GitLab

optimize the softmax operator

Improvements/Additions to be made:

changes for the quant_convolution,
changes for the deconvolution,
Macros for MIOpen status checks

This PR allows for other values of epsilon to be matched when finding layernorm. Similarly, the calculation now uses the variable for epsilon.

* Implement concat using jit compilation

* Remove unused headers

* Add in changes for onnx Mod operator

Initial operator for mod implementation and test cases for integer and floating based types.

Need to use fmod from stdlib for floating point types. half_float::half thankfully is specced to the use the existing std::fmod() call when looking at the half.hpp implementation.

fmod_flag should mirror the onnx fmod attribute. Right now using a floating point type without setting that on the user side to true will result in an exception.

Ref ticket #1283 

* Add softmax kernel

* Add mlir c api

* Formatting

* Create a type attribute

* Formatting

* Parse module

* Formatting

* Add mlir dump function

* Add test case

* Formatting

* Fix tidy issues

* Update mlit version

* Update to newer mlir

* Format

* Move mlir to the gpu and update the test

* Formatting

* Fix bug when appending module

* Format

* Remove old cmake flag

* Update message

* Add return

* Format

* Add mlir_compile

* Format

* Register dialect

* Handle unsinged integers

* Dont provide output for return instruction

* Format

* Add code to insert memrefs

* Format

* Add mlir verification

* Formatting

* Enable pointwise_fusion

* Disable eliminate_data_type

* Set kernal name

* Format

* Fix device name

* Formatting

* Fix output arg

* Format

* Updates

* Upate hash

* Add fuse_mlir pass

* Format

* Add fuse mlir

* Format

* Update mlir

* Sort parameter names

* Format

* Reenable disabled passes

* Remove old mlir conv

* Remove asym default padding

* Add more verbose tracing

* Format

* Fix compilation errors

* Format

* Whitelist operators

* Format

* Add namespace

* Format

* Update triple

* Format

* Use func dialect

* Format

* Use func.return

* Format

* Upgrade mlir version

* Add comment

* Handle symetrical padding

* Format

* Cleanup debug output

* Format

* List failed tests

* Move mlir compile to jit pipeline

* Format

* Update version

* Add source locations

* Format

* Correctly add module

* Format

* Update failed tests

* Fix failures when mlir is disabled

* Format

* Update mlir version

* Check type for fp32

* Format

* Remove failed test

* Update mlir in driver

* Tidy fixes

* Foramt

* Tidy fixes

* Format

* Fix const

* Remove from requirements

* Fix cmake version

* Fix tidy warning

* Use another ifdef

* Fix tidy

* Other tidy fix

* Format

* Update hash

* Add missing license files

* Format

* Format

* Fix fnction name

* Jit contiguous

Updated each source file in the repo with the existing license.

Consolidate the vectorize and preload
Add vectorization to reduction
Co-authored-by: kahmed10 <15948690+kahmed10@users.noreply.github.com>

Remove std references in runtime compilation since these are not available when using hiprtc and the headers may not be available on the system

For clarity on kernel names found when profiling. The new names are set to the order of the ops being compiled. For example: add + relu = add_relu_kernel.

Improves performance for add_gelu.  In bert it is 4x faster and for mul_add it is 50% faster than what we current have.

Move to CI containers to rocm 5.0.2
upgrade to 20.04
free up some more file space in github action environments

Add ref and gpu implementations for ONNX op GatherND

Resolves #1032

With reductions such as {2048, 2, 1456} on axes 1, this is 23x faster than using our new block_reduce, and its even over 100x faster than our original reduce_sum:

# lane
gpu::code_object[code_object=13736,symbol_name=kernel,global=2981888,local=1024,]: 0.0672928ms
# block
gpu::code_object[code_object=13800,symbol_name=kernel,global=39321600,local=64,]: 1.46072ms
# original
gpu::reduce_sum[axes={1}]: 6.73456ms
There is some basic logic to pick between lane and block reduce automatically.

There is significant improvement on larger tensors with half almost 50% faster:

lens: [1024, 384, 768]
gpu::code_object[code_object=13832,symbol_name=kernel,global=39321600,local=256,]: 1.16685ms
gpu::reduce_sum[axes={2}]: 1.73126ms
Also for non-trivial layouts this can sometimes be over 2x faster:

lens: [64, 1024, 768, 4]
gpu::code_object[code_object=13832,symbol_name=kernel,global=39321600,local=256,]: 1.1706ms
gpu::reduce_sum[axes={1}]: 2.63375ms
Of course if the stride becomes larger this speed improvement diminishes due to poor memory access patterns. A lane_reduce instead of a block_reduce is needed for such type of kernels. I plan to address that in a future PR.

Finally, this also includes a MIGRAPHX_GPU_DUMP_ASM env variable which will print out the assembly when the kernel compiles.

This adds the infrastructure so we can compile everything in parallel, whereas before only pointwise kernels were compiled in parallel. This will also directly integrate with lowering and the gpu-driver. The kernels for pointwise and roialign are using this infrastructure. Scatternd is not since it does require standard shape.

This also makes it easier to add new runtime compiled kernels in the future.