void evg_compute_unit_map_work_group(struct evg_compute_unit_t *compute_unit, struct evg_work_group_t *work_group)
{
struct evg_ndrange_t *ndrange = work_group->ndrange;
struct evg_wavefront_t *wavefront;
int wavefront_id;
/* Map work-group */
assert(compute_unit->work_group_count < evg_gpu->work_groups_per_compute_unit);
assert(!work_group->id_in_compute_unit);
while (work_group->id_in_compute_unit < evg_gpu->work_groups_per_compute_unit
&& compute_unit->work_groups[work_group->id_in_compute_unit])
work_group->id_in_compute_unit++;
assert(work_group->id_in_compute_unit < evg_gpu->work_groups_per_compute_unit);
compute_unit->work_groups[work_group->id_in_compute_unit] = work_group;
compute_unit->work_group_count++;
/* If compute unit reached its maximum load, remove it from 'ready' list.
* Otherwise, move it to the end of the 'ready' list. */
assert(DOUBLE_LINKED_LIST_MEMBER(evg_gpu, ready, compute_unit));
DOUBLE_LINKED_LIST_REMOVE(evg_gpu, ready, compute_unit);
if (compute_unit->work_group_count < evg_gpu->work_groups_per_compute_unit)
DOUBLE_LINKED_LIST_INSERT_TAIL(evg_gpu, ready, compute_unit);
/* If this is the first scheduled work-group, insert to 'busy' list. */
if (!DOUBLE_LINKED_LIST_MEMBER(evg_gpu, busy, compute_unit))
DOUBLE_LINKED_LIST_INSERT_TAIL(evg_gpu, busy, compute_unit);
/* Assign wavefronts identifiers in compute unit */
EVG_FOREACH_WAVEFRONT_IN_WORK_GROUP(work_group, wavefront_id)
{
wavefront = ndrange->wavefronts[wavefront_id];
wavefront->id_in_compute_unit = work_group->id_in_compute_unit *
ndrange->wavefronts_per_work_group + wavefront->id_in_work_group;
}
void si_compute_unit_map_work_group(struct si_compute_unit_t *compute_unit,
struct si_work_group_t *work_group)
{
struct si_ndrange_t *ndrange = work_group->ndrange;
struct si_wavefront_t *wavefront;
int wavefront_id;
int ib_id;
assert(compute_unit->work_group_count <
si_gpu->work_groups_per_compute_unit);
assert(!work_group->id_in_compute_unit);
/* Find an available slot */
while (work_group->id_in_compute_unit <
si_gpu->work_groups_per_compute_unit &&
compute_unit->work_groups[work_group->id_in_compute_unit])
{
work_group->id_in_compute_unit++;
}
assert(work_group->id_in_compute_unit <
si_gpu->work_groups_per_compute_unit);
compute_unit->work_groups[work_group->id_in_compute_unit] = work_group;
compute_unit->work_group_count++;
/* If compute unit reached its maximum load, remove it from
* 'compute_unit_ready' list. Otherwise, move it to the end of
* the 'compute_unit_ready' list. */
assert(DOUBLE_LINKED_LIST_MEMBER(si_gpu, compute_unit_ready,
compute_unit));
DOUBLE_LINKED_LIST_REMOVE(si_gpu, compute_unit_ready, compute_unit);
if (compute_unit->work_group_count <
si_gpu->work_groups_per_compute_unit)
{
DOUBLE_LINKED_LIST_INSERT_TAIL(si_gpu, compute_unit_ready,
compute_unit);
}
/* If this is the first scheduled work-group, insert to
* 'compute_unit_busy' list. */
if (!DOUBLE_LINKED_LIST_MEMBER(si_gpu, compute_unit_busy, compute_unit))
{
DOUBLE_LINKED_LIST_INSERT_TAIL(si_gpu, compute_unit_busy,
compute_unit);
}
/* Assign wavefronts identifiers in compute unit */
SI_FOREACH_WAVEFRONT_IN_WORK_GROUP(work_group, wavefront_id)
{
wavefront = ndrange->wavefronts[wavefront_id];
wavefront->id_in_compute_unit = work_group->id_in_compute_unit *
ndrange->wavefronts_per_work_group +
wavefront->id_in_work_group;
}
static struct gpu_wavefront_t *gpu_schedule_round_robin(struct gpu_compute_unit_t *compute_unit)
{
struct gpu_wavefront_t *wavefront, *temp_wavefront;
struct linked_list_t *wavefront_pool = compute_unit->wavefront_pool;
/* Select current position in pool as initial candidate wavefront */
if (!linked_list_get(wavefront_pool))
linked_list_head(wavefront_pool);
wavefront = linked_list_get(wavefront_pool);
temp_wavefront = wavefront;
/* Look for a valid candidate */
for (;;)
{
/* Wavefront must be running,
* and the corresponding slot in fetch buffer must be free. */
assert(wavefront->id_in_compute_unit < gpu->wavefronts_per_compute_unit);
if (DOUBLE_LINKED_LIST_MEMBER(wavefront->work_group, running, wavefront) &&
!compute_unit->cf_engine.fetch_buffer[wavefront->id_in_compute_unit])
break;
/* Current candidate is not valid - go to next.
* If we went through the whole pool, no fetch. */
linked_list_next_circular(wavefront_pool);
wavefront = linked_list_get(wavefront_pool);
if (wavefront == temp_wavefront)
return NULL;
}
/* Wavefront found, remove from pool and return. */
assert(wavefront->clause_kind == GPU_CLAUSE_CF);
linked_list_remove(wavefront_pool);
return wavefront;
}
/* Enqueue access in port wait list. */
void mod_stack_wait_in_port(struct mod_stack_t *stack,
struct mod_port_t *port, int event)
{
assert(port == stack->port);
assert(!DOUBLE_LINKED_LIST_MEMBER(port, waiting, stack));
stack->waiting_list_event = event;
DOUBLE_LINKED_LIST_INSERT_TAIL(port, waiting, stack);
}
/* Enqueue access in module wait list. */
void mod_stack_wait_in_mod(struct mod_stack_t *stack,
struct mod_t *mod, int event)
{
assert(mod == stack->mod);
assert(!DOUBLE_LINKED_LIST_MEMBER(mod, waiting, stack));
stack->waiting_list_event = event;
DOUBLE_LINKED_LIST_INSERT_TAIL(mod, waiting, stack);
}
static void evg_cf_engine_fetch(struct evg_compute_unit_t *compute_unit)
{
struct evg_ndrange_t *ndrange = evg_gpu->ndrange;
struct evg_wavefront_t *wavefront;
char str[MAX_LONG_STRING_SIZE];
char str_trimmed[MAX_LONG_STRING_SIZE];
struct evg_inst_t *inst;
struct evg_uop_t *uop;
struct evg_work_item_uop_t *work_item_uop;
struct evg_work_item_t *work_item;
int work_item_id;
/* Schedule wavefront */
wavefront = evg_compute_unit_schedule(compute_unit);
if (!wavefront)
return;
/* Emulate CF instruction */
evg_wavefront_execute(wavefront);
inst = &wavefront->cf_inst;
/* Create uop */
uop = evg_uop_create();
uop->wavefront = wavefront;
uop->work_group = wavefront->work_group;
uop->compute_unit = compute_unit;
uop->id_in_compute_unit = compute_unit->gpu_uop_id_counter++;
uop->alu_clause_trigger = wavefront->clause_kind == EVG_CLAUSE_ALU;
uop->tex_clause_trigger = wavefront->clause_kind == EVG_CLAUSE_TEX;
uop->no_clause_trigger = wavefront->clause_kind == EVG_CLAUSE_CF;
uop->last = DOUBLE_LINKED_LIST_MEMBER(wavefront->work_group, finished, wavefront);
uop->wavefront_last = uop->last && uop->no_clause_trigger;
uop->global_mem_read = wavefront->global_mem_read;
uop->global_mem_write = wavefront->global_mem_write;
uop->active_mask_update = wavefront->active_mask_update;
uop->active_mask_push = wavefront->active_mask_push;
uop->active_mask_pop = wavefront->active_mask_pop;
uop->active_mask_stack_top = wavefront->stack_top;
uop->vliw_slots = 1;
/* If debugging active mask, store active state for work-items */
if (debug_status(evg_stack_debug_category))
evg_uop_save_active_mask(uop);
/* If instruction is a global memory write, record addresses */
if (uop->global_mem_write)
{
assert((inst->info->flags & EVG_INST_FLAG_MEM_WRITE));
EVG_FOREACH_WORK_ITEM_IN_WAVEFRONT(wavefront, work_item_id)
{
work_item = ndrange->work_items[work_item_id];
work_item_uop = &uop->work_item_uop[work_item->id_in_wavefront];
work_item_uop->global_mem_access_addr = work_item->global_mem_access_addr;
work_item_uop->global_mem_access_size = work_item->global_mem_access_size;
}
/* Enqueue access in stack wait list. */
void mod_stack_wait_in_stack(struct mod_stack_t *stack,
struct mod_stack_t *master_stack, int event)
{
assert(master_stack != stack);
assert(!DOUBLE_LINKED_LIST_MEMBER(master_stack, waiting, stack));
stack->waiting_list_event = event;
DOUBLE_LINKED_LIST_INSERT_TAIL(master_stack, waiting, stack);
}
void si_ndrange_free(struct si_ndrange_t *ndrange)
{
/* Set event status to complete if an event was set. */
if(ndrange->event)
ndrange->event->status = SI_OPENCL_EVENT_STATUS_COMPLETE;
int i;
/* Clear task from command queue */
if (ndrange->command_queue && ndrange->command)
{
si_opencl_command_queue_complete(ndrange->command_queue, ndrange->command);
si_opencl_command_free(ndrange->command);
}
/* Clear all states that affect lists. */
si_ndrange_clear_status(ndrange, si_ndrange_pending);
si_ndrange_clear_status(ndrange, si_ndrange_running);
si_ndrange_clear_status(ndrange, si_ndrange_finished);
/* Extract from ND-Range list in Southern Islands emulator */
assert(DOUBLE_LINKED_LIST_MEMBER(si_emu, ndrange, ndrange));
DOUBLE_LINKED_LIST_REMOVE(si_emu, ndrange, ndrange);
/* Free lists */
list_free(ndrange->uav_list);
/* Free work-groups */
for (i = 0; i < ndrange->work_group_count; i++)
si_work_group_free(ndrange->work_groups[i]);
free(ndrange->work_groups);
/* Free wavefronts */
for (i = 0; i < ndrange->wavefront_count; i++)
{
si_wavefront_free(ndrange->wavefronts[i]);
si_work_item_free(ndrange->scalar_work_items[i]);
}
free(ndrange->wavefronts);
free(ndrange->scalar_work_items);
/* Free work-items */
for (i = 0; i < ndrange->work_item_count; i++)
si_work_item_free(ndrange->work_items[i]);
free(ndrange->work_items);
/* Free instruction histogram */
if (ndrange->inst_histogram)
free(ndrange->inst_histogram);
/* Free ndrange */
free(ndrange->name);
free(ndrange);
}
/* Enqueue access in stack wait list. */
void mod_stack_wait_in_stack(struct mod_stack_t *stack,
struct mod_stack_t *master_stack, int event)
{
assert(master_stack != stack);
assert(!DOUBLE_LINKED_LIST_MEMBER(master_stack, waiting, stack));
stack->waiting_list_event = event;
DOUBLE_LINKED_LIST_INSERT_TAIL(master_stack, waiting, stack);
if((stack->addr >= 0x2F20 && stack->addr<= 0x2F2C))
;// fprintf(stderr, " wait master %x, %x, %d,%lld\n", master_stack->addr, stack->addr , event,esim_time);
}
void X86ContextDestroy(X86Context *self)
{
X86Emu *emu = self->emu;
/* If context is not finished/zombie, finish it first.
* This removes all references to current freed context. */
if (!X86ContextGetState(self, X86ContextFinished | X86ContextZombie))
X86ContextFinish(self, 0);
/* Remove context from finished contexts list. This should
* be the only list the context is in right now. */
assert(!DOUBLE_LINKED_LIST_MEMBER(emu, running, self));
assert(!DOUBLE_LINKED_LIST_MEMBER(emu, suspended, self));
assert(!DOUBLE_LINKED_LIST_MEMBER(emu, zombie, self));
assert(DOUBLE_LINKED_LIST_MEMBER(emu, finished, self));
DOUBLE_LINKED_LIST_REMOVE(emu, finished, self);
/* Free private structures */
x86_regs_free(self->regs);
x86_regs_free(self->backup_regs);
x86_signal_mask_table_free(self->signal_mask_table);
spec_mem_free(self->spec_mem);
bit_map_free(self->affinity);
/* Unlink shared structures */
x86_loader_unlink(self->loader);
x86_signal_handler_table_unlink(self->signal_handler_table);
x86_file_desc_table_unlink(self->file_desc_table);
mem_unlink(self->mem);
/* Remove context from contexts list and free */
DOUBLE_LINKED_LIST_REMOVE(emu, context, self);
X86ContextDebug("inst %lld: context %d freed\n",
asEmu(emu)->instructions, self->pid);
/* Static instruction */
delete_static(&self->inst);
}
void si_ndrange_free(struct si_ndrange_t *ndrange)
{
int i;
/* Run free notify call-back */
if (ndrange->free_notify_func)
ndrange->free_notify_func(ndrange->free_notify_data);
/* Clear all states that affect lists. */
si_ndrange_clear_status(ndrange, si_ndrange_pending);
si_ndrange_clear_status(ndrange, si_ndrange_running);
si_ndrange_clear_status(ndrange, si_ndrange_finished);
/* Extract from ND-Range list in Southern Islands emulator */
assert(DOUBLE_LINKED_LIST_MEMBER(si_emu, ndrange, ndrange));
DOUBLE_LINKED_LIST_REMOVE(si_emu, ndrange, ndrange);
/* Free work-groups */
for (i = 0; i < ndrange->work_group_count; i++)
si_work_group_free(ndrange->work_groups[i]);
free(ndrange->work_groups);
/* Free wavefronts */
for (i = 0; i < ndrange->wavefront_count; i++)
{
si_wavefront_free(ndrange->wavefronts[i]);
si_work_item_free(ndrange->scalar_work_items[i]);
}
free(ndrange->wavefronts);
free(ndrange->scalar_work_items);
/* Free work-items */
for (i = 0; i < ndrange->work_item_count; i++)
si_work_item_free(ndrange->work_items[i]);
free(ndrange->work_items);
/* Free instruction histogram */
if (ndrange->inst_histogram)
free(ndrange->inst_histogram);
/* Free instruction buffer */
if (ndrange->inst_buffer)
free(ndrange->inst_buffer);
/* Free ndrange */
free(ndrange->name);
free(ndrange);
}
int x86_emu_list_member(enum x86_emu_list_kind_t list, struct x86_ctx_t *ctx)
{
switch (list) {
case x86_emu_list_context: return DOUBLE_LINKED_LIST_MEMBER(x86_emu, context, ctx);
case x86_emu_list_running: return DOUBLE_LINKED_LIST_MEMBER(x86_emu, running, ctx);
case x86_emu_list_finished: return DOUBLE_LINKED_LIST_MEMBER(x86_emu, finished, ctx);
case x86_emu_list_zombie: return DOUBLE_LINKED_LIST_MEMBER(x86_emu, zombie, ctx);
case x86_emu_list_suspended: return DOUBLE_LINKED_LIST_MEMBER(x86_emu, suspended, ctx);
case x86_emu_list_alloc: return DOUBLE_LINKED_LIST_MEMBER(x86_emu, alloc, ctx);
}
return 0;
}
int MIPSEmuListMember(MIPSEmu *self, enum mips_emu_list_kind_t list,
struct mips_ctx_t *ctx)
{
switch (list) {
case mips_emu_list_context: return DOUBLE_LINKED_LIST_MEMBER(self, context, ctx);
case mips_emu_list_running: return DOUBLE_LINKED_LIST_MEMBER(self, running, ctx);
case mips_emu_list_finished: return DOUBLE_LINKED_LIST_MEMBER(self, finished, ctx);
case mips_emu_list_zombie: return DOUBLE_LINKED_LIST_MEMBER(self, zombie, ctx);
case mips_emu_list_suspended: return DOUBLE_LINKED_LIST_MEMBER(self, suspended, ctx);
case mips_emu_list_alloc: return DOUBLE_LINKED_LIST_MEMBER(self, alloc, ctx);
}
return 0;
}
void evg_ndrange_free(struct evg_ndrange_t *ndrange)
{
int i;
/* Clear task from command queue */
if (ndrange->command_queue && ndrange->command)
{
evg_opencl_command_queue_complete(ndrange->command_queue, ndrange->command);
evg_opencl_command_free(ndrange->command);
}
/* Clear all states that affect lists. */
evg_ndrange_clear_status(ndrange, evg_ndrange_pending);
evg_ndrange_clear_status(ndrange, evg_ndrange_running);
evg_ndrange_clear_status(ndrange, evg_ndrange_finished);
/* Extract from ND-Range list in Evergreen emulator */
assert(DOUBLE_LINKED_LIST_MEMBER(evg_emu, ndrange, ndrange));
DOUBLE_LINKED_LIST_REMOVE(evg_emu, ndrange, ndrange);
/* Free work-groups */
for (i = 0; i < ndrange->work_group_count; i++)
evg_work_group_free(ndrange->work_groups[i]);
free(ndrange->work_groups);
/* Free wavefronts */
for (i = 0; i < ndrange->wavefront_count; i++)
evg_wavefront_free(ndrange->wavefronts[i]);
free(ndrange->wavefronts);
/* Free work-items */
for (i = 0; i < ndrange->work_item_count; i++)
evg_work_item_free(ndrange->work_items[i]);
free(ndrange->work_items);
/* Free instruction histogram */
if (ndrange->inst_histogram)
free(ndrange->inst_histogram);
/* Free ND-Range */
free(ndrange->name);
free(ndrange);
}
static struct gpu_wavefront_t *gpu_schedule_greedy(struct gpu_compute_unit_t *compute_unit)
{
struct gpu_wavefront_t *wavefront, *temp_wavefront;
struct linked_list_t *wavefront_pool = compute_unit->wavefront_pool;
/* Check all candidates */
temp_wavefront = NULL;
LINKED_LIST_FOR_EACH(wavefront_pool)
{
/* Get wavefront from list */
wavefront = linked_list_get(wavefront_pool);
/* Wavefront must be running,
* and the corresponding slot in fetch buffer must be free. */
assert(wavefront->id_in_compute_unit < gpu->wavefronts_per_compute_unit);
if (!DOUBLE_LINKED_LIST_MEMBER(wavefront->work_group, running, wavefront) ||
compute_unit->cf_engine.fetch_buffer[wavefront->id_in_compute_unit])
continue;
/* Select current wavefront temporarily */
if (!temp_wavefront || temp_wavefront->sched_when < wavefront->sched_when)
temp_wavefront = wavefront;
}
/* No wavefront found */
wavefront = NULL;
if (!temp_wavefront)
return NULL;
/* Wavefront found, remove from pool and return. */
assert(temp_wavefront->clause_kind == GPU_CLAUSE_CF);
linked_list_find(wavefront_pool, temp_wavefront);
assert(!wavefront_pool->error_code);
linked_list_remove(wavefront_pool);
temp_wavefront->sched_when = gpu->cycle;
return temp_wavefront;
}
/* Execute one instruction in the wavefront */
void si_wavefront_execute(struct si_wavefront_t *wavefront)
{
struct si_ndrange_t *ndrange;
struct si_work_group_t *work_group;
struct si_work_item_t *work_item;
struct si_inst_t *inst;
char inst_dump[MAX_INST_STR_SIZE];
unsigned int pc;
ndrange = wavefront->ndrange;
int work_item_id;
/* Get current work-group */
ndrange = wavefront->ndrange;
work_group = wavefront->work_group;
work_item = NULL;
inst = NULL;
assert(!DOUBLE_LINKED_LIST_MEMBER(work_group, finished, wavefront));
/* Reset instruction flags */
wavefront->vector_mem_write = 0;
wavefront->vector_mem_read = 0;
wavefront->scalar_mem_read = 0;
wavefront->local_mem_write = 0;
wavefront->local_mem_read = 0;
wavefront->pred_mask_update = 0;
wavefront->mem_wait = 0;
wavefront->barrier = 0;
assert(!wavefront->finished);
/* Grab the next instruction and update the pointer */
wavefront->inst_size = si_inst_decode(wavefront->wavefront_pool, &wavefront->inst, 0);
/* Stats */
si_emu->inst_count++;
wavefront->emu_inst_count++;
wavefront->inst_count++;
/* Set the current instruction */
inst = &wavefront->inst;
pc = wavefront->wavefront_pool - wavefront->wavefront_pool_start;
/*MIAOW start - Print the debug message to stdout, stderr to the open file stream*/
si_isa_debug("\n###%d_%d_%d", kernel_config_count - 1, wavefront->work_group->id, wavefront->id_in_work_group, pc, wavefront->inst_size);
/*MIAOW stop*/
/* Execute the current instruction */
switch (inst->info->fmt)
{
/* Scalar ALU Instructions */
case SI_FMT_SOP1:
{
/* Dump instruction string when debugging */
if (debug_status(si_isa_debug_category))
{
si_inst_dump_sop1(inst, wavefront->inst_size, pc, wavefront->wavefront_pool, inst_dump,
MAX_INST_STR_SIZE);
si_isa_debug("\n%s", inst_dump);
}
/* Stats */
si_emu->scalar_alu_inst_count++;
wavefront->scalar_alu_inst_count++;
/* Only one work item executes the instruction */
work_item = wavefront->scalar_work_item;
(*si_isa_inst_func[inst->info->inst])(work_item, inst);
if (debug_status(si_isa_debug_category))
{
si_isa_debug("\n");
}
break;
}
case SI_FMT_SOP2:
{
/* Dump instruction string when debugging */
if (debug_status(si_isa_debug_category))
{
si_inst_dump_sop2(inst, wavefront->inst_size, pc, wavefront->wavefront_pool, inst_dump,
MAX_INST_STR_SIZE);
si_isa_debug("\n%s", inst_dump);
}
/* Stats */
si_emu->scalar_alu_inst_count++;
wavefront->scalar_alu_inst_count++;
/* Only one work item executes the instruction */
work_item = wavefront->scalar_work_item;
(*si_isa_inst_func[inst->info->inst])(work_item, inst); //Calling a function pointer in machine.c
if (debug_status(si_isa_debug_category))
{
si_isa_debug("\n");
}
break;
}
case SI_FMT_SOPP:
{
/* Dump instruction string when debugging */
if (debug_status(si_isa_debug_category))
{
si_inst_dump_sopp(inst, wavefront->inst_size, pc, wavefront->wavefront_pool, inst_dump,
MAX_INST_STR_SIZE);
si_isa_debug("\n%s", inst_dump);
}
/* Stats */
if (wavefront->inst.micro_inst.sopp.op > 1 &&
wavefront->inst.micro_inst.sopp.op < 10)
{
si_emu->branch_inst_count++;
wavefront->branch_inst_count++;
} else
{
si_emu->scalar_alu_inst_count++;
wavefront->scalar_alu_inst_count++;
}
/* Only one work item executes the instruction */
work_item = wavefront->scalar_work_item;
(*si_isa_inst_func[inst->info->inst])(work_item, inst);
if (debug_status(si_isa_debug_category))
{
si_isa_debug("\n");
}
break;
}
case SI_FMT_SOPC:
{
/* Dump instruction string when debugging */
if (debug_status(si_isa_debug_category))
{
si_inst_dump_sopc(inst, wavefront->inst_size, pc, wavefront->wavefront_pool, inst_dump,
MAX_INST_STR_SIZE);
si_isa_debug("\n%s", inst_dump);
}
/* Stats */
si_emu->scalar_alu_inst_count++;
wavefront->scalar_alu_inst_count++;
/* Only one work item executes the instruction */
work_item = wavefront->scalar_work_item;
(*si_isa_inst_func[inst->info->inst])(work_item, inst);
if (debug_status(si_isa_debug_category))
{
si_isa_debug("\n");
}
break;
}
case SI_FMT_SOPK:
{
/* Dump instruction string when debugging */
if (debug_status(si_isa_debug_category))
{
si_inst_dump_sopk(inst, wavefront->inst_size, pc, wavefront->wavefront_pool, inst_dump,
MAX_INST_STR_SIZE);
si_isa_debug("\n%s", inst_dump);
}
/* Stats */
si_emu->scalar_alu_inst_count++;
wavefront->scalar_alu_inst_count++;
/* Only one work item executes the instruction */
work_item = wavefront->scalar_work_item;
(*si_isa_inst_func[inst->info->inst])(work_item, inst);
if (debug_status(si_isa_debug_category))
{
si_isa_debug("\n");
}
break;
}
/* Scalar Memory Instructions */
case SI_FMT_SMRD:
{
/* Dump instruction string when debugging */
if (debug_status(si_isa_debug_category))
{
si_inst_dump_smrd(inst, wavefront->inst_size, pc, wavefront->wavefront_pool, inst_dump,
MAX_INST_STR_SIZE);
si_isa_debug("\n%s", inst_dump);
}
/* Stats */
si_emu->scalar_mem_inst_count++;
wavefront->scalar_mem_inst_count++;
/* Only one work item executes the instruction */
work_item = wavefront->scalar_work_item;
(*si_isa_inst_func[inst->info->inst])(work_item, inst);
if (debug_status(si_isa_debug_category))
{
si_isa_debug("\n");
}
break;
}
/* Vector ALU Instructions */
case SI_FMT_VOP2:
{
/* Dump instruction string when debugging */
if (debug_status(si_isa_debug_category))
{
si_inst_dump_vop2(inst, wavefront->inst_size, pc, wavefront->wavefront_pool, inst_dump,
MAX_INST_STR_SIZE);
si_isa_debug("\n%s", inst_dump);
}
/* Stats */
si_emu->vector_alu_inst_count++;
wavefront->vector_alu_inst_count++;
/* Execute the instruction */
SI_FOREACH_WORK_ITEM_IN_WAVEFRONT(wavefront, work_item_id)
{
work_item = ndrange->work_items[work_item_id];
if(si_wavefront_work_item_active(wavefront, work_item->id_in_wavefront))
(*si_isa_inst_func[inst->info->inst])(work_item, inst);
}
if (debug_status(si_isa_debug_category))
{
si_isa_debug("\n");
}
break;
}
case SI_FMT_VOP1:
{
/* Dump instruction string when debugging */
if (debug_status(si_isa_debug_category))
{
si_inst_dump_vop1(inst, wavefront->inst_size, pc, wavefront->wavefront_pool, inst_dump,
MAX_INST_STR_SIZE);
si_isa_debug("\n%s", inst_dump);
}
/* Stats */
si_emu->vector_alu_inst_count++;
wavefront->vector_alu_inst_count++;
if (inst->micro_inst.vop1.op == 2)
{
/* Instruction ignores execution mask and is only executed on one work item.
* Execute on the first active work item from the least significant bit in EXEC.
* (if exec is 0, execute work item 0) */
work_item = ndrange->work_items[wavefront->work_item_id_first];
if (si_isa_read_sreg(work_item, SI_EXEC) == 0 && si_isa_read_sreg(work_item, SI_EXEC + 1) == 0)
{
(*si_isa_inst_func[inst->info->inst])(work_item, inst);
}
else {
SI_FOREACH_WORK_ITEM_IN_WAVEFRONT(wavefront, work_item_id)
{
work_item = ndrange->work_items[work_item_id];
if(si_wavefront_work_item_active(wavefront, work_item->id_in_wavefront))
{
(*si_isa_inst_func[inst->info->inst])(work_item, inst);
break;
}
}
}
}
static void X86ContextUpdateState(X86Context *self, X86ContextState state)
{
X86Emu *emu = self->emu;
X86ContextState status_diff;
char state_str[MAX_STRING_SIZE];
/* Remove contexts from the following lists:
* running, suspended, zombie */
if (DOUBLE_LINKED_LIST_MEMBER(emu, running, self))
DOUBLE_LINKED_LIST_REMOVE(emu, running, self);
if (DOUBLE_LINKED_LIST_MEMBER(emu, suspended, self))
DOUBLE_LINKED_LIST_REMOVE(emu, suspended, self);
if (DOUBLE_LINKED_LIST_MEMBER(emu, zombie, self))
DOUBLE_LINKED_LIST_REMOVE(emu, zombie, self);
if (DOUBLE_LINKED_LIST_MEMBER(emu, finished, self))
DOUBLE_LINKED_LIST_REMOVE(emu, finished, self);
/* If the difference between the old and new state lies in other
* states other than 'x86_ctx_specmode', a reschedule is marked. */
status_diff = self->state ^ state;
if (status_diff & ~X86ContextSpecMode)
emu->schedule_signal = 1;
/* Update state */
self->state = state;
if (self->state & X86ContextFinished)
self->state = X86ContextFinished
| (state & X86ContextAlloc)
| (state & X86ContextMapped);
if (self->state & X86ContextZombie)
self->state = X86ContextZombie
| (state & X86ContextAlloc)
| (state & X86ContextMapped);
if (!(self->state & X86ContextSuspended) &&
!(self->state & X86ContextFinished) &&
!(self->state & X86ContextZombie) &&
!(self->state & X86ContextLocked))
self->state |= X86ContextRunning;
else
self->state &= ~X86ContextRunning;
/* Insert context into the corresponding lists. */
if (self->state & X86ContextRunning)
DOUBLE_LINKED_LIST_INSERT_HEAD(emu, running, self);
if (self->state & X86ContextZombie)
DOUBLE_LINKED_LIST_INSERT_HEAD(emu, zombie, self);
if (self->state & X86ContextFinished)
DOUBLE_LINKED_LIST_INSERT_HEAD(emu, finished, self);
if (self->state & X86ContextSuspended)
DOUBLE_LINKED_LIST_INSERT_HEAD(emu, suspended, self);
/* Dump new state (ignore 'x86_ctx_specmode' state, it's too frequent) */
if (debug_status(x86_context_debug_category) && (status_diff & ~X86ContextSpecMode))
{
str_map_flags(&x86_context_state_map, self->state, state_str, sizeof state_str);
X86ContextDebug("inst %lld: ctx %d changed state to %s\n",
asEmu(emu)->instructions, self->pid, state_str);
}
/* Start/stop x86 timer depending on whether there are any contexts
* currently running. */
if (emu->running_list_count)
m2s_timer_start(asEmu(emu)->timer);
else
m2s_timer_stop(asEmu(emu)->timer);
}