boost cmake version for Tensile 4.2 #961

ce1adon · 2021-06-01T23:33:18Z

To issue #960

2021/6/11 blocker #980

junliume

@jerryyin could you comment on the selection of cmake version for libMLIRMIOpen?

junliume · 2021-06-02T18:53:56Z

Dockerfile

@@ -80,7 +80,7 @@ RUN cget -p $PREFIX init --cxx /opt/rocm/llvm/bin/clang++ --std=c++14 -DAMDGPU_T
 # Install dependencies
 RUN cget -p $PREFIX install pfultz2/rocm-recipes
 # Install a newer version of cmake for libMLIRMIOpen
-RUN cget -p $PREFIX install kitware/cmake@v3.13.4
+RUN if [ "$USE_TARGETID" = "ON" ] ; then cget -p $PREFIX install kitware/cmake@v3.15.1; else cget -p $PREFIX install kitware/cmake@v3.13.4; fi


@ce1adon It's more a question to discuss how do we determine the versions for Common Development Environment (CDE) and Common Runtime Environment (CRE)?
https://pkgs.org/download/cmake and https://cliutils.gitlab.io/modern-cmake/chapters/intro/installing.html as references for cmake versions:
Ubuntu 18.04 LTS has 3.10.2
Ubuntu 20.04 LTS has 3.16.3
RHEL/CentOS well being RHEL/CentOS 2.8.12.2 (7)

The two versions mentioned in this change, however:
3.13.4 seems to be with Ubuntu 19.04, which EOL on Jan 23rd 2021 or Debian 10
Why 3.15.1? Is it selected based on a specific feature, or in other words, why not 3.16.3 or even 3.20.3 directly?

3.15.1 is the minimum version for the latest tensile to build.
3.20.x is not working, too new for latest tensile build.
3.16.x may work but there may be concern about its compatibility with ubuntu 18.04.

Okay. No objections to this PR, just curious how we determine these dependencies and their long term effects on support.

Hopefully @pfultz2 can give some comment on this question.

@junliume Sorry I was busy with the TF/MLIR triaging and didn't got a chance to look at this. I'm a bit concerned at this comment:

3.20.x is not working, too new for latest tensile build.

The cmake required to build the latest LLVM usually bump to the latest version. I can forsee this causing conflict between libMLIRMIOpen and Tensile.

There's a possibility that we have to bump it to use 3.15.1 at all scenarios for next MLIR PRs, we'll see.

Yes. I do not see anything that enforces us to stay with 3.13.4. Let's switch to 3.15 globally.

#978 (that bumps CMake to 3.15.1 unconditionally) is merged in.

junliume

If the rest of the stack are all using cmake 3.15.1 then it is fine with us. Consistency and stability are my only concerns.

ce1adon · 2021-06-03T18:10:01Z

If the rest of the stack are all using cmake 3.15.1 then it is fine with us. Consistency and stability are my only concerns.

@pfultz2 Hey Paul, I created issue #960 and PR #961 for switching cmake version. Do you have any idea about the impact on MIOpen?

pfultz2 · 2021-06-03T22:40:34Z

Ideally, all of our components should use the cmake version in 18.04(with the exception of LLVM), and most of our components follow that. It is concerning that tensile requires such a new version. The entire rocm stack needs to decide the minimum version of cmake we will use and then there should be a process in place which includes all component owners to consider a newer version.

For now, we have no choice but to upgrade the cmake version in order to build tensile, but there should be a larger discussion about what minimum version of cmake we plan to support in rocm.

junliume

Comments from MLIR, blocking this PR for now

jerryyin · 2021-06-04T16:48:36Z

@junliume I think this PR is good to go with the result here. @krzysz00 confirmed that 3.15.1 successfully builds/installs libMLIRMIOpen.a so we are good.

On the other hand, long term wise, we will plan on exporting artifact from MLIR CI so MIOpen doesn't have to rebuild every time.

junliume

Unblock it for comments from MLIR.
3.15.1 is still a little bit awkward, while again as long as overall consistency is kept.

P.S. this page is outdated: https://github.com/ROCmSoftwarePlatform/Tensile/wiki/Dependencies

krzysz00 · 2021-06-04T17:56:08Z

Y'all could go for any 3.15.x

We're good through 3.20

atamazov · 2021-06-07T14:57:28Z

CI errors.

junliume · 2021-06-07T17:20:27Z

CI errors.

Strange that boost cmake version can cause Memory access fault

Memory access fault by GPU node-2 (Agent handle: 0x278a6d0) on address 0x7f6d32400000. Reason: Page not present or supervisor privilege.

ce1adon · 2021-06-07T21:41:57Z

CI errors.

@atamazov For some reason the CI is no longer print out the configures that failed in conv2d and immed_con2d, which makes the tracking of the issue very difficult. Is it possible to enable printing out of the driver cmd in CI?

atamazov · 2021-06-07T22:24:40Z

@ce1adon

@atamazov For some reason the CI is no longer print out the configures that failed in conv2d and immed_con2d, which makes the tracking of the issue very difficult. Is it possible to enable printing out of the driver cmd in CI?

WRT "Fp16 Hip Tensile All gfx908" that failed at run #3. I think that adding the following:

-DMIOPEN_TEST_FLAGS='--verbose --disable-verification-cache'

to this line: https://github.com/ROCmSoftwarePlatform/MIOpen/blob/01ff3da68b68b2ffbc31cb681184a8082ac7ce26/Jenkinsfile#L843

should add necessary printing. The line have been added in #556.

🔴 Please note that we should use --disable-verification-cache in all conv* tests (I do not remember exact reasoning, but otherwise CI can run into false verification failures).

atamazov · 2021-06-08T16:32:39Z

Please review.

junliume · 2021-06-10T04:16:19Z

@ce1adon @atamazov could you take a look at the CI failure issues?

ce1adon · 2021-06-10T17:10:50Z

@ce1adon @atamazov could you take a look at the CI failure issues?

I've updated the branch to develop with verbose flag. Hopefully it will give some clue.

atamazov · 2021-06-10T17:16:20Z

test_immed_conv2d was created some time ago by copy-pasting from test_conv2d. At that time there was no valid support for logging. Later, I added acceptable logging to test_conv2d, but test_immed_conv2d still lacks it.

Therefore I recommend using MIOPEN_LOG_LEVEL=6 for triaging the issue. Is it reproducible locally?

ce1adon · 2021-06-10T17:49:13Z

Please keep both 7 and 8 build alive. They are testing different stages.

atamazov · 2021-06-10T17:55:11Z

@ce1adon I suspect ther ecould be some problem with some other algorithm, Maybe it's worth disabling ALL algorithms except GEMM for MIOpenTensile "Full" stages? This way, the problems in other algorithms won't affect these stages.

I am looking into ways of implementing this.

ce1adon · 2021-06-10T18:09:13Z

@ce1adon I suspect ther ecould be some problem with some other algorithm, Maybe it's worth disabling ALL algorithms except GEMM for MIOpenTensile "Full" stages? This way, the problems in other algorithms won't affect these stages.

I am looking into ways of implementing this.

Good thoughts. Will try to prove this theory locally.

ce1adon · 2021-06-10T22:13:41Z

As @atamazov said, cmd "MIOPEN_DEBUG_HIP_KERNELS=0 MIOPEN_LOG_LEVEL=5 CTEST_PARALLEL_LEVEL=4 MIOPEN_CONV_PRECISE_ROCBLAS_TIMING=0 MIOPEN_DEBUG_CONV_FFT=0 MIOPEN_DEBUG_CONV_DIRECT=0 MIOPEN_DEBUG_CONV_WINOGRAD=0 MIOPEN_DEBUG_CONV_IMPLICIT_GEMM=0 MIOPEN_DEBUG_CONV_SCGEMM=0 make -j check" actually passed the test so the issue is not in GEMM.

I ran the tests with log level at 6 and got the following error msg. I will share the full log through email.

MIOpen(HIP): Info [GetWrwSolutions] 
MIOpen(HIP): Info2 [FindRecordUnsafe] Looking for key 1-26-26-5x5-128-56-56-1-0x0-2x2-1x1-0-NCHW-FP16-W in file /dockerx/MIOpen/src/kernels/gfx90878.HIP.2_12_0.ufdb.txt
MIOpen(HIP): Info2 [FindRecordUnsafe] Key match: 1-26-26-5x5-128-56-56-1-0x0-2x2-1x1-0-NCHW-FP16-W
MIOpen(HIP): Info2 [FindRecordUnsafe] Contents found: miopenConvolutionBwdWeightsAlgoImplicitGEMM:ConvAsmImplicitGemmGTCDynamicWrwXdlops,0.058538,12800,miopenConvolutionBwdWeightsAlgoImplicitGEMM,<unused>;miopenConvolutionBwdWeightsAlgoGEMM:GemmWrwUniversal,0.315882,4326400,miopenConvolutionBwdWeightsAlgoGEMM,<unused>;miopenConvolutionBwdWeightsAlgoDirect:ConvDirectNaiveConvWrw,2.67181,0,miopenConvolutionBwdWeightsAlgoDirect,<unused>
MIOpen(HIP): Info2 [Measure] Db::FindRecord time: 0.485527 ms
MIOpen(HIP): Info2 [GetWrwSolutionWorkspaceSize] solver_id = ConvAsmImplicitGemmGTCDynamicWrwXdlops
MIOpen(HIP): Info [CompileWrwSolution] solver_id = ConvAsmImplicitGemmGTCDynamicWrwXdlops
MIOpen(HIP): Info2 [GetInvoker] Returning an invoker for problem 1x26x26x5x5x128x56x56x1xNCHWxFP16x0x0x2x2x1x1x1xW and solver ConvAsmImplicitGemmGTCDynamicWrwXdlops
MIOpen(HIP): Info [ConvolutionWrwImmediate] solver_id = ConvAsmImplicitGemmGTCDynamicWrwXdlops, workspace = 12800
MIOpen(HIP): Info2 [GetInvoker] Returning an invoker for problem 1x26x26x5x5x128x56x56x1xNCHWxFP16x0x0x2x2x1x1x1xW and solver ConvAsmImplicitGemmGTCDynamicWrwXdlops
Memory access fault by GPU node-2 (Agent handle: 0x14faf10) on address 0x7f6436e05000. Reason: Page not present or supervisor privilege.
CMake Error at test_test_immed_conv2d.cmake:7 (message):
  Test failed

atamazov · 2021-06-11T14:06:49Z

@ce1adon I will provide you with the necessary patch (that disables all but GEMM for conv tests) soon.

atamazov · 2021-06-11T14:25:10Z

@ce1adon Please let me know if you would like me to update this branch directly. I will give you a .diff file otherwise.

atamazov · 2021-06-11T14:41:21Z

@ce1adon Here it is: test_conv_all_miotensile_only_gemm.diff.txt. Currently it affects test_conv2d/3d and test_immed_conv2d/3d only.

ce1adon · 2021-06-11T17:04:55Z

@ce1adon Here it is: test_conv_all_miotensile_only_gemm.diff.txt. Currently it affects test_conv2d/3d and test_immed_conv2d/3d only.

@atamazov This is not going to work since some configures requires workspace that exceeds the memory limitation in GEMM path. Let's pin relevant author to solve the root cause in ConvAsmImplicitGemmGTCDynamicWrwXdlops, unless there are other concerns.

atamazov · 2021-06-11T19:19:46Z

@ce1adon

@atamazov This is not going to work since some configures requires workspace that exceeds the memory limitation in GEMM path.

Please use MIOPEN_USE_MIOPENTENSILE from config.h and skip configures that not applicable for GEMM. See how it is done under #if TEST_DIRECT_SUPPORTED_CONFIG_ONLY and under #if WORKAROUND_MI100_ROM37_HIP_COMPILER_CRASH in conv_common.hpp. I can do that for you, if you wish.

ce1adon · 2021-06-11T19:30:03Z

@ce1adon

@atamazov This is not going to work since some configures requires workspace that exceeds the memory limitation in GEMM path.

Please use MIOPEN_USE_MIOPENTENSILE from config.h and skip configures that not applicable for GEMM. See how it is done under #if TEST_DIRECT_SUPPORTED_CONFIG_ONLY and under #if WORKAROUND_MI100_ROM37_HIP_COMPILER_CRASH in conv_common.hpp. I can do that for you, if you wish.

Thank you Artem. These are all good thoughts. But I'm afraid I'm not planning to add workarounds to let this test pass. I would prefer to wait for solution to the blocker. Thanks for understanding!

atamazov · 2021-06-11T19:34:13Z

No problem. But please note that "my" solution improves the quality of miopentensile testing. It makes sure that GEMM will be always fully validated. Without it, the "best" solver will be verified.

ce1adon · 2021-06-11T20:16:16Z

No problem. But please note that "my" solution improves the quality of miopentensile testing. It makes sure that GEMM will be always fully validated. Without it, the "best" solver will be verified.

Sure. This is in fact my original proposal to run miopentensile test one year ago. Since the test rule has been settled, let's follow it.

atamazov · 2021-06-20T13:30:01Z

Merge conflicts.

boost cmake version for Tensile 4.2

85176b9

ce1adon requested review from pfultz2, junliume and atamazov June 1, 2021 23:33

ce1adon added this to the ROCm 4.4 milestone Jun 1, 2021

ce1adon added the urgency_high label Jun 1, 2021

This comment has been minimized.

Sign in to view

junliume reviewed Jun 2, 2021

View reviewed changes

junliume mentioned this pull request Jun 3, 2021

Update cmake version to 3.15.1 #960

Closed

junliume approved these changes Jun 3, 2021

View reviewed changes

junliume requested changes Jun 4, 2021

View reviewed changes

jerryyin self-requested a review June 4, 2021 16:46

jerryyin previously approved these changes Jun 4, 2021

View reviewed changes

junliume previously approved these changes Jun 4, 2021

View reviewed changes

update miopentensile version

bddca04

ce1adon dismissed stale reviews from junliume and jerryyin via bddca04 June 7, 2021 23:21

atamazov added the TESTING_CI_PASSED label Jun 8, 2021

ce1adon added 2 commits June 10, 2021 10:02

add verbose flag to miopentensile tests

3946e8c

Merge branch 'develop' into cmake_ver

e584b70

Merge branch 'develop' into cmake_ver

3af191b

Merge branch 'develop' into cmake_ver

101ae53

atamazov approved these changes Jun 20, 2021

View reviewed changes

atamazov merged commit fca9fed into develop Jun 21, 2021

atamazov deleted the cmake_ver branch June 23, 2021 13:17

atamazov pushed a commit that referenced this pull request Jul 22, 2021

Bump cmake version for Tensile 4.2 (#961)

0678f8f

boost cmake version for Tensile 4.2 #961

boost cmake version for Tensile 4.2 #961

Conversation

ce1adon commented Jun 1, 2021 • edited Loading

This comment has been minimized.

junliume left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

junliume left a comment

Choose a reason for hiding this comment

ce1adon commented Jun 3, 2021

pfultz2 commented Jun 3, 2021

junliume left a comment

Choose a reason for hiding this comment

jerryyin commented Jun 4, 2021

junliume left a comment • edited Loading

Choose a reason for hiding this comment

krzysz00 commented Jun 4, 2021

atamazov commented Jun 7, 2021

junliume commented Jun 7, 2021

ce1adon commented Jun 7, 2021

atamazov commented Jun 7, 2021

atamazov commented Jun 8, 2021

junliume commented Jun 10, 2021

ce1adon commented Jun 10, 2021

atamazov commented Jun 10, 2021

ce1adon commented Jun 10, 2021 • edited Loading

atamazov commented Jun 10, 2021

ce1adon commented Jun 10, 2021

ce1adon commented Jun 10, 2021

atamazov commented Jun 11, 2021

atamazov commented Jun 11, 2021

atamazov commented Jun 11, 2021

ce1adon commented Jun 11, 2021

atamazov commented Jun 11, 2021

ce1adon commented Jun 11, 2021

atamazov commented Jun 11, 2021

ce1adon commented Jun 11, 2021

atamazov commented Jun 20, 2021

ce1adon commented Jun 1, 2021 •

edited

Loading

junliume left a comment •

edited

Loading

ce1adon commented Jun 10, 2021 •

edited

Loading