GitLab

Support nonstandard shapes like slice, broadcast and transpose for the unsqueeze op

* add submodule test
* remove for loop
* simplify reshape test

There is now a MIGRAPHX_DISABLE_POINTWISE_FUSION to disable it

Support slice, broadcast and transpose shapes for the squeeze op.

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

Currently, MIGRAPHX_TRACE_EVAL=2 prints out the entire output buffer, but this can produce a lot of output. To make it easier to inspect and debug, using MIGRAPHX_TRACE_EVAL=2 now only prints 10 elements from the buffer(the first 5 and last 5) and shows any fp classifications found in the buffer(ie nans, infinity, etc). The previous behavior can still be enabled with MIGRAPHX_TRACE_EVAL=3.

* use the parse_resize to parse the upsample operator

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

* Add mean op onnx parser and unit tests
* Refactor parse_mean to use add_broadcastable_binary_op

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

Add HardSwish to HardSigmoid parser

HardSwish formula is y = x * HardSigmoid<alpha=1/6, beta=0.5>(x)
HardSigmoid parser sets alpha to 1/6 and adds the mul instruction if op name is HardSwish

Resolves #1062

Add onnx parser for operator GreaterOrEqual

Add onnx parser and unit tests for Softsign

* Add onnx parser and unit test

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

Make clip a pointwise op

Add HardSigmoid onnx parser and unit tests
Produces mathematical equivalent to ONNX operator through combination of existing pointwise ops.
Resolves #1028

* Add matcher for conv_bias pointwise
* Add fusion op

Fix bug caused by casting time seed to float

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.

Fuse last instruction in fuse_pointwise
This is also fixes a bug with using an invalid iterator.

simple variable rename

Fix pointwise compile error with half sqrt 

Fix whitespace bug in fusable_conv matcher and add unit test

Resolves a problem in parsing the ssd-10 model.

The problem is, after inserting contiguous in the auto_contiguous pass, standard output shape of some operators becomes non-standard. Then, if the next operator requires standard input shape, an exception is throw.

For example, if we pass the following model:
Input (standard shape) -> transpose (transposed) -> softmax (transposed) -> transpose (standard) -> gather.
It works fine, and no contiguous is required.

In the auto_contiguous pass, a contiguous is inserted after the first transpose. Then we need to replace the first transpose with the contiguous and recompute all shapes. When it comes to the gather operator, its input is a transposed shape, and an exception is thrown.

The solution is in the recompute_shape() function. If it is called by the auto_contiguous pass and shape of an instruction is changed, and the shape is non_standard, we do not recompute shape of its output. The reason is: since its output shape is non_standard, a contiguous op will be added after the instruction, which will recompute shape for later operators.

* Check jit kernels files with clang-tidy

Allows --fp16 to be used in the driver to compare the target fp16 result and the ref fp32 result.

Do compilation in parallel

Currently, eliminate_contiguous will never remove contiguous for operators that use module inputs due to the fact that it doesn't pass the module inputs to compute_shape.

- Update to pass the module inputs correctly to compute_shape
- Fix the overloads of compute_shape so that when passed an empty vector of module inputs it will call the overload without module inputs
- Add tests with contiguous and pointwise module function.
- Move add_pointwise function to a seperate header to reuse across different tests

Currently we have the option of passing in --batch to the driver to change the batch size when the model has a dynamic dim value. We can use this flag to adjust the perf report's rate.