88 lines
1.7 KiB
Markdown
88 lines
1.7 KiB
Markdown
rknn_benchmark是用来测试rknn模型性能。请确保测试前将CPU/DDR/NPU频率锁定到最高频率。
|
||
用法：
|
||
|
||
./rknn_benchmark xxx.rknn [input_data] [loop_count] [core_mask]
|
||
|
||
core_mask: 0: auto, 1: npu core1, 2: npu core2, 4:npu core3,
|
||
|
||
3: npu core1&2,
|
||
|
||
7: npu core1&2&3
|
||
|
||
仅RK3588支持 core mask。
|
||
|
||
例如:
|
||
|
||
```
|
||
./rknn_benchmark mobilenet_v1.rknn
|
||
./rknn_benchmark mobilenet_v1.rknn dog.jpg 10 3
|
||
./rknn_benchmark mobilenet_v1.rknn dog.npy 10 7
|
||
./rknn_benchmark xxx.rknn input1.npy#input2.npy
|
||
```
|
||
|
||
|
||
以下 <TARGET_PLATFORM> 表示RK3566_RK3568、RK3562或RK3588。
|
||
|
||
# Aarch64 Linux 示例
|
||
## 编译
|
||
|
||
将`build-linux_<TARGET_PLATFORM>.sh`中的`GCC_COMPILER`修改成交叉编译器路径, 然后执行
|
||
|
||
```
|
||
./build-linux_<TARGET_PLATFORM>.sh
|
||
```
|
||
|
||
## 安装
|
||
|
||
将 install/rknn_benchmark_Linux 拷贝到设备上.
|
||
|
||
- 如果使用Rockchip的EVB板，可以使用以下命令：
|
||
|
||
连接设备并将程序和模型传输到`/userdata`
|
||
```
|
||
adb push install/rknn_benchmark_Linux /userdata/
|
||
```
|
||
|
||
- 如果你的板子有sshd服务，可以使用scp命令或者其他方式将程序和模型传输到板子上。
|
||
|
||
## 运行
|
||
|
||
```
|
||
adb shell
|
||
cd /userdata/rknn_benchmark_Linux/
|
||
```
|
||
|
||
```
|
||
export LD_LIBRARY_PATH=./lib
|
||
./rknn_benchmark xxx.rknn
|
||
```
|
||
|
||
# Android 示例
|
||
## 编译
|
||
|
||
modify `ANDROID_NDK_PATH` on `build-android_<TARGET_PLATFORM>.sh` for target platform, then execute
|
||
|
||
```
|
||
./build-android_<TARGET_PLATFORM>.sh
|
||
```
|
||
|
||
## 安装
|
||
|
||
connect device and push build output into `/data`
|
||
|
||
```
|
||
adb push install/rknn_benchmark_Android /data/
|
||
```
|
||
|
||
## 运行
|
||
|
||
```
|
||
adb shell
|
||
cd /data/rknn_benchmark_Android/
|
||
```
|
||
|
||
```
|
||
export LD_LIBRARY_PATH=./lib
|
||
./rknn_benchmark xxx.rknn
|
||
```
|