int main(int argc, char *argv[])
{ // export OMP_NUM_THREADS=1
float **A, **B, **C; // matrices
int d1, d2, d3; // dimensions of matrices
int i, j, k; // loop variables
double start, end;
start = omp_get_wtime();
/* print user instruction */
if (argc != 4)
{
printf ("Matrix multiplication: C = A x B\n");
printf ("Usage: %s <NumRowA> <NumColA> <NumColB>\n", argv[0]);
return 0;
}
/* read user input */
d1 = atoi(argv[1]); // rows of A and C
d2 = atoi(argv[2]); // cols of A and rows of B
d3 = atoi(argv[3]); // cols of B and C
printf("Matrix sizes C[%d][%d] = A[%d][%d] x B[%d][%d]\n", d1, d3, d1, d2, d2, d3);
/* prepare matrices */
A = alloc_mat(d1, d2);
init_mat(A, d1, d2);
B = alloc_mat(d2, d3);
init_mat(B, d2, d3);
C = alloc_mat(d1, d3); // no initialisation of C, because it gets filled by matmult
/* serial version of matmult */
printf("Perform matrix multiplication...\n");
/* spezielle collapse-Schleife, um über ein mehrdimensionales Array zu iterieren
Schleifen müssen sehr einfach gehalten sein, damit Parallelisierung erfolgen kann
Alle Schleifenvariablen müssen völlig unabhängig voneinander sein.
Sind es aber nicht, die Ausgabe war Fehlerhaft. */
double sum;
// #pragma omp parallel for collapse(3) schedule(dynamik)
for (i = 0; i < d1; i++)
for (j = 0; j < d3; j++)
#pragma omp parallel for private(sum)// Rechenintensive Operation wird parallelisiert.
for (k = 0; k < d2; k++)
{ // Nur hier darf beliebiger Code stehen! wenn collaps verwendet würde
sum = A[i][k] * B[k][j];
#pragma omp atomic
C[i][j] += sum;
}
/* test output */
print_mat(A, d1, d2, "A");
print_mat(B, d2, d3, "B");
print_mat(C, d1, d3, "C");
printf ("\nDone.\n");
end = omp_get_wtime();
printf("This task took %f seconds\n", end-start);
return 0;
}
int main(int argc, char **argv){
MPI_Init(argc, argv);
MPI_Barrier(MPI_COMM_WORLD);
int nb;
int rank = -1;
double t,start,stop;
double* mat_A;
double* mat_B;
double* mat_res = alloc_mat();
// Allocations
mat_A = init_mat();
mat_B = init_mat();
// Init
printf("Nb.threads\tTps.\n");
for(nb=1;nb<=PROC;nb++){
MPI_Comm_size(MPI_COMM_WORLD, &nb);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
start = get_time();
multiply_mat(mat_A, mat_B, mat_res);
stop=get_time();
t=stop-start;
printf("%d\t%f\n",nb,t);
}
// Memory free
free(mat_A);
free(mat_B);
free(mat_res);
return EXIT_SUCCESS;
MPI_Finalize(void);
}
int main(int argc, char *argv[])
{
float **A, **B, **C; // matrices
int d1, d2, d3; // dimensions of matrices
int i, j, k; // loop variables
/* print user instruction */
if (argc != 4)
{
printf ("Matrix multiplication: C = A x B\n");
printf ("Usage: %s <NumRowA> <NumColA> <NumColB>\n", argv[0]);
return 0;
}
/* read user input */
d1 = atoi(argv[1]); // rows of A and C
d2 = atoi(argv[2]); // cols of A and rows of B
d3 = atoi(argv[3]); // cols of B and C
printf("Matrix sizes C[%d][%d] = A[%d][%d] x B[%d][%d]\n", d1, d3, d1, d2, d2, d3);
/* prepare matrices */
A = alloc_mat(d1, d2);
init_mat(A, d1, d2);
B = alloc_mat(d2, d3);
init_mat(B, d2, d3);
C = alloc_mat(d1, d3); // no initialisation of C, because it gets filled by matmult
/* serial version of matmult */
printf("Perform matrix multiplication...\n");
for (i = 0; i < d1; i++)
for (j = 0; j < d3; j++)
for (k = 0; k < d2; k++)
C[i][j] += A[i][k] * B[k][j];
/* test output */
print_mat(A, d1, d2, "A");
print_mat(B, d2, d3, "B");
print_mat(C, d1, d3, "C");
printf ("\nDone.\n");
return 0;
}
double* init_mat(){
double* m = alloc_mat();
int i, j;
for(i = 0; i < SIZE; i++){
for(j = 0; j < SIZE; j++){
m[i*SIZE + j] = (double)rand()/(double)RAND_MAX;
}
}
return m;
}
/** **/
int main (int argc, char* argv[])
{
int WORK_DIM = 2; // Wie viele Dimensionen hat der Indexraum?
std::chrono::time_point<std::chrono::system_clock> s_start, s_end, p_start, p_end;
// Lese den Kernel dynamisch ein: (uebernommen von Foliensatz 9, Folie 20)
FILE *fp;
const char *FileName = "matmult.cl";
char *KernelSource;
fp = fopen(FileName, "r");
if (!fp) {
printf("Can't open kernel source: %s", FileName); exit(1);
}
KernelSource = (char *)malloc(MAX_SOURCE_SIZE);
size_t kernel_s_size = fread(KernelSource, 1, MAX_SOURCE_SIZE, fp);
fclose(fp);
cl_int err;
cl_platform_id* platforms = NULL;
char platform_name[1024];
cl_device_id device_id = NULL;
cl_uint num_of_platforms = 0,
num_of_devices = 0;
cl_context context;
cl_kernel kernel;
cl_command_queue command_queue;
cl_program program;
err = clGetPlatformIDs(0, NULL, &num_of_platforms);
if (err != CL_SUCCESS) {
printf("No platforms found. Error: %d\n", err);
return 0;
}
// Liefert Plattformen
platforms = (cl_platform_id *)malloc(num_of_platforms);
err = clGetPlatformIDs(num_of_platforms, platforms, NULL);
if (err != CL_SUCCESS) {
printf("No platforms found. Error: %d\n", err);
return 0;
} else {
int nvidia_platform = 0; // Speichert den Rang der letzten NVIDIA-Plattform
for (unsigned int i=0; i<num_of_platforms; i++) // Fuer jede Plattform:
{
clGetPlatformInfo(platforms[i], CL_PLATFORM_NAME, sizeof(platform_name), platform_name, NULL);
if (err != CL_SUCCESS) {
printf("Could not get information about platform. Error: %d\n", err);
return 0;
}
if (strstr(platform_name, "NVIDIA") != NULL) { // Falls die Plattform eine NVIDIA-Plattform ist: Speichere ihren Rang
nvidia_platform = i;
break;
}
}
// Gibt die ID des Devices der NVIDIA-Plattform zurueck
err = clGetDeviceIDs(platforms[nvidia_platform], CL_DEVICE_TYPE_GPU, 1, &device_id, &num_of_devices);
if (err != CL_SUCCESS) {
printf("Could not get device in platform. Error: %d\n", err);
return 0;
}
}
// Erschaffe einen OpenCl-context, in dem OpenCl-Datenobjekte verwaltet werden koennen
context = clCreateContext(0, 1, &device_id, NULL, NULL, &err);
if (err != CL_SUCCESS) {
printf("Unable to create context. Error: %d\n", err);
return 0;
}
// Initialisiere eine Befehlswarteschleife, die Befehle fuer OpenCl-Objekte speichern kann
command_queue = clCreateCommandQueue(context, device_id, 0, &err);
if (err != CL_SUCCESS) {
printf("Unable to create command queue. Error: %d\n", err);
return 0;
}
// Initialisiere ein Programm und spezifiziere, aus welchem Code dieses kompiliert werden soll
program = clCreateProgramWithSource(context, 1, (const char **)&KernelSource, (const size_t *)& kernel_s_size, &err);
if (err != CL_SUCCESS) {
printf("Unable to create program. Error: %d\n", err);
return 0;
}
// Kompiliere das Programm zur Laufzeit
err = clBuildProgram(program, 0, NULL, NULL, NULL, NULL);
if (err != CL_SUCCESS) {
// Zeige Compilermeldungen an: (uebernommen von Foliensatz 9, Folie 23)
char *log;
size_t size;
clGetProgramBuildInfo(program, device_id, CL_PROGRAM_BUILD_LOG, 0, NULL, &size);
log = (char *)malloc(size+1);
if (log) {
clGetProgramBuildInfo(program, device_id, CL_PROGRAM_BUILD_LOG, size, log, NULL);
log[size] = '\0';
printf("%s", log);
free(log);
}
printf("Error building program. Error: %d\n", err);
return 0;
}
// Erschaffe einen Kernel und lade oben kompiliertes Programm ein
kernel = clCreateKernel(program, "matmult", &err);
if (err != CL_SUCCESS) {
printf("Error setting kernel. Error: %d\n", err);
return 0;
}
float **A, **B, **C; // Matrizen
int dim1, dim2, dim3; // Matrixdimensionen
dim1 = D1; // Zeilen von A, Zeilen von C
dim2 = D2; // Spalten von A, Zeilen von B
dim3 = D3; // Spalten von B, Spalten von C
A = alloc_mat(dim1, dim2);
B = alloc_mat(dim2, dim3);
C = alloc_mat(dim1, dim3);
init_mat(A, dim1, dim2);
init_mat(B, dim2, dim3);
cl_mem in_A, in_B, output;
// float data[DATA_SIZE] = {1, 2, 3, 4, 5, 6, 7, 8, 9, 10};
size_t global[1] = {dim1*dim3}; // Dimensionen von C
size_t global_two[2] = {dim1, dim3};
in_A = clCreateBuffer (context, CL_MEM_READ_ONLY, sizeof(float)*dim1*dim2, NULL, &err);
in_B = clCreateBuffer (context, CL_MEM_READ_ONLY, sizeof(float)*dim2*dim3, NULL, &err);
output = clCreateBuffer (context, CL_MEM_WRITE_ONLY, sizeof(float)*dim1*dim3, NULL, &err);
clEnqueueWriteBuffer(command_queue, in_A, CL_TRUE, 0, sizeof(float)*dim1*dim2, *A, 0, NULL, NULL);
clEnqueueWriteBuffer(command_queue, in_B, CL_TRUE, 0, sizeof(float)*dim2*dim3, *B, 0, NULL, NULL);
clSetKernelArg(kernel, 0, sizeof(cl_mem), &in_A);
clSetKernelArg(kernel, 1, sizeof(cl_mem), &in_B);
clSetKernelArg(kernel, 2, sizeof(cl_mem), &output);
// clSetKernelArg(kernel, 3, sizeof(int), &dim2);
// clSetKernelArg(kernel, 4, sizeof(int), &dim3);
clEnqueueNDRangeKernel(command_queue, kernel, 1, NULL, global, NULL, 0, NULL, NULL);
if (WORK_DIM == 2) {
clEnqueueNDRangeKernel(command_queue, kernel, 2, NULL, global_two, NULL, 0, NULL, NULL);
}
// Zeitmessung fuer parallele Version
p_start = std::chrono::system_clock::now();
err = clFinish(command_queue);
p_end = std::chrono::system_clock::now();
std::chrono::duration<double> p_duration = p_end - p_start;
if (err == CL_INVALID_COMMAND_QUEUE ) {
printf("CL_INVALID_COMMAND_QUEUE: %d\n", err);
return 0;
}
clEnqueueReadBuffer(command_queue, output, CL_TRUE, 0, sizeof(float)*dim1*dim3, *C, 0, NULL, NULL);
// Ueberpruefe, ob serielle Version und parallele gleich sind:
float **correct_matrix;
correct_matrix = alloc_mat(dim1, dim3);
s_start = std::chrono::system_clock::now(); // Zeitmessung fuer serielle Version
correct_matrix = mult_mat(A, B, dim1, dim2, dim3);
s_end = std::chrono::system_clock::now();
std::chrono::duration<double> s_duration = s_end - s_start;
is_correct(C, correct_matrix, dim1, dim3); // Numerischer Korrektheitsbeweis
print_mat(C, dim1, dim3, "C = ");
print_mat(correct_matrix, dim1, dim3, "correct_matrix = ");
// printf("Kernel execution time: %f\n", t_end-t_start);
clReleaseMemObject(in_A);
clReleaseMemObject(in_B);
clReleaseMemObject(output);
clReleaseProgram(program);
clReleaseKernel(kernel);
err = clReleaseCommandQueue(command_queue); //!!
if (err != CL_SUCCESS) {
printf("Error releasing command queue: %d\n", err);
return 0;
}
clReleaseContext(context);
printf("Dauer der seriellen Version: %.2f Millisekunden\n", s_duration.count() * 1000);
printf("Dauer der parallelen Version: %.2f Millisekunden\n", p_duration.count() * 1000);
printf("Erhaltenes Speed Up: %.2f \n", p_duration.count() / p_duration.count());
return 0;
}
int main(int argc, char** argv)
{
double serial_time, openCL_time, start_time;
cl_int err;
cl_platform_id* platforms = NULL;
char platform_name[1024];
cl_device_id device_id = NULL;
cl_uint num_of_platforms = 0;
cl_uint num_of_devices = 0;
cl_context context;
cl_kernel kernel;
cl_command_queue command_queue;
cl_program program;
cl_mem input1, input2, input3, output;
float **A, **B, **C, **serialC; // matrices
int d1, d2, d3; // dimensions of matrices
/* print user instruction */
if (argc != 4)
{
printf("Matrix multiplication: C = A x B\n");
printf("Usage: %s <NumRowA> <NumColA> <NumColB>\n", argv[0]);
return 0;
}
/* read user input */
d1 = 1000; // rows of A and C
d2 = 1000; // cols of A and rows of B
d3 = 1000; // cols of B and C
int d[4] = { 0, d1, d2, d3 };
size_t global[1] = { (size_t)d1*d3 };
printf("Matrix sizes C[%d][%d] = A[%d][%d] x B[%d][%d]\n", d1, d3, d1, d2, d2, d3);
/* prepare matrices */
A = alloc_mat(d1, d2);
init_mat(A, d1, d2);
B = alloc_mat(d2, d3);
init_mat(B, d2, d3);
C = alloc_mat(d1, d3);
serialC = alloc_mat(d1, d3);
err = clGetPlatformIDs(0, NULL, &num_of_platforms);
if (err != CL_SUCCESS) {
printf("No platforms found. Error: %d\n", err);
return 0;
}
platforms = (cl_platform_id *)malloc(num_of_platforms);
err = clGetPlatformIDs(num_of_platforms, platforms, NULL);
if (err != CL_SUCCESS) {
printf("No platforms found. Error: %d\n", err);
return 0;
}
else {
int nvidia_platform = 0;
for (unsigned int i = 0; i<num_of_platforms; i++) {
clGetPlatformInfo(platforms[i], CL_PLATFORM_NAME, sizeof(platform_name), platform_name, NULL);
if (err != CL_SUCCESS) {
printf("Could not get information about platform. Error: %d\n", err);
return 0;
}
if (strstr(platform_name, "NVIDIA") != NULL) {
nvidia_platform = i;
break;
}
}
err = clGetDeviceIDs(platforms[nvidia_platform], CL_DEVICE_TYPE_GPU, 1, &device_id, &num_of_devices);
if (err != CL_SUCCESS) {
printf("Could not get device in platform. Error: %d\n", err);
return 0;
}
}
context = clCreateContext(0, 1, &device_id, NULL, NULL, &err);
if (err != CL_SUCCESS) {
printf("Unable to create context. Error: %d\n", err);
return 0;
}
command_queue = clCreateCommandQueue(context, device_id, 0, &err);
if (err != CL_SUCCESS) {
printf("Unable to create command queue. Error: %d\n", err);
return 0;
}
program = clCreateProgramWithSource(context, 1, (const char **)&KernelSource, NULL, &err);
if (err != CL_SUCCESS) {
printf("Unable to create program. Error: %d\n", err);
return 0;
}
if (clBuildProgram(program, 0, NULL, NULL, NULL, NULL) != CL_SUCCESS) {
char *log;
size_t size;
clGetProgramBuildInfo(program, device_id, CL_PROGRAM_BUILD_LOG, 0, NULL, &size); // 1. Länge des Logbuches?
log = (char *)malloc(size + 1);
if (log) {
clGetProgramBuildInfo(program, device_id, CL_PROGRAM_BUILD_LOG, size, log, NULL); // 2. Hole das Logbuch ab
log[size] = '\0';
printf("%s", log);
free(log);
}
return 1;
}
err = clBuildProgram(program, 0, NULL, NULL, NULL, NULL);
if (err != CL_SUCCESS) {
printf("Error building program. Error: %d\n", err);
return 0;
}
kernel = clCreateKernel(program, "matmult_ocl", &err);
if (err != CL_SUCCESS) {
printf("Error setting kernel. Error: %d\n", err);
return 0;
}
input1 = clCreateBuffer(context, CL_MEM_READ_ONLY, d1*d2*sizeof(float), NULL, &err);
input2 = clCreateBuffer(context, CL_MEM_READ_ONLY, d2*d3*sizeof(float), NULL, &err);
input3 = clCreateBuffer(context, CL_MEM_READ_ONLY, 4 * sizeof(int), NULL, &err);
output = clCreateBuffer(context, CL_MEM_READ_WRITE, d1*d3*sizeof(float), NULL, &err);
start_time = omp_get_wtime();
clEnqueueWriteBuffer(command_queue, input1, CL_TRUE, 0, d1*d2*sizeof(float), *A, 0, NULL, NULL);
clEnqueueWriteBuffer(command_queue, input2, CL_TRUE, 0, d2*d3*sizeof(float), *B, 0, NULL, NULL);
clEnqueueWriteBuffer(command_queue, input3, CL_TRUE, 0, 4 * sizeof(int), d, 0, NULL, NULL);
clSetKernelArg(kernel, 0, sizeof(cl_mem), &input1);
clSetKernelArg(kernel, 1, sizeof(cl_mem), &input2);
clSetKernelArg(kernel, 2, sizeof(cl_mem), &input3);
clSetKernelArg(kernel, 3, sizeof(cl_mem), &output);
clEnqueueNDRangeKernel(command_queue, kernel, 1, NULL, global, NULL, 0, NULL, NULL);
clFinish(command_queue);
clEnqueueReadBuffer(command_queue, output, CL_TRUE, 0, d1*d3*sizeof(float), *C, 0, NULL, NULL);
// for (unsigned int i = 0; i < (unsigned int) d1*d3; i++)
// printf("%f\n", C[0][i]);
openCL_time = omp_get_wtime() - start_time;
clReleaseMemObject(input1);
clReleaseMemObject(input2);
clReleaseMemObject(input3);
clReleaseMemObject(output);
clReleaseProgram(program);
clReleaseKernel(kernel);
clReleaseCommandQueue(command_queue);
clReleaseContext(context);
printf("Running serial algorithm...\n");
start_time = omp_get_wtime();
serialC = mult_mat(A, B, d1, d2, d3);
serial_time = omp_get_wtime() - start_time;
printf("Checking results... ");
is_correct(C, serialC, d1, d3);
printf("Showing stats...\n");
printf(" serial runtime = %f\n", serial_time);
printf(" OpenCL runtime = %f\n", openCL_time);
printf(" Speedup = %f\n", serial_time / openCL_time);
return 0;
}
