AMDMIGraphX issues

Inconsistent and incorrect use of operator descriptor in onnx parsing

There are onnx `parse() ` functions which use `opd.op_name` inside `MIGRAPHX_THROW` macro, although this operator descriptor field is not initialized. List of such files: - src/onnx/parse_instancenorm.cpp - src/onnx/parse_randomnormal_ops.cpp - src/onnx/parse_randomuniform_ops.cpp...

nives-vukovic

Prevent collapsing batch dims in dot ops with constants

4

This simplifies many `reshape -> dot -> reshape` patterns that are not handled by the `find_reshape_reshape_dot` pass (ie. in gemms where one input is a constant). This also simplifies the...

shivadbhavsar

Perf Improve

[Issue]: MIGraphX Python module cannot find attribute

### Problem Description I've compiled mgx from source with the following CMake config: ```bash cmake -S . -B build -G Ninja \ -DCMAKE_BUILD_TYPE=Release \ -DCMAKE_PREFIX_PATH="$ROCM_INSTALL_PREFIX;$CONDA_PREFIX/lib/python3.11/site-packages/pybind11" \ -DCMAKE_INSTALL_PREFIX=$ROCM_INSTALL_PREFIX \ -DCMAKE_CXX_COMPILER="$ROCM_INSTALL_PREFIX/bin/clang++" \...

xuantengh

[Issue]: ~20% perf difference between MIGraphX and rocblas-bench

7

### Problem Description Given this program: ``` # test.py p = migraphx.program() m = p.get_main_module() p_a = m.add_parameter("inputA",migraphx.shape(type="half_type", lens=[2,1024,1280])) p_b = m.add_parameter("inputB",migraphx.shape(type="half_type", lens=[2,1280,10240])) p_c = m.add_parameter("inputC",migraphx.shape(type="half_type", lens=[2,1024,10240])) p_dot = m.add_instruction(migraphx.op("dot"),...

giuseros

normalize_compute_shape: CONVOLUTION: mismatched channel numbers

4

- MIGraphX commit: [84fc9f0](https://github.com/ROCm/AMDMIGraphX/commit/84fc9f0) - GPU: MI210 (gfx90a) - ROCm version: 6.0.2 Model(s) to reproduce: - https://github.com/onnx/models/blob/main/validated/vision/classification/shufflenet/model/shufflenet-v2-10.onnx - https://github.com/onnx/models/blob/main/validated/vision/classification/shufflenet/model/shufflenet-v2-12.onnx Fails at: `/code/AMDMIGraphX/src/include/migraphx/op/convolution.hpp:100: normalize_compute_shape: CONVOLUTION: mismatched channel numbers`

gyulaz-htec

Accuracy issue in quantized resnet50 after match_qlinear_reused

The accuracy test in torch_migraphx for asymmetricly quantized resnet50 is failed after adding #2613. Here is a simpler repro program for the issue: ``` import migraphx import numpy as np...

shivadbhavsar

bug