CUDA：是否可以将一个内核视为“主”来执行 memory malloc，并运行其他“逻辑代码”？

Question

I'm porting a C++ program to CUDA, the calculations are all about matrices/vectors. The first ported function is matrix's FFT. After porting matrix's FFT to CUDA, I found: data transter between CPU and GPU take almost all the time.

// interface: do shift and inverse FFT on a matrix
extern "C" int cu_inv_fft_shift(std::complex<double>* ptrDest, int nRows, int nCols) {

    #ifdef ENABLE_DEBUG_TIME_MEASURE
    float ms1, ms2 = 0.f, ms3 = 0.f, ms4 = 0.f;
    cudaEvent_t startEvent, stopEvent;
    cudaEventCreate(&startEvent); cudaEventCreate(&stopEvent);
    #endif

    // step1: cpu -> gpu, and column-major -> row-major
    #ifdef ENABLE_DEBUG_TIME_MEASURE
    cudaEventRecord(startEvent, 0);
    #endif

    cufftDoubleComplex* ptr_data = matrix_to_cu_data(ptrDest, nRows, nCols);

    #ifdef ENABLE_DEBUG_TIME_MEASURE2
    cudaEventRecord(stopEvent, 0); 
    cudaEventSynchronize(startEvent);cudaEventSynchronize(stopEvent);
    cudaEventElapsedTime(&ms1, startEvent, stopEvent);
    #endif

    // step2: do shift on gpu buffer
    #ifdef ENABLE_DEBUG_TIME_MEASURE2
    cudaEventRecord(startEvent, 0);
    #endif

    ptr_data = fft_shift_cd(ptr_data, nRows, nCols);

    #ifdef ENABLE_DEBUG_TIME_MEASURE2
    cudaEventRecord(stopEvent, 0); 
    cudaEventSynchronize(startEvent);cudaEventSynchronize(stopEvent);
    cudaEventElapsedTime(&ms2, startEvent, stopEvent);
    #endif

    // step3: do FFT on gpu buffer
    #ifdef ENABLE_DEBUG_TIME_MEASURE2
    cudaEventRecord(startEvent, 0);
    #endif

    ptr_data = do_fft_cd(ptr_data, nRows, nCols, CUFFT_INVERSE);

    #ifdef ENABLE_DEBUG_TIME_MEASURE2
    cudaEventRecord(stopEvent, 0); 
    cudaEventSynchronize(startEvent);cudaEventSynchronize(stopEvent);
    cudaEventElapsedTime(&ms3, startEvent, stopEvent);
    #endif

    // step4: row-major -> column-major, and gpu -> cpu
    #ifdef ENABLE_DEBUG_TIME_MEASURE2
    cudaEventRecord(startEvent, 0);
    #endif

    ptr_data = cu_data_to_matrix_inv(ptrDest, nRows, nCols, ptr_data);

    #ifdef ENABLE_DEBUG_TIME_MEASURE
    cudaEventRecord(stopEvent, 0); 
    cudaEventSynchronize(startEvent);cudaEventSynchronize(stopEvent);
    cudaEventElapsedTime(&ms4, startEvent, stopEvent);
    #endif

    #ifdef ENABLE_DEBUG_TIME_MEASURE
    cudaEventDestroy(startEvent); cudaEventDestroy(stopEvent);
    //std::cout << __func__ << " called.."<< std::endl;
    printf("%s: %.4fms, %.4fms, %.4fms, %.4fms\n", __func__, ms1, ms2, ms3, ms4);
    #endif

    cudaFree(ptr_data);
    return 0;
}

The measured result when the matrix is 8192x8192:

cu_fwd_fft_shift: 4.2841ms, 0.7394ms, 0.0492ms, 4.2857ms

It means that(It is verified):

• CPU->GPU: 4.2ms.
• forward FFT: 0.7ms.
• FFT shift: 0.05ms.
• GPU->CPU: 4.2ms.

The problem I encountered is that: in a CPU function, there are some "code snippet" (just like the FFT) could be ported to CUDA, but thre are some if/else code, and intermediate memory malloc between them.

I want to reduce data transfer CPU<-->GPU.My optinion is that porting a whole CPU function to CUDA(GPU side), But there are many "logic code" like if/else, intermediate memory malloc.

So my question are:

1. Does it possible to set one core as master(just like CPU) to process these malloc / "logic code" and dispache subsequest calculation to all other cores?
2. Are there any other CUDA projects can I study from? Or
3. Is this solution impossible?

Answer 1

1. Does it possible to set one core as master(just like CPU) to process these malloc / "logic code" and dispache subsequest [sic] calculation to all other cores?

CUDA doesn't expose that level of granularity in its execution model, so that isn't possible. There is dynamic parallelism , which can allow one kernel to dispatch other kernels and offers a very minimal subset of the CUDA runtime API. You might be able to adapt that paradigm to your application.

2. Are there any other CUDA projects can I study from? Or

If you search and read the various material NVIDIA have made available on dynamic parallelism, you might find something you can learn from and make assessment of whether that might work for your use case.

3. Is this solution impossible?

Probably, yes.

In general when you start a GPU programming question or proposition with "I'm porting a C++ program to CUDA", and you mean porting in the most literal sense, you usually are doing something wrong. It is exceedingly rare that a conventional codebase or serial algorithm can be blindly "ported" and either be correct, fast, or both correct and fast. The GPU programming paradigm is rather different to conventional single and multithreaded CPU coding and if you try and treat it like a CPU, you will fail.