Add AMD support via ort alternative to trt

src/inference/backend_config.hpp

@@ -0,0 +1,35 @@
+#pragma once
+#include <string>
+
+// ── BackendConfig ─────────────────────────────────────────────────────────────
+// Backend-neutral tuning knobs for the inference backends. Carried inside Config
+// (as Config::trt) and handed to the make_face_detector / make_face_embedder
+// factories. The core application sets these fields without knowing which
+// backend (ONNX Runtime or TensorRT) will consume them.
+//
+// fp16:      FP16 Tensor Core kernels — safe for both SCRFD and ArcFace.
+// int8:      INT8 quantisation — fast but UNSAFE for ArcFace without a
+//            calibration table (embedding cosine space will shift, breaking
+//            similarity thresholds).  Safe for the SCRFD detector.
+// cache_dir: TRT engines are compiled once and cached here.  First run is
+//            slow (~30–60 s per model); every subsequent run loads instantly.
+//
+// Shape profile (optional, set input_name + profile_{min,opt,max} to enable):
+//   Shape strings are trtexec-style, e.g. "1x3x112x112".  When left empty the
+//   backend derives a sensible default from the model.
+//
+// ort_cache_dir: ORT optimized-model cache. On first load ORT writes a
+//   pre-optimized .ort file here; subsequent loads skip graph optimization.
+//   Empty string = disabled.  Applies to all ORT providers (ROCm, CUDA, CPU).
+struct BackendConfig {
+    bool        fp16      = true;
+    bool        int8      = false;
+    std::string cache_dir = "./trt_cache";
+    // Per-tensor optimisation profile.  All four fields must be set together.
+    std::string input_name;     // e.g. "input.1"
+    std::string profile_min;    // e.g. "1x3x112x112"
+    std::string profile_opt;    // e.g. "4x3x112x112"
+    std::string profile_max;    // e.g. "8x3x112x112"
+
+    std::string ort_cache_dir = "./ort_cache";
+};

src/inference/face_detector.hpp

@@ -0,0 +1,27 @@
+#pragma once
+#include "types.hpp"
+
+#include <memory>
+#include <vector>
+
+// ── IFaceDetector ─────────────────────────────────────────────────────────────
+// Backend-agnostic face detector interface. The core application detects faces
+// through this interface without knowing whether the implementation is ONNX
+// Runtime (SCRFD via ORT) or raw TensorRT (a pre-built SCRFD engine).
+//
+// The concrete implementation is selected at compile time by CMake
+// (SAE_INFERENCE_BACKEND): exactly one of backends/ort_backend.cpp or
+// backends/trt_backend.cpp is compiled and provides make_face_detector().
+
+struct Config;
+
+struct IFaceDetector {
+    virtual ~IFaceDetector() = default;
+    // Detect all faces in a BGR image. Thread-safety is backend-defined; callers
+    // in this project drive a detector from a single pipeline thread.
+    virtual std::vector<DetectedFace> detect(const cv::Mat& img) = 0;
+};
+
+// Construct the detector for the compiled-in backend. Reads cfg.detector_model,
+// cfg.detector_engine, cfg.detector_conf, cfg.detector_nms and cfg.trt.
+std::unique_ptr<IFaceDetector> make_face_detector(const Config& cfg);

src/inference/face_embedder.hpp

@@ -0,0 +1,32 @@
+#pragma once
+#include "types.hpp"
+
+#include <memory>
+#include <vector>
+
+// ── IFaceEmbedder ─────────────────────────────────────────────────────────────
+// Backend-agnostic ArcFace embedder interface. The core application produces
+// 512-d L2-normalised embeddings through this interface without knowing whether
+// the implementation is ONNX Runtime or raw TensorRT.
+//
+// The concrete implementation is selected at compile time by CMake
+// (SAE_INFERENCE_BACKEND): exactly one of backends/ort_backend.cpp or
+// backends/trt_backend.cpp is compiled and provides make_face_embedder().
+
+struct Config;
+
+struct IFaceEmbedder {
+    virtual ~IFaceEmbedder() = default;
+    // Embed a batch of 112×112 BGR crops → one L2-normalised 512-d embedding
+    // each, parallel to the input.
+    virtual std::vector<Embedding> embed(const std::vector<cv::Mat>& crops) = 0;
+    // Largest batch the backend accepts in a single embed() call. TRT engines
+    // are capped by their build profile; ORT reports the configured batch size.
+    virtual int max_batch() const = 0;
+
+    Embedding embed_one(const cv::Mat& crop) { return embed({crop}).front(); }
+};
+
+// Construct the embedder for the compiled-in backend. Reads cfg.arcface_model,
+// cfg.arcface_engine, cfg.embed_batch_size and cfg.trt.
+std::unique_ptr<IFaceEmbedder> make_face_embedder(const Config& cfg);

src/inference/similarity.hpp

@@ -0,0 +1,35 @@
+#pragma once
+#include <memory>
+#include <vector>
+
+// ── ISimilarityEngine ─────────────────────────────────────────────────────────
+// Backend-agnostic gallery similarity engine for the identity matcher.
+//
+// The full reference gallery (n_gallery × 512 L2-normalised embeddings) is
+// uploaded to the GPU once at construction and stays resident. Per frame, the
+// small query matrix (n_faces × 512) is uploaded and a single SGEMM produces the
+// similarity matrix S (n_gallery × n_faces, column-major) — i.e. S[g + f*n_gal]
+// is cosine_similarity(gallery[g], query[f]).
+//
+// The GPU math backend (cuBLAS/CUDA or rocBLAS/HIP) is selected at compile time
+// by CMake (SAE_GEMM_BACKEND); backends/gemm_backend.cpp provides
+// make_similarity_engine(). The core matcher node sees only this interface and
+// holds no CUDA/HIP/BLAS headers.
+
+struct ISimilarityEngine {
+    virtual ~ISimilarityEngine() = default;
+
+    // Largest n_faces accepted by compute() per call (bounds GPU buffer sizes).
+    virtual int max_faces() const = 0;
+
+    // Compute similarities for n_faces query embeddings.
+    // query_row_major: n_faces × 512, row fi at query + fi*512.
+    // Returns a pointer to host memory holding S column-major: the gallery
+    // similarities for face fi start at result + fi*n_gallery. The pointer is
+    // owned by the engine and valid until the next compute() call.
+    virtual const float* compute(const float* query_row_major, int n_faces) = 0;
+};
+
+// gallery_row_major: n_gallery × 512, embedding i at gallery + i*512.
+std::unique_ptr<ISimilarityEngine> make_similarity_engine(
+    const float* gallery_row_major, int n_gallery, int max_faces);