AMDMIGraphX

AMDMIGraphX copied to clipboard

Fuse reductions with MLIR with multi-outputs

5

This issue has two parts. The first part is to fuse reductions(including split reductions) with MLIR, including any pointwise. The second part is to use multiple outputs when fusing, to...

pfultz2

Bump CI to 6.2 once released

1

Update our [Dockerfile ](https://github.com/ROCm/AMDMIGraphX/blob/develop/Dockerfile) and [hip-clang.docker](https://github.com/ROCm/AMDMIGraphX/blob/develop/hip-clang.docker) Additional files may also be needed https://github.com/ROCm/AMDMIGraphX/tree/develop/tools/docker

causten

Add weight streaming

11

Add weight streaming to allow running of large models on GPUs with low memory. Closes #3156.

eddieliao

Fuse convolution + reshape + transpose + sigmoid

4

``` @404 = gpu::code_object[code_object=6464,symbol_name=mlir_convolution_add,global=102400,local=256,](@402,@293,@400,@403) -> half_type, {1, 255, 80, 80}, {1632000, 6400, 80, 1}, target_id=0: 0.0226192ms, 1% @405 = reshape_lazy[dims={1, 3, 85, 80, 80}](@404) -> half_type, {1, 3, 85, 80,...

umangyadav

Load and save problem_cache from a file

3

pfultz2

Add weight streaming at runtime

7

Figure out a way to have weight streaming at runtime i.e. be able to fit large models on gpu without needing to know literal size ahead of time - [x]...

eddieliao

There appears to be an occasional issue in which we try to allocate a buffer to the gpu that seems to be an overflow of an UInt64. @kahmed10 has reportedly...

eddieliao

Fix googleFnet for HF model

1

causten

Concat + Convert Fusion

3

``` 440 = gpu::code_object[code_object=9544,symbol_name=concat_kernel,global=714000,local=1024,](@439,@436,@437,@438) -> half_type, {1, 25200, 85}, {2142000, 85, 1}, target_id=0: 0.0305542ms, 2% main:#output_0 = @param:main:#output_0 -> float_type, {1, 25200, 85}, {2142000, 85, 1}, target_id=0: 0.0007664ms, 1% @442...

umangyadav

Reuse inner buffers when generating kernels for blockwise reductions

2

Due to the use of `__syncthreads` in the `reduce` methods registers are not reused. We can reuse them directly by assigning to them with `r.inner([](auto& y, auto x) { y...

pfultz2