AMDMIGraphX
AMDMIGraphX copied to clipboard
ROCmSoftwarePlatform
AMD's graph optimization engine.
 This pattern is from YOLOv5. Notice that Multiplication is between output of horizontally convolution and output of sigmoid. Multiplication can actually be horizontally fused as well because it is...
https://github.com/ROCm/AMDMIGraphX/blob/develop/docs/migraphx-driver.rst Docs requires an update for time command. Was introduced recently but docs weren't updated.
When benchmarking kernels during the Quick tune (exhaustive as well), the algorithm is to take the average of 10 runs per tried Kernel and then compare to the other configs....
Each PR gets a perf and accuracy check. Changes needed start with https://github.com/rocm/migraphx-benchmark and https://github.com/ROCm/migraphx-benchmark-utils will be to build a new base image, run all the models with --exaustive-tune and...
Now that the MIGX IR can be extracted in to files PR (https://github.com/ROCm/AMDMIGraphX/pull/3550) , the next step is to convert the IR to Linalg. The rocMLIR team has a tool...
### Desciption As a MIGraphX performance developer I want to break down models in to fused portions and run the single fusion in isolation using different graph optimizers. Now that...
### Problem Description Once the mul_add_kernel, add_kernel and add_kernel are horizontally fused with the mlir_dot_add and mlir_slice_mul_reshape_transpose we can make this case more generic. For example if in place of...