GitLab

* Improve applicable batched gemms for bert

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

* Fuse gemm add with pointwise fusions

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.

Updated variable names according to #1193

Fuse layernorm and added triadd_layernorm fusion.  This is a prep performance booster

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

Fixes the #error when using cppcheck. This no longer suppresses cppcheck errors when including those errors. This fixes the cppcheck errors that was there already.

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.

Issue 1127 Updates the math.hpp header file to perform overloads of various standard functions (ops) for the hip half2 type. The half2 type is two 16-bit floats packed into a 32-bit number and therefore the overloads act on vectors of sizes that are multiples of 2. They are invoked in runtime compilation any time one of the ops is called on a tensor declared with the data type shape::half_type.

Defined new template, made instances of the template for those math operations that the hip library contains, added verify tests for the sqrt operator for three cases:

tensor size not divisible by 2
tensor size divisible by 2 but not by 4
tensor size divisible by 4

out-of-bounds access when generate uses nonpacked tensors and add some additional asserts for gpu memory.

Change the "scatter" struct and op to a base/child set of three: scatter_none, scatter_add, scatter_mul to mirror Onnx' ScatterElements op. and its three reduction options. (Onnx Scatter op is deprecated and is equivalent to scatter_none.)

Provides both a reference op. and update to Onnx parsing. Tests updated and new test case added.

When create a tensor_view with vector date type, the last dimension of the shape should be divided by the vec_size.

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.

* Use ccache for runtime compilation

Add exclusive and reverse modes to gpu implementation of prefix_scan_sum, which completes support for ONNX op CumSum

This adds iterators to tensor_view, which can allow kernels to work with non-standard shapes like for roialign.

To improve the performance of indexing when using the iterators, the shape class was updated to use integral_constants since the compiler doesn't always fold the const values. An integral_constant will at least enforce that in the AST.

Finally, since index calculations with single integers are improved, I also updated pointwise to use single index rather than multi index. There is about 4% improvement in some cases.

change max number of groups in a kernel to 1B for greater performance

Changed the pooling values for two structures from strings to specialized enum classes. Many test and operator parsing changes to support this. Introduces one new source file, op_enums.cpp.

Boost the max number of workgroups for pointwise ops by matching what we are doing in launch.hpp

better performance doing it this way

Add onnx parser and ref and gpu implementations of ONNX op ScatterND

Implements the IsNaN operator, ref, gpu, and onnx parser.

Update the base version of clang-format from 5.0 to 10.0

wrapped in a any_ptr class so the type can be checked at runtime for a mismatch.

Make doc/CMakeLists.txt standalone
Switch to use rocm-cmake modules for document generation
Add CONFIGURE_DEPENDS to file(GLOB) so it will update without an explicit cmake run
Add STRINGS property for build type to make it easier to switch build types with ccmake
Various fixes and improvements

There is now a MIGRAPHX_DISABLE_POINTWISE_FUSION to disable it

This causes incorrect memory coloring, which was causing the accuracy failures in the vision model when enabling the pointwise fusions. Resnet50, inceptionv3 and inceptionv4 do verify now in the driver.

Enforce types to avoid compilation error in pointwise fusions
This fixes compile failure: gpt-2, fp16 on Navi

* Enable auto vectorization
* Handle vector types with convert function
* Dont vectorize when it will cause problems with preload

allow nonstd shape for the arg ops, non-standard shapes include broadcast, slice and transpose

* Improve handling of generator expressions when getting the flags for hip

* Add matcher for conv_bias pointwise
* Add fusion op

Changed the number of threads in a block from 256 to 128
Increased the max number of blocks in the kernel from 256 to 1M.
For the case that the axis is the last dimension, we removed the computation of index since it is not required.

With these change, we can get about 2x speedup compared to the develop branch for the softmax op used in the BertSquad model.

simple variable rename

Fix pointwise compile error with half sqrt