/* Tom Gall (tom.gall@linaro.org / tom_gall@mac.com)
 * Copyright 2014 Linaro 
 * Released under the terms and conditions of the license documented in the LICENSE file
 */

int opencl_select(opencl_sqlite_context *s, int work_units_per_kernel) {
//	cl_event prof_event;
//	cl_ulong time_start, time_end, total_time=0;
	int stride; 
	int totalRows, err;
	totalRows = s->data_cpu->rows;
	stride = s->data_cpu->stride/sizeof(int);

//	printf("%d total Rows %d stride\n", totalRows, stride);

	err=clSetKernelArg(s->kernel, 0, sizeof(int), &totalRows);
	err|=clSetKernelArg(s->kernel, 1, sizeof(int), &(stride));
	err|=clSetKernelArg(s->kernel, 2, sizeof(cl_mem), &(s->data_gpu->d));
	err|=clSetKernelArg(s->kernel, 3, sizeof(cl_mem), &(s->results_gpu->r));
	if (s->results_gpu->roffsetResults_cl)
		err|=clSetKernelArg(s->kernel, 4, sizeof(cl_mem), &(s->results_gpu->roffsetResults_cl));
	if(err < 0) {
		perror("Couldn't set a kernel arg");
		exit(1);   
	}

	err = clEnqueueNDRangeKernel(s->queue, s->kernel, 
		1,  // 1 dimention of data
		NULL, &work_units_per_kernel, 
		NULL, 0, NULL, NULL);
	if(err < 0) {
		perror("Couldn't enqueue the kernel execution command");
		exit(1);   
	}

	clFinish(s->queue);
/*    clGetEventProfilingInfo(prof_event, CL_PROFILING_COMMAND_START,
            sizeof(time_start), &time_start, NULL);
    clGetEventProfilingInfo(prof_event, CL_PROFILING_COMMAND_END,
            sizeof(time_end), &time_end, NULL);
    total_time = time_end - time_start;

	printf("dispatch time: %llu microseconds\n", total_time/1000);
*/
	return OPENCL_SUCCESS;;
}

int opencl_rowcolumn_select(opencl_sqlite_context *s, int work_units_per_kernel) {
//	cl_event prof_event;
//	cl_ulong time_start, time_end, total_time=0;
//	int stride; 
	int totalRows, err;
	totalRows = s->data_cpu->rows;
	//stride = s->data_cpu->stride/sizeof(int);

//	printf("%d total Rows %d stride\n", totalRows, stride);

	err=clSetKernelArg(s->kernel, 0, sizeof(int), &totalRows);
	err|=clSetKernelArg(s->kernel, 1, sizeof(cl_mem), &(s->data_gpu->d));
	err|=clSetKernelArg(s->kernel, 2, sizeof(cl_mem), &(s->results_gpu->r));
	if (s->results_gpu->roffsetResults_cl)
		err|=clSetKernelArg(s->kernel, 3, sizeof(cl_mem), &(s->results_gpu->roffsetResults_cl));
	if(err < 0) {
		perror("Couldn't set a kernel arg");
		exit(1);   
	}

	err = clEnqueueNDRangeKernel(s->queue, s->kernel, 
		1,  // 1 dimention of data
		NULL, &work_units_per_kernel, 
		NULL, 0, NULL, NULL);
	if(err < 0) {
		perror("Couldn't enqueue the kernel execution command");
		exit(1);   
	}

	clFinish(s->queue);
/*    clGetEventProfilingInfo(prof_event, CL_PROFILING_COMMAND_START,
            sizeof(time_start), &time_start, NULL);
    clGetEventProfilingInfo(prof_event, CL_PROFILING_COMMAND_END,
            sizeof(time_end), &time_end, NULL);
    total_time = time_end - time_start;

	printf("dispatch time: %llu microseconds\n", total_time/1000);
*/
	return OPENCL_SUCCESS;
}

int opencl_vector_select(opencl_sqlite_context *s, opencl_data_shard *col1_shard, opencl_data_shard *col2_shard,
		int work_units_per_kernel) {
//	cl_event prof_event;
//	cl_ulong time_start, time_end, total_time=0;
	int totalRows, err;
	totalRows = s->data_cpu->rows;
//	totalRows = 12;
//	work_units_per_kernel = 1;
//	s->data_cpu->rows = 12;

	err=clSetKernelArg(s->kernel, 0, sizeof(int), &totalRows);
	err|=clSetKernelArg(s->kernel, 1, sizeof(cl_mem), &(col1_shard->d));
	err|=clSetKernelArg(s->kernel, 2, sizeof(cl_mem), &(col2_shard->d));
	err|=clSetKernelArg(s->kernel, 3, sizeof(cl_mem), &(s->results_gpu->r));
	if(err < 0) {
		perror("Couldn't set a kernel arg");
		exit(1);   
	}

	err = clEnqueueNDRangeKernel(s->queue, s->kernel, 
		1,  // 1 dimention of data
		NULL,  &work_units_per_kernel, 
		NULL, 0, NULL, NULL);
	if(err < 0) {
		perror("Couldn't enqueue the kernel execution command");
		exit(1);   
	}

	clFinish(s->queue);
/*    clGetEventProfilingInfo(prof_event, CL_PROFILING_COMMAND_START,
            sizeof(time_start), &time_start, NULL);
    clGetEventProfilingInfo(prof_event, CL_PROFILING_COMMAND_END,
            sizeof(time_end), &time_end, NULL);
    total_time = time_end - time_start;

	printf("dispatch time: %llu microseconds\n", total_time/1000);
*/
	return OPENCL_SUCCESS;
}