LWP100-1465 generalize common GEMM graph

[rocm-devops]

Original author: @yoonchoi

PR overview

For LWP100-1465

Common GEMM graph is used for Address calculation test.
Changes in common GEMM graph from #1144 are extracted and subsequent changes are added.

Testing

Many existing tests using GEMM common graphs.
Tested if the existing tests passes with the minimal changes to match jamming factors.

Commit message

Refactor and generalize common GEMM graph

• Removed templatized datatype from common GEMM graph
• datatypes are now per matrix and member variables.
• Workgroupsizes are flattened similarly to client code.

[rocm-devops]

Original commenter: @a1-mlselibci-npi

Generated Documentation

• Link to view Generated Docs.

[rocm-devops]

Original commenter: @a1-mlselibci-npi

CodeQL report

Results Summary

recommendation
6

Full table of results

Type	Description	File Path	Start Line	Start Column	End Line	End Column
recommendation	Variant access from getElement; use graph.get<> instead.	/lib/source/CodeGen/LoadStoreTileGenerator.cpp	334	30	334.1	48
recommendation	Variant access from getElement; use graph.get<> instead.	/lib/source/CodeGen/LoadStoreTileGenerator.cpp	338	30	338	48
recommendation	Variant access from getElement; use graph.get<> instead.	/test/unit/KernelGraphTest.cpp	1052	17	1052	35
recommendation	Variant access from getElement; use graph.get<> instead.	/test/unit/KernelGraphTest.cpp	1060	17	1060	35
recommendation	Variant access from getElement; use graph.get<> instead.	/test/unit/KernelGraphTest.cpp	984	17	984	35
recommendation	Variant access from getElement; use graph.get<> instead.	/test/unit/KernelGraphTest.cpp	992	17	992	35

Links

• HTML
• Sarif (for download and usage in conjunction with SARIF viewers)

[rocm-devops]

Original commenter: @a1-mlselibci-npi

Code Coverage Report for gfx942

Summary

Type	Total	Missed	Master Missed	Missed Change	Coverage	Master Coverage	Coverage Change
Lines	43480	5967	5422	545	86.28%	87.34%	-1.06%
Functions	4340	658	636	22	84.84%	85.32%	-.48%
Regions	23265	3719	5536	-1817	84.01%	78.74%	5.27%
Branches	13430	3392	3477	-85	74.74%	74.46%	.28%

This PR adds/edits 12743 newly uncovered lines.

Artifacts

• HTML Coverage Report and Diff
• File Coverage Summary
• Diff Text File
• Full Text Coverage Report
• Python Coverage Report

Commit Hashes

• 1f31f5b

[rocm-devops]

Original commenter: @yoonchoi

@maemmett, @psandhu, I changed the codes in GEMM common graph and restored original checks. Only changed TileSizes of Y to make jamming factors maintained.

[rocm-devops]

Original commenter: @yoonchoi

Not sure why following CI errors occurred. Build parameters for gfx950 were given when I when to Jenkins page at http://math-ci.amd.com/job/enterprise/job/precheckin/job/rocRoller/job/PR-1157/build?delay=0sec

3133/8422 Test #3135: NewArithmeticTests/IntegralArithmeticTest.GPU_Int32/(gfx950,VGPR) ............................................................................................................***Failed    1.28 sec
Running main() from ./build/_deps/googletest-src/googletest/src/gtest_main.cc
Note: Google Test filter = NewArithmeticTests/IntegralArithmeticTest.GPU_Int32/22
[==========] Running 1 test from 1 test suite.
[----------] Global test environment set-up.
[----------] 1 test from NewArithmeticTests/IntegralArithmeticTest
[ RUN      ] NewArithmeticTests/IntegralArithmeticTest.GPU_Int32/22
clang++: error: invalid target ID 'gfx950'; format is a processor name followed by an optional colon-delimited list of features followed by an enable/disable sign (e.g., 'gfx908:sramecc+:xnack-')
unknown file: Failure
C++ exception with description "lib/source/Assemblers/SubprocessAssembler.cpp:85: FatalError(retCode == 0)
	command = /opt/rocm/llvm/bin/clang++ -x assembler -target amdgcn-amd-amdhsa -mcode-object-version=5 -mcpu=gfx950 -mwavefrontsize64 -c -o /tmp/rocroller-1z71q5/kernel.o /tmp/rocroller-1z71q5/kernel.s
	retCode = 256
	output = 
" thrown in the test body.

[rocm-devops]

Original commenter: @yoonchoi

Original review comment on CMakeLists.txt at line 1

Remove changes in this file, which will be separately merged by Lauren's PR.

[rocm-devops]

Original commenter: @yoonchoi

Original review comment on cmake/Dependencies.cmake at line 1

Remove changes in this file, which will be separately merged by Lauren's PR.

[rocm-devops]

Original commenter: @yoonchoi

Original review comment on test/catch/KernelGraphRemoveDuplicatesTest.cpp at line 47

To match the original jamming factors. WorkgroupsizeY is half of the original code.

[rocm-devops]

Original commenter: @yoonchoi

Original review comment on test/unit/KernelGraphTest.cpp at line 1258

To match the original jamming factors. WorkgroupsizeY is half of the original code.

[rocm-devops]

Original commenter: @yoonchoi

Original review comment on test/common/CommonGraphs.cpp at line 1

This is refactoring from CommonGraphs_impl.hpp.

[rocm-devops]

Original commenter: @nhenders

Original review comment on test/common/CommonGraphs.cpp at line 313

We should generalize this past MFMA. Maybe setMITileSize. (MI for Matrix Instruction)

[rocm-devops]

Original commenter: @nhenders

Original review comment on test/common/CommonGraphs.cpp at line 368

We also need to parameterize the wavefrontSize (e.g GEMM::setWavefrontSize); gfx12's default is 32 but the underlying GEMMProblem defaults to 64.

We could also have an API for GEMM::setWavefrontSize that takes in a GPUArchitecture and queries the GPUCapability::DefaultWavefrontSize.

[rocm-devops]

Original commenter: @nhenders

Original review comment on test/common/common/CommonGraphs.hpp at line 219

GEMMProblem has switches for controlling optimizations. Some of these are on by default e.g. fuseLoops, tailLoops, LDS loads. I think we should consider switching all these off by default, and have enablement be explicit.

[rocm-devops]

Original commenter: @yoonchoi

Original review comment on test/common/common/CommonGraphs.hpp at line 212

These became member variables to allow the capability to reassign new argument values. (Used in PR #1144).

[rocm-devops]

Original commenter: @jolabega

Original review comment on test/common/CommonGraphs.cpp at line 313

Yes, that makes more sense as we try to move away for explicitly refer to MFMA where code should be generic.