pstarkcdpr
pstarkcdpr
I have a workaround for this issue. The problem shows up particularly with `IRFFT` since `linalg_ext.fft` gives back both a real and imaginary tensors, but the imaginary part is ignored...
Sorry for the confusion @hanhanW. Thanks to your original comment, the original issue that I logged here is fixed pending a merge (see issue https://github.com/iree-org/iree/issues/22695). This fix exposed another, related...
This issue is fixed in PR https://github.com/iree-org/iree/issues/22695
Thanks @MaheshRavishankar. I should be able adapt my tests to use the same mechanism as the existing FFT tests (`check.expect_almost_eq_const`). I'll do that and then make a PR with my...
Thanks. I doubt that I know how to vectorize it since I've never looked at any of this before, but I'm still curious to look. Thanks for the pointers. Yes,...
I started to put together the e2e tests for FFTs. My regular tests work on LLVM and Vulkan, but converting them to an MLIR test leads to a compile error....
Actually, is the above error a problem with the vmvx backend? I made a simple test that had nothing to do with FFTs but used complex numbers and get the...
Okay, I added new e2e tests in a new mlir file and excluded it on vmvx. I also excluded it on Vulkan because there is sometimes a compile error, depending...