GitLab

Enable with MIGRAPHX_ENABLE_CK=1 and --exhaustive-tune tune flag

* Track kernel time

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

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.

Co-authored-by: Chris Austen <causten@users.noreply.github.com>

* Add method to compile pointwise

* Formatting

* Add lambda

* Add semicolon

* Rename variable

* Add driver to run jit kernels

* Formatting

* Add context

* Formatting

* Make seperate driver folder

* Add more general gpu driver

* Formatting

* Print out wll time

* Formatting

* Run multiple times and skip first run

* Formatting

* Seperate time_op

* Run an op for comparison

* Formatting

* Add debug asserts

* Formatting

* Change parameer name

* Formatting

* Fix argument order

* Formatting

* Add preloading

* Formatting

* Allow a different data type

* Formatting

* Pipeline transformations

* Formatting

* Add vectorization

* Formatting

* Reduce dims

* Formatting

* Compile with launch params as constant

* Formatting

* Make sure buffer can be vecotrized

* Formatting

* Enable vectorization and preloading

* Formatting

* Add print header

* Formatting

* Avoid allocating to large of LDS

* Formatting

* Add some vec functions to a seperate header

* Formatting

* Add stride loops

* Formatting

* Improve the transform pipeline

* Formatting

* Add const

* Fix shape check

* Formatting

* Just check stride axis is zero

* Remove extra finc_vector_axis overload

* Simplify some mroe functions

* Formatting

* Remove some more extra functions

* Formatting

* Simplify more decltypes

* Add another const

* Fix test

* Get buffer pointer different for older compilers
Co-authored-by: Shucai Xiao <shucai@gmail.com>
Co-authored-by: Chris Austen <causten@users.noreply.github.com>