Add DDR tooling and reduce RGA inference copies

configs/sample_person_shoe_two_stage.json

@@ -40,6 +40,7 @@
           "use_rga": true,
           "rga_gate": "person_shoe_two_stage",
           "rga_max_inflight": 4,
+          "dst_packed": true,
           "infer_fps": 2,
           "model_path": "./models/yolov8n-640.rknn",
           "model_version": "v8",
@@ -80,6 +81,7 @@
           "use_rga": true,
           "rga_gate": "person_shoe_two_stage",
           "rga_max_inflight": 4,
+          "dst_packed": true,
           "infer_fps": 1,
           "model_path": "./models/shoe_detector_openimages_ppe_v1.rknn",
           "model_w": 640,

configs/sample_person_shoe_two_stage_balanced.json

@@ -40,6 +40,7 @@
           "use_rga": true,
           "rga_gate": "person_shoe_two_stage_balanced",
           "rga_max_inflight": 4,
+          "dst_packed": true,
           "infer_fps": 2,
           "model_path": "./models/yolov8n-640.rknn",
           "model_version": "v8",
@@ -80,6 +81,7 @@
           "use_rga": true,
           "rga_gate": "person_shoe_two_stage_balanced",
           "rga_max_inflight": 4,
+          "dst_packed": true,
           "infer_fps": 2,
           "model_path": "./models/shoe_detector_openimages_ppe_v1.rknn",
           "model_w": 640,

configs/sample_person_shoe_two_stage_recall.json

@@ -40,6 +40,7 @@
           "use_rga": true,
           "rga_gate": "person_shoe_two_stage_recall",
           "rga_max_inflight": 4,
+          "dst_packed": true,
           "infer_fps": 3,
           "model_path": "./models/yolov8n-640.rknn",
           "model_version": "v8",
@@ -80,6 +81,7 @@
           "use_rga": true,
           "rga_gate": "person_shoe_two_stage_recall",
           "rga_max_inflight": 4,
+          "dst_packed": true,
           "infer_fps": 2,
           "model_path": "./models/shoe_detector_openimages_ppe_v1.rknn",
           "model_w": 640,

docs/deployment.md

@@ -123,6 +123,32 @@ sudo ./scripts/deploy.sh install
 sudo ./scripts/deploy.sh status
 ```
 
+### 3.1 DDR 固频排障
+
+当出现 `CPU/NPU` 利用率不高，但 RTSP 输出帧率周期性跌落时，优先检查 DDR devfreq 是否在频繁降频。
+
+项目内置了 DDR 模式切换脚本：
+
+```bash
+sudo bash ./scripts/ddr_mode.sh status
+sudo bash ./scripts/ddr_mode.sh performance
+sudo bash ./scripts/ddr_mode.sh restore
+```
+
+也可以从运维入口调用：
+
+```bash
+sudo ./scripts/ops.sh ddr-status
+sudo ./scripts/ops.sh ddr-performance
+sudo ./scripts/ops.sh ddr-restore
+```
+
+说明：
+
+- `performance`：保存当前 governor/min/max 后，将 DDR 切到高性能模式
+- `restore`：恢复到切换前保存的 governor/min/max
+- 默认节点为 `/sys/class/devfreq/dmc`
+
 ### 4. 卸载
 
 ```bash

plugins/ai_shoe_det/ai_shoe_det_node.cpp

@@ -571,6 +571,7 @@ private:
 
         const uint8_t* model_input_data = nullptr;
         size_t model_input_size = 0;
+        int model_input_dma_fd = -1;
 
         if (image_processor_ && win_w > 0 && win_h > 0) {
             Frame crop_frame;
@@ -595,7 +596,11 @@ private:
                                                ? resized_frame.planes[0].stride
                                                : (resized_frame.stride > 0 ? resized_frame.stride : model_w_ * 3);
                 const size_t packed_size = static_cast<size_t>(model_w_) * static_cast<size_t>(model_h_) * 3;
-                if (resized_data && resized_stride > 0) {
+                if (resized_frame.DmaFd() >= 0 && resized_data) {
+                    model_input_data = resized_data;
+                    model_input_size = resized_frame.data_size > 0 ? resized_frame.data_size : packed_size;
+                    model_input_dma_fd = resized_frame.DmaFd();
+                } else if (resized_data && resized_stride > 0) {
                     if (packed_size > input_buf_.size()) input_buf_.resize(packed_size);
                     for (int row = 0; row < model_h_; ++row) {
                         memcpy(input_buf_.data() + static_cast<size_t>(row) * static_cast<size_t>(model_w_) * 3,
@@ -622,6 +627,7 @@ private:
         input.is_nhwc = true;
         input.data = model_input_data;
         input.size = model_input_size;
+        input.dma_fd = model_input_dma_fd;
         input.type = RKNN_TENSOR_UINT8;
 
         auto r = infer_backend_->InferBorrowed(model_handle_, input);

plugins/ai_yolo/ai_yolo_node.cpp

@@ -929,31 +929,50 @@ private:
                     const int resized_stride = resized_frame.planes[0].stride > 0
                                                    ? resized_frame.planes[0].stride
                                                    : (resized_frame.stride > 0 ? resized_frame.stride : static_cast<int>(input_row));
-                    if (resized_data && resized_stride > 0) {
+                    if (resized_frame.DmaFd() >= 0 && resized_data) {
+                        input.data = resized_data;
+                        input.size = resized_frame.data_size > 0 ? resized_frame.data_size : input_size;
+                        input.width = model_input_w_;
+                        input.height = model_input_h_;
+                        input.dma_fd = resized_frame.DmaFd();
+                        input.dma_offset = 0;
+                    } else if (resized_data && resized_stride > 0) {
                         for (int y = 0; y < model_input_h_; ++y) {
                             memcpy(resized_input_.data() + static_cast<size_t>(y) * input_row,
                                    resized_data + static_cast<size_t>(y) * static_cast<size_t>(resized_stride),
                                    input_row);
                         }
+                        input.data = resized_input_.data();
+                        input.size = input_size;
+                        input.width = model_input_w_;
+                        input.height = model_input_h_;
                     } else {
                         ResizeRgbBilinear(src, w, h, src_stride,
                                           resized_input_.data(), model_input_w_, model_input_h_,
                                           static_cast<int>(input_row));
+                        input.data = resized_input_.data();
+                        input.size = input_size;
+                        input.width = model_input_w_;
+                        input.height = model_input_h_;
                     }
                 } else {
                     ResizeRgbBilinear(src, w, h, src_stride,
                                       resized_input_.data(), model_input_w_, model_input_h_,
                                       static_cast<int>(input_row));
+                    input.data = resized_input_.data();
+                    input.size = input_size;
+                    input.width = model_input_w_;
+                    input.height = model_input_h_;
                 }
             } else {
                 ResizeRgbBilinear(src, w, h, src_stride,
                                   resized_input_.data(), model_input_w_, model_input_h_,
                                   static_cast<int>(input_row));
+                input.data = resized_input_.data();
+                input.size = input_size;
+                input.width = model_input_w_;
+                input.height = model_input_h_;
             }
-            input.data = resized_input_.data();
-            input.size = input_size;
-            input.width = model_input_w_;
-            input.height = model_input_h_;
         }
         input.is_nhwc = true;
 

scripts/ddr_mode.sh

@@ -0,0 +1,102 @@
+#!/bin/bash
+set -euo pipefail
+
+DEV=${DDR_DEV:-/sys/class/devfreq/dmc}
+STATE_DIR=${DDR_STATE_DIR:-/tmp/orangepi-ddr-mode}
+
+usage() {
+    cat <<'EOF'
+Usage:
+  sudo ./scripts/ddr_mode.sh status
+  sudo ./scripts/ddr_mode.sh performance
+  sudo ./scripts/ddr_mode.sh ondemand
+  sudo ./scripts/ddr_mode.sh restore
+EOF
+}
+
+require_dev() {
+    if [[ ! -d "$DEV" ]]; then
+        echo "DDR devfreq node not found: $DEV" >&2
+        exit 1
+    fi
+}
+
+show_status() {
+    require_dev
+    echo "DDR devfreq node: $DEV"
+    [[ -f "$DEV/name" ]] && echo "name: $(cat "$DEV/name")"
+    [[ -f "$DEV/governor" ]] && echo "governor: $(cat "$DEV/governor")"
+    [[ -f "$DEV/available_governors" ]] && echo "available_governors: $(cat "$DEV/available_governors")"
+    [[ -f "$DEV/min_freq" ]] && echo "min_freq: $(cat "$DEV/min_freq")"
+    [[ -f "$DEV/max_freq" ]] && echo "max_freq: $(cat "$DEV/max_freq")"
+    [[ -f "$DEV/cur_freq" ]] && echo "cur_freq: $(cat "$DEV/cur_freq")"
+    [[ -f "$DEV/available_frequencies" ]] && echo "available_frequencies: $(cat "$DEV/available_frequencies")"
+}
+
+save_state() {
+    mkdir -p "$STATE_DIR"
+    [[ -f "$DEV/governor" ]] && cat "$DEV/governor" > "$STATE_DIR/governor"
+    [[ -f "$DEV/min_freq" ]] && cat "$DEV/min_freq" > "$STATE_DIR/min_freq"
+    [[ -f "$DEV/max_freq" ]] && cat "$DEV/max_freq" > "$STATE_DIR/max_freq"
+}
+
+set_performance() {
+    require_dev
+    save_state
+    if [[ -f "$DEV/available_governors" ]] && grep -qw performance "$DEV/available_governors"; then
+        echo performance > "$DEV/governor"
+    elif [[ -f "$DEV/available_frequencies" ]]; then
+        max=$(tr ' ' '\n' < "$DEV/available_frequencies" | sort -n | tail -n1)
+        echo "$max" > "$DEV/min_freq"
+        echo "$max" > "$DEV/max_freq"
+    else
+        echo "No performance governor or available_frequencies for $DEV" >&2
+        exit 1
+    fi
+    show_status
+}
+
+set_ondemand() {
+    require_dev
+    if [[ -f "$DEV/available_governors" ]] && grep -qw dmc_ondemand "$DEV/available_governors"; then
+        echo dmc_ondemand > "$DEV/governor"
+    else
+        echo "dmc_ondemand governor not available for $DEV" >&2
+        exit 1
+    fi
+    show_status
+}
+
+restore_state() {
+    require_dev
+    if [[ ! -d "$STATE_DIR" ]]; then
+        echo "No saved DDR state found in $STATE_DIR" >&2
+        exit 1
+    fi
+    [[ -f "$STATE_DIR/governor" ]] && cat "$STATE_DIR/governor" > "$DEV/governor"
+    [[ -f "$STATE_DIR/min_freq" ]] && cat "$STATE_DIR/min_freq" > "$DEV/min_freq"
+    [[ -f "$STATE_DIR/max_freq" ]] && cat "$STATE_DIR/max_freq" > "$DEV/max_freq"
+    show_status
+}
+
+case "${1:-status}" in
+    status)
+        show_status
+        ;;
+    performance)
+        set_performance
+        ;;
+    ondemand)
+        set_ondemand
+        ;;
+    restore)
+        restore_state
+        ;;
+    -h|--help|help)
+        usage
+        ;;
+    *)
+        usage
+        exit 1
+        ;;
+esac

scripts/ops.sh

@@ -2,7 +2,7 @@
 # RK3588 Media Server 运维脚本
 # 安装后位于 /opt/rk3588-media-server/script/ops.sh
 #
-# 用法: sudo ./ops.sh [status|start|stop|restart|logs|hw]
+# 用法: sudo ./ops.sh [status|start|stop|restart|logs|hw|ddr-status|ddr-performance|ddr-ondemand|ddr-restore]
 
 set -e
 
@@ -118,6 +118,22 @@ cmd_metrics() {
     /opt/rk3588-media-server/scripts/monitor_hw.sh
 }
 
+cmd_ddr_status() {
+    bash /opt/rk3588-media-server/scripts/ddr_mode.sh status
+}
+
+cmd_ddr_performance() {
+    bash /opt/rk3588-media-server/scripts/ddr_mode.sh performance
+}
+
+cmd_ddr_ondemand() {
+    bash /opt/rk3588-media-server/scripts/ddr_mode.sh ondemand
+}
+
+cmd_ddr_restore() {
+    bash /opt/rk3588-media-server/scripts/ddr_mode.sh restore
+}
+
 # 主入口
 case "${1:-status}" in
     status)
@@ -138,6 +154,18 @@ case "${1:-status}" in
     hw)
         cmd_metrics
         ;;
+    ddr-status)
+        cmd_ddr_status
+        ;;
+    ddr-performance)
+        cmd_ddr_performance
+        ;;
+    ddr-ondemand)
+        cmd_ddr_ondemand
+        ;;
+    ddr-restore)
+        cmd_ddr_restore
+        ;;
     *)
         echo "RK3588 Media Server 运维脚本"
         echo ""
@@ -150,6 +178,10 @@ case "${1:-status}" in
         echo "  restart  重启服务"
         echo "  logs     查看实时日志"
         echo "  hw       查看硬件监控"
+        echo "  ddr-status       查看 DDR 频率策略"
+        echo "  ddr-performance  DDR 切到高性能模式"
+        echo "  ddr-ondemand     DDR 切回 dmc_ondemand"
+        echo "  ddr-restore      恢复上次保存的 DDR 配置"
         echo ""
         echo "示例:"
         echo "  sudo $(basename $0) status"