GitLab

use this to find a reshape->contiguous and then determine if aliasing can be done
thus use the proper reshape or reshape lazy operator

Bug found due to failing torch benchmark. Added test case to reproduce issue causing the model to error out on compile.
Original logic results in the following error:
AMDMIGraphX/src/include/migraphx/op/unsqueeze.hpp:128: normalize_compute_shape: UNSQUEEZE: Axis dimenstion is not divisible by step

This PR resolves the bug addressed in #1496. 

Improves performance for 4/6 GEMMs used by huggingface BERT models with batch_size>1 by using a non-batched rocBLAS call for GEMMs where the B input has a broadcasted batch dimension.
The four verify tests added reflect the actual configurations used by bert-base-cased, with varied batch sizes.

Also adds a matcher to simplify_reshapes to move multibroadcasts after concats.

Using not and or improves readability. The cppcheck rule will help ensure we are doing it consistently.

* Horizontally fuse contiguous

Instead of just unsqueezing to an axis of 1 a step can be set to use instead. So instead of unsqueezing {3, 12} to {3, 1, 12} a step of 2 will unsqeeze to {3, 2, 6} instead

This reorders the transposes across slice to improve horizontal fusion for contiguous. This also improves eliminate_contiguous to remove contiguous better across splits.

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

Updated variable names according to #1193

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

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

In migraphx, DepthToSpace (d2s) is implemented as reshape --> transpose --> contiguous --> reshape.

If there is trailing binary pointwise operator after depthToSpace then, migraphx can move binary operator before contiguous and reshape of the depthtospce.

So, it becomes reshape-->transpose-->binary_op-->contiguous-->reshape.

Explicit contiguous wouldn't be required since binary_op outputs standard shape. So, it becomes reshape-->transpose-->binary-->reshape.

simplify_reshapes already has matcher that can do this transformation. This PR adds test for cases like depthtospace +binary op.

solves #905

In migraphx, DepthToSpace (d2s) is implemented as reshape --> transpose --> contiguous --> reshape.

This PR adds matcher to find d2s + unary pointwise ops.

Application of the matcher moves the pointwise unary operation before the contiguous and reshape of the d2s.
So it becomes
reshape --> transpose --> unary --> contiguous --> reshape.

Motivation is that, later pointwise module would be created out of unary --> contiguous --> reshape. Codegen for this pointwise module can write out buffer such that explicit contiguous and reshape wouldn't be required.

This transformation is not always guaranteed to improve performance, since unary op will operate on non-standard shape. So, we would need some tuning mechanism to make decision.

#905 pending PR for binary operations.

* rename broadcast and multibroadcast output_lens attribute to out_lens attribute, and change tests and source code to reflect the same

* change the reshape attribute from dims to out_lens

* change transpose attribute's name from dims to perm to reflect better meaning

* use permutation instead of perm for transpose

clang formaating

* use dims instead of out_lens for reshape

clang formatting

* Create lazy range

* Formatting

* Use lazy iota

* code backup

* clang format

* add a matcher related to the special resize case for optimization

* clang format

* code backup

* clang format

* code backup

* remove unnecessary code

* add optimization for the where op

* clang format

* fix cppcheck error

* add a unit test for optimize resize

* clang format

* remove unnecessary header include

* code backup

* clang format

* add unit tests for optimizing resize

* clang format

* add more unit test for optimizing where op

* clang format

* remove unnecessary code

* add one more optimzation to remove contiguous

* clang format

* add a pointwise requirement

* clang format

* fix cppcheck error

* add one more unit test

* fixed a bug

* clang format

* remove unnecessary code

* clang format

* fix a build error

* fix review comments

* clang format

* fix a review comments

* clang format

* code refinement

* clang format

* refine more code

* refine more code

* fix a bug related to reshape_cont optimization

* clang format

* fix a review comment

* removed an unnecessary comment

* refine code according to comments

* clang format
Co-authored-by: mvermeulen <5479696+mvermeulen@users.noreply.github.com>

* Add eliminate_data_type pass

* Formatting

* Auto convert quant ops

* Formatting

* Flip the order of decompose

* Compute max size differently

* Formatting

* Clamp values in convert

* Formatting

* Fix loss of precision in reduce

* Formatting

* Fix bugs in reduction

* Fix accumulator type in reference softmax implementation

* Formatting

* Update convert test

* Remove unused variables

* Remove unnecessary quant_dot check

* Formatting

* Add tests

* Formatting

* Remove unused code

* Remove duplicate ops

* Remove blaze dependency

* Use set since shape::type_t is no hashable on gcc 5

* Formatting
Co-authored-by: Shucai Xiao <shucai@gmail.com>
Co-authored-by: mvermeulen <5479696+mvermeulen@users.noreply.github.com>

* initial testing

* initial testing

* add dequantize

* formatting

* add tests

* formatting

* revert file

* add parse files

* formatting

* add axis tuning and fix tests

* formatting

* add tests and fix int8

* formatting

* fix tidy

* test with int32

* add default name and change string to upper

* formatting

* remove boost call

* refactor to use tune_axis)

* formatting

* Load op when serializing

* Formatting

* Add missing clip field

* Use make_op almost everywhere

* Formatting

* More make ops for rnns

* Get rid of spaces

* Formatting

* Remove operators headers

* Formatting

* Remove unused op headers

* Increase line threshold

* code backup

* clang format

* change corresponding tool files

* clang format
Co-authored-by: mvermeulen <5479696+mvermeulen@users.noreply.github.com>

* Fix fusions in bert model

* Formatting

* Add unit tests

* Formatting

* Fix one_half matcher

* Workaround ICE on gcc

* Formatting

* Tidy fixes
Co-authored-by: mvermeulen <5479696+mvermeulen@users.noreply.github.com>

* fix a bug related to concat transpose.

* clang format

* use return instruction to replace the fake instruction
Co-authored-by: mvermeulen <5479696+mvermeulen@users.noreply.github.com>

* Add initial optimization when using a mul over a sliced convolution

* Formatting

* Add more tests

* Formatting

* Convert to an assert

* Check if used once

* Formatting

* Add test with horiz fusion

* Formatting

* Optimize nested slice

* Formatting

* Fix test

* Add const refs

* Remove unnecessary assert
Co-authored-by: mvermeulen <5479696+mvermeulen@users.noreply.github.com>

* Flatten nested concats

* Formatting

* Rename tests

* Add more shape operators that can be nops

* Dont remove pooling

* Add env to trace nary device functions

* Formatting

* Improve contiguous and concat performance

* Formatting

* Remove unused variable

* Formatting

* Fix gpu tests

* Formatting

* Add more test for transposed concat

* Formatting

* Compute offset and not index

* Compute multi-index once

* Formatting

* Fix transposed inputs

* Formatting

* Use product order for comparisons of hip_array

* Formatting

* Add missing s parameter

* Formatting

* Dont invert permutation

* Fix tidy warnings

* Formatting

* Remove incorrect license

* Use a single integer for stride

* Formatting

* Fix tidy issue

* Fix compiler crash in TF inceptionv4

* Formatting

* Remove else