docker/rocm/CMakeLists.txt

@@ -0,0 +1,36 @@
+cmake_minimum_required(VERSION 3.1)
+
+set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CXX_STANDARD_REQUIRED ON)
+set(CMAKE_CXX_EXTENSIONS OFF)
+
+if(NOT CMAKE_BUILD_TYPE)
+  set(CMAKE_BUILD_TYPE Release)
+endif()
+
+project(migraphx_py)
+
+find_package(pybind11 REQUIRED)
+pybind11_add_module(migraphx_py migraphx_py.cpp)
+
+include_directories(
+  /opt/rocm/include
+  /opt/rocm/lib/migraphx/include
+)
+
+target_link_directories(migraphx_py PRIVATE
+  "/opt/rocm/lib"
+  "/opt/rocm/lib/migraphx/lib"
+)
+target_link_libraries(migraphx_py PRIVATE
+    migraphx
+    migraphx_tf
+    migraphx_onnx
+)
+SET(CMAKE_INSTALL_RPATH_USE_LINK_PATH TRUE)
+
+set_target_properties(migraphx_py PROPERTIES OUTPUT_NAME "migraphx")
+install(TARGETS migraphx_py
+  COMPONENT python
+  LIBRARY DESTINATION lib
+)

docker/rocm/Dockerfile

@@ -58,6 +58,8 @@ RUN cd /opt/rocm-$ROCM/lib && cp -dpr \
     migraphx \
     # NOTE: Include rocRoller library (only ROCm 7.0+)
     librocroller*.so* \
+    # NOTE: Include libdnnl for migraphx CPU device type support
+    libdnnl*.so* \
     /opt/rocm-dist/opt/rocm-$ROCM/lib/
 
 # Copy ROCm HIP include (only ROCm 7.2+)
@@ -74,6 +76,18 @@ RUN mkdir -p /opt/rocm-dist/etc/ld.so.conf.d/
 RUN printf "/opt/rocm/lib\n/opt/rocm/lib/migraphx/lib" | tee /opt/rocm-dist/etc/ld.so.conf.d/rocm.conf
 RUN printf "/usr/lib/llvm-14/lib" | tee /opt/rocm-dist/etc/ld.so.conf.d/llvm.conf
 
+# Build migraphx python bindings for our python version
+# NOTE: AMDMIGraphX repo does not contain a tag for all ROCm versions, hence the fallback to lower patch versions
+WORKDIR /tmp/migraphx_py
+RUN apt install -qq -y --no-install-recommends g++ python3-dev python3-pybind11 cmake make
+COPY docker/rocm/CMakeLists.txt .
+RUN for i in 0 1 2; do \
+    VERSION=$(echo $ROCM | awk -F. -v d=$i '{OFS="."; $3-=d; print}') && \
+    wget https://raw.githubusercontent.com/ROCm/AMDMIGraphX/refs/tags/rocm-$VERSION/src/py/migraphx_py.cpp && break; \
+    done
+RUN cmake . -DCMAKE_INSTALL_PREFIX=/opt/rocm-dist/opt/rocm-$ROCM && \
+    make install
+
 #######################################################################
 FROM deps AS deps-prelim
 
@@ -83,6 +97,8 @@ RUN apt update && \
     apt install -qq -y -t bookworm-backports mesa-va-drivers mesa-vulkan-drivers && \
     # Install C++ standard library headers for HIPRTC kernel compilation fallback
     apt install -qq -y libstdc++-12-dev libnuma1 && \
+    # Install libomp for migraphx CPU device type support
+    apt install -qq -y libomp-dev && \
     rm -rf /var/lib/apt/lists/*
 
 WORKDIR /opt/frigate

frigate/detectors/detector_utils.py

@@ -85,18 +85,10 @@ def apply_amd_compatibility_env_vars():
     logger.info(f"Setting AMD environment variables for {gfx_id} compatibility...")
 
     configs = {
-        ("gfx902", "gfx909", "gfx90c", "gfx1035", "gfx1103"): {
+        ("gfx902", "gfx909", "gfx90c"): {
             "HSA_ENABLE_SDMA": "0", # Disable System Direct Memory Access for APU compatibility
-            "MIGRAPHX_DISABLE_MIOPEN_FUSION": "1", # Disable unsupported fusion optimization
-        },
-        ("gfx902", "gfx909", "gfx90c"): { # Vega
             "HSA_OVERRIDE_GFX_VERSION": "9.0.0", # Force compatible GFX version
-        },
+            "MIGRAPHX_DISABLE_MIOPEN_FUSION": "1", # Disable unsupported fusion optimization
-        ("gfx1035"): { # 680M
-            "HSA_OVERRIDE_GFX_VERSION": "10.3.0", # Force compatible GFX version
-        },
-        ("gfx1103"): { # 780M
-            "HSA_OVERRIDE_GFX_VERSION": "11.0.0", # Force compatible GFX version
         }
     }
     for gfx_ids, vars in configs.items():

frigate/detectors/plugins/migraphx.py

@@ -0,0 +1,263 @@
+import logging
+
+import ctypes
+import sys
+from pathlib import Path
+from typing import Literal
+
+import numpy as np
+from pydantic import Field
+
+from frigate.const import MODEL_CACHE_DIR
+from frigate.detectors.detection_api import DetectionApi
+from frigate.detectors.detector_config import BaseDetectorConfig
+from frigate.detectors.detector_utils import apply_amd_compatibility_env_vars
+
+logger = logging.getLogger(__name__)
+
+DETECTOR_KEY = "migraphx"
+
+class migraphxDetectorConfig(BaseDetectorConfig):
+    type: Literal[DETECTOR_KEY]
+    device: str = Field(default="gpu", title="Device Type (gpu or cpu)")
+    conserve_cpu: bool = Field(default=True, title="Conserve CPU at the expense of latency")
+    fast_math: bool = Field(default=True, title="Optimize math functions to use faster approximate versions")
+    exhaustive_tune: bool = Field(default=False, title="Use exhaustive search to find the fastest generated kernels")
+
+class migraphxDetector(DetectionApi):
+    type_key = DETECTOR_KEY
+
+    def __init__(self, detector_config: migraphxDetectorConfig):
+        super().__init__(detector_config)
+
+        apply_amd_compatibility_env_vars()
+
+        try:
+            sys.path.append("/opt/rocm/lib")
+            import migraphx
+            logger.info(f"migraphx: loaded migraphx module")
+        except ModuleNotFoundError:
+            logger.error("migraphx: module loading failed, missing migraphx")
+            raise
+
+        assert detector_config.model.path is not None, "No model.path configured, please configure model.path"
+        assert detector_config.model.labelmap_path is not None, "No model.labelmap_path configured, please configure model.labelmap_path"
+
+        device = detector_config.device.lower()
+        assert device in ["gpu", "cpu"], "Invalid device set, must be gpu (default) or cpu"
+
+        try:
+            if device == "gpu":
+                if detector_config.conserve_cpu:
+                    logger.info(f"migraphx: enabling hipDeviceScheduleYield to forcefully conserve CPU (conserve_cpu=true)")
+                    ctypes.CDLL('/opt/rocm/lib/libamdhip64.so').hipSetDeviceFlags(4)
+                else:
+                    # Default to hipDeviceScheduleAuto
+                    ctypes.CDLL('/opt/rocm/lib/libamdhip64.so').hipSetDeviceFlags(0)
+        except Exception as e:
+            logger.warning(f"migraphx: could not set hipSetDeviceFlags: {e}")
+
+        cache_dir = Path(MODEL_CACHE_DIR) / "migraphx"
+        cache_dir.mkdir(parents=True, exist_ok=True)
+
+        path = Path(detector_config.model.path)
+        filename = path.stem + ('_cpu' if device == 'cpu' else ('_tune' if detector_config.exhaustive_tune else ''))
+        mxr_cache_path = cache_dir / f"{filename}.mxr"
+        mxr_path = path.parent / f"{filename}.mxr"
+
+        if mxr_cache_path.exists():
+            logger.info(f"migraphx: loading compiled model from cache {mxr_cache_path}")
+            self.model = migraphx.load(str(mxr_cache_path))
+
+        elif mxr_path.exists():
+            logger.info(f"migraphx: loading compiled model from {mxr_path}")
+            self.model = migraphx.load(str(mxr_path))
+
+        else:
+            logger.info(f"migraphx: loading model from {path}")
+            if path.suffix == '.onnx':
+                self.model = migraphx.parse_onnx(str(path))
+            else:
+                raise Exception(f"migraphx: unknown model format {path}")
+
+            self.model_output_shapes = self.model.get_output_shapes()
+            self.model_is_nms = self.model_output_shapes[0].lens()[2] <= 7
+            assert self.model_is_nms is False, "migraphx does not currently support NMS models"
+
+            logger.info(f"migraphx: compiling the model... (fast_math: {detector_config.fast_math}, exhaustive_tune: {detector_config.exhaustive_tune})")
+            self.model.compile(
+                migraphx.get_target(device),
+                offload_copy=True,
+                fast_math=detector_config.fast_math,
+                exhaustive_tune=detector_config.exhaustive_tune
+            )
+
+            logger.info(f"migraphx: saving compiled model into cache {mxr_cache_path}")
+            migraphx.save(self.model, str(mxr_cache_path))
+
+        self.model_param_name = self.model.get_parameter_names()[0]
+        self.model_input_shape_obj = self.model.get_parameter_shapes()[self.model_param_name]
+        self.model_input_shape = tuple(self.model_input_shape_obj.lens())
+        self.model_input_arg = migraphx.generate_argument(self.model_input_shape_obj)
+        self.model_input_array = np.frombuffer(self.model_input_arg, dtype=np.float32).reshape(self.model_input_shape)
+
+        logger.info(f"migraphx: model loaded (input: {self.model_input_shape})")
+
+    def preprocess(self, tensor_input):
+
+        # Ensure we have a 4D tensor
+        if tensor_input.ndim == 3:
+            tensor_input = np.expand_dims(tensor_input, axis=0)
+
+        target_n, target_c, target_h, target_w = self.model_input_shape
+
+        # Ensure input shape is NCHW
+        if tensor_input.shape[1] == target_c:
+            # Already (N, C, H, W), do nothing
+            pass
+        elif tensor_input.shape[3] == target_c:
+            # From (N, H, W, C) to (N, C, H, W)
+            tensor_input = tensor_input.transpose(0, 3, 1, 2)
+        elif tensor_input.shape[2] == target_c:
+            # From (N, H, C, W) to (N, C, H, W)
+            tensor_input = tensor_input.transpose(0, 2, 1, 3)
+
+        np.copyto(self.model_input_array, tensor_input)
+
+    def detect_raw(self, tensor_input):
+        self.preprocess(tensor_input)
+
+        outputs = self.model.run({self.model_param_name: self.model_input_arg})
+
+        tensor_output = np.frombuffer(
+            outputs[0],
+            dtype=np.float32
+        ).reshape(outputs[0].get_shape().lens())
+
+        return self.optimized_process_yolo(tensor_output)
+
+    def optimized_process_yolo(
+        self,
+        tensor_output,
+        confidence_threshold=0.4,
+        intersection_over_union_threshold=0.4,
+        top_k=100
+    ):
+        # Transpose the raw output so each row represents one candidate detection
+        # Typical YOLO format: [batch, features, candidates] -> [candidates, features]
+        candidate_detections = tensor_output[0].T
+
+        # Extract class probability scores
+        class_probability_matrix = candidate_detections[:, 4:]
+
+        # Identify the highest class score and its index for every candidate box
+        max_class_scores = np.max(class_probability_matrix, axis=1)
+
+        # Create a boolean mask to filter out low-confidence predictions immediately
+        confidence_mask = max_class_scores > confidence_threshold
+
+        # Early exit if no detections meet the minimum confidence requirements
+        if not np.any(confidence_mask):
+            return np.zeros((20, 6), dtype=np.float32)
+
+        # Filter the detections, scores, and class IDs using the boolean mask
+        filtered_detections = candidate_detections[confidence_mask]
+        filtered_scores = max_class_scores[confidence_mask]
+
+        # Limit detections to top_k, reducing the set
+        if len(filtered_scores) > top_k:
+            partition_idx = np.argpartition(-filtered_scores, top_k)[:top_k]
+            filtered_detections = filtered_detections[partition_idx]
+            filtered_scores = filtered_scores[partition_idx]
+
+        # Obtain class IDs of reduced set
+        filtered_class_ids = np.argmax(filtered_detections[:, 4:], axis=1)
+
+        # YOLO boxes are typically [center_x, center_y, width, height]
+        center_coordinates_and_dimensions = filtered_detections[:, :4]
+
+        # Pre-calculate an inversion scale to convert pixels to 0.0-1.0 range
+        # This replaces multiple division operations with a single multiplication later
+        normalization_scale_vector = np.array([
+            1.0 / self.width,
+            1.0 / self.height,
+            1.0 / self.width,
+            1.0 / self.height
+        ], dtype=np.float32)
+
+        # Calculate half-widths and half-heights to find box corners
+        half_dimensions = center_coordinates_and_dimensions[:, 2:4] * 0.5
+
+        # Initialize a buffer for the corner-format boxes [x1, y1, x2, y2]
+        corner_format_boxes = np.empty_like(center_coordinates_and_dimensions)
+
+        # Calculate Top-Left corners (x1, y1)
+        corner_format_boxes[:, :2] = (center_coordinates_and_dimensions[:, :2] - half_dimensions)
+
+        # Calculate Bottom-Right corners (x2, y2)
+        corner_format_boxes[:, 2:4] = (center_coordinates_and_dimensions[:, :2] + half_dimensions)
+
+        # Transform pixel coordinates into normalized values (0.0 to 1.0)
+        normalized_boxes = corner_format_boxes * normalization_scale_vector
+
+        # Apply Non-Maximum Suppression (NMS) to remove redundant, overlapping boxes
+        # returns indices of detections to keep
+        nms_kept_indices = self.optimized_nms(
+            normalized_boxes, 
+            filtered_scores, 
+            intersection_over_union_threshold
+        )
+
+        # Limit the number of detections to the top 20
+        number_of_detections_to_save = min(len(nms_kept_indices), 20)
+        top_detection_indices = nms_kept_indices[:number_of_detections_to_save]
+
+        # Pre-allocate a fixed-size result array (20 detections, 6 columns each)
+        # Column structure: [class_id, confidence, y1, x1, y2, x2]
+        results = np.zeros((20, 6), dtype=np.float32)
+
+        # Bulk assign data into results using the NMS indices
+        # We explicitly swap X and Y during assignment to meet the [y1, x1, y2, x2] requirement
+        results[:number_of_detections_to_save, 0] = filtered_class_ids[top_detection_indices] # class_id
+        results[:number_of_detections_to_save, 1] = filtered_scores[top_detection_indices]    # confidence
+
+        # Coordinate Mapping
+        results[:number_of_detections_to_save, 2] = normalized_boxes[top_detection_indices, 1] # y1
+        results[:number_of_detections_to_save, 3] = normalized_boxes[top_detection_indices, 0] # x1
+        results[:number_of_detections_to_save, 4] = normalized_boxes[top_detection_indices, 3] # y2
+        results[:number_of_detections_to_save, 5] = normalized_boxes[top_detection_indices, 2] # x2
+
+        return results
+
+    def optimized_nms(self, boxes, scores, iou_threshold):
+        x1 = boxes[:, 0]
+        y1 = boxes[:, 1]
+        x2 = boxes[:, 2]
+        y2 = boxes[:, 3]
+        areas = (x2 - x1) * (y2 - y1)
+        order = scores.argsort()[::-1]
+        keep = []
+
+        while order.size > 0:
+            i = order[0]
+            keep.append(i)
+            if order.size == 1: break
+
+            # Calculate intersection
+            xx1 = np.maximum(x1[i], x1[order[1:]])
+            yy1 = np.maximum(y1[i], y1[order[1:]])
+            xx2 = np.minimum(x2[i], x2[order[1:]])
+            yy2 = np.minimum(y2[i], y2[order[1:]])
+
+            w = np.maximum(0.0, xx2 - xx1)
+            h = np.maximum(0.0, yy2 - yy1)
+            inter = w * h
+
+            # IoU = Inter / (Area1 + Area2 - Inter)
+            ovr = inter / (areas[i] + areas[order[1:]] - inter)
+
+            # Filter indices
+            inds = np.where(ovr <= iou_threshold)[0]
+            order = order[inds + 1]
+
+        return keep

frigate/detectors/plugins/onnx.py

@@ -39,7 +39,6 @@ class ONNXDetector(DetectionApi):
         path = detector_config.model.path
         logger.info(f"ONNX: loading {detector_config.model.path}")
 
-        # Import get_optimized_runner AFTER setting AMD compatibility env vars
         from frigate.detectors.detection_runners import get_optimized_runner
 
         self.runner = get_optimized_runner(