Nvidia Jetson ffmpeg + TensorRT support (#6458)

* Non-Jetson changes

Required for later commits:
- Allow base image to be overridden (and don't assume its WORKDIR)
- Ensure python3.9
- Map hwaccel decode presets as strings instead of lists
Not required:
- Fix existing documentation
- Simplify hwaccel scale logic

* Prepare for multi-arch tensorrt build

* Add tensorrt images for Jetson boards

* Add Jetson ffmpeg hwaccel

* Update docs

* Add CODEOWNERS

* CI

* Change default model from yolov7-tiny-416 to yolov7-320

In my experience the tiny models perform markedly worse without being
much faster

* fixup! Update docs

docker/tensorrt/detector/build_python_tensorrt.sh

@@ -0,0 +1,28 @@
+#!/bin/bash
+
+set -euxo pipefail
+
+mkdir -p /trt-wheels
+
+if [[ "${TARGETARCH}" == "arm64" ]]; then
+
+  # NVIDIA supplies python-tensorrt for python3.8, but frigate uses python3.9,
+  # so we must build python-tensorrt ourselves.
+
+  # Get python-tensorrt source
+  mkdir /workspace
+  cd /workspace
+  git clone -b ${TENSORRT_VER} https://github.com/NVIDIA/TensorRT.git --depth=1
+
+  # Collect dependencies
+  EXT_PATH=/workspace/external && mkdir -p $EXT_PATH
+  pip3 install pybind11 && ln -s /usr/local/lib/python3.9/dist-packages/pybind11 $EXT_PATH/pybind11
+  ln -s /usr/include/python3.9 $EXT_PATH/python3.9
+  ln -s /usr/include/aarch64-linux-gnu/NvOnnxParser.h /workspace/TensorRT/parsers/onnx/
+
+  # Build wheel
+  cd /workspace/TensorRT/python
+  EXT_PATH=$EXT_PATH PYTHON_MAJOR_VERSION=3 PYTHON_MINOR_VERSION=9 TARGET_ARCHITECTURE=aarch64 /bin/bash ./build.sh
+  mv build/dist/*.whl /trt-wheels/
+
+fi

docker/tensorrt/detector/rootfs/etc/s6-overlay/s6-rc.d/trt-model-prepare/run

@@ -2,27 +2,35 @@
 # shellcheck shell=bash
 # Generate models for the TensorRT detector
 
+# One or more comma-separated models may be specified via the YOLO_MODELS env.
+# Append "-dla" to the model name to generate a DLA model with GPU fallback;
+# otherwise a GPU-only model will be generated.
+
 set -o errexit -o nounset -o pipefail
 
 MODEL_CACHE_DIR=${MODEL_CACHE_DIR:-"/config/model_cache/tensorrt"}
+TRT_VER=${TRT_VER:-$(cat /etc/TENSORRT_VER)}
 OUTPUT_FOLDER="${MODEL_CACHE_DIR}/${TRT_VER}"
 
 # Create output folder
 mkdir -p ${OUTPUT_FOLDER}
 
 FIRST_MODEL=true
+MODEL_DOWNLOAD=""
 MODEL_CONVERT=""
 
 for model in ${YOLO_MODELS//,/ }
 do
     # Remove old link in case path/version changed
     rm -f ${MODEL_CACHE_DIR}/${model}.trt
-    
+
     if [[ ! -f ${OUTPUT_FOLDER}/${model}.trt ]]; then
         if [[ ${FIRST_MODEL} = true ]]; then
+            MODEL_DOWNLOAD="${model%-dla}";
             MODEL_CONVERT="${model}"
             FIRST_MODEL=false;
         else
+            MODEL_DOWNLOAD+=",${model%-dla}";
             MODEL_CONVERT+=",${model}";
         fi
     else
@@ -35,19 +43,49 @@ if [[ -z ${MODEL_CONVERT} ]]; then
     exit 0
 fi
 
+# On Jetpack 4.6, the nvidia container runtime will mount several host nvidia libraries into the
+# container which should not be present in the image - if they are, TRT model generation will
+# fail or produce invalid models. Thus we must request the user to install them on the host in
+# order to run libyolo here.
+# On Jetpack 5.0, these libraries are not mounted by the runtime and are supplied by the image.
+if [[ "$(arch)" == "aarch64" ]]; then
+    if [[ ! -e /usr/lib/aarch64-linux-gnu/tegra ]]; then
+        echo "ERROR: Container must be launched with nvidia runtime"
+        exit 1
+    elif [[ ! -e /usr/lib/aarch64-linux-gnu/libnvinfer.so.8 ||
+            ! -e /usr/lib/aarch64-linux-gnu/libnvinfer_plugin.so.8 ||
+            ! -e /usr/lib/aarch64-linux-gnu/libnvparsers.so.8 ||
+            ! -e /usr/lib/aarch64-linux-gnu/libnvonnxparser.so.8 ]]; then
+        echo "ERROR: Please run the following on the HOST:"
+        echo "  sudo apt install libnvinfer8 libnvinfer-plugin8 libnvparsers8 libnvonnxparsers8 nvidia-container"
+        exit 1
+    fi
+fi
+
 echo "Generating the following TRT Models: ${MODEL_CONVERT}"
 
 # Build trt engine
 cd /usr/local/src/tensorrt_demos/yolo
 
-# Download yolo weights
-./download_yolo.sh $MODEL_CONVERT > /dev/null
+echo "Downloading yolo weights"
+./download_yolo.sh $MODEL_DOWNLOAD 2> /dev/null
 
 for model in ${MODEL_CONVERT//,/ }
 do
-    echo "Converting ${model} model"
-    python3 yolo_to_onnx.py -m ${model} > /dev/null
-    python3 onnx_to_tensorrt.py -m ${model} > /dev/null
-    cp ${model}.trt ${OUTPUT_FOLDER}/${model}.trt
+    python3 yolo_to_onnx.py -m ${model%-dla} > /dev/null
+
+    echo -e "\nGenerating ${model}.trt. This may take a few minutes.\n"; start=$(date +%s)
+    if [[ $model == *-dla ]]; then
+        cmd="python3 onnx_to_tensorrt.py -m ${model%-dla} --dla_core 0"
+    else
+        cmd="python3 onnx_to_tensorrt.py -m ${model}"
+    fi
+    $cmd > /tmp/onnx_to_tensorrt.log || { cat /tmp/onnx_to_tensorrt.log && continue; }
+
+    mv ${model%-dla}.trt ${OUTPUT_FOLDER}/${model}.trt;
     ln -s ${OUTPUT_FOLDER}/${model}.trt ${MODEL_CACHE_DIR}/${model}.trt
+    echo "Generated ${model}.trt in $(($(date +%s)-start)) seconds"
 done
+
+echo "Available tensorrt models:"
+cd ${OUTPUT_FOLDER} && ls *.trt;

docker/tensorrt/detector/tensorrt_libyolo.sh

@@ -8,7 +8,10 @@ SCRIPT_DIR="/usr/local/src/tensorrt_demos"
 git clone --depth 1 https://github.com/NateMeyer/tensorrt_demos.git -b conditional_download
 
 # Build libyolo
-cd ./tensorrt_demos/plugins && make all
+if [ ! -e /usr/local/cuda ]; then
+    ln -s /usr/local/cuda-* /usr/local/cuda
+fi
+cd ./tensorrt_demos/plugins && make all -j$(nproc)
 cp libyolo_layer.so /usr/local/lib/libyolo_layer.so
 
 # Store yolo scripts for later conversion