Skip to content

CLAUDE.md

Latest commit

 

History

History
119 lines (91 loc) · 5.23 KB

CLAUDE.md

File metadata and controls

119 lines (91 loc) · 5.23 KB
Raw

MIVisionX

AMD MIVisionX toolkit - version 4.0.0. A streamlined computer vision toolkit built on AMD OpenVX. As of 4.0.0, the project is reduced to its core: the AMD OpenVX engine, the AMD RPP OpenVX extension, and the RunVX graph executor, across CPU, HIP, and OpenCL backends.

Project Structure

MIVisionX/
├── amd_openvx/                    # Core AMD OpenVX implementation (v1.3.2)
│   └── openvx/
│       ├── ago/                   # AMD Graph Optimizer (AGO) engine - CPU & GPU kernels
│       ├── api/                   # OpenVX API (vx_api.cpp, vx_nodes.cpp, vxu.cpp)
│       ├── hipvx/                 # HIP GPU backend kernels
│       └── include/               # Khronos OpenVX 1.3.2 standard headers (VX/) + AMD extensions
├── amd_openvx_extensions/         # OpenVX extension modules
│   └── amd_rpp/                   # ROCm Performance Primitives extension (CPU/HIP; CPU-only with OpenCL core)
├── utilities/                     # CLI tools
│   └── runvx/                     # OpenVX graph executor
├── tests/                         # Test suites
│   ├── openvx_conformance_tests/  # Khronos OpenVX CTS runner (runConformanceTests.py)
│   ├── openvx_api_tests/          # Individual API test programs
│   ├── amd_openvx_gdfs/           # GDF (Graph Description Format) test scripts
│   ├── vision_tests/              # Python-driven vision node tests
│   └── vx_rpp_tests/              # AMD RPP extension GDF tests
├── apps/                          # Sample CV apps (bubble_pop, optical_flow) - OpenVX + OpenCV, built separately
├── samples/                       # Sample GDF graphs and c_samples (canny)
├── cmake/                         # CMake find modules
├── docs/                          # Documentation (Sphinx-based)
├── .github/workflows/             # GitHub Actions workflows
│   ├── conformance.yml            # OpenVX conformance tests (CPU backend)
│   └── codeql-analysis.yml        # CodeQL static analysis
└── CMakeLists.txt                 # Top-level CMake (project: mivisionx)

Build System

CMake >= 3.10, C++17 required. Default compiler: amdclang++ (from ROCm), falls back to system compiler.

Key CMake Options

Option Values Default Description
BACKEND HIP, OPENCL/OCL, CPU/host HIP GPU backend (Windows defaults to OpenCL, macOS forces CPU)
GPU_SUPPORT ON/OFF ON Enable GPU support (macOS forces OFF)
CODE_COVERAGE ON/OFF OFF LLVM source-based coverage instrumentation

Build Commands

CPU-only (HOST backend):

mkdir build && cd build
cmake .. -DGPU_SUPPORT=OFF
make -j$(nproc)

HIP backend:

mkdir build && cd build
cmake .. -DBACKEND=HIP
make -j$(nproc)

OpenCL backend:

mkdir build && cd build
cmake .. -DBACKEND=OCL
make -j$(nproc)

Default install prefix: ${ROCM_PATH} (defaults to /opt/rocm).

Build Output

  • Libraries go to build/lib/ (libopenvx.so, libvxu.so, libvx_rpp.so)
  • Binaries go to build/bin/ (runvx)

OpenVX Conformance Tests

The Khronos OpenVX CTS is cloned from https://github.com/KhronosGroup/OpenVX-cts.git (branch openvx_1.3.2).

Running Conformance Locally

Use the provided Python script:

python tests/openvx_conformance_tests/runConformanceTests.py --backend_type HOST

Options: --backend_type (ALL/HOST/HIP/OCL), --jobs N, --skip-mivisionx-build, --skip-cts-build, --skip-cts-run, --directory, --cts-repo, --cts-branch.

CTS Build Configuration

The CTS needs these CMake variables pointing to the pre-built MIVisionX:

  • OPENVX_INCLUDES → amd_openvx/openvx/include
  • OPENVX_LIBRARIES → libopenvx.so;libvxu.so;pthread;dl;m;rt
  • OPENVX_CONFORMANCE_VISION=ON
  • CMAKE_POLICY_VERSION_MINIMUM=3.5

Runtime Environment

  • LD_LIBRARY_PATH must include the directory containing libopenvx.so
  • VX_TEST_DATA_PATH must point to OpenVX-cts/test_data/
  • AGO_DEFAULT_TARGET can be set to CPU or GPU (for OCL/HIP backends)

Git Workflow

  • Primary development branch: develop
  • CI triggers on: master, main, develop
  • Conformance workflow runs parallel test suites: baseline (required), graph, data-objects, image-ops, vision-color, vision-filters, vision-arithmetic, vision-geometric, vision-features, vision-statistics, vision-pyramid

Dependencies

MIVisionX 4.0+ is built on the ROCm Core SDK and requires ROCm 7.13 or later (install via the ROCm install guide, e.g. amdrocm-core-sdk7.13-gfx<arch>). The Core SDK provides HIP, OpenCL, amdclang++, OpenMP, half, and RPP.

Core (CPU-only): CMake >= 3.10, C++17 compiler, SSE4.2 support HIP backend: ROCm (hip::host, hip::device) — from the ROCm Core SDK OpenCL backend: OpenCL libraries/headers — from the ROCm Core SDK amd_rpp extension: RPP >= 3.1.0 (CPU/HIP; CPU-only with an OpenCL core) — from the ROCm Core SDK Optional: OpenCV 3.x/4.x (RunVX display only) Linux link libs: dl, m (core); pthread, dl, m, rt (conformance tests)