#define M 3000 // 第一个矩阵中的行数
#define K 2000 // 第一个矩阵中的列数等于第二个矩阵的行数
#define N 3000 // 第二个矩阵中的列数
//+------------------------------------------------------------------+
const string clSrc=
"#define N "+IntegerToString(N)+" \r\n"
"#define K "+IntegerToString(K)+" \r\n"
" \r\n"
"__kernel void matricesMul( __global float *in1, \r\n"
" __global float *in2, \r\n"
" __global float *out ) \r\n"
"{ \r\n"
" int m = get_global_id( 0 ); \r\n"
" int n = get_global_id( 1 ); \r\n"
" float sum = 0.0; \r\n"
" for( int k = 0; k < K; k ++ ) \r\n"
" sum += in1[ m * K + k ] * in2[ k * N + n ]; \r\n"
" out[ m * N + n ] = sum; \r\n"
"} \r\n";
//+------------------------------------------------------------------+
//| 脚本程序起始函数 |
//+------------------------------------------------------------------+
void OnStart()
{
//--- 初始化随机数生成器
MathSrand((int)TimeCurrent());
//--- 用随机值填写给定大小的矩阵
matrixf mat1(M, K, MatrixRandom) ; // 第一个矩阵
matrixf mat2(K, N, MatrixRandom); //第二个矩阵
//--- 使用原始方法计算矩阵的乘积
uint start=GetTickCount();
matrixf matrix_naive=matrixf::Zeros(M, N);// 在这里设置两个矩阵相乘的结果
for(int m=0; m<M; m++)
for(int k=0; k<K; k++)
for(int n=0; n<N; n++)
matrix_naive[m][n]+=mat1[m][k]*mat2[k][n];
uint time_naive=GetTickCount()-start;
//--- 通过MatMull计算矩阵的乘积
start=GetTickCount();
matrixf matrix_matmul=mat1.MatMul(mat2);
uint time_matmul=GetTickCount()-start;
//--- 在OpenCL中计算矩阵的乘积
matrixf matrix_opencl=matrixf::Zeros(M, N);
int cl_ctx; // 环境句柄
if((cl_ctx=CLContextCreate(CL_USE_GPU_ONLY))==INVALID_HANDLE)
{
Print("OpenCL not found, leaving");
return;
}
int cl_prg; // 程序句柄
int cl_krn; // 内核句柄
int cl_mem_in1; // 第一个(输入)缓冲区句柄
int cl_mem_in2; // 第二个(输入)缓冲区句柄
int cl_mem_out; // 第三个(输出)缓冲区句柄
//--- 创建程序和内核
cl_prg = CLProgramCreate(cl_ctx, clSrc);
cl_krn = CLKernelCreate(cl_prg, "matricesMul");
//--- 为三个矩阵创建所有三个缓冲区
cl_mem_in1=CLBufferCreate(cl_ctx, M*K*sizeof(float), CL_MEM_READ_WRITE);
cl_mem_in2=CLBufferCreate(cl_ctx, K*N*sizeof(float), CL_MEM_READ_WRITE);
//--- 第三个矩阵 - 输出
cl_mem_out=CLBufferCreate(cl_ctx, M*N*sizeof(float), CL_MEM_READ_WRITE);
//--- 设置内核参数
CLSetKernelArgMem(cl_krn, 0, cl_mem_in1);
CLSetKernelArgMem(cl_krn, 1, cl_mem_in2);
CLSetKernelArgMem(cl_krn, 2, cl_mem_out);
//--- 将矩阵写入设备缓冲区
CLBufferWrite(cl_mem_in1, 0, mat1);
CLBufferWrite(cl_mem_in2, 0, mat2);
CLBufferWrite(cl_mem_out, 0, matrix_opencl);
//--- OpenCL代码执行时间开始
start=GetTickCount();
//--- 设置任务工作区参数,并执行OpenCL程序
uint offs[2] = {0, 0};
uint works[2] = {M, N};
start=GetTickCount();
bool ex=CLExecute(cl_krn, 2, offs, works);
//--- 计算矩阵结果
if(CLBufferRead(cl_mem_out, 0, matrix_opencl))
PrintFormat("[%d x %d] matrix read: ", matrix_opencl.Rows(), matrix_opencl.Cols());
else
Print("CLBufferRead(cl_mem_out, 0, matrix_opencl failed. Error ",GetLastError());
uint time_opencl=GetTickCount()-start;
Print("Compare calculation time using each method");
PrintFormat("Naive product time = %d ms",time_naive);
PrintFormat("MatMul product time = %d ms",time_matmul);
PrintFormat("OpenCl product time = %d ms",time_opencl);
//--- 释放所有OpenCL环境
CLFreeAll(cl_ctx, cl_prg, cl_krn, cl_mem_in1, cl_mem_in2, cl_mem_out);
//--- 比较所有获得的结果矩阵
Print("How many discrepancy errors are there between result matrices?");
ulong errors=matrix_naive.Compare(matrix_matmul,(float)1e-12);
Print("matrix_direct.Compare(matrix_matmul,1e-12)=",errors);
errors=matrix_matmul.Compare(matrix_opencl,float(1e-12));
Print("matrix_matmul.Compare(matrix_opencl,1e-12)=",errors);
/*
结果:
[3000 x 3000] matrix read:
比较每种方法的计算时间
Naive product time = 54750 ms
MatMul product time = 4578 ms
OpenCl product time = 922 ms
结果矩阵之间有多少误差?
matrix_direct.Compare(matrix_matmul,1e-12)=0
matrix_matmul.Compare(matrix_opencl,1e-12)=0
*/
}
//+------------------------------------------------------------------+
//| 用随机值填写矩阵 |
//+------------------------------------------------------------------+
void MatrixRandom(matrixf& m)
{
for(ulong r=0; r<m.Rows(); r++)
{
for(ulong c=0; c<m.Cols(); c++)
{
m[r][c]=(float)((MathRand()-16383.5)/32767.);
}
}
}
//+------------------------------------------------------------------+
//| 释放所有OpenCL环境 |
//+------------------------------------------------------------------+
void CLFreeAll(int cl_ctx, int cl_prg, int cl_krn,
int cl_mem_in1, int cl_mem_in2, int cl_mem_out)
{
//--- 以相反顺序删除OpenCL创建的所有环境
CLBufferFree(cl_mem_in1);
CLBufferFree(cl_mem_in2);
CLBufferFree(cl_mem_out);
CLKernelFree(cl_krn);
CLProgramFree(cl_prg);
CLContextFree(cl_ctx);
}
|