static int opencl_abi_si_ndrange_finish_impl(X86Context *ctx)
{
driver_state.wait_for_ndrange_completion = 1;
if (!list_count(si_emu->running_work_groups) &&
!list_count(si_emu->waiting_work_groups))
{
opencl_debug("\tndrange is complete\n");
/* Reset driver state */
si_ndrange_free(driver_state.ndrange);
driver_state.ndrange = NULL;
driver_state.kernel = NULL;
driver_state.wait_for_ndrange_completion = 0;
driver_state.ndrange_complete = 0;
driver_state.ready_for_work = 0;
si_emu->ndrange = NULL;
}
else
{
opencl_debug("\tsuspending driver thread\n");
/* Suspend x86 context until simulation completes */
X86ContextSuspend(ctx, opencl_abi_si_ndrange_finish_can_wakeup,
NULL, opencl_abi_si_ndrange_finish_wakeup, NULL);
}
return 0;
}
/* Finalize GPU kernel */
void si_emu_done()
{
/* GPU report */
if (si_emu_report_file)
fclose(si_emu_report_file);
/* Free ND-Ranges */
while (si_emu->ndrange_list_count)
si_ndrange_free(si_emu->ndrange_list_head);
/* Free OpenCL objects */
si_opencl_repo_free_all_objects(si_emu->opencl_repo);
si_opencl_repo_free(si_emu->opencl_repo);
/* Finalize disassembler */
si_disasm_done();
/* Finalize ISA */
si_isa_done();
/* Finalize GPU kernel */
mem_free(si_emu->global_mem);
m2s_timer_free(si_emu->timer);
free(si_emu);
}
static void opengl_abi_si_ndrange_finish_wakeup(X86Context *ctx,
void *user_data)
{
assert(!user_data);
/* Reset driver state */
si_ndrange_free(driver_state.ndrange);
driver_state.ndrange = NULL;
driver_state.shader = NULL;
driver_state.wait_for_ndrange_completion = 0;
driver_state.ndrange_complete = 0;
driver_state.ready_for_work = 0;
si_emu->ndrange = NULL;
}
static void opencl_abi_si_ndrange_finish_wakeup(X86Context *ctx,
void *user_data)
{
assert(!user_data);
opencl_debug("waking up after finish");
/* Reset driver state */
si_ndrange_free(driver_state.ndrange);
driver_state.ndrange = NULL;
driver_state.kernel = NULL;
driver_state.wait_for_ndrange_completion = 0;
driver_state.ndrange_complete = 0;
driver_state.ready_for_work = 0;
si_emu->ndrange = NULL;
}
/* Run one iteration of the Southern Islands GPU emulation loop.
* Return FALSE if there is no more emulation to perform. */
int si_emu_run(void)
{
struct si_ndrange_t *ndrange;
struct si_ndrange_t *ndrange_next;
struct si_work_group_t *work_group;
struct si_work_group_t *work_group_next;
struct si_wavefront_t *wavefront;
struct si_wavefront_t *wavefront_next;
/* For efficiency when no Southern Islands emulation is selected, exit here
* if the list of existing ND-Ranges is empty. */
if (!si_emu->ndrange_list_count)
return 0;
/* Start any ND-Range in state 'pending' */
while ((ndrange = si_emu->pending_ndrange_list_head))
{
/* Set all ready work-groups to running */
while ((work_group = ndrange->pending_list_head))
{
si_work_group_clear_status(work_group, si_work_group_pending);
si_work_group_set_status(work_group, si_work_group_running);
}
/* Set is in state 'running' */
si_ndrange_clear_status(ndrange, si_ndrange_pending);
si_ndrange_set_status(ndrange, si_ndrange_running);
}
/* Run one instruction of each wavefront in each work-group of each
* ND-Range that is in status 'running'. */
for (ndrange = si_emu->running_ndrange_list_head; ndrange; ndrange = ndrange_next)
{
/* Save next ND-Range in state 'running'. This is done because the state
* might change during the execution of the ND-Range. */
ndrange_next = ndrange->running_ndrange_list_next;
/* Execute an instruction from each work-group */
for (work_group = ndrange->running_list_head; work_group; work_group = work_group_next)
{
/* Save next running work-group */
work_group_next = work_group->running_list_next;
/* Run an instruction from each wavefront */
for (wavefront = work_group->running_list_head; wavefront; wavefront = wavefront_next)
{
/* Save next running wavefront */
wavefront_next = wavefront->running_list_next;
/* Execute instruction in wavefront */
si_wavefront_execute(wavefront);
}
}
}
/* Free ND-Ranges that finished */
while ((ndrange = si_emu->finished_ndrange_list_head))
{
/* Dump ND-Range report */
si_ndrange_dump(ndrange, si_emu_report_file);
/* Stop if maximum number of kernels reached */
if (si_emu_max_kernels && si_emu->ndrange_count >= si_emu_max_kernels)
esim_finish = esim_finish_si_max_kernels;
/* Extract from list of finished ND-Ranges and free */
si_ndrange_free(ndrange);
}
/* Return TRUE */
return 1;
}
/* Run one iteration of the Southern Islands GPU timing simulation loop. */
int si_gpu_run(void)
{
struct si_ndrange_t *ndrange;
struct si_compute_unit_t *compute_unit;
struct si_compute_unit_t *compute_unit_next;
/* For efficiency when no Southern Islands emulation is selected, exit here
* if the list of existing ND-Ranges is empty. */
if (!si_emu->ndrange_list_count)
return 0;
/* Start one ND-Range in state 'pending' */
while ((ndrange = si_emu->pending_ndrange_list_head))
{
/* Currently not supported for more than 1 ND-Range */
if (si_gpu->ndrange)
fatal("%s: Southern Islands GPU timing simulation not supported "
"for multiple ND-Ranges", __FUNCTION__);
/* Set ND-Range status to 'running' */
si_ndrange_clear_status(ndrange, si_ndrange_pending);
si_ndrange_set_status(ndrange, si_ndrange_running);
/* Trace */
si_trace("si.new_ndrange id=%d wg_first=%d wg_count=%d\n", ndrange->id,
ndrange->work_group_id_first, ndrange->work_group_count);
/* Map ND-Range to GPU */
si_gpu_map_ndrange(ndrange);
si_calc_plot();
}
/* Mapped ND-Range */
ndrange = si_gpu->ndrange;
assert(ndrange);
/* Allocate work-groups to compute units */
while (si_gpu->compute_unit_ready_list_head && ndrange->pending_list_head)
si_compute_unit_map_work_group(si_gpu->compute_unit_ready_list_head,
ndrange->pending_list_head);
/* One more cycle */
si_gpu->cycle++;
/* Stop if maximum number of GPU cycles exceeded */
if (si_emu_max_cycles && si_gpu->cycle >= si_emu_max_cycles)
esim_finish = esim_finish_si_max_cycles;
/* Stop if maximum number of GPU instructions exceeded */
if (si_emu_max_inst && si_emu->inst_count >= si_emu_max_inst)
esim_finish = esim_finish_si_max_inst;
/* Stop if any reason met */
if (esim_finish)
return 1;
/* Free instructions in trash */
si_gpu_uop_trash_empty();
/* Run one loop iteration on each busy compute unit */
for (compute_unit = si_gpu->compute_unit_busy_list_head; compute_unit;
compute_unit = compute_unit_next)
{
/* Store next busy compute unit, since this can change
* during the compute unit simulation loop iteration. */
compute_unit_next = compute_unit->compute_unit_busy_list_next;
/* Run one cycle */
si_compute_unit_run(compute_unit);
}
/* If ND-Range finished execution in all compute units, free it. */
if (!si_gpu->compute_unit_busy_list_count)
{
/* Dump ND-Range report */
si_ndrange_dump(ndrange, si_emu_report_file);
/* Stop if maximum number of kernels reached */
if (si_emu_max_kernels && si_emu->ndrange_count >= si_emu_max_kernels)
esim_finish = esim_finish_si_max_kernels;
/* Finalize and free ND-Range */
assert(si_ndrange_get_status(ndrange, si_ndrange_finished));
si_gpu_uop_trash_empty();
si_gpu_unmap_ndrange();
si_ndrange_free(ndrange);
}
/* Return true */
return 1;
}
