Crash when compiling miniqmc

[ye-luo]

aomp-0.7 has this issue. 0.6.5 was fine.

clang-9: /home/yeluo/git/aomp/llvm-project/llvm/lib/IR/Instructions.cpp:1349: void llvm::StoreInst::AssertOK(): Assertion `getOperand(0)->getType() == cast<PointerType>(getOperand(1)->getType())->getElementType() && "Ptr must be a pointer to Val type!"' failed.

reproducer

git clone https://github.com/ye-luo/miniqmc
cd miniqmc/build
cmake -DCMAKE_CXX_COMPILER=/home/yeluo/rocm/aomp_0.7-0/bin/clang++ \
-DENABLE_OFFLOAD=1 -DOFFLOAD_TARGET=amdgcn-amd-amdhsa \
-DCMAKE_CXX_FLAGS="-Xopenmp-target=amdgcn-amd-amdhsa -march=gfx906" \
..
make -j15

[ronlieb]

Greg pushes up a fix for this issue a few days ago to llvm-project Pull and rebuild should produce a compiler with a resolution Let us know if you still see the issue Thx Sent from my iPhone.

…

On Aug 1, 2019, at 12:00 AM, Ye Luo ***@***.***> wrote: aomp-0.7 has this issue. 0.6.5 was fine. clang-9: /home/yeluo/git/aomp/llvm-project/llvm/lib/IR/Instructions.cpp:1349: void llvm::StoreInst::AssertOK(): Assertion `getOperand(0)->getType() == cast<PointerType>(getOperand(1)->getType())->getElementType() && "Ptr must be a pointer to Val type!"' failed. reproducer git clone https://github.com/ye-luo/miniqmc cd miniqmc/build cmake -DCMAKE_CXX_COMPILER=/home/yeluo/rocm/aomp_0.7-0/bin/clang++ \ -DENABLE_OFFLOAD=1 -DOFFLOAD_TARGET=amdgcn-amd-amdhsa \ -DCMAKE_CXX_FLAGS="-Xopenmp-target=amdgcn-amd-amdhsa -march=gfx906" \ .. make -j15 — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or mute the thread.

[ye-luo]

I built everything from scratch using the build script provided in this repo last night. Should I checkout a specific branch of llvm-project?

[gregrodgers]

not yet. I am able to recreate the compile fails in miniqmc. I was hoping they would be easy to fix before we release 0.7-0. As 0.7-0 is now, miniqmc fails. So no need to rebuild. I see the compile fails. There are many places in clang that are missing addrspacecasts. Your code found another one.

[gregrodgers]

Ye-luo, I was able to find the bug causing the compiler to crash for miniqmc. I pushed the update into llvm-project a few minutes ago. Till we release 0.7-0, you need to build it from source. If you have already done this you could pull the update to llvm-project which is in branch AOMP-190715. You should get that branch and all the other correct 0.7 branches if you checked out branch 0.7 in this aomp repo, and then ran clone_aomp.sh. If you did that already, just rerun clone_aomp.sh and you will get the fix. It was to file CodeGenFunction.h. Then run "build_project.sh install" which will pick up the changed file and rebuild and install the affected clang components.

Greg

[ye-luo]

@gregrodgers now miniqmc can be built correctly by the latest compiler with your fix.

[ronlieb]

this cmake command will fully enable the device offload in our compiler .

cmake -DCMAKE_CXX_COMPILER=/usr/lib/aomp_11.7-1/bin/clang++ -DENABLE_OFFLOAD=1 -DOFFLOAD_TARGET=amdgcn-amd-amdhsa -DCMAKE_CXX_FLAGS="-target x86_64-pc-linux-gnu -fopenmp -fopenmp-targets=amdgcn-amd-amdhsa -Xopenmp-target=amdgcn-amd-amdhsa -march=gfx900" ..

[ronlieb]

with that i can see the following error on 11.7-1
OMP_NUM_THREADS=4 ./bin/check_spo -n 10
miniqmc git branch: OMP_offload
miniqmc git commit: fe7367fc3d5d768db264d6613dd40d97e24cab03

Number of orbitals/splines = 192
Tile size = 192
Number of tiles = 1
Rmax = 1.7
Iterations = 10
OpenMP threads = 4

SPO coefficients size = 98304000 bytes (93.75 MB)
[GPU Memory Error] Addr: 0x7f19b4579000 Reason: Page not present or supervisor privilege.
Memory access fault by GPU node-1 (Agent handle: 0x1171ce0) on address 0x7f19b4579000. Reason: Page not present or supervisor privilege.
Aborted (core dumped)