How to run in CUDA mode?

[twmht]

Hi,

the code use cpu mode, how to run in cuda mode with MNN?

Thank you.

[twmht]

see also https://github.com/alibaba/MNN/issues/1879

[Qengineering]

@twmht

Please see our guide. You may run into version issues as we haven't updated tutorial for a while.

[twmht]

@Qengineering

I have seen the tutorial, however, I can built latest MNN on jetson nano without any errors. Except when running with cuda forward the estimated running time was too fast to believe.

So you could not build MNN without applying those two PRs (1616 and 1530)? Or you can built but found some errors when running

[Qengineering]

@twmht Just installed the latest version of MNN 1.2.7. I didn't have to apply the two PR (#1616 and #1530) any more. Could not get it compiled with TensorRT. Run ./benchmark without specifying the forward type as MNN_FORWARD_CUDA is not a type recognized by the benchmark.cpp (see line 190)

static inline std::string forwardType(MNNForwardType type) {
    switch (type) {
        case MNN_FORWARD_CPU:
            return "CPU";
        case MNN_FORWARD_VULKAN:
            return "Vulkan";
        case MNN_FORWARD_OPENCL:
            return "OpenCL";
        case MNN_FORWARD_METAL:
            return "Metal";
        default:
            break;
    }
    return "N/A";
}

That's why you have the output "N/A" with your run.

Here's my run.

[Qengineering]

See https://github.com/alibaba/MNN/pull/1880

[twmht]

@Qengineering

thanks

but I got the following errors when applying https://github.com/alibaba/MNN/pull/1880

there should be a fix for nano with cuda in MNN.

[rlczddl]

any progress?