

Nov 3 2017, 1:48 PM
gtbercea updated the diff for D38976: [OpenMP] Add implicit data sharing support when offloading to NVIDIA GPUs using OpenMP device offloading.
Nov 3 2017, 1:25 PM

Oct 18 2017

gtbercea updated the diff for D39061: [buildbot] Increase timeout for libomp-clang-ppc64le-linux-debian builder.
Oct 18 2017, 11:26 AM
gtbercea updated the diff for D39061: [buildbot] Increase timeout for libomp-clang-ppc64le-linux-debian builder.
Oct 18 2017, 11:06 AM
gtbercea added a reviewer for D39061: [buildbot] Increase timeout for libomp-clang-ppc64le-linux-debian builder: sfantao.
Oct 18 2017, 10:57 AM
gtbercea created D39061: [buildbot] Increase timeout for libomp-clang-ppc64le-linux-debian builder.
Oct 18 2017, 10:55 AM
gtbercea added a comment to D39005: [OpenMP] Clean up variable and function names for NVPTX backend.

I'd be interested to get the ball rolling in regard to coming up with a fix for this. I see some suggestions in past patches. Some help/clarification would be much appreciated.

Happy to help, but I'm not sure what to offer beyond the link in Art's previous comment.

Oct 18 2017, 8:17 AM

Oct 17 2017

gtbercea added a comment to D39005: [OpenMP] Clean up variable and function names for NVPTX backend.

Hi Artem, Justin,

Oct 17 2017, 2:19 PM
gtbercea updated the diff for D38978: [OpenMP] Enable the lowering of implicitly shared variables in OpenMP GPU-offloaded target regions to the GPU shared memory.

Eliminate variable and function name clean-up. That has been moved into a separate patch: D39005

Oct 17 2017, 8:27 AM
gtbercea created D39005: [OpenMP] Clean up variable and function names for NVPTX backend.
Oct 17 2017, 8:27 AM

Oct 16 2017

gtbercea updated the summary of D38976: [OpenMP] Add implicit data sharing support when offloading to NVIDIA GPUs using OpenMP device offloading.
Oct 16 2017, 2:30 PM
gtbercea created D38978: [OpenMP] Enable the lowering of implicitly shared variables in OpenMP GPU-offloaded target regions to the GPU shared memory.
Oct 16 2017, 2:29 PM
gtbercea created D38976: [OpenMP] Add implicit data sharing support when offloading to NVIDIA GPUs using OpenMP device offloading.
Oct 16 2017, 2:21 PM
gtbercea added a comment to D38883: [CMake][OpenMP] Customize default offloading arch.

LGTM

Oct 16 2017, 11:56 AM

Oct 13 2017

gtbercea added inline comments to D38883: [CMake][OpenMP] Customize default offloading arch.
Oct 13 2017, 11:41 AM
gtbercea added inline comments to D38883: [CMake][OpenMP] Customize default offloading arch.
Oct 13 2017, 11:39 AM
gtbercea added inline comments to D38883: [CMake][OpenMP] Customize default offloading arch.
Oct 13 2017, 11:19 AM
gtbercea added inline comments to D38883: [CMake][OpenMP] Customize default offloading arch.
Oct 13 2017, 11:17 AM
gtbercea added inline comments to D38883: [CMake][OpenMP] Customize default offloading arch.
Oct 13 2017, 11:16 AM
gtbercea added inline comments to D38883: [CMake][OpenMP] Customize default offloading arch.
Oct 13 2017, 11:05 AM
gtbercea added inline comments to D38883: [CMake][OpenMP] Customize default offloading arch.
Oct 13 2017, 11:04 AM
gtbercea added inline comments to D38883: [CMake][OpenMP] Customize default offloading arch.
Oct 13 2017, 11:02 AM

Sep 27 2017

gtbercea accepted D38258: [OpenMP] Fix passing of -m arguments to device toolchain.

LGTM

Sep 27 2017, 7:46 AM
gtbercea accepted D38259: [OpenMP] Fix translation of target args.

LGTM

Sep 27 2017, 7:42 AM
gtbercea added inline comments to D38258: [OpenMP] Fix passing of -m arguments to device toolchain.
Sep 27 2017, 7:40 AM
gtbercea accepted D38257: [OpenMP] Fix memory leak when translating arguments.

LGTM

Sep 27 2017, 7:37 AM
gtbercea closed D38040: [OpenMP] Add an additional test for D34888.
Sep 27 2017, 7:32 AM

Sep 26 2017

gtbercea added a reviewer for D38040: [OpenMP] Add an additional test for D34888: ABataev.
Sep 26 2017, 6:59 PM
gtbercea updated the diff for D38040: [OpenMP] Add an additional test for D34888.

Fix test.

Sep 26 2017, 6:58 PM
gtbercea reopened D38040: [OpenMP] Add an additional test for D34888.

Open

Sep 26 2017, 5:59 PM
gtbercea reopened D29660: [OpenMP] Add flag for overwriting default PTX version for OpenMP targets.

Open

Sep 26 2017, 3:30 PM
gtbercea closed D38040: [OpenMP] Add an additional test for D34888.
Sep 26 2017, 3:30 PM
gtbercea updated the diff for D38040: [OpenMP] Add an additional test for D34888.

Add nocudalib flag.

Sep 26 2017, 3:30 PM
gtbercea closed D37914: [OpenMP] Don't throw cudalib not found error if only front-end is required..
Sep 26 2017, 3:30 PM
gtbercea reopened D38040: [OpenMP] Add an additional test for D34888.

Open

Sep 26 2017, 3:30 PM
gtbercea closed D38040: [OpenMP] Add an additional test for D34888.
Sep 26 2017, 3:30 PM
gtbercea updated the diff for D38040: [OpenMP] Add an additional test for D34888.

Fix test.

Sep 26 2017, 3:30 PM

Sep 25 2017

gtbercea reopened D37914: [OpenMP] Don't throw cudalib not found error if only front-end is required..

Open.

Sep 25 2017, 2:59 PM
gtbercea closed D37913: [OpenMP] Enable the existing nocudalib flag for OpenMP offloading toolchain..
Sep 25 2017, 2:58 PM
gtbercea updated the diff for D37913: [OpenMP] Enable the existing nocudalib flag for OpenMP offloading toolchain..

Split line.

Sep 25 2017, 2:54 PM
gtbercea closed D37912: [OpenMP] Bugfix: output file name drops the absolute path where full path is needed..
Sep 25 2017, 2:27 PM
gtbercea closed D37914: [OpenMP] Don't throw cudalib not found error if only front-end is required..
Sep 25 2017, 2:08 PM
gtbercea added a comment to D38040: [OpenMP] Add an additional test for D34888.

The test is verifying whether the parameter is passed to the kernel correctly. I believe it was not passed as a reference before the patch.

Ah, right: This isn't checked anywhere before. Maybe add a comment about what's tested here?
Do we want to check the rest of the codegen with a focus that the variable is passed as a reference?

In addition to that, something that was in my previous patch is related to this code:

DSAStack->checkMappableExprComponentListsForDeclAtLevel(
        D, Level, [&](OMPClauseMappableExprCommon::MappableExprComponentListRef

In particular with the Level variable. Should the Level variable actually be Level + 1 in this case?

I'm not sure, the current public clang-ykt has Level: https://github.com/clang-ykt/clang/blob/d181aed/lib/Sema/SemaOpenMP.cpp#L1361

Sep 25 2017, 5:02 AM

Sep 21 2017

gtbercea added a comment to D38040: [OpenMP] Add an additional test for D34888.

Hi Doru,

if I remember correctly I submitted D34888 for a crash when mapping a scalar value with nested regions.
I've marked another test in this file that the codegen for tofrom is correct. So I don't know if this test checks some other conditions?

Jonas

Sep 21 2017, 6:08 PM

Sep 19 2017

gtbercea updated the diff for D37914: [OpenMP] Don't throw cudalib not found error if only front-end is required..
Sep 19 2017, 6:24 PM
gtbercea updated the diff for D38040: [OpenMP] Add an additional test for D34888.
Sep 19 2017, 6:12 PM
gtbercea updated the diff for D37913: [OpenMP] Enable the existing nocudalib flag for OpenMP offloading toolchain..

Don't take into account unknown CUDA archs not even for testing purposes.

Sep 19 2017, 6:06 PM
gtbercea updated the diff for D37912: [OpenMP] Bugfix: output file name drops the absolute path where full path is needed..

Address comment.

Sep 19 2017, 5:55 PM
gtbercea created D38040: [OpenMP] Add an additional test for D34888.
Sep 19 2017, 8:51 AM
gtbercea added a reviewer for D37912: [OpenMP] Bugfix: output file name drops the absolute path where full path is needed.: tra.
Sep 19 2017, 8:46 AM
gtbercea added a reviewer for D37913: [OpenMP] Enable the existing nocudalib flag for OpenMP offloading toolchain.: tra.
Sep 19 2017, 8:46 AM

Sep 18 2017

gtbercea added a reviewer for D37913: [OpenMP] Enable the existing nocudalib flag for OpenMP offloading toolchain.: hfinkel.
Sep 18 2017, 12:12 PM
gtbercea added inline comments to D37914: [OpenMP] Don't throw cudalib not found error if only front-end is required..
Sep 18 2017, 11:57 AM
gtbercea updated the diff for D37914: [OpenMP] Don't throw cudalib not found error if only front-end is required..

Only check for -S.

Sep 18 2017, 11:51 AM
gtbercea updated the diff for D37913: [OpenMP] Enable the existing nocudalib flag for OpenMP offloading toolchain..

Add test.

Sep 18 2017, 11:33 AM
gtbercea updated the diff for D37912: [OpenMP] Bugfix: output file name drops the absolute path where full path is needed..
Sep 18 2017, 9:26 AM

Sep 15 2017

gtbercea updated the diff for D37914: [OpenMP] Don't throw cudalib not found error if only front-end is required..

Fix diff.

Sep 15 2017, 2:40 PM
gtbercea updated the diff for D37914: [OpenMP] Don't throw cudalib not found error if only front-end is required..

Add test.

Sep 15 2017, 2:36 PM
gtbercea updated the diff for D37912: [OpenMP] Bugfix: output file name drops the absolute path where full path is needed..

Fix tests.

Sep 15 2017, 2:32 PM
gtbercea updated the diff for D37914: [OpenMP] Don't throw cudalib not found error if only front-end is required..

Fix condition.

Sep 15 2017, 1:45 PM
gtbercea updated the diff for D37914: [OpenMP] Don't throw cudalib not found error if only front-end is required..

Fix parantheses.

Sep 15 2017, 11:58 AM
gtbercea updated the diff for D37914: [OpenMP] Don't throw cudalib not found error if only front-end is required..

Contract check.

Sep 15 2017, 11:55 AM
gtbercea added a comment to D37912: [OpenMP] Bugfix: output file name drops the absolute path where full path is needed..
In D37912#872294, @tra wrote:

Shouldn't this temp .cubin file go into the temporary directory, as opposed to the same directory as the input file?

Sep 15 2017, 11:48 AM
gtbercea created D37914: [OpenMP] Don't throw cudalib not found error if only front-end is required..
Sep 15 2017, 11:43 AM
gtbercea created D37913: [OpenMP] Enable the existing nocudalib flag for OpenMP offloading toolchain..
Sep 15 2017, 11:36 AM
gtbercea created D37912: [OpenMP] Bugfix: output file name drops the absolute path where full path is needed..
Sep 15 2017, 11:29 AM

Aug 12 2017

gtbercea added a comment to D29660: [OpenMP] Add flag for overwriting default PTX version for OpenMP targets.

Couldn't fix/find the actual error so for now, just moving the flag patch tests to openmp-offload-gpu.c which is a disabled test.

310765

Bad news, the bot is still red: http://lab.llvm.org:8011/builders/sanitizer-x86_64-linux-fast/builds/7114

Disabled openmp-offload.c on Linux again: https://reviews.llvm.org/rL310772

Aug 12 2017, 11:03 AM

Aug 11 2017

gtbercea added a comment to D29660: [OpenMP] Add flag for overwriting default PTX version for OpenMP targets.
Aug 11 2017, 2:19 PM
gtbercea added a comment to D29660: [OpenMP] Add flag for overwriting default PTX version for OpenMP targets.

I have re-enabled the previous offloading tests and moved the new GPU offloading tests to a new file which is disabled for linux (for now).

310718

Alex thanks so much for the logs, they have been very useful to understand what's going on.

Aleksey, I have since tried to install a Clang version with the address sanitizer enabled but without much success. Apart from turning on the sanitizer in the cmake using the -DLLVM_USE_SANITIZER="Address" flag is there any other flag that I need to pass to cmake?
I am trying to run this on my macbook x86_64 and OS X 10.11. I am getting the following error when building the compiler:

[2966/4254] Linking CXX shared library lib/libc++abi.1.0.dylib
FAILED: lib/libc++abi.1.0.dylib
Undefined symbols for architecture x86_64:

"___asan_after_dynamic_init", referenced from:
    __GLOBAL__sub_I_cxa_default_handlers.cpp in cxa_default_handlers.cpp.o
"___asan_before_dynamic_init", referenced from:
    __GLOBAL__sub_I_cxa_default_handlers.cpp in cxa_default_handlers.cpp.o

[...]
ld: symbol(s) not found for architecture x86_64

Actually, you can run our bot, it is in zorg (http://llvm.org/git/zorg.git), zorg/buildbot/builders/sanitizers/buildbot_fast.sh (the one I linked the last time).

Create a temp folder and from that folder run:
BUILDBOT_REVISION= BUILDBOT_CLOBBER= $PATH_YOUR_PROJECTS$/zorg/zorg/buildbot/builders/sanitizers/buildbot_fast.sh

Aug 11 2017, 1:38 PM
gtbercea added a comment to D29660: [OpenMP] Add flag for overwriting default PTX version for OpenMP targets.

I have re-enabled the previous offloading tests and moved the new GPU offloading tests to a new file which is disabled for linux (for now).

Aug 11 2017, 9:08 AM

Aug 10 2017

gtbercea added a comment to D29660: [OpenMP] Add flag for overwriting default PTX version for OpenMP targets.

First of all, I apologize if I've upset you with my previous post. I am actively working on understanding what is causing these issues. It is not my intention to write tests that work on local configurations only. I am upset to see that these tests keep failing for your and maybe other configurations. Without knowing the actual reason of the failures I can only speculate what is going wrong with them hence the flurry of changes.

Thank you, apology accepted. That was exactly my point, not to start a fight, but to emphasize that depending on local configuration is never going to work, you will never be able to see and test all of them. Please disable the test ASAP and until the better way to handle it is determined.

Aug 10 2017, 10:00 AM
gtbercea added a comment to D29660: [OpenMP] Add flag for overwriting default PTX version for OpenMP targets.

The failures were very widespread, e.g. there's a linux buildbot that was red until the revert: http://bb.pgr.jp/builders/test-clang-i686-linux-RA. If you have access to a linux machine you should be able to reproduce the failures that the bot experienced by using the same cmake arguments (I don't know the exact ones, but judging from the bot you should be able to reproduce them using 32 bit release build with assertions enabled). I don't know what GPU that buildbot has.

Aug 10 2017, 9:24 AM
gtbercea added a comment to D29660: [OpenMP] Add flag for overwriting default PTX version for OpenMP targets.

First of all, I apologize if I've upset you with my previous post. I am actively working on understanding what is causing these issues. It is not my intention to write tests that work on local configurations only. I am upset to see that these tests keep failing for your and maybe other configurations. Without knowing the actual reason of the failures I can only speculate what is going wrong with them hence the flurry of changes.

Aug 10 2017, 8:35 AM
gtbercea added a comment to D29660: [OpenMP] Add flag for overwriting default PTX version for OpenMP targets.
Aug 10 2017, 6:44 AM

Aug 9 2017

gtbercea added a comment to D29660: [OpenMP] Add flag for overwriting default PTX version for OpenMP targets.

310549 should solve this problem by using a default architecture that is supported by the underlying device version.

Aug 9 2017, 10:04 PM
gtbercea added a comment to D29660: [OpenMP] Add flag for overwriting default PTX version for OpenMP targets.

Thanks for running the test on your machine! This is very useful.

Aug 9 2017, 8:15 PM
gtbercea added a comment to D29660: [OpenMP] Add flag for overwriting default PTX version for OpenMP targets.

I've removed that test. Let's see if the other two tests pass or not. (310537)

Aug 9 2017, 4:50 PM
gtbercea added a comment to D29660: [OpenMP] Add flag for overwriting default PTX version for OpenMP targets.

Even after r310505, openmp-offload.c continues to haunt our bots, for example http://lab.llvm.org:8011/builders/sanitizer-x86_64-linux-bootstrap/builds/2012. Can you please fix this test?

Aug 9 2017, 1:51 PM
gtbercea closed D36537: [OpenMP] Enable executable lookup into driver directory..
Aug 9 2017, 12:53 PM
gtbercea updated the summary of D36537: [OpenMP] Enable executable lookup into driver directory..
Aug 9 2017, 12:03 PM
gtbercea updated the diff for D36537: [OpenMP] Enable executable lookup into driver directory..

Add comment.

Aug 9 2017, 12:02 PM
gtbercea created D36537: [OpenMP] Enable executable lookup into driver directory..
Aug 9 2017, 11:50 AM
gtbercea added a comment to D29660: [OpenMP] Add flag for overwriting default PTX version for OpenMP targets.

Revision 310505 fixes the tests for this patch.

Aug 9 2017, 11:30 AM
gtbercea added a comment to D29660: [OpenMP] Add flag for overwriting default PTX version for OpenMP targets.

Looks like this test is failing on macOS again after this change:

http://lab.llvm.org:8080/green/job/clang-stage1-cmake-RA-incremental_check/39231/testReport/Clang/Driver/openmp_offload_c/

Can you please take a look?

Aug 9 2017, 10:50 AM
gtbercea closed D29905: [OpenMP] Pass argument to device kernel by reference when map is used. .

Already covered by D34888

Aug 9 2017, 9:18 AM
gtbercea closed D29660: [OpenMP] Add flag for overwriting default PTX version for OpenMP targets.
Aug 9 2017, 8:57 AM
gtbercea closed D29659: [OpenMP] Add flag for disabling the default generation of relocatable OpenMP target code for NVIDIA GPUs..
Aug 9 2017, 8:28 AM

Aug 8 2017

gtbercea added a comment to D29654: [OpenMP] Integrate OpenMP target region cubin into host binary.

I have just pushed a fix, revision 310433.

Aug 8 2017, 6:05 PM