clfunctions keep using cpu device

[junyuancat1]

` // funtion one void ImageProcess::GpuPreProcessConfigure(const uint32_t width_src, const uint32_t height_src, const uint32_t width_dst, const uint32_t height_dst) { std::cout << "TESTING GPU PREPROCESS CONFIGURE" << std::endl; std::cout << "CONFIGURE INFO width_src: " << width_src << " height_src: " << height_src << " width_dst: " << width_dst << " height_dst: " << height_dst << std::endl; // 0. cl shedule && init input tensor // CLScheduler::get().default_init(); CLBackendType backend_type = CLBackendType::Native; auto ctx_dev_err = create_opencl_context_and_device(backend_type); // print device info form tuple std::cout << "OpenCL device info: " << std::endl; std::cout << "-------------------" << std::endl; std::cout << "Device name: " << std::get<1>(ctx_dev_err).getInfo() << std::endl; CLScheduler::get().default_init_with_context(std::get<1>(ctx_dev_err), std::get<0>(ctx_dev_err), nullptr); inputtensor.allocator()->init(TensorInfo(TensorShape(3U, width_src, height_src), 1, DataType::U8).set_data_layout(DataLayout::NHWC)); inputtensor.info()->set_format(Format::U8);

// 1. configure image padding, padding rectangle to square [width_src, width_src]
pad_output_tensor_.allocator()->init(TensorInfo(TensorShape(3U, width_src, width_src), 1, DataType::U8).set_data_layout(DataLayout::NHWC));
pad_output_tensor_.info()->set_format(Format::U8);
PaddingList padding_list = {{0, 0}, {0, 0}, {(width_src - height_src) / 2, (width_src - height_src) / 2}};  // nhwc 第一个维度是width，第二个维度是height，第三个维度是channel, 在这里padding的是height
pad_layer_.configure(&input_tensor_, &pad_output_tensor_, padding_list, PixelValue(127), PaddingMode::CONSTANT);

// 2. configure resize
rescale_output_tensor_.allocator()->init(TensorInfo(TensorShape(3U, width_dst, height_dst), 1, DataType::U8).set_data_layout(DataLayout::NHWC));
rescale_output_tensor_.info()->set_format(Format::U8);
rescale_.configure(&pad_output_tensor_, &rescale_output_tensor_, ScaleKernelInfo(InterpolationPolicy::BILINEAR, BorderMode::CONSTANT));

// 3. configure nhwc to nchw
nchw_cl_tensor_.allocator()->init(TensorInfo(TensorShape(width_dst, width_dst, 3U), 1, DataType::U8).set_data_layout(DataLayout::NCHW));
nchw_cl_tensor_.info()->set_format(Format::U8);
const arm_compute::PermutationVector nhwc_to_nchw_permutation_vector = {1U, 2U, 0U};
permute_nhwc_to_nchw_.configure(&rescale_output_tensor_, &nchw_cl_tensor_, nhwc_to_nchw_permutation_vector);

// 4. allocate memory
input_tensor_.allocator()->allocate();
rescale_output_tensor_.allocator()->allocate();
nchw_cl_tensor_.allocator()->allocate();
pad_output_tensor_.allocator()->allocate();

std::cout << "GpuPreProcessConfigure end" << std::endl;

}

void ImageProcess::GpuPreProcessXBGR(const char src, const uint32_t width_src, const uint32_t height_src, char dest, const uint32_t width_dst, const uint32_t height_dst) { // 0. cl shedule std::cout << "GpuPreProcessXBGR start" << std::endl; std::cout << "PROCESS INFO: width_src: " << width_src << ", height_src: " << height_src << ", width_dst: " << width_dst << ", height_dst: " << height_dst << std::endl; std::cout << "GpuPreProcessXBGR default init cl shedule" << std::endl; // CLScheduler::get().default_init();

// 1. 开一个临时数组，并将xbgr转为rgb
// 这里不建议将这部分内存零散的一个个拷贝到tensor显存中，而是开一个临时buffer在内存中先处理好后再一次性拷贝到tensor显存中，实测这样cpu占用率会降低5%左右
std::cout << "GpuPreProcessXBGR convert xbgr to rgb" << std::endl;
int size = width_src * height_src * 3;
std::unique_ptr<uint8_t[]> rgb_image(new uint8_t[size]);
for (uint32_t i = 0; i < height_src; ++i) {
    for (uint32_t j = 0; j < width_src; ++j) {
        rgb_image[i * width_src * 3 + j * 3 + 0] = src[i * width_src * 4 + j * 4 + 2];
        rgb_image[i * width_src * 3 + j * 3 + 1] = src[i * width_src * 4 + j * 4 + 1];
        rgb_image[i * width_src * 3 + j * 3 + 2] = src[i * width_src * 4 + j * 4 + 0];
    }
}

// 2. copy data to input_tensor_
std::cout << "GpuPreProcessXBGR memcpy to input_tensor_" << std::endl;
input_tensor_.map();
memcpy(input_tensor_.buffer(), rgb_image.get(), size);
input_tensor_.unmap();

pad_layer_.run();
// DrawTensorImage(&pad_output_tensor_, width_src, width_src, "pad_output_tensor_.jpg");
rescale_.run();
// DrawTensorImage(&rescale_output_tensor_, width_dst, height_dst, "rescale_output_tensor_.jpg");
permute_nhwc_to_nchw_.run();

// 5. 取出permute_nhwc_to_nchw_的结果作为返回
std::cout << "GpuPreProcessXBGR get result" << std::endl;
nchw_cl_tensor_.map();
memcpy(dest, nchw_cl_tensor_.buffer(), width_dst * height_dst * 3);
nchw_cl_tensor_.unmap();
return;

} ` platform rk3399， Gpu：maliT860 branch v22.08

This program runs on some rk3399 boards and can normally call the gpu. You can see that the range occupied by the gpu is 20-100. And the recognition thread using about 10% cpu.

However, when the same program is running on some other rk3399 boards, the thread that called the function has a high CPU usage, but a low gpu usage.

gpu usage remains between 0 and 1

And It is strange that all boards can print device Mali-T860 by the code std::cout << "Device name: " << std::get<1>(ctx_dev_err).getInfo() << std::endl;

[morgolock]

Hi @junyuancat1

I''d advice using one of the examples in ACL to implement your test, a good candidate would be: https://github.com/ARM-software/ComputeLibrary/blob/main/examples/cl_sgemm.cpp

In ACL, the computation for CL operators and functions is performaned on the GPU so you should not see high CPU usage.

Without the full source code of your test it's difficult to speculate about the reasons you see this behavior on some boards. I would suspect the software configuration on these boards is not identical (mali drivers, power managerment. frequencies, etc)

Hope this helps.