GitLab

Add mlir quant_dot operator support

Co-authored-by: causten <causten@users.noreply.github.com>
Co-authored-by: Ted Themistokleous <107195283+TedThemistokleous@users.noreply.github.com>

* Fix convert for the NaNs

* NaNs can't be compared, use std::isnan()

* formatting

* formatting

* formatting

* add extra tests

* initial

* Added tests and new functionality

* Update optimals handling

* Simplify conditionals

* Ref test, update docs

* Remove comment, suggestion unclear

---------
Co-authored-by: Umang Yadav <29876643+umangyadav@users.noreply.github.com>

* fix parse_instancenorm to create broadcast and multibroadcast instructions with two dynamic shape arguments instead of 1.  Their make_op() functions don't support dynamic shapes when called with one input.  This caused an error when parsing an ONNX 3duunet model

* Use add_common_op() to create multibroadcast op.

* add verification and parsing test for instance_norm with dynamic input.  Parse test doesn't pass.

* fix for test; still doesn't pass

* another fix for test; still doesn't pass

* work in progress, instance_norm_dyn_batch_test works but instance_norm_test doesn't

* fix onnx instancenorm tests to match parser changes.  Passes all check tests

* Updated comments explaining usage of add_common_op()

* hand-merged conflicts with develop

* fix instance_norm_half_test after merge

* add Onnx test instance_norm_dyn_batch_half_test

* add shape test cases broadcast_1in_dyn_error and multibroadcast_1in_dyn_error_0

Changed the doc for find_permutation(shape) to be more clear that it is finding the permutation that would make the shape standard

By converting to fp32 : fp16 3d-unet model accuracy comes out the same as FP32 accuracy.

By using reduce_sum method on Fp16 : accuracy comes out ~0.9% lower compared to fp32 while keeping entire model in fp16.

partially solves #1656
This PR only handles compilation part of multitarget.

Use generate_argument instead of generate_literal for python output as generate_literal doesnt exists
Shorten the names for variables from the main module
Use prefix p_ for parameters
Use shorter variable m for main module in python

Use std::numeric_limits::min/max() functions plus the appropriate value to encode -inf/inf 

* use half hip functions to compute max and min
* add verify test for min and max

Co-authored-by: Paul Fultz II <pfultz2@yahoo.com>

Move CI to support the rocm5.5 release

Adding support for broadcasted scalars to unsqueeze op.

Specifying steps other than 1 is disallowed in this implementation since we want the output the always be a tensor. We can support varying step sizes if we allow a broadcasted scalar output from this op.

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

Python API with documentation updates

When multiplying either the input or output across the K dimensions then the multiple can be applied to the constant which can then be folded with propagate_const.

Exposed the mlir_enabled() call the decide for lowering pipeline's enablement
Disabled the rewrite quantization pipeline in mlir compilation
Added quant convolution as anchor ops
Fixed the return type expectations
Added the fall back hip implementation for quantizelinear and dequantizelinear
Will need advises to improve the implementation for quantizelinear

Relies on Removed split_single_dyn_dim compile flag #1711
Exposes dynamic_dimension as a opaque object with dynamic_dimensions and optimals
Exposes ONNX dyn_input_dims and default_dyn_dim to run with dynamic batch
Updates api.py to be able to create objects from aggregate initialization (used for dynamic_dimension)
Uses offload copy for now

Improves the constant propagation for bert models. Larger batch size no longer use as large of constants.  Also improves the speed of model compilation 

This fixes #1700

Solves #1311

Expose get_shape and get_operator methods for instruction_ref object in the python API.

Ensure that we don't have empty inputs when computing shape for pointwise function

Fixes the above behavior
This needs to be changed to allow for setting static shapes with map_dyn_input_dims since you cannot also use map_input_dims

Expose the shape::type_t values to be used by the python api and is required by torch_migraphx to support torchbench models.

Adds a matcher to split_single_dyn_dim to find all broadcast or multibroadcast with two static shape inputs and replaces the instruction with the one input version.
Sorts the get_output_parameters() list to ensure the correct ordering. (Was getting an error for some models.)

Converts can be inserted when the scales and input differ in the onnx file(we are already doing this implicit conversion in the ref implementation). This will also improve the compile-time of quantizelinear.hpp since we can remove the nested visit method.

Examples..

bin/driver verify /codes/onnx_models/resnet50-v1-7/resnet50-v1-7.onnx --split-single-dyn-dim --batch 3 --dyn-input-dim @data "[{min:1, max:4}, 3, 224, 224]"

bin/driver compile /codes/onnx_models/resnet50-v1-7/resnet50-v1-7.onnx --split-single-dyn-dim --default-dyn-dim "{min:1, max:10}" --output resnet50_batch1-10.mxr

bin/driver perf resnet50_batch1-10.mxr --batch 4

Automatically fuse multiple reductions and pointwise operations.

This will replace conv(x+a, w) with conv(x, w) + conv(a, w) where a is a constant so conv(a, w) can be replaced with a constant.

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