Open
Description
Hi,
In https://reviews.llvm.org/D143514 I updated some SparseCompiler integration tests to use SVE code-gen path. In two cases the generated output disagrees with the expected output.
1. sparse_tanh.mlir
Expected output:
// CHECK-NEXT: {{( -0.761[0-9]*, 0.761[0-9]*, 0.96[0-9]*, 0.99[0-9]*, 0.99[0-9]*, 0.99[0-9]*, 0.99[0-9]*, 0.99[0-9]*, 1 )}}
Output generated by the SVE RUN line
( -0.761594, 0, 0, 0.761594, 0, 0, 0, 0, 0, 0, 0, 0.964028, 0, 0, 0, 0, 0, 0.995055, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1 )
Note the non-zero values at the end in one case and zeros in the other.
2. sparse_matmul.mlir
Output generated by non-sve RUN lines:
[30.5, 4.2, 4.6, 7, 8, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
Output generated by the SVE RUN line:
[30.5, 4.2, 4.6, 7, 8, 4.32995e-304, 5.24135e-304, 8.48805e-314, 3.95253e-322, 3.95253e-322, 9.2689e-310, 9.27115e-310, 0, 2, 0, 0]
All values are effectively correct except 2 at the end. The test passes nonetheless as the CHECK lines only verifiy the first 5 values:
// CHECK: [30.5, 4.2, 4.6, 7, 8{{.*}}]
SET-UP
Tested with ToT on AWS Graviton 3 (9c7g.12xlarge).
The results with qemu-aarch64 are also incorrect, but the actual values are different.
CC @aartbik
-Andrzej