Test regression in scheduling_eval.spmvGPU, scheduling_eval.ttvGPU

[Infinoid]

The recent fix for precompute with non-matching indexes, #459, has caused two new CUDA test failures.

You will need to build TACO with -DCUDA=ON to run these tests:

$ ctest --output-on-failure -R 'scheduling_eval.(spmv|ttv)GPU'
Test project /home/infinoid/workspace/taco/build
    Start 87: scheduling_eval.spmvGPU
1/2 Test #87: scheduling_eval.spmvGPU ..........***Failed    7.90 sec
Note: Google Test filter = scheduling_eval.spmvGPU
[==========] Running 1 test from 1 test case.
[----------] Global test environment set-up.
[----------] 1 test from scheduling_eval
[ RUN      ] scheduling_eval.spmvGPU
Compiler bug at /home/infinoid/workspace/taco/git/src/lower/lower.cpp:50 in lower
Please report it to developers
 Condition failed: isLowerable(stmt, &reason)
 Not lowerable, because the index statement is not in concrete index notation, because all variables in concrete notation must be bound by a forall statement: suchthat(forall(block, forall(warp, forall(thread, where(forall(thread_nz, y(i) += precomputed(thread_nz)), forall(thread_nz_pre, precomputed(thread_nz_pre) = A(i,j) * x(j))), GPUThread, Atomics), GPUWarp, IgnoreRaces), GPUBlock, IgnoreRaces), fuse(i, j, f) and pos(f, fpos, A(i,j)) and split(fpos, block, fpos1, 2048) and split(fpos1, warp, fpos2, 256) and split(fpos2, thread, thread_nz, 8))
[  FAILED  ] scheduling_eval.spmvGPU (3799 ms)
[----------] 1 test from scheduling_eval (3799 ms total)

[----------] Global test environment tear-down
[==========] 1 test from 1 test case ran. (3799 ms total)
[  PASSED  ] 0 tests.
[  FAILED  ] 1 test, listed below:
[  FAILED  ] scheduling_eval.spmvGPU

 1 FAILED TEST

    Start 92: scheduling_eval.ttvGPU
2/2 Test #92: scheduling_eval.ttvGPU ...........***Failed    8.11 sec
Note: Google Test filter = scheduling_eval.ttvGPU
[==========] Running 1 test from 1 test case.
[----------] Global test environment set-up.
[----------] 1 test from scheduling_eval
[ RUN      ] scheduling_eval.ttvGPU
Compiler bug at /home/infinoid/workspace/taco/git/src/lower/lower.cpp:50 in lower
Please report it to developers
 Condition failed: isLowerable(stmt, &reason)
 Not lowerable, because the index statement is not in concrete index notation, because all variables in concrete notation must be bound by a forall statement: suchthat(forall(block, forall(warp, forall(thread, where(forall(thread_nz, A(i,j) += precomputed(thread_nz)), forall(thread_nz_pre, precomputed(thread_nz_pre) = B(i,j,k) * c(k))), GPUThread, Atomics), GPUWarp, IgnoreRaces), GPUBlock, IgnoreRaces), fuse(j, k, jk) and fuse(i, jk, f) and pos(f, fpos, B(i,j,k)) and split(fpos, block, fpos1, 2048) and split(fpos1, warp, fpos2, 256) and split(fpos2, thread, thread_nz, 8))
[  FAILED  ] scheduling_eval.ttvGPU (3949 ms)

When I revert back to ffd7155, these tests pass again.

$ ctest --output-on-failure -R 'scheduling_eval.(spmv|ttv)GPU'
Test project /home/infinoid/workspace/taco/build
    Start 87: scheduling_eval.spmvGPU
1/2 Test #87: scheduling_eval.spmvGPU ..........   Passed   13.80 sec
    Start 92: scheduling_eval.ttvGPU
2/2 Test #92: scheduling_eval.ttvGPU ...........   Passed   14.61 sec

100% tests passed, 0 tests failed out of 2

Total Test time (real) =  28.41 sec

This simple command-line reproducer also exposes the issue:

$ bin/taco "y(i) = A(i,j) * x(j)" -f=A:ds -s="fuse(i,j,f),pos(f,fpos,A),precompute(A(i,j)*x(j),fpos,fpre)"
terminate called after throwing an instance of 'taco::TacoException'
  what():  Compiler bug at /home/infinoid/workspace/taco/git/src/lower/lower.cpp:50 in lower
Please report it to developers
 Condition failed: isLowerable(stmt, &reason)
 Not lowerable, because the index statement is not in concrete index notation, because all variables in concrete notation must be bound by a forall statement: suchthat(where(forall(fpos, y(i) += workspace(fpos)), forall(fpre, workspace(fpre) = A(i,j) * x(j))), fuse(i, j, f) and pos(f, fpos, A(i,j)))

Changing fpre to fpos in the precompute allows it to succeed. The above command also works after reverting to ffd7155.

Cc: @nirvikbaruah. Apologies for the late feedback on this failure; hopefully #457 will close the feedback loop.

[weiya711]

Working with @nirvikbaruah on fixing this, a sample fix is in the https://github.com/tensor-compiler/taco/test-pr470 branch. The tests for scheduling_eval.spmvGPU, scheduling_eval.ttvGPU now fail at a different point

Error at /home/infinoid/taco-cuda-ci-runner/_work/taco/taco/src/lower/lowerer_impl.cpp:926 in lowerForallCloned:
 Unable to vectorize or unroll loop over unbound variable thread_nz_pre

[weiya711]

Should be fixed with #474