Feature: batchnorm gpu test

[harz05]

Closes #4

This implements the 3 GPU/Alpaka methods for ROperator_BatchNormalization and adds a test in the Alpaka CUDA test suite.

ROperator_BatchNormalization.hxx

• Generate_GPU_Kernel_ALPAKA : elementwise kernel struct. Initialize() already pre-expands per-channel scale/bias/mean tensors to the full [N,C,H,W] shape and folds 1/sqrt(var + eps) into scale, so the per-element formula reduces to y[i] = (x[i] - mean[i]) * scale[i] + bias[i]. Kernel pattern is same like other operators
• Generate_GPU_Kernel_Definitions_ALPAKA : declares the kernel instance used at launch.
• Generate_GPU_ALPAKA : sets up work division via alpaka::getValidWorkDiv and dispatches the kernel with alpaka::exec. Uses the deviceBuf_<name> buffers for all weight tensors consistent with how RModel_ALPAKA.cxx allocates them.

Files for testing:

• BatchNormModelGenerator.py : creates BatchNorm.onnx from nn.BatchNorm2d(2) in eval mode with fixed weights. Input shape (1, 2, 2, 2), scale [1, 2], bias [0, 0.5], running mean [0.5, 3], running var [1, 4]. Also prints the reference floats used in the .ref.hxx.
• input_models/BatchNorm.onnx : the exported ONNX model (opset 13).
• input_models/references/BatchNorm.ref.hxx : 8 reference output values computed by PyTorch.
• TestCustomModelsFromONNXForAlpakaCuda.cxx : added SofieAlpakaTest.BatchNorm which loads the model on GPU, runs inference, and checks each output element against the reference within tolerance.

Testing:

Tested on Google Colab (NVIDIA T4, CUDA 12.x). New test passes and all 26 tests in the suite pass with no regressions.

[sanjibansg]

Hi @harz05,
Thanks for the PR. A question regarding the onnx file for testing.

[sanjibansg]

How will this file be executed? Also, since you provide the .onnx file anyway, how will this script be useful?

[harz05]

Thank you for reviewing,
this file is standalone and I included it for reproducibility so that the model and ref values can be regenerated and verified independently. python3 BatchNormModelGenerator.py executes the file and we get the .onnx along with the ref values.