x86 CPU
Use the CPU SDK with a general model pack such as Pikachu or Megatron on Windows x64, Linux x86_64 and Intel macOS. Image preparation, model inference and feature comparison all contribute to frame time. Keep their costs separate when tuning an application.
| Platform | SDK setup | Build from source |
|---|---|---|
| Windows x64 | C/C++ and Python | Windows build |
| Linux x86_64 | C API · C++ · Python | Linux build |
| Intel macOS | macOS integration · Python | macOS build |
Match the library to the process architecture. Apple Silicon running a native arm64 application uses the ARM path; an x86_64 application under Rosetta needs an x86_64 library.
Image processing and preprocessing
InspireCV provides CPU paths for preparing images before inference. Recent source changes expand SSE/AVX2 coverage across Image operations and Task preprocessing. Existing API calls can use the available optimized paths without changes to application code.
| Stage | Work covered by optimized paths | Typical use |
|---|---|---|
| Geometry | Resize, affine sampling, rotation and flips | Preview scaling, camera orientation and face alignment |
| Pixel operations | Channel conversion, arithmetic, thresholds and blending | Prepare channel order and pixel values |
| Tensor output | Float conversion, normalization and layout conversion | Write model inputs into application-owned storage |
The available path depends on the operation, pixel type, channel count and CPU. See the Image and Task examples for input formats, transforms and buffer ownership.
Source version
These additions are available in the latest standalone InspireCV source covered here. The current InspireFace dependency has not yet incorporated them. Check the bundled dependency when building an InspireFace SDK; its version number alone does not establish which image-processing paths it includes.
AVX2 build options
For standalone InspireCV, leave INSPIRECV_ENABLE_AVX2 at its default OFF when the library will run on several different CPUs.
| Setting | Behavior |
|---|---|
OFF — default | Keeps AVX2 kernels separate from the rest of the library. Supported operations can still select them when runtime checks confirm CPU and OS support. |
ON | Compiles all C++ sources with AVX2 enabled; GCC and Clang also enable FMA. Every target machine must support the instructions used by that build. |
OFF does not mean that AVX2 is disabled everywhere. Runtime dispatch was already available; the recent changes extend the operations it covers. Other x86 paths still have their own instruction requirements, including SSE4.1 in Task, so test on the oldest CPU and OS you intend to support. Do not assume a build with ON will run under Rosetta.
The InspireCV build instructions show where to set this option. A prebuilt SDK's compilation options cannot be changed by an application at runtime.
Measure the application workload
- Reuse the session and buffers. Keep a session for each ordered camera stream, and reuse a Task pipeline and destination storage when the shape stays the same.
- Enable only the outputs you need. Detection, recognition and optional analysis have separate costs. Measure the same workload when comparing builds.
- Include the full frame path. Camera conversion, memory copies, queuing and display can add latency beyond the SDK call itself.
The image-processing benchmarks include a runner for measuring public Image and Task calls on your machine. The SDK performance guide measures detection, tracking and feature extraction. Compare the same input, compiler settings and warm-up procedure; image-operation timings alone do not predict the frame rate of a complete application.