mnn 模型，Android 平台 benchmark 测试，推理使用 openCL, vulkan 都比 CPU 慢，可能是什么原因？

[dearwind153]

MNN 版本：MNN-2.9.0 平台：Android 小米9 硬件 NDK: NDK-r20b 编译时，同时编译 Android 平台 CPU， openCL, vulkan 后端

mnn 模型，Android 平台 benchmark 测试，推理使用 openCL, vulkan 都比 CPU 慢，可能是什么原因？

CPU backend 测试： max = 297.433 ms min = 294.566 ms avg = 296.252 ms openCL backend 测试： max = 720.318 ms min = 661.634 ms avg = 699.978 ms vulkan backend 测试： max = 2201.615 ms min = 2060.933 ms avg = 2118.065 ms

mnn model 已发送到 ：[email protected]

详细测试情况如下：

CPU backend 测试 ————————————————————————————————————————————————

./benchmark.out ./mnn_models/ 15 5 0 4 0 MNN benchmark Forward type: CPU thread=4 precision=0 sparsity=0 sparseBlockOC=1 testQuantizedModel=0 --------> Benchmarking... loop = 15, warmup = 5 [-INFO-]: precision!=2, use fp32 inference. parsed /proc/cpuinfo Hardware = "Qualcomm Technologies, Inc SM8150" The device support i8sdot:1, support fp16:1, support i8mm: 0 [ - ] onnxsim.mnn max = 297.433 ms min = 294.566 ms avg = 296.252 ms

openCL backend 测试 ———————————————————————————————————————————————— ./benchmark.out ./mnn_models/ 15 5 3 4 0 MNN benchmark Forward type: OpenCL thread=4 precision=0 sparsity=0 sparseBlockOC=1 testQuantizedModel=0 --------> Benchmarking... loop = 15, warmup = 5 [-INFO-]: precision!=2, use fp32 inference. parsed /proc/cpuinfo Hardware = "Qualcomm Technologies, Inc SM8150" The device support i8sdot:1, support fp16:1, support i8mm: 0 [ - ] onnxsim.mnn max = 720.318 ms min = 661.634 ms avg = 699.978 ms

vulkan backend 测试 ———————————————————————————————————————————————— ./benchmark.out ./mnn_models/ 15 5 7 4 0 MNN benchmark Forward type: Vulkan thread=4 precision=0 sparsity=0 sparseBlockOC=1 testQuantizedModel=0 --------> Benchmarking... loop = 15, warmup = 5 [-INFO-]: precision!=2, use fp32 inference. parsed /proc/cpuinfo Hardware = "Qualcomm Technologies, Inc SM8150" The device support i8sdot:1, support fp16:1, support i8mm: 0 [ - ] onnxsim.mnn max = 2201.615 ms min = 2060.933 ms avg = 2118.065 ms

[jxt1234]

可能是有尺寸超出。换用 buffer 试下吧。 OpenCL: numberThread 设成 68 Vulkan ：编译时加上 -DMNN_VULKAN_IMAGE=OFF

[jxt1234]

另外模型文件我好像没收到，重新发一下?

[lacoboi]

@dearwind153 请教一下，cpu 和 gpu 利用率的这个图是怎么画出来的

[github-actions[bot]]

Marking as stale. No activity in 60 days.