starpu-max/socl/src/cl_enqueuendrangekernel.c

/* StarPU --- Runtime system for heterogeneous multicore architectures.
 *
 * Copyright (C) 2011-2012                                Inria
 * Copyright (C) 2012,2014,2016-2017                      CNRS
 * Copyright (C) 2010-2011,2013,2016-2018                 Université de Bordeaux
 *
 * StarPU is free software; you can redistribute it and/or modify
 * it under the terms of the GNU Lesser General Public License as published by
 * the Free Software Foundation; either version 2.1 of the License, or (at
 * your option) any later version.
 *
 * StarPU is distributed in the hope that it will be useful, but
 * WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
 *
 * See the GNU Lesser General Public License in COPYING.LGPL for more details.
 */

#include "socl.h"
#include "event.h"

void soclEnqueueNDRangeKernel_task(void *descr[], void *args)
{
	command_ndrange_kernel cmd = (command_ndrange_kernel)args;

	cl_command_queue cq;
	int wid;
	cl_int err;

	cl_event ev = command_event_get(cmd);
	ev->prof_start = _socl_nanotime();
	gc_entity_release(ev);

	wid = starpu_worker_get_id_check();
	starpu_opencl_get_queue(wid, &cq);

	DEBUG_MSG("[worker %d] [kernel %d] Executing kernel...\n", wid, cmd->kernel->id);

	int range = starpu_worker_get_range();

	/* Set arguments */
	{
		unsigned int i;
		int buf = 0;
		for (i=0; i<cmd->num_args; i++)
		{
			switch (cmd->arg_types[i])
			{
			case Null:
				err = clSetKernelArg(cmd->kernel->cl_kernels[range], i, cmd->arg_sizes[i], NULL);
				break;
			case Buffer:
			{
				cl_mem mem;
				mem = (cl_mem)STARPU_VARIABLE_GET_PTR(descr[buf]);
				err = clSetKernelArg(cmd->kernel->cl_kernels[range], i, cmd->arg_sizes[i], &mem);
				buf++;
			}
			break;
			case Immediate:
				err = clSetKernelArg(cmd->kernel->cl_kernels[range], i, cmd->arg_sizes[i], cmd->args[i]);
				break;
			}
			if (err != CL_SUCCESS)
			{
				DEBUG_CL("clSetKernelArg", err);
				DEBUG_ERROR("Aborting\n");
			}
		}
	}

	/* Calling Kernel */
	cl_event event;
	err = clEnqueueNDRangeKernel(cq, cmd->kernel->cl_kernels[range], cmd->work_dim, cmd->global_work_offset, cmd->global_work_size, cmd->local_work_size, 0, NULL, &event);

	if (err != CL_SUCCESS)
	{
		ERROR_MSG("Worker[%d] Unable to Enqueue kernel (error %d)\n", wid, err);
		DEBUG_CL("clEnqueueNDRangeKernel", err);
		DEBUG_MSG("Workdim %u, global_work_offset %p, global_work_size %p, local_work_size %p\n",
			  cmd->work_dim, cmd->global_work_offset, cmd->global_work_size, cmd->local_work_size);
		DEBUG_MSG("Global work size: %ld %ld %ld\n", (long)cmd->global_work_size[0],
			  (long)(cmd->work_dim > 1 ? cmd->global_work_size[1] : 1), (long)(cmd->work_dim > 2 ? cmd->global_work_size[2] : 1));
		if (cmd->local_work_size != NULL)
			DEBUG_MSG("Local work size: %ld %ld %ld\n", (long)cmd->local_work_size[0],
				  (long)(cmd->work_dim > 1 ? cmd->local_work_size[1] : 1), (long)(cmd->work_dim > 2 ? cmd->local_work_size[2] : 1));
	}
	else
	{
		/* Waiting for kernel to terminate */
		clWaitForEvents(1, &event);
		clReleaseEvent(event);
	}
}

/**
 * Real kernel enqueuing command
 */
cl_int command_ndrange_kernel_submit(command_ndrange_kernel cmd)
{
	starpu_task task = task_create();
	task->cl = &cmd->codelet;
	task->cl->model = cmd->kernel->perfmodel;
	task->cl_arg = cmd;
	task->cl_arg_size = sizeof(cmd);

	/* Execute the task on a specific worker? */
	if (cmd->_command.event->cq->device != NULL)
	{
		task->execute_on_a_specific_worker = 1;
		task->workerid = cmd->_command.event->cq->device->worker_id;
	}

	struct starpu_codelet * codelet = task->cl;

	/* We need to detect which parameters are OpenCL's memory objects and
	 * we retrieve their corresponding StarPU buffers */
	cmd->num_buffers = 0;
	cmd->buffers = malloc(sizeof(cl_mem) * cmd->num_args);

	unsigned int i;
	for (i=0; i<cmd->num_args; i++)
	{
		if (cmd->arg_types[i] == Buffer)
		{
			cl_mem buf = *(cl_mem*)cmd->args[i];

			gc_entity_store(&cmd->buffers[cmd->num_buffers], buf);
			task->handles[cmd->num_buffers] = buf->handle;

			/* Determine best StarPU buffer access mode */
			int mode;
			if (buf->mode == CL_MEM_READ_ONLY)
				mode = STARPU_R;
			else if (buf->mode == CL_MEM_WRITE_ONLY)
			{
				mode = STARPU_W;
				buf->scratch = 0;
			}
			else if (buf->scratch)
			{ //RW but never accessed in RW or W mode
				mode = STARPU_W;
				buf->scratch = 0;
			}
			else
			{
				mode = STARPU_RW;
				buf->scratch = 0;
			}
			codelet->modes[cmd->num_buffers] = mode;

			cmd->num_buffers += 1;
		}
	}
	codelet->nbuffers = cmd->num_buffers;

	task_submit(task, cmd);

	return CL_SUCCESS;
}

CL_API_SUFFIX__VERSION_1_1
CL_API_ENTRY cl_int CL_API_CALL
soclEnqueueNDRangeKernel(cl_command_queue cq,
			 cl_kernel        kernel,
			 cl_uint          work_dim,
			 const size_t *   global_work_offset,
			 const size_t *   global_work_size,
			 const size_t *   local_work_size,
			 cl_uint          num_events,
			 const cl_event * events,
			 cl_event *       event)
{
	if (kernel->split_func != NULL && !STARPU_PTHREAD_MUTEX_TRYLOCK(&kernel->split_lock))
	{
		cl_event beforeEvent, afterEvent, totalEvent;

		totalEvent = event_create();
		gc_entity_store(&totalEvent->cq, cq);

		command_marker cmd = command_marker_create();
		beforeEvent = command_event_get(cmd);
		command_queue_enqueue(cq, cmd, num_events, events);

		cl_uint iter = 1;
		cl_uint split_min = CL_UINT_MAX;
		cl_uint split_min_iter = 1;
		while (iter < kernel->split_space && kernel->split_perfs[iter] != 0)
		{
			if (kernel->split_perfs[iter] < split_min)
			{
				split_min = kernel->split_perfs[iter];
				split_min_iter = iter;
			}
			iter++;
		}

		if (iter == kernel->split_space)
		{
			iter = split_min_iter;
		}

		cl_int ret = kernel->split_func(cq, iter, kernel->split_data, beforeEvent, &afterEvent);

		if (ret == CL_SUCCESS)
		{
			//FIXME: blocking call
			soclWaitForEvents(1, &afterEvent);

			/* Store perf */
			cl_ulong start,end;
			soclGetEventProfilingInfo(beforeEvent, CL_PROFILING_COMMAND_END, sizeof(cl_ulong), &start, NULL);
			soclGetEventProfilingInfo(afterEvent, CL_PROFILING_COMMAND_END, sizeof(cl_ulong), &end, NULL);
			soclReleaseEvent(afterEvent);

			kernel->split_perfs[iter] = end-start;

			STARPU_PTHREAD_MUTEX_UNLOCK(&kernel->split_lock);

			event_complete(totalEvent);

			totalEvent->prof_start = start;
			totalEvent->prof_submit = start;
			totalEvent->prof_queued = start;
			totalEvent->prof_end = end;

			RETURN_EVENT(totalEvent,event);
		}
		else
		{
			STARPU_PTHREAD_MUTEX_UNLOCK(&kernel->split_lock);
			soclReleaseEvent(totalEvent);
		}

		return ret;
	}
	else
	{
		command_ndrange_kernel cmd = command_ndrange_kernel_create(kernel, work_dim,
									   global_work_offset, global_work_size, local_work_size);

		cl_event ev = command_event_get(cmd);

		command_queue_enqueue(cq, cmd, num_events, events);

		RETURN_EVENT(ev, event);
	}

	return CL_SUCCESS;
}