CUDA test regression scheduling.parallelizeAtomicReduction (and one other)

[Infinoid]

the test applies this schedule:

    stmt = stmt.split(i, i0, i1, 2)
            .split(i0, block, thread, 2)
            .parallelize(block, ParallelUnit::GPUBlock, OutputRaceStrategy::Temporary)
            .parallelize(thread, ParallelUnit::GPUThread, OutputRaceStrategy::Temporary);

Apparently the second parallelize call is ignored. Later on, the cuda codegen fails because it sees block parallism but no thread parallelism.

git bisect says this was caused by c004447. This commit is in the middle of the assembly-v2 effort, and some debug messages are present.

In c004447:

$ bin/taco-test --gtest_filter=scheduling.parallelizeAtomicReduction
[…]
temps:tthreadC : double
results:C : double
checking: where(C += tthreadC, forall(thread, forall(i1, tthreadC += A(i) * B(i))))
checking: where(C += tthreadC, forall(thread, forall(i1, tthreadC += A(i) * B(i))))
temps:tthreadC : double
results:C : double
checking: where(C += tthreadC, forall(thread, forall(i1, tthreadC += A(i) * B(i))))
checking: where(C += tthreadC, forall(thread, forall(i1, tthreadC += A(i) * B(i))))
Compiler bug at /home/infinoid/workspace/taco/git/src/codegen/codegen_cuda.cpp:373 in visit
Please report it to developers
 Condition failed: blockIDVars.size() == threadIDVars.size()
 No matching GPUThread parallelize for GPUBlock
[  FAILED  ] scheduling.parallelizeAtomicReduction (3736 ms)

In c004447^ (the commit preceding it):

$ bin/taco-test --gtest_filter=scheduling.parallelizeAtomicReduction
[…]
temps:tthreadC : double, ti1tthreadC : double
results:C : double
checking: where(C += tthreadC, forall(thread, where(tthreadC += ti1tthreadC, forall(i1, ti1tthreadC += A(i) * B(i))), GPUThread, Atomics))
checking: where(C += tthreadC, forall(thread, where(tthreadC += ti1tthreadC, forall(i1, ti1tthreadC += A(i) * B(i))), GPUThread, Atomics))
checking: where(tthreadC += ti1tthreadC, forall(i1, ti1tthreadC += A(i) * B(i)))
checking: where(tthreadC += ti1tthreadC, forall(i1, ti1tthreadC += A(i) * B(i)))
temps:tthreadC : double, ti1tthreadC : double
results:C : double
checking: where(C += tthreadC, forall(thread, where(tthreadC += ti1tthreadC, forall(i1, ti1tthreadC += A(i) * B(i))), GPUThread, Atomics))
checking: where(C += tthreadC, forall(thread, where(tthreadC += ti1tthreadC, forall(i1, ti1tthreadC += A(i) * B(i))), GPUThread, Atomics))
checking: where(tthreadC += ti1tthreadC, forall(i1, ti1tthreadC += A(i) * B(i)))
checking: where(tthreadC += ti1tthreadC, forall(i1, ti1tthreadC += A(i) * B(i)))
temps:
results:expected : double
temps:
results:expected : double
temps:
results:A1337 : double
temps:
results:
[       OK ] scheduling.parallelizeAtomicReduction (9290 ms)

A second test, scheduling.parallelizeTemporaryReduction fails in the same way, though the debug output is a bit bigger. It also passes in the preceding commit.

I am building on linux with -DCMAKE_BUILD_TYPE=Debug -DCUDA=ON -DOPENMP=ON.

[stephenchouca]

I believe this should be fixed on master now. @Infinoid It'd be great if you can check and see if that's the case.

[Infinoid]

e94e68e fixed these errors, but introduced some new failures. I will log a separate issue for those.

We really need CI testing for CUDA...