Question

我在理解比较结果时遇到了一些问题。

配备i7 / intel hd4000的笔记本电脑和配备8Xeon 5400/7970 HDRadeon的服务器。

我在成倍增长：

int M =1024*2, N = 1024*6, P = 1024*2;
    //      N               P      
    //|-----------|   |-----------|
    //|           |   |           |
    //|M          | * |N          |
    //|           |   |           |
    //|-----------|   |-----------|

这是内核：

/*
 * Copyright 1993-2010 NVIDIA Corporation.  All rights reserved.
 *
 * Please refer to the NVIDIA end user license agreement (EULA) associated
 * with this source code for terms and conditions that govern your use of
 * this software. Any use, reproduction, disclosure, or distribution of
 * this software and related documentation outside the terms of the EULA
 * is strictly prohibited.
 *
 */

/* Matrix multiplication: C = A * B.
 * Device code.
 */
#ifndef BLOCK_SIZE
#define BLOCK_SIZE 16
#endif

#define AS(i, j) As[j + i * BLOCK_SIZE]
#define BS(i, j) Bs[j + i * BLOCK_SIZE]

///////////////////////////////////////////////////////////////////////////////
//! Matrix multiplication on the device: C = A * B
//! uiWA is A's width and uiWB is B's width
////////////////////////////////////////////////////////////////////////////////
__kernel void
m_m_mul( __global float* A, __global float* B, __global float* C, 
       /*__local float* As, __local float* Bs,*/ int uiWA, int uiWB, int trueLocalSize1)
{

   __local float As[BLOCK_SIZE*BLOCK_SIZE];
   __local float Bs[BLOCK_SIZE*BLOCK_SIZE];

    // Block index
    int bx = get_group_id(0);
    int by = get_group_id(1);

    // Thread index
    int tx = get_local_id(0);
    int ty = get_local_id(1);

    // Index of the first sub-matrix of A processed by the block
    int aBegin = uiWA * BLOCK_SIZE * by;

    // Index of the last sub-matrix of A processed by the block
    int aEnd   = aBegin + uiWA - 1;

    // Step size used to iterate through the sub-matrices of A
    int aStep  = BLOCK_SIZE;

    // Index of the first sub-matrix of B processed by the block
    int bBegin = BLOCK_SIZE * bx;

    // Step size used to iterate through the sub-matrices of B
    int bStep  = BLOCK_SIZE * uiWB;

    // Csub is used to store the element of the block sub-matrix
    // that is computed by the thread
    float Csub = 0.0f;

    // Loop over all the sub-matrices of A and B
    // required to compute the block sub-matrix
    for (int a = aBegin, b = bBegin;
             a <= aEnd;
             a += aStep, b += bStep) {

        // Load the matrices from device memory
        // to shared memory; each thread loads
        // one element of each matrix
        AS(ty, tx) = A[a + uiWA * ty + tx];
        BS(ty, tx) = B[b + uiWB * ty + tx];

        // Synchronize to make sure the matrices are loaded
        barrier(CLK_LOCAL_MEM_FENCE);

        // Multiply the two matrices together;
        // each thread computes one element
        // of the block sub-matrix        
        #pragma unroll
        for (int k = 0; k < BLOCK_SIZE; ++k)
            Csub += AS(ty, k) * BS(k, tx);

        // Synchronize to make sure that the preceding
        // computation is done before loading two new
        // sub-matrices of A and B in the next iteration
        barrier(CLK_LOCAL_MEM_FENCE);
    }

    if (get_global_id(1) < trueLocalSize1)
    // Write the block sub-matrix to device memory;
    // each thread writes one element
    C[get_global_id(1) * get_global_size(0) + get_global_id(0)] = Csub;

}

我与Eigen::Matrix<float,-1,-1,Eigen::RowMajor> m4 = m1 * m2比较;

在服务器上：

Creating matrices on GPU....... Done [0ms]
Creating matrices on CPU....... Done [0ms]
Filling GPU with random numbers....... Done [19ms]
M3 = M1 * M2... on GPU (Loading Kernels)... Done [240ms]
M3 = M1 * M2... on GPU (3 times)... Done [211ms]
Loading M1, M2 on GPU... Done [93ms]
M4 = M1 * M2 on CPU... Done [7775ms] Error:3.78049e-008
Press any key to continue . . .
Matlab: Elapsed time is 3.010626 seconds.

在笔记本电脑上：

Creating matrices on GPU....... Done [22ms]
Creating matrices on CPU....... Done [0ms]
Filling GPU with random numbers....... Done [35ms]
M3 = M1 * M2... on GPU (Loading Kernels)... Done [2975ms]
M3 = M1 * M2... on GPU (3 times)... Done [6891ms]
Loading M1, M2 on GPU... Done [80ms]
M4 = M1 * M2 on CPU... Done [5966ms] Error:3.78049e-008
Press any key to continue . . .
Matlab: Elapsed time is 2.310626 seconds.

我现在的问题。 1）为什么笔记本电脑的特性比xeon的8核更快。可能是因为本征只在两个系统上使用一个核心而i7具有更高的时钟速度？2.0 vs 2.4？

2）在labtop上使用Intel HD4000与Eigen的速度提高了近3倍，但Matlab在相同的乘法中需要2.3秒。这与HD4000上的内核相同。（我能做些什么让Eigen以与Matlab相同的速度运行吗？）

Answer 1

http://eigen.tuxfamily.org/dox/TopicMultiThreading.html

在visual studio中启用OpenMP使我的代码以8核运行并且速度显着降低。运行时间80％的matlabs速度。

Cores: 8
M: 4096 N:12288 P:4096
Creating matrices on GPU....... Done [0ms]
Creating matrices on CPU....... Done [0ms]
Filling GPU with random numbers....... Done [44ms]
M3 = M1 * M2... on GPU (Loading Kernels)... Done [850ms]
M3 = M1 * M2... on GPU (3 times)... Done [2063ms]
Loading M1, M2 on GPU... Done [355ms]
M4 = M1 * M2 on CPU... Done [22263ms] Error:5.70124e-007
Press any key to continue . . .

Radeon HD 7970和Intel HD 4000 GPU上的OpenCL矩阵乘法与CPU上的Eigen相比

1 个答案: