Max191
Max191
This PR adds filtering functionality to Winograd conv rewrites based on a `__winograd_conv` attribute. Convolutions are expected to be annotated with this attribute before the pass invocation, most likely by...
There are numerical errors in unet of 1.0-3.0 max error compared with pytorch on 71a994559068929f206ca60b764416ebfe65daa2. Also, the results of `iree-run-module` are different with each invocation. ### Steps to Reproduce ###...
The Conv2DToWinograd pass only matches on static channel dimensions right now because there are no cases where dynamic shapes are needed. This issue is just a reminder to test dynamic...
When trying to fuse `tensor.pad` with producers, reshapes can be blocking fusion unnecessarily. The following IR is an example of this from VAE: ``` %168 = linalg.generic {indexing_maps = [affine_map...
Some `tensor.pack` and `tensor.unpack` ops become significantly slower with different `outer_dims_perm` values. The following gist has a bad `tensor.pack` and `tensor.unpack` case, as well as a good `tensor.pack` case: https://gist.github.com/Max191/a32c07b72272e74cf625cd810ae09c0a...
This PR changes the winograd op semantics to allow specifying different positions for the input tile dimensions. This enables faster reads and writes for winograd transform ops, since with inner...
https://github.com/llvm/llvm-project/commit/137a7451f458cf7d8e1d88df93dbd8da6888886d enables support for n-D `arith.trunci` narrow type emulation. This causes iree-compile failures in the benchmark suite, so it is reverted in IREE by https://github.com/iree-org/iree/pull/17770. The compiler failure can be...
https://github.com/iree-org/iree/pull/17770 exposed some failures in `iree-compile` related to unit dim folding caused by https://github.com/llvm/llvm-project/commit/2c06fb899966b49ff0fe4adf55fceb7d1941fbca. The failure can be seen in this example: ```mlir func.func @unit_dim_folding(%arg0: vector) -> vector { %cst...
The VMVX backend does not support the `arith.maxnumf` op, which causes compiler failures for softmax. https://github.com/llvm/llvm-project/commit/fa0666876cdf11162af341911b99311a56be2274 changes the softmax decomposition to produce this op, which we are carrying a revert...