PR #12107 Add hipified SkipLayerNorm code for ROCmEP

First attempt for half2 vectorized memory access in SkipLayerNorm

hubertlu-tw committed 3 years ago

Add some functions for debugging

hubertlu-tw committed 3 years ago

Clean up the code

hubertlu-tw committed 3 years ago

Clean up the code

hubertlu-tw committed 3 years ago

Generalize the vectorized kernels with aligned_vector and remove cudaDeviceProp

hubertlu-tw committed 3 years ago

Add a unit test for a larger input size

hubertlu-tw committed 3 years ago

Fix some Lint C++ warnings

hubertlu-tw committed 3 years ago

Use ILP = 4 for the vectorized kernels

hubertlu-tw committed 3 years ago

Rewrite the vectorized kernel and templatize ComputeSkipLayerNorm

hubertlu-tw committed 3 years ago

Use conditional operator for input_v

hubertlu-tw committed 3 years ago

Refactor LaunchSkipLayerNormKernel and replace the original SkipLayerNormKernelSmall with the vectorized kernel

hubertlu-tw committed 3 years ago

Clean some comments and rename the layernorm function

hubertlu-tw committed 3 years ago

Use ComputeSkipLayerNorm to replace LaunchSkipLayerNormKernel

hubertlu-tw committed 3 years ago

Resolve a Lint C++ warning

hubertlu-tw committed 3 years ago

Merge remote-tracking branch 'upstream/master' into hubertlu/skiplayernorm_half2

hubertlu-tw committed 3 years ago

Fix SkipLayerNormBatch1_Float16_vec output data

hubertlu-tw committed 3 years ago

Add hipified code of bert SkipLayerNorm for ROCmEP

hubertlu-tw committed 3 years ago

Resolve some Lint C++ warnings

hubertlu-tw committed 3 years ago

Merge remote-tracking branch 'upstream/master' into hubertlu/hipify_skiplayernorm

hubertlu-tw committed 3 years ago

Resolve some Lint C++ warnings

hubertlu-tw committed 3 years ago

Resolve some Lint C++ warnings

hubertlu-tw committed 3 years ago

Resolve Python formatting issue

hubertlu-tw committed 3 years ago

onnxruntime Add hipified SkipLayerNorm code for ROCmEP #12107 Merged