What if the gpu doesn't support support_int8_storage？

[tumuyan]

Some derive doesn't support support_int8_storage, just print “fp16p/s/a=1/1/1 int8-p/s/a=1/0/1” and "vkQueueSubmit failed -4"

I find this

realsr-ncnn-vulkan/src/realsr.cpp

Line 52 in 50e1055

net.opt.use_int8_storage = true;

and this

realsr-ncnn-vulkan/src/realsr.cpp

Line 142 in 50e1055

if (net.opt.use_fp16_storage && net.opt.use_int8_storage)

          if (net.opt.use_fp16_storage && net.opt.use_int8_storage)
                realsr_preproc->create(realsr_preproc_int8s_spv_data, sizeof(realsr_preproc_int8s_spv_data), specializations);
            else if (net.opt.use_fp16_storage)
                realsr_preproc->create(realsr_preproc_fp16s_spv_data, sizeof(realsr_preproc_fp16s_spv_data), specializations);
            else
                realsr_preproc->create(realsr_preproc_spv_data, sizeof(realsr_preproc_spv_data), specializations);

            if (net.opt.use_fp16_storage && net.opt.use_int8_storage)
                realsr_postproc->create(realsr_postproc_int8s_spv_data, sizeof(realsr_postproc_int8s_spv_data), specializations);
            else if (net.opt.use_fp16_storage)
                realsr_postproc->create(realsr_postproc_fp16s_spv_data, sizeof(realsr_postproc_fp16s_spv_data), specializations);
            else
                realsr_postproc->create(realsr_postproc_spv_data, sizeof(realsr_postproc_spv_data), specializations);

(net.opt.use_fp16_storage && net.opt.use_int8_storage) always be true.

So I fixed the code as:

RealSR::RealSR(int gpuid, bool _tta_mode)
{
    const ncnn::GpuInfo& gpuInfo =  ncnn::get_gpu_info(gpuid);

    net.opt.use_vulkan_compute = true;
    net.opt.use_fp16_packed = gpuInfo.support_int8_packed();
    net.opt.use_fp16_storage = gpuInfo.support_fp16_storage();
    net.opt.use_fp16_arithmetic = false;
    net.opt.use_int8_storage = gpuInfo.support_int8_storage();
    net.opt.use_int8_arithmetic = false;

    net.set_vulkan_device(gpuid);

    realsr_preproc = 0;
    realsr_postproc = 0;
    bicubic_4x = 0;
    tta_mode = _tta_mode;
}

Does the modify is unnecessary?

[ArchieMeng]

It's a known issue.
I run realsr on snapdragon 865 with -t 64 option, and it works well. The default tilesize doesn't work for smartphones.
Tencent/ncnn#2636
#20

Update:

I used waifu2x_srmd_termux to run realsr on Android devices. If I didn't make it wrong, I guess you were using your RealSR NCNN Android to run it.
I have read a part of the source code in the RealSR NCNN CLI, I guess the reason may be that it used an incorrect CMakeList.txt configuration to build it.

[tumuyan]

It's a known issue.
I run realsr on snapdragon 865 with -t 64 option, and it works well. The default tilesize doesn't work for smartphones.

Thank for your reply. Same result but may have different reason.

In Tencent/ncnn#2636 , the patch is turn off both net.opt.use_fp16_storage and net.opt.use_int8_storage ;
If it works well, define net.opt.use_fp16_storage and net.opt.use_int8_storage by GPUinfo is a fine solution.

In #20 , the print info is fp16p/s/a=1/1/0 int8-p/s/a=1/1/1
but net.opt.use_fp16_arithmetic and net.opt.use_int8_arithmetic are defined to false, so it may cause of VRAM/RAM not enough. But how many VRAM/RAM does it used when the error happened? We don't know. In fact, when I ran it on my PC, I didn't see a very high usage. I guess there is another bug

if the root reason is tilesize, the default value should be fixed.

realsr-ncnn-vulkan/src/main.cpp

Line 753 in 50e1055

if (heap_budget > 1900)

I used waifu2x_srmd_termux to run realsr on Android devices. If I didn't make it wrong, I guess you were using your RealSR NCNN Android to run it.

The error occurred on someone else's device, everything fine my device.

[ArchieMeng]

I guess these options are just "prefer" options rather than "force" options. I read the document, and it seems like (when they are set as true) they are enabled only if the end device is capable of it. I am not sure, please ask @nihui .

My evidence is that, my waifu2x ncnn android didn't change the waifu2x.cpp source code and leave net.opt.use_int8_storage as true. None of my smartphones support int8_storage function, but they run the app without any problems. I think this explains a lot.

Additionally, I suggest you to modified the CMakeList.txt according to the existing ncnn android demos. Such as https://github.com/nihui/ncnn-android-nanodet and https://github.com/nihui/ncnn-android-scrfd. And rebuild the app to see if the problem is solved.