#include "LinearTests.h"
#include <Common.h>

Large::Large(cl::Context* _context, cl::CommandQueue* _queue, cl::Program* _program)
{
	context = _context;
	queue = _queue;
	program = _program;
	name = "Large";
}

cl_ulong Large::dewIt(int n, int m, float* y, const float* A, const float* x, const float* b)
{
	cl_int err = CL_SUCCESS;
	cl::Event _event;

	cl::Kernel kernel = cl::Kernel(*program, "largeMV", &err);
	CheckCLError(err);

	cl::Buffer ABuffer(*context, CL_MEM_READ_ONLY, sizeof(float) * n * m, NULL, &err);
	CheckCLError(err);
	queue->enqueueWriteBuffer(ABuffer, true, 0, sizeof(float) * n * m, A);

	cl::Buffer XBuffer(*context, CL_MEM_READ_ONLY, sizeof(float) * m, NULL, &err);
	CheckCLError(err);
	queue->enqueueWriteBuffer(XBuffer, true, 0, sizeof(float) * m, x);

	cl::Buffer YBuffer(*context, CL_MEM_WRITE_ONLY, sizeof(float) * n, NULL, &err);
	CheckCLError(err);
	queue->enqueueWriteBuffer(YBuffer, true, 0, sizeof(float) * n, y);

	cl::Buffer BBuffer(*context, CL_MEM_READ_ONLY, sizeof(float) * n, NULL, &err);
	CheckCLError(err);
	queue->enqueueWriteBuffer(BBuffer, true, 0, sizeof(float) * n, b);

	//void largeMV(const int n, const int m, __global float* y, __global float* A,
	//             __global float* x, __global float* b, const int T, const int Z, __local float* Q)
	int T = 2; // kimenet T hosszu darabokra
	int Z = 2; // Bemenet Z hosszu darabokra

	kernel.setArg(0, n);
	kernel.setArg(1, m);
	kernel.setArg(2, YBuffer);
	kernel.setArg(3, ABuffer);
	kernel.setArg(4, XBuffer);
	kernel.setArg(5, BBuffer);
	kernel.setArg(6, T);
	kernel.setArg(7, Z);
	kernel.setArg(8, sizeof(float) * T * Z, NULL);

	queue->enqueueNDRangeKernel(kernel, cl::NullRange, cl::NDRange(Z*T, 1), cl::NDRange(T * Z, 1), NULL, &_event);
	_event.wait();
	cl_ulong execStart, execEnd;
	execStart = _event.getProfilingInfo<CL_PROFILING_COMMAND_START>(&err);
	execEnd = _event.getProfilingInfo<CL_PROFILING_COMMAND_END>(&err);

	queue->enqueueReadBuffer(YBuffer, true, 0, sizeof(float) * n, y);
	return (execEnd - execStart);
}