increase performance of opencl - c++

Question

I am trying to implement some image processing algorithm using opencl. But as i see when i use opencl it is taking around 0.5 ms to complet one process i.e one frame. Isn't there any way than i initialize the opencl parameters only once using class object declaration than only call a function run the main kernel? I tried like this by creating class but as i find context, device can't be declared and used seperately and needs to be created each time.
#include <CL/cl.hpp>
#include <chrono>
#include <iostream>
using namespace std::chrono;
using namespace std;
namespace Color {
enum Code {
FG_RED = 31,
FG_GREEN = 32,
FG_BLUE = 34,
FG_DEFAULT = 39,
BG_RED = 41,
BG_GREEN = 42,
BG_BLUE = 44,
BG_DEFAULT = 49
};
class Modifier {
Code code;
public:
Modifier(Code pCode) : code(pCode) {}
friend std::ostream& operator<<(std::ostream& os, const Modifier& mod) {
return os << "\033[" << mod.code << "m";
}
};
} // namespace Color
class useOpenCL {
public:
int size = 294400;
std::vector<cl::Platform> all_platforms;
std::vector<cl::Device> all_devices;
cl::Platform default_platform;
cl::Device default_device;
cl::Program::Sources sources;
std::string kernel_code;
cl::Kernel kernel_add;
cl::Buffer buffer_A;
useOpenCL();
~useOpenCL() {}
void backgroundSub();
};
useOpenCL::useOpenCL() {
Color::Modifier green(Color::FG_GREEN);
Color::Modifier red(Color::FG_RED);
Color::Modifier def(Color::FG_DEFAULT);
// get all platforms (drivers)
cl::Platform::get(&all_platforms);
if (all_platforms.size() == 0) {
std::cout << red << " No platforms found. Check OpenCL installation!" << def
<< endl;
exit(1);
}
default_platform = all_platforms[0];
std::cout << green << "Using platform: " << def
<< default_platform.getInfo<CL_PLATFORM_NAME>() << std::endl;
// get default device of the default platform
default_platform.getDevices(CL_DEVICE_TYPE_ALL, &all_devices);
if (all_devices.size() == 0) {
std::cout << red << " No devices found. Check OpenCL installation!" << def
<< endl;
exit(1);
}
default_device = all_devices[0];
std::cout << green << "Using device: " << def
<< default_device.getInfo<CL_DEVICE_NAME>() << std::endl;
// kernel calculates for each element C=A+B
kernel_code =
" void kernel simple_add(global const int* A, global const int* B, "
"global int* C){ "
" C[get_global_id(0)]=A[get_global_id(0)]+B[get_global_id(0)]; "
" "
" } "
" ";
sources.push_back({kernel_code.c_str(), kernel_code.length()});
}
void useOpenCL::backgroundSub() {
int A[size], B[size];
for (int i = 0; i < size; i++) {
A[i] = i;
B[i] = i + 1;
}
auto start1 = high_resolution_clock::now();
cl::Context context({default_device});
cl::Program program(context, sources);
if (program.build({default_device}) != CL_SUCCESS) {
std::cout << " Error building: "
<< program.getBuildInfo<CL_PROGRAM_BUILD_LOG>(default_device)
<< "\n";
exit(1);
}
// create buffers on the device
cl::Buffer buffer_A(context, CL_MEM_READ_WRITE, sizeof(int) * size);
cl::Buffer buffer_B(context, CL_MEM_READ_WRITE, sizeof(int) * size);
cl::Buffer buffer_C(context, CL_MEM_READ_WRITE, sizeof(int) * size);
// create queue to which we will push commands for the device.
cl::CommandQueue queue(context, default_device);
// write arrays A and B to the device
queue.enqueueWriteBuffer(buffer_A, CL_TRUE, 0, sizeof(int) * size, A);
queue.enqueueWriteBuffer(buffer_B, CL_TRUE, 0, sizeof(int) * size, B);
// run the kernel
/*cl::KernelFunctor
simple_add(cl::Kernel(program,"simple_add"),queue,cl::NullRange,cl::NDRange(10),cl::NullRange);
simple_add(buffer_A,buffer_B,buffer_C);*/
// alternative way to run the kernel
kernel_add.setArg(0, buffer_A);
kernel_add.setArg(1, buffer_B);
kernel_add.setArg(2, buffer_C);
queue.enqueueNDRangeKernel(kernel_add, cl::NullRange, cl::NDRange(size),
cl::NullRange);
queue.finish();
int C[size];
// read result C from the device to array C
queue.enqueueReadBuffer(buffer_C, CL_TRUE, 0, sizeof(int) * size, C);
/*std::cout<<" result: \n";
for(int i=0;i<size;i++){
std::cout<<C[i]<<"\t";
}*/
auto stop1 = high_resolution_clock::now();
auto duration1 = duration_cast<microseconds>(stop1 - start1);
auto FPS = 1000000.0 / duration1.count();
cout << "Segmentation FPS=" << FPS << "\t"
<< "Execution Time(sec)=" << duration1.count() / 1000000.0 << endl;
}
int main() {
useOpenCL img;
while (true) {
img.backgroundSub();
}
return 0;
}
It is giving me below results:
Segmentation FPS=13.2557 Execution Time(sec)=0.075439
Segmentation FPS=15.7602 Execution Time(sec)=0.063451
Segmentation FPS=14.3872 Execution Time(sec)=0.069506
Segmentation FPS=12.7525 Execution Time(sec)=0.078416
Which is not good since fps is only 12, 13 fps. So how can i make this program faster?

Answer 1

Put the initialization part that you only need to call once in the beginning in the constructor. This initialization should contain ALL memory allocation, OpenCL C code compilation and any initial memory transfers from host to device:
useOpenCL::useOpenCL() {
Color::Modifier green(Color::FG_GREEN);
Color::Modifier red(Color::FG_RED);
Color::Modifier def(Color::FG_DEFAULT);
// get all platforms (drivers)
cl::Platform::get(&all_platforms);
if (all_platforms.size() == 0) {
std::cout << red << " No platforms found. Check OpenCL installation!" << def
<< endl;
exit(1);
}
default_platform = all_platforms[0];
std::cout << green << "Using platform: " << def
<< default_platform.getInfo<CL_PLATFORM_NAME>() << std::endl;
// get default device of the default platform
default_platform.getDevices(CL_DEVICE_TYPE_ALL, &all_devices);
if (all_devices.size() == 0) {
std::cout << red << " No devices found. Check OpenCL installation!" << def
<< endl;
exit(1);
}
default_device = all_devices[0];
std::cout << green << "Using device: " << def
<< default_device.getInfo<CL_DEVICE_NAME>() << std::endl;
// kernel calculates for each element C=A+B
kernel_code =
" void kernel simple_add(global const int* A, global const int* B, "
"global int* C){ "
" C[get_global_id(0)]=A[get_global_id(0)]+B[get_global_id(0)]; "
" "
" } "
" ";
sources.push_back({kernel_code.c_str(), kernel_code.length()});
context = cl::Context({default_device});
program = cl::Program(context, sources);
if (program.build({default_device}) != CL_SUCCESS) {
std::cout << " Error building: "
<< program.getBuildInfo<CL_PROGRAM_BUILD_LOG>(default_device)
<< "\n";
exit(1);
}
// create queue to which we will push commands for the device.
queue = cl::CommandQueue(context, default_device);
// create buffers on host
int A[size], B[size];
int C[size];
for (int i = 0; i < size; i++) {
A[i] = i;
B[i] = i + 1;
}
// create buffers on the device
buffer_A = cl::Buffer(context, CL_MEM_READ_WRITE, sizeof(int) * size);
buffer_B = cl::Buffer(context, CL_MEM_READ_WRITE, sizeof(int) * size);
buffer_C = cl::Buffer(context, CL_MEM_READ_WRITE, sizeof(int) * size);
// write arrays A and B to the device
queue.enqueueWriteBuffer(buffer_A, CL_TRUE, 0, sizeof(int) * size, A);
queue.enqueueWriteBuffer(buffer_B, CL_TRUE, 0, sizeof(int) * size, B);
// alternative way to run the kernel
kernel_add.setArg(0, buffer_A);
kernel_add.setArg(1, buffer_B);
kernel_add.setArg(2, buffer_C);
}
Therefore make context, program, queue, buffer_A, buffer_B, buffer_C member variables of your class useOpenCL. Especially the memory allocation and compilation take a long time, so do them only once and reuse the buffers.
class useOpenCL {
public:
int size = 294400;
std::vector<cl::Platform> all_platforms;
std::vector<cl::Device> all_devices;
cl::Platform default_platform;
cl::Device default_device;
cl::Program::Sources sources;
std::string kernel_code;
cl::Kernel kernel_add;
cl::Buffer buffer_A;
cl::Buffer buffer_B;
cl::Buffer buffer_C;
cl::Context context;
cl::Program program;
cl::CommandQueue queue;
useOpenCL();
~useOpenCL() {}
void backgroundSub();
};
Then only the kernel call and eventually memory transfers host<->device remain for every frame calculation:
void useOpenCL::backgroundSub() {
auto start1 = high_resolution_clock::now();
// write arrays A and B to the device (ONLY IF NECESSARY FOR EVERY FRAME)
//queue.enqueueWriteBuffer(buffer_A, CL_TRUE, 0, sizeof(int) * size, A);
//queue.enqueueWriteBuffer(buffer_B, CL_TRUE, 0, sizeof(int) * size, B);
// run the kernel
queue.enqueueNDRangeKernel(kernel_add, cl::NullRange, cl::NDRange(size),
cl::NullRange);
// read result C from the device to array C
queue.enqueueReadBuffer(buffer_C, CL_TRUE, 0, sizeof(int) * size, C);
queue.finish();
auto stop1 = high_resolution_clock::now();
auto duration1 = duration_cast<microseconds>(stop1 - start1);
auto FPS = 1000000.0 / duration1.count();
cout << "Segmentation FPS=" << FPS << "\t"
<< "Execution Time(sec)=" << duration1.count() / 1000000.0 << endl;
}
The latter code can be called over and over again and should be much faster than if you re-initialize everything over and over again. Also make sure that size is large enough, otherwise the GPU might not be utilized at its full potential and the latencies for host<->device memory transfers will make every frame disproportunately slower.