Nirvedh Meshram
Nirvedh Meshram
Currently some parts of the XRT C++ API require use of RTTI which down stream projects may have disabled. To see the issue you may build this snippet. https://gist.github.com/nirvedhmeshram/eb051397606439aa50016780ead8cf00 With...
With this custom dispatch should be able to be demonstarted with just iree-opt command but in my experiment no matching happened?
I tried running an example similar to the checked in test [here](https://github.com/iree-org/iree/blob/main/tools/test/iree-run-module-multi.mlir) with the exception that I would like to run across GPU and CPU. Here is the sample mlir...
PadAndVectorDistribute has failures on certain shapes as shown in https://github.com/iree-org/iree/issues/18601 There were some fixes needed to support such cases for TileAndFuse. This also adds TileAndFuse as a preferred pipeline over...
TileandFuse is the modernized pipeline that we would want to use over the older SIMT pipeline when possible.
For this elementwise + pad dispatch ``` func.func @main(%8 : tensor, %9 : tensor) -> tensor { %c0_f16 = arith.constant 0.0 : f16 %13 = tensor.empty() : tensor %14 =...
Failure is seen in following tests LLVMCPU/test/pipeline_pad_tests.mlir An example of IR from `pipeline_tile_and_fuse.mlir` is here https://gist.github.com/nirvedhmeshram/3349f2739dfb529fa4800040bf1c8490 It needs to be verified that the IR generated is what we want and...
Pack ops can affect tiling decisions and hence it is beneficial to generalize them, for e.g for below IR ``` %5 = linalg.generic {indexing_maps = [affine_map (d0, d1)>], iterator_types =...
We keep SLP vectorization off because it can mask perf issues or create regressions. However, on ROCM what we have noticed is that we are hitting issues in several untested...