AMDMIGraphX issues

Analyze runtime as a function of tuning iterations and see when GPU clicks throttle down

Pick a kernel, and collect runtime as a function of repetitions. Set warm up to zero. The goal is to figure out when exactly the GPU clicks throttle down...

ahsan-ca

Add GPU onnx support for com.microsoft.SparseAttention

1

- Should be reviewed after https://github.com/ROCm/AMDMIGraphX/pull/3866 is merged. - Partially resolves https://github.com/migraphx-benchmark/AMDMIGraphX/issues/200 - Adds GPU support for SparseAttention using simple handmade kernels, will be replaced by an implementation that uses...

music-dino

Exhaustive tune reduce operators

4

This will tune the block size and algorithm chosen. It also fixes the benchmarking.

pfultz2

Adding skip-layer-normalization operator

1

Based off changes via- https://github.com/ROCm/AMDMIGraphX/pull/3865 Split these out from that branch

ahsan-ca

enhancement

Onnx Operators

change (cerr -> cout) for consistency across all error messages

5

~~A few error messages are getting piped to `cerr`, while most go to `cout`. Thus, making it consistent.~~ Channeling warnings that stream to cout, instead into cerr. (Ideally the verify...

lakhinderwalia

Fuse multiple outputs for pointwise ops

3

pfultz2

Add bitonic topk

2

This implements a faster GPU topk. * Update the ref version of topk to take a parameter for the indices, and also updated to handle any layout. * Added a...

pfultz2

Horiz fuse after pointwise

1

Updates `find_splits` matcher: * Skip trying to fuse instructions that have inter-group dependencies (where a split group depends on another such as the case `sigmoid(x) + x` * Allow fusing...

CharlieL7