关于mali-T860这款GPU的带宽问题

Question

问题遇到的现象和发生背景

我有一块开发板使用的是rk3399这款瑞芯微的芯片，带有一个mali-T860的GPU，烧入了ubuntu18.04系统，系统中有用OpenCL。
我是用OpenCL测试了GPU带宽性能，发现Device2Host和Host2Device这两个方向的带宽是4000MB/s，而Device2Device这个方向的带宽是2000MB/s。
请问为何Device2Host和Host2Device这两个方向的带宽会比Device2Device这个方向的带宽多一倍？个人认为因为mali-gpu和cpu是共享内存的，难道不应该是三个方向都差不多吗？

问题相关代码，请勿粘贴截图

///////////////////////////////////////////////////////////////////////////////
//  test the bandwidth of a device to host memcopy of a specific size
///////////////////////////////////////////////////////////////////////////////
double testHostToDeviceTransfer(unsigned int memSize, accessMode accMode, memoryMode memMode)
{
    double elapsedTimeInSec = 0.0;
    double bandwidthInMBs = 0.0;
    unsigned char* h_data = NULL;
    cl_mem cmPinnedData = NULL;
    cl_mem cmDevData = NULL;
    cl_int ciErrNum = CL_SUCCESS;

    // Allocate and init host memory, pinned or conventional
    if(memMode == PINNED)
   { 
        // Create a host buffer
        cmPinnedData = clCreateBuffer(cxGPUContext, CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, memSize, NULL, &ciErrNum);
        oclCheckError(ciErrNum, CL_SUCCESS);

        // Get a mapped pointer
        h_data = (unsigned char*)clEnqueueMapBuffer(cqCommandQueue, cmPinnedData, CL_TRUE, CL_MAP_WRITE, 0, memSize, 0, NULL, NULL, &ciErrNum);
        oclCheckError(ciErrNum, CL_SUCCESS);

        //initialize 
        for(unsigned int i = 0; i < memSize/sizeof(unsigned char); i++)
        {
            h_data[i] = (unsigned char)(i & 0xff);
        }
    
        // unmap and make data in the host buffer valid
        ciErrNum = clEnqueueUnmapMemObject(cqCommandQueue, cmPinnedData, (void*)h_data, 0, NULL, NULL);
        oclCheckError(ciErrNum, CL_SUCCESS);
        h_data = NULL;  // buffer is unmapped
    }
    else 
    {
        // standard host alloc
        h_data = (unsigned char *)malloc(memSize);

        //initialize 
        for(unsigned int i = 0; i < memSize/sizeof(unsigned char); i++)
        {
            h_data[i] = (unsigned char)(i & 0xff);
        }
    }

    // allocate device memory 
    cmDevData = clCreateBuffer(cxGPUContext, CL_MEM_READ_WRITE, memSize, NULL, &ciErrNum);
    oclCheckError(ciErrNum, CL_SUCCESS);

    // Sync queue to host, start timer 0, and copy data from Host to GPU
    clFinish(cqCommandQueue);
    shrDeltaT(0);
    if(accMode == DIRECT)
    { 
        if(memMode == PINNED) 
        {
            // Get a mapped pointer
            h_data = (unsigned char*)clEnqueueMapBuffer(cqCommandQueue, cmPinnedData, CL_TRUE, CL_MAP_READ, 0, memSize, 0, NULL, NULL, &ciErrNum);
            oclCheckError(ciErrNum, CL_SUCCESS);
        }

        // DIRECT:  API access to device buffer 
        for(unsigned int i = 0; i < MEMCOPY_ITERATIONS; i++)
        {
                ciErrNum = clEnqueueWriteBuffer(cqCommandQueue, cmDevData, CL_FALSE, 0, memSize, h_data, 0, NULL, NULL);
                oclCheckError(ciErrNum, CL_SUCCESS);
        }
        ciErrNum = clFinish(cqCommandQueue);
        oclCheckError(ciErrNum, CL_SUCCESS);
    } 
    else 
    {
        // MAPPED: mapped pointers to device buffer and conventional pointer access
        void* dm_idata = clEnqueueMapBuffer(cqCommandQueue, cmDevData, CL_TRUE, CL_MAP_WRITE, 0, memSize, 0, NULL, NULL, &ciErrNum);
        oclCheckError(ciErrNum, CL_SUCCESS);
        if(memMode == PINNED ) 
        {
            h_data = (unsigned char*)clEnqueueMapBuffer(cqCommandQueue, cmPinnedData, CL_TRUE, CL_MAP_READ, 0, memSize, 0, NULL, NULL, &ciErrNum); 
            oclCheckError(ciErrNum, CL_SUCCESS); 
        } 
        for(unsigned int i = 0; i < MEMCOPY_ITERATIONS; i++)
        {
            memcpy(dm_idata, h_data, memSize);
        }
        ciErrNum = clEnqueueUnmapMemObject(cqCommandQueue, cmDevData, dm_idata, 0, NULL, NULL);
        oclCheckError(ciErrNum, CL_SUCCESS);
    }
    
    //get the the elapsed time in seconds
    elapsedTimeInSec = shrDeltaT(0);
    
    //calculate bandwidth in MB/s
    bandwidthInMBs = ((double)memSize * (double)MEMCOPY_ITERATIONS)/(elapsedTimeInSec * (double)(1 << 20));

    //clean up memory
    if(cmDevData)clReleaseMemObject(cmDevData);
    if(cmPinnedData) 
    {
        clEnqueueUnmapMemObject(cqCommandQueue, cmPinnedData, (void*)h_data, 0, NULL, NULL);
        clReleaseMemObject(cmPinnedData);
    }
    h_data = NULL;

    return bandwidthInMBs;
}
///////////////////////////////////////////////////////////////////////////////
//  test the bandwidth of a device to host memcopy of a specific size
///////////////////////////////////////////////////////////////////////////////
double testDeviceToDeviceTransfer(unsigned int memSize)
{
    double elapsedTimeInSec = 0.0;
    double bandwidthInMBs = 0.0;
    unsigned char* h_idata = NULL;
    cl_int ciErrNum = CL_SUCCESS;
    
    //allocate host memory
    h_idata = (unsigned char *)malloc( memSize );
        
    //initialize the memory
    for(unsigned int i = 0; i < memSize/sizeof(unsigned char); i++)
    {
        h_idata[i] = (unsigned char) (i & 0xff);
    }

    // allocate device input and output memory and initialize the device input memory
    cl_mem d_idata = clCreateBuffer(cxGPUContext, CL_MEM_READ_ONLY, memSize, NULL, &ciErrNum);
    oclCheckError(ciErrNum, CL_SUCCESS);
    cl_mem d_odata = clCreateBuffer(cxGPUContext, CL_MEM_WRITE_ONLY, memSize, NULL, &ciErrNum);         
    oclCheckError(ciErrNum, CL_SUCCESS);
    ciErrNum = clEnqueueWriteBuffer(cqCommandQueue, d_idata, CL_TRUE, 0, memSize, h_idata, 0, NULL, NULL);
    oclCheckError(ciErrNum, CL_SUCCESS);

    // Sync queue to host, start timer 0, and copy data from one GPU buffer to another GPU bufffer
    clFinish(cqCommandQueue);
    shrDeltaT(0);
    for(unsigned int i = 0; i < MEMCOPY_ITERATIONS; i++)
    {
        ciErrNum = clEnqueueCopyBuffer(cqCommandQueue, d_idata, d_odata, 0, 0, memSize, 0, NULL, NULL);                
        oclCheckError(ciErrNum, CL_SUCCESS);
    }    

    // Sync with GPU
    clFinish(cqCommandQueue);
    
    //get the the elapsed time in seconds
    elapsedTimeInSec = shrDeltaT(0);
    
    // Calculate bandwidth in MB/s 
    //      This is for kernels that read and write GMEM simultaneously 
    //      Obtained Throughput for unidirectional block copies will be 1/2 of this #
    bandwidthInMBs = 2.0 * ((double)memSize * (double)MEMCOPY_ITERATIONS)/(elapsedTimeInSec * (double)(1 << 20));

    //clean up memory on host and device
    free(h_idata);
    clReleaseMemObject(d_idata);
    clReleaseMemObject(d_odata);

    return bandwidthInMBs;
}

运行结果及报错内容

我的解答思路和尝试过的方法

我想要达到的结果

Answer 1

你好，我是有问必答小助手，非常抱歉，本次您提出的有问必答问题，技术专家团超时未为您做出解答


本次提问扣除的有问必答次数，将会以问答VIP体验卡（1次有问必答机会、商城购买实体图书享受95折优惠）的形式为您补发到账户。


因为有问必答VIP体验卡有效期仅有1天，您在需要使用的时候【私信】联系我，我会为您补发。