IUS-2026 short-course: Advanced Ultrasound Signal Processing on GPUs

The course will consists of the following sections:

Name	Instructor
0. Introduction	Marcin Lewandowski
1. GPU Programming Refresher	Marcin Lewandowski, Piotr Jarosik
2. GPU Performance Guidelines and Patterns	Piotr Jarosik
3. CUDA in practive: ray-tracing	Piotr Karwat
4. Software Defined Ultrasound Architecture with GPU Processing	Piotr Jarosik

Running the examples

Requirements

• CMake 3.18 or newer (3.21 or newer for the HIP backend).
• A C++17 host compiler.
• One of the following GPU toolchains:

| backend | toolchain | | --- | --- | --- | | CUDA | NVIDIA CUDA Toolkit (nvcc) | | HIP | AMD ROCm (hipcc, or the ROCm Clang) |

Optional:

• A profiler (nsys and ncu, or rocprofv3) if metrics are to be collected.

Running locally

Run the complete suite with the arguments quoted in the slides:

./scripts/run_benchmarks.sh

The results are written to a directory named after the GPU: run.log, host.txt, csv/<example>.csv and prof/*.txt. Useful options:

./scripts/run_benchmarks.sh --no-profile        # benchmarks only
./scripts/run_benchmarks.sh --no-build          # reuse an existing build
./scripts/run_benchmarks.sh --backend=hip
./scripts/run_benchmarks.sh --tag=h100 --out=/tmp/h100

A single benchmark may also be run directly:

./build/fir_filter_cuda --n=33554432 --taps=64 --csv=fir.csv

To verify correctness only:

ctest --test-dir build

On AMD hardware, define GPUSC_NO_TEXTURE when building example 05: CDNA GPUs provide no texture hardware, and the texture variants do not compile there.

Running in Docker

Build the images from the examples/ directory. Each image builds every example and runs the full suite on start.

# NVIDIA (requires the NVIDIA Container Toolkit)
docker build -f docker/Dockerfile.cuda -t gpusc-cuda .
docker run --rm --gpus all gpusc-cuda

# AMD
docker build -f docker/Dockerfile.hip -t gpusc-hip .
docker run --rm --device=/dev/kfd --device=/dev/dri --group-add video \
           --security-opt seccomp=unconfined gpusc-hip


To run a single benchmark, pass it as the command:

```bash
docker run --rm --gpus all gpusc-cuda ./build/fir_filter_cuda --n=33554432 --taps=64

Target architectures are build arguments, for example --build-arg CUDA_ARCHITECTURES="90" or --build-arg HIP_ARCHITECTURES="gfx950". To collect the CSV files, mount a directory and set GPUSC_CSV_DIR:

docker run --rm --gpus all -v "$PWD/results:/results" -e GPUSC_CSV_DIR=/results gpusc-cuda