void X86CpuDumpSummary(Timing *self, FILE *f) {
X86Cpu *cpu = asX86Cpu(self);
double inst_per_cycle;
double uinst_per_cycle;
double branch_acc;
/* Calculate statistics */
inst_per_cycle = asTiming(cpu)->cycle
? (double)cpu->num_committed_inst / asTiming(cpu)->cycle
: 0.0;
uinst_per_cycle =
asTiming(cpu)->cycle
? (double)cpu->num_committed_uinst / asTiming(cpu)->cycle
: 0.0;
branch_acc =
cpu->num_branch_uinst
? (double)(cpu->num_branch_uinst - cpu->num_mispred_branch_uinst) /
cpu->num_branch_uinst
: 0.0;
/* Print statistics */
fprintf(f, "FastForwardInstructions = %lld\n", cpu->num_fast_forward_inst);
fprintf(f, "CommittedInstructions = %lld\n", cpu->num_committed_inst);
fprintf(f, "CommittedInstructionsPerCycle = %.4g\n", inst_per_cycle);
fprintf(f, "CommittedMicroInstructions = %lld\n", cpu->num_committed_uinst);
fprintf(f, "CommittedMicroInstructionsPerCycle = %.4g\n", uinst_per_cycle);
fprintf(f, "BranchPredictionAccuracy = %.4g\n", branch_acc);
/* Call parent */
TimingDumpSummary(self, f);
}
void ARMCpuCreate(ARMCpu *self) {
/* Parent */
TimingCreate(asTiming(self));
/* Virtual functions */
asObject(self)->Dump = ARMCpuDump;
asTiming(self)->DumpSummary = ARMCpuDumpSummary;
asTiming(self)->Run = ARMCpuRun;
}
void frm_lds_decode(struct frm_lds_t *lds) {
struct frm_uop_t *uop;
int instructions_processed = 0;
int list_entries;
int list_index = 0;
int i;
list_entries = list_count(lds->issue_buffer);
/* Sanity check the issue buffer */
assert(list_entries <= frm_gpu_lds_issue_buffer_size);
for (i = 0; i < list_entries; i++) {
uop = list_get(lds->issue_buffer, list_index);
assert(uop);
instructions_processed++;
/* Uop not ready yet */
if (asTiming(frm_gpu)->cycle < uop->issue_ready) {
list_index++;
continue;
}
/* Stall if the issue width has been reached. */
if (instructions_processed > frm_gpu_lds_width) {
frm_trace(
"si.inst id=%lld cu=%d wf=%d uop_id=%lld "
"stg=\"s\"\n",
uop->id_in_sm, lds->sm->id, uop->warp->id, uop->id_in_warp);
list_index++;
continue;
}
/* Sanity check the decode buffer */
assert(list_count(lds->decode_buffer) <= frm_gpu_lds_decode_buffer_size);
/* Stall if the decode buffer is full. */
if (list_count(lds->decode_buffer) == frm_gpu_lds_decode_buffer_size) {
frm_trace(
"si.inst id=%lld cu=%d wf=%d uop_id=%lld "
"stg=\"s\"\n",
uop->id_in_sm, lds->sm->id, uop->warp->id, uop->id_in_warp);
list_index++;
continue;
}
uop->decode_ready = asTiming(frm_gpu)->cycle + frm_gpu_lds_decode_latency;
list_remove(lds->issue_buffer, uop);
list_enqueue(lds->decode_buffer, uop);
frm_trace(
"si.inst id=%lld cu=%d wf=%d uop_id=%lld "
"stg=\"lds-d\"\n",
uop->id_in_sm, lds->sm->id, uop->warp->id, uop->id_in_warp);
}
}
void EvgGpuDumpSummary(Timing *self, FILE *f)
{
double inst_per_cycle;
/* Call parent */
TimingDumpSummary(asTiming(self), f);
/* Additional statistics */
inst_per_cycle = asTiming(evg_gpu)->cycle ?
(double) asEmu(evg_emu)->instructions
/ asTiming(evg_gpu)->cycle : 0.0;
fprintf(f, "IPC = %.4g\n", inst_per_cycle);
}
void EvgGpuCreate(EvgGpu *self)
{
struct evg_compute_unit_t *compute_unit;
int compute_unit_id;
/* Parent */
TimingCreate(asTiming(self));
/* Frequency */
asTiming(self)->frequency = evg_gpu_frequency;
asTiming(self)->frequency_domain = esim_new_domain(evg_gpu_frequency);
/* Initialize */
self->trash_uop_list = linked_list_create();
self->compute_units = xcalloc(evg_gpu_num_compute_units, sizeof(void *));
EVG_GPU_FOREACH_COMPUTE_UNIT(compute_unit_id)
{
self->compute_units[compute_unit_id] = evg_compute_unit_create();
compute_unit = self->compute_units[compute_unit_id];
compute_unit->id = compute_unit_id;
DOUBLE_LINKED_LIST_INSERT_TAIL(self, ready, compute_unit);
}
/* Virtual functions */
asObject(self)->Dump = EvgGpuDump;
asTiming(self)->DumpSummary = EvgGpuDumpSummary;
asTiming(self)->Run = EvgGpuRun;
asTiming(self)->MemConfigCheck = EvgGpuMemConfigCheck;
asTiming(self)->MemConfigDefault = EvgGpuMemConfigDefault;
asTiming(self)->MemConfigParseEntry = EvgGpuMemConfigParseEntry;
}
void frm_sm_update_fetch_visualization(
struct frm_sm_t *sm, int non_active_fb)
{
struct frm_uop_t *uop;
int list_entries;
int i;
/* Update visualization states for all instructions not issued */
list_entries = list_count(sm->fetch_buffers[non_active_fb]);
for (i = 0; i < list_entries; i++)
{
uop = list_get(sm->fetch_buffers[non_active_fb], i);
assert(uop);
/* Skip all uops that have not yet completed the fetch */
if (asTiming(frm_gpu)->cycle < uop->fetch_ready)
{
continue;
}
frm_trace("si.inst id=%lld cu=%d wf=%d uop_id=%lld stg=\"s\"\n",
uop->id_in_sm, sm->id,
uop->warp->id, uop->id_in_warp);
}
}
static int X86ThreadIssueSQ(X86Thread *self, int quantum) {
X86Cpu *cpu = self->cpu;
X86Core *core = self->core;
struct x86_uop_t *store;
struct linked_list_t *sq = self->sq;
struct mod_client_info_t *client_info;
/* Process SQ */
linked_list_head(sq);
while (!linked_list_is_end(sq) && quantum) {
/* Get store */
store = linked_list_get(sq);
assert(store->uinst->opcode == x86_uinst_store);
/* Only committed stores issue */
if (store->in_rob) break;
/* Check that memory system entry is ready */
if (!mod_can_access(self->data_mod, store->phy_addr)) break;
/* Remove store from store queue */
X86ThreadRemoveFromSQ(self);
/* create and fill the mod_client_info_t object */
client_info = mod_client_info_create(self->data_mod);
client_info->prefetcher_eip = store->eip;
/* Issue store */
mod_access(self->data_mod, mod_access_store, store->phy_addr, NULL,
core->event_queue, store, client_info);
/* The cache system will place the store at the head of the
* event queue when it is ready. For now, mark "in_event_queue" to
* prevent the uop from being freed. */
store->in_event_queue = 1;
store->issued = 1;
store->issue_when = asTiming(cpu)->cycle;
/* Statistics */
core->num_issued_uinst_array[store->uinst->opcode]++;
core->lsq_reads++;
core->reg_file_int_reads += store->ph_int_idep_count;
core->reg_file_fp_reads += store->ph_fp_idep_count;
self->num_issued_uinst_array[store->uinst->opcode]++;
self->lsq_reads++;
self->reg_file_int_reads += store->ph_int_idep_count;
self->reg_file_fp_reads += store->ph_fp_idep_count;
cpu->num_issued_uinst_array[store->uinst->opcode]++;
if (store->trace_cache) self->trace_cache->num_issued_uinst++;
/* One more instruction, update quantum. */
quantum--;
/* MMU statistics */
if (*mmu_report_file_name)
mmu_access_page(store->phy_addr, mmu_access_write);
}
return quantum;
}
static int X86ThreadCanFetch(X86Thread *self) {
X86Cpu *cpu = self->cpu;
X86Context *ctx = self->ctx;
unsigned int phy_addr;
unsigned int block;
/* Context must be running */
if (!ctx || !X86ContextGetState(ctx, X86ContextRunning))
return 0;
/* Fetch stalled or context evict signal activated */
if (self->fetch_stall_until >= asTiming(cpu)->cycle || ctx->evict_signal)
return 0;
/* Fetch queue must have not exceeded the limit of stored bytes
* to be able to store new macro-instructions. */
if (self->fetchq_occ >= x86_fetch_queue_size)
return 0;
/* If the next fetch address belongs to a new block, cache system
* must be accessible to read it. */
block = self->fetch_neip & ~(self->inst_mod->block_size - 1);
if (block != self->fetch_block) {
phy_addr = mmu_translate(self->ctx->address_space_index,
self->fetch_neip);
if (!mod_can_access(self->inst_mod, phy_addr))
return 0;
}
/* We can fetch */
return 1;
}
void frm_lds_write(struct frm_lds_t *lds) {
struct frm_uop_t *uop;
int instructions_processed = 0;
int list_entries;
int list_index = 0;
int i;
list_entries = list_count(lds->mem_buffer);
/* Sanity check the mem buffer */
assert(list_entries <= frm_gpu_lds_max_inflight_mem_accesses);
for (i = 0; i < list_entries; i++) {
uop = list_get(lds->mem_buffer, list_index);
assert(uop);
instructions_processed++;
/* Uop is not ready yet */
if (uop->lds_witness) {
list_index++;
continue;
}
/* Stall if the width has been reached */
if (instructions_processed > frm_gpu_lds_width) {
frm_trace(
"si.inst id=%lld cu=%d wf=%d uop_id=%lld "
"stg=\"s\"\n",
uop->id_in_sm, lds->sm->id, uop->warp->id, uop->id_in_warp);
list_index++;
continue;
}
/* Sanity check the write buffer */
assert(list_count(lds->write_buffer) <= frm_gpu_lds_write_buffer_size);
/* Stop if the write buffer is full */
if (list_count(lds->write_buffer) >= frm_gpu_lds_write_buffer_size) {
frm_trace(
"si.inst id=%lld cu=%d wf=%d uop_id=%lld "
"stg=\"s\"\n",
uop->id_in_sm, lds->sm->id, uop->warp->id, uop->id_in_warp);
list_index++;
continue;
}
/* Access complete, remove the uop from the queue */
uop->write_ready = asTiming(frm_gpu)->cycle + frm_gpu_lds_write_latency;
list_remove(lds->mem_buffer, uop);
list_enqueue(lds->write_buffer, uop);
instructions_processed++;
frm_trace(
"si.inst id=%lld cu=%d wf=%d uop_id=%lld "
"stg=\"lds-w\"\n",
uop->id_in_sm, lds->sm->id, uop->warp->id, uop->id_in_warp);
}
}
void X86CpuDumpUopReport(X86Cpu *self, FILE *f, long long *uop_stats,
char *prefix, int peak_ipc) {
long long uinst_int_count = 0;
long long uinst_logic_count = 0;
long long uinst_fp_count = 0;
long long uinst_mem_count = 0;
long long uinst_ctrl_count = 0;
long long uinst_total = 0;
char *name;
enum x86_uinst_flag_t flags;
int i;
for (i = 0; i < x86_uinst_opcode_count; i++) {
name = x86_uinst_info[i].name;
flags = x86_uinst_info[i].flags;
fprintf(f, "%s.Uop.%s = %lld\n", prefix, name, uop_stats[i]);
if (flags & X86_UINST_INT) uinst_int_count += uop_stats[i];
if (flags & X86_UINST_LOGIC) uinst_logic_count += uop_stats[i];
if (flags & X86_UINST_FP) uinst_fp_count += uop_stats[i];
if (flags & X86_UINST_MEM) uinst_mem_count += uop_stats[i];
if (flags & X86_UINST_CTRL) uinst_ctrl_count += uop_stats[i];
uinst_total += uop_stats[i];
}
fprintf(f, "%s.Integer = %lld\n", prefix, uinst_int_count);
fprintf(f, "%s.Logic = %lld\n", prefix, uinst_logic_count);
fprintf(f, "%s.FloatingPoint = %lld\n", prefix, uinst_fp_count);
fprintf(f, "%s.Memory = %lld\n", prefix, uinst_mem_count);
fprintf(f, "%s.Ctrl = %lld\n", prefix, uinst_ctrl_count);
fprintf(f, "%s.WndSwitch = %lld\n", prefix,
uop_stats[x86_uinst_call] + uop_stats[x86_uinst_ret]);
fprintf(f, "%s.Total = %lld\n", prefix, uinst_total);
fprintf(f, "%s.IPC = %.4g\n", prefix,
asTiming(self)->cycle ? (double)uinst_total / asTiming(self)->cycle
: 0.0);
fprintf(f, "%s.DutyCycle = %.4g\n", prefix,
asTiming(self)->cycle && peak_ipc
? (double)uinst_total / asTiming(self)->cycle / peak_ipc
: 0.0);
fprintf(f, "\n");
}
void X86CpuCreate(X86Cpu *self, X86Emu *emu) {
X86Core *core;
X86Thread *thread;
char name[MAX_STRING_SIZE];
int i;
int j;
/* Parent */
TimingCreate(asTiming(self));
/* Frequency */
asTiming(self)->frequency = x86_cpu_frequency;
asTiming(self)->frequency_domain = esim_new_domain(x86_cpu_frequency);
/* Initialize */
self->emu = emu;
self->uop_trace_list = linked_list_create();
/* Create cores */
self->cores = xcalloc(x86_cpu_num_cores, sizeof(X86Core *));
for (i = 0; i < x86_cpu_num_cores; i++) self->cores[i] = new (X86Core, self);
/* Assign names and IDs to cores and threads */
for (i = 0; i < x86_cpu_num_cores; i++) {
core = self->cores[i];
snprintf(name, sizeof name, "c%d", i);
X86CoreSetName(core, name);
core->id = i;
for (j = 0; j < x86_cpu_num_threads; j++) {
thread = core->threads[j];
snprintf(name, sizeof name, "c%dt%d", i, j);
X86ThreadSetName(thread, name);
thread->id_in_core = j;
thread->id_in_cpu = i * x86_cpu_num_threads + j;
}
}
/* Virtual functions */
asObject(self)->Dump = X86CpuDump;
asTiming(self)->DumpSummary = X86CpuDumpSummary;
asTiming(self)->Run = X86CpuRun;
asTiming(self)->MemConfigCheck = X86CpuMemConfigCheck;
asTiming(self)->MemConfigDefault = X86CpuMemConfigDefault;
asTiming(self)->MemConfigParseEntry = X86CpuMemConfigParseEntry;
/* Trace */
x86_trace_header("x86.init version=\"%d.%d\" num_cores=%d num_threads=%d\n",
X86_TRACE_VERSION_MAJOR, X86_TRACE_VERSION_MINOR,
x86_cpu_num_cores, x86_cpu_num_threads);
}
void frm_vector_mem_complete(struct frm_vector_mem_unit_t *vector_mem)
{
struct frm_uop_t *uop = NULL;
int list_entries;
int i;
int list_index = 0;
/* Process completed memory instructions */
list_entries = list_count(vector_mem->write_buffer);
/* Sanity check the write buffer */
assert(list_entries <= frm_gpu_vector_mem_width);
for (i = 0; i < list_entries; i++)
{
uop = list_get(vector_mem->write_buffer, list_index);
assert(uop);
/* Uop is not ready */
if (asTiming(frm_gpu)->cycle < uop->write_ready)
{
list_index++;
continue;
}
/* Access complete, remove the uop from the queue */
list_remove(vector_mem->write_buffer, uop);
assert(uop->warp_inst_queue_entry->lgkm_cnt > 0);
uop->warp_inst_queue_entry->lgkm_cnt--;
frm_trace("si.end_inst id=%lld cu=%d\n", uop->id_in_sm,
uop->sm->id);
/* Free uop */
frm_uop_free(uop);
/* Statistics */
vector_mem->inst_count++;
frm_gpu->last_complete_cycle = asTiming(frm_gpu)->cycle;
}
}
void frm_sm_spatial_report_dump(struct frm_sm_t *sm)
{
FILE *f = spatial_report_file;
fprintf(f,"CU,%d,MemAcc,%lld,MappedWGs,%lld,Cycles,%lld\n",
sm->id,
sm->vector_mem_unit.inflight_mem_accesses,
sm->interval_mapped_thread_blocks,
asTiming(frm_gpu)->cycle);
}
void evg_cu_spatial_report_dump(struct evg_compute_unit_t *compute_unit) {
FILE *f = spatial_report_file;
fprintf(f,
"CU,%d,CFInst,%lld,MemAcc,%lld,TEXInstn,%lld,ALUInstn,%lld,Cycles,%"
"lld \n",
compute_unit->id, compute_unit->cf_engine.interval_inst_count,
compute_unit->inflight_mem_accesses,
compute_unit->tex_engine.interval_inst_count,
compute_unit->alu_engine.interval_inst_count,
asTiming(evg_gpu)->cycle);
}
void si_simd_complete(struct si_simd_t *simd)
{
struct si_uop_t *uop;
int list_entries;
int list_index = 0;
int i;
list_entries = list_count(simd->exec_buffer);
assert(list_entries <= si_gpu_simd_exec_buffer_size);
for (i = 0; i < list_entries; i++)
{
uop = list_get(simd->exec_buffer, list_index);
assert(uop);
if (asTiming(si_gpu)->cycle < uop->execute_ready)
{
list_index++;
continue;
}
/* Access complete, remove the uop from the queue */
list_remove(simd->exec_buffer, uop);
si_trace("si.end_inst id=%lld cu=%d\n", uop->id_in_compute_unit,
uop->compute_unit->id);
/* Free uop */
si_uop_free(uop);
/* Statistics */
simd->inst_count++;
si_gpu->last_complete_cycle = asTiming(si_gpu)->cycle;
}
}
void evg_cu_interval_update(struct evg_compute_unit_t *compute_unit) {
/* If interval - reset the counters in all the engines */
compute_unit->interval_cycle++;
if (!(asTiming(evg_gpu)->cycle % spatial_profiling_interval)) {
evg_cu_spatial_report_dump(compute_unit);
compute_unit->cf_engine.interval_inst_count = 0;
compute_unit->alu_engine.interval_inst_count = 0;
compute_unit->tex_engine.interval_inst_count = 0;
/* This counter is not reset since memory accesses could still be in flight
* in the hierarchy*/
/* compute_unit->inflight_mem_accesses = 0; */
compute_unit->interval_cycle = 0;
}
}
void si_cu_spatial_report_dump(struct si_compute_unit_t *compute_unit)
{
FILE *f = spatial_report_file;
fprintf(f,
"CU,%d,MemAcc,%lld,MappedWGs,%lld,UnmappedWGs,%lld,ALUIssued,%lld,LDSIssued,%lld,Cycles,%lld\n",
compute_unit->id,
compute_unit->vector_mem_unit.inflight_mem_accesses,
compute_unit->interval_mapped_work_groups,
compute_unit->interval_unmapped_work_groups,
compute_unit->interval_alu_issued,
compute_unit->interval_lds_issued,
asTiming(si_gpu)->cycle);
}
void frm_sm_interval_update(struct frm_sm_t *sm)
{
/* If interval - reset the counters in all the engines */
sm->interval_cycle ++;
if (!(asTiming(frm_gpu)->cycle % spatial_profiling_interval))
{
frm_sm_spatial_report_dump(sm);
/*
* This counter is not reset since memory accesses could still
* be in flight in the hierarchy
* sm->inflight_mem_accesses = 0;
*/
sm->interval_cycle = 0;
sm->interval_mapped_thread_blocks = 0;
}
}
int X86ThreadCacheMissInEventQueue(X86Thread *self)
{
X86Cpu *cpu = self->cpu;
X86Core *core = self->core;
struct linked_list_t *event_queue = core->event_queue;
struct x86_uop_t *uop;
LINKED_LIST_FOR_EACH(event_queue)
{
uop = linked_list_get(event_queue);
if (uop->thread != self || uop->uinst->opcode != x86_uinst_load)
continue;
if (asTiming(cpu)->cycle - uop->issue_when > 5)
return 1;
}
return 0;
}
int X86ThreadLongLatencyInEventQueue(X86Thread *self)
{
X86Cpu *cpu = self->cpu;
X86Core *core = self->core;
struct linked_list_t *event_queue = core->event_queue;
struct x86_uop_t *uop;
LINKED_LIST_FOR_EACH(event_queue)
{
uop = linked_list_get(event_queue);
if (uop->thread != self)
continue;
if (asTiming(cpu)->cycle - uop->issue_when > 20)
return 1;
}
return 0;
}
void si_cu_interval_update(struct si_compute_unit_t *compute_unit)
{
/* If interval - reset the counters in all the engines */
compute_unit->interval_cycle ++;
if (!(asTiming(si_gpu)->cycle % spatial_profiling_interval))
{
si_cu_spatial_report_dump(compute_unit);
/*
* This counter is not reset since memory accesses could still
* be in flight in the hierarchy
* compute_unit->inflight_mem_accesses = 0;
*/
compute_unit->interval_cycle = 0;
compute_unit->interval_mapped_work_groups = 0;
compute_unit->interval_unmapped_work_groups = 0;
compute_unit->interval_alu_issued = 0;
compute_unit->interval_lds_issued = 0;
}
}
static struct evg_wavefront_t *evg_schedule_greedy(
struct evg_compute_unit_t *compute_unit) {
struct evg_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 <
evg_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 == EVG_CLAUSE_CF);
linked_list_find(wavefront_pool, temp_wavefront);
assert(!wavefront_pool->error_code);
linked_list_remove(wavefront_pool);
temp_wavefront->sched_when = asTiming(evg_gpu)->cycle;
return temp_wavefront;
}
void frm_vector_mem_mem(struct frm_vector_mem_unit_t *vector_mem)
{
struct frm_uop_t *uop;
struct frm_thread_uop_t *thread_uop;
struct frm_thread_t *thread;
int thread_id;
int instructions_processed = 0;
int list_entries;
int i;
enum mod_access_kind_t access_kind;
int list_index = 0;
list_entries = list_count(vector_mem->read_buffer);
/* Sanity check the read buffer */
assert(list_entries <= frm_gpu_vector_mem_read_buffer_size);
for (i = 0; i < list_entries; i++)
{
uop = list_get(vector_mem->read_buffer, list_index);
assert(uop);
instructions_processed++;
/* Uop is not ready yet */
if (asTiming(frm_gpu)->cycle < uop->read_ready)
{
list_index++;
continue;
}
/* Stall if the width has been reached. */
if (instructions_processed > frm_gpu_vector_mem_width)
{
frm_trace("si.inst id=%lld cu=%d wf=%d uop_id=%lld "
"stg=\"s\"\n", uop->id_in_sm,
vector_mem->sm->id,
uop->warp->id, uop->id_in_warp);
list_index++;
continue;
}
/* Sanity check mem buffer */
assert(list_count(vector_mem->mem_buffer) <=
frm_gpu_vector_mem_max_inflight_mem_accesses);
/* Stall if there is not room in the memory buffer */
if (list_count(vector_mem->mem_buffer) ==
frm_gpu_vector_mem_max_inflight_mem_accesses)
{
frm_trace("si.inst id=%lld cu=%d wf=%d uop_id=%lld "
"stg=\"s\"\n", uop->id_in_sm,
vector_mem->sm->id,
uop->warp->id, uop->id_in_warp);
list_index++;
continue;
}
/* Set the access type */
if (uop->vector_mem_write && !uop->glc)
access_kind = mod_access_nc_store;
else if (uop->vector_mem_write && uop->glc)
access_kind = mod_access_store;
else if (uop->vector_mem_read)
access_kind = mod_access_load;
else
fatal("%s: invalid access kind", __FUNCTION__);
/* Access global memory */
assert(!uop->global_mem_witness);
for (thread_id = uop->warp->threads[0]->id_in_warp;
thread_id < uop->warp->thread_count;
thread_id++)
{
thread = uop->warp->threads[thread_id];
thread_uop =
&uop->thread_uop[thread->id_in_warp];
mod_access(vector_mem->sm->global_memory,
access_kind,
thread_uop->global_mem_access_addr,
&uop->global_mem_witness, NULL, NULL, NULL);
uop->global_mem_witness--;
}
if(frm_spatial_report_active)
{
if (uop->vector_mem_write)
{
uop->num_global_mem_write +=
uop->global_mem_witness;
frm_report_global_mem_inflight(uop->sm,
uop->num_global_mem_write);
}
else if (uop->vector_mem_read)
{
uop->num_global_mem_read +=
uop->global_mem_witness;
frm_report_global_mem_inflight(uop->sm,
uop->num_global_mem_read);
}
else
fatal("%s: invalid access kind", __FUNCTION__);
}
/* Increment outstanding memory access count */
uop->warp_inst_queue_entry->lgkm_cnt++;
/* Transfer the uop to the mem buffer */
list_remove(vector_mem->read_buffer, uop);
list_enqueue(vector_mem->mem_buffer, uop);
frm_trace("si.inst id=%lld cu=%d wf=%d uop_id=%lld "
"stg=\"mem-m\"\n", uop->id_in_sm,
vector_mem->sm->id, uop->warp->id,
uop->id_in_warp);
}
}
void frm_lds_mem(struct frm_lds_t *lds) {
struct frm_uop_t *uop;
struct frm_thread_uop_t *thread_uop;
struct frm_thread_t *thread;
int thread_id;
int instructions_processed = 0;
int list_entries;
int i, j;
enum mod_access_kind_t access_type;
int list_index = 0;
list_entries = list_count(lds->read_buffer);
/* Sanity check the read buffer */
assert(list_entries <= frm_gpu_lds_read_buffer_size);
for (i = 0; i < list_entries; i++) {
uop = list_get(lds->read_buffer, list_index);
assert(uop);
instructions_processed++;
/* Uop is not ready yet */
if (asTiming(frm_gpu)->cycle < uop->read_ready) {
list_index++;
continue;
}
/* Stall if the width has been reached. */
if (instructions_processed > frm_gpu_lds_width) {
frm_trace(
"si.inst id=%lld cu=%d wf=%d uop_id=%lld "
"stg=\"s\"\n",
uop->id_in_sm, lds->sm->id, uop->warp->id, uop->id_in_warp);
list_index++;
continue;
}
/* Sanity check uop */
assert(uop->lds_read || uop->lds_write);
/* Sanity check mem buffer */
assert(list_count(lds->mem_buffer) <=
frm_gpu_lds_max_inflight_mem_accesses);
/* Stall if there is no room in the memory buffer */
if (list_count(lds->mem_buffer) == frm_gpu_lds_max_inflight_mem_accesses) {
frm_trace(
"si.inst id=%lld cu=%d wf=%d uop_id=%lld "
"stg=\"s\"\n",
uop->id_in_sm, lds->sm->id, uop->warp->id, uop->id_in_warp);
list_index++;
continue;
}
/* Access local memory */
for (thread_id = uop->warp->threads[0]->id_in_warp;
thread_id < uop->warp->thread_count; thread_id++) {
thread = uop->warp->threads[thread_id];
thread_uop = &uop->thread_uop[thread->id_in_warp];
for (j = 0; j < thread_uop->lds_access_count; j++) {
if (thread->lds_access_type[j] == 1) {
access_type = mod_access_load;
} else if (thread->lds_access_type[j] == 2) {
access_type = mod_access_store;
} else {
fatal("%s: invalid lds access type", __FUNCTION__);
}
mod_access(lds->sm->lds_module, access_type,
thread_uop->lds_access_addr[j], &uop->lds_witness, NULL,
NULL, NULL);
uop->lds_witness--;
}
}
/* Increment outstanding memory access count */
uop->warp_inst_queue_entry->lgkm_cnt++;
/* Transfer the uop to the mem buffer */
list_remove(lds->read_buffer, uop);
list_enqueue(lds->mem_buffer, uop);
frm_trace(
"si.inst id=%lld cu=%d wf=%d uop_id=%lld "
"stg=\"lds-m\"\n",
uop->id_in_sm, lds->sm->id, uop->warp->id, uop->id_in_warp);
}
}
void si_simd_execute(struct si_simd_t *simd)
{
struct si_uop_t *uop;
int list_entries;
int list_index = 0;
int instructions_processed = 0;
int i;
list_entries = list_count(simd->decode_buffer);
/* Sanity check the decode buffer */
assert(list_entries <= si_gpu_simd_decode_buffer_size);
for (i = 0; i < list_entries; i++)
{
uop = list_get(simd->decode_buffer, list_index);
assert(uop);
instructions_processed++;
/* Uop is not ready yet */
if (asTiming(si_gpu)->cycle < uop->decode_ready)
{
list_index++;
continue;
}
/* Stall if the width has been reached */
if (instructions_processed > si_gpu_simd_width)
{
si_trace("si.inst id=%lld cu=%d wf=%d uop_id=%lld "
"stg=\"s\"\n", uop->id_in_compute_unit,
simd->compute_unit->id, uop->wavefront->id,
uop->id_in_wavefront);
list_index++;
continue;
}
/* Sanity check exec buffer */
assert(list_count(simd->exec_buffer) <=
si_gpu_simd_exec_buffer_size);
/* Stall if SIMD unit is full */
if (list_count(simd->exec_buffer) ==
si_gpu_simd_exec_buffer_size)
{
si_trace("si.inst id=%lld cu=%d wf=%d uop_id=%lld "
"stg=\"s\"\n", uop->id_in_compute_unit,
simd->compute_unit->id, uop->wavefront->id,
uop->id_in_wavefront);
list_index++;
continue;
}
/* Includes time for pipelined read-exec-write of
* all subwavefronts */
uop->execute_ready = asTiming(si_gpu)->cycle +
si_gpu_simd_exec_latency;
/* Transfer the uop to the outstanding execution buffer */
list_remove(simd->decode_buffer, uop);
list_enqueue(simd->exec_buffer, uop);
uop->wavefront_pool_entry->ready_next_cycle = 1;
si_trace("si.inst id=%lld cu=%d wf=%d uop_id=%lld "
"stg=\"simd-e\"\n", uop->id_in_compute_unit,
simd->compute_unit->id, uop->wavefront->id,
uop->id_in_wavefront);
}
}
void frm_vector_mem_write(struct frm_vector_mem_unit_t *vector_mem)
{
struct frm_uop_t *uop;
int instructions_processed = 0;
int list_entries;
int list_index = 0;
int i;
list_entries = list_count(vector_mem->mem_buffer);
/* Sanity check the mem buffer */
assert(list_entries <= frm_gpu_vector_mem_max_inflight_mem_accesses);
for (i = 0; i < list_entries; i++)
{
uop = list_get(vector_mem->mem_buffer, list_index);
assert(uop);
instructions_processed++;
/* Uop is not ready yet */
if (uop->global_mem_witness)
{
list_index++;
continue;
}
/* Stall if the width has been reached. */
if (instructions_processed > frm_gpu_vector_mem_width)
{
frm_trace("si.inst id=%lld cu=%d wf=%d uop_id=%lld "
"stg=\"s\"\n", uop->id_in_sm,
vector_mem->sm->id,
uop->warp->id, uop->id_in_warp);
list_index++;
continue;
}
/* Sanity check write buffer */
assert(list_count(vector_mem->write_buffer) <=
frm_gpu_vector_mem_write_buffer_size);
/* Stop if the write buffer is full. */
if (list_count(vector_mem->write_buffer) ==
frm_gpu_vector_mem_write_buffer_size)
{
frm_trace("si.inst id=%lld cu=%d wf=%d uop_id=%lld "
"stg=\"s\"\n", uop->id_in_sm,
vector_mem->sm->id,
uop->warp->id, uop->id_in_warp);
list_index++;
continue;
}
/* Access complete, remove the uop from the queue */
uop->write_ready = asTiming(frm_gpu)->cycle +
frm_gpu_vector_mem_write_latency;
/* In the above context, access means any of the
* mod_access calls in frm_vector_mem_mem. Means all
* inflight accesses for uop are done */
if(frm_spatial_report_active)
{
if (uop->vector_mem_write)
{
frm_report_global_mem_finish(uop->sm,
uop->num_global_mem_write);
}
else if (uop->vector_mem_read)
{
frm_report_global_mem_finish(uop->sm,
uop->num_global_mem_read);
}
else
{
fatal("%s: invalid access kind", __FUNCTION__);
}
}
list_remove(vector_mem->mem_buffer, uop);
list_enqueue(vector_mem->write_buffer, uop);
frm_trace("si.inst id=%lld cu=%d wf=%d uop_id=%lld "
"stg=\"mem-w\"\n", uop->id_in_sm,
vector_mem->sm->id, uop->warp->id,
uop->id_in_warp);
}
}
void si_simd_decode(struct si_simd_t *simd)
{
struct si_uop_t *uop;
int instructions_processed = 0;
int list_entries;
int list_index = 0;
int i;
list_entries = list_count(simd->issue_buffer);
/* Sanity check the issue buffer */
assert(list_entries <= si_gpu_simd_issue_buffer_size);
for (i = 0; i < list_entries; i++)
{
uop = list_get(simd->issue_buffer, list_index);
assert(uop);
instructions_processed++;
/* Uop not ready yet */
if (asTiming(si_gpu)->cycle < uop->issue_ready)
{
list_index++;
continue;
}
/* Stall if the issue width has been reached. */
if (instructions_processed > si_gpu_simd_width)
{
si_trace("si.inst id=%lld cu=%d wf=%d uop_id=%lld "
"stg=\"s\"\n", uop->id_in_compute_unit,
simd->compute_unit->id, uop->wavefront->id,
uop->id_in_wavefront);
list_index++;
continue;
}
/* Sanity check the decode buffer */
assert(list_count(simd->decode_buffer) <=
si_gpu_simd_decode_buffer_size);
/* Stall if the decode buffer is full. */
if (list_count(simd->decode_buffer) ==
si_gpu_simd_decode_buffer_size)
{
si_trace("si.inst id=%lld cu=%d wf=%d uop_id=%lld "
"stg=\"s\"\n", uop->id_in_compute_unit,
simd->compute_unit->id, uop->wavefront->id,
uop->id_in_wavefront);
list_index++;
continue;
}
uop->decode_ready = asTiming(si_gpu)->cycle + si_gpu_simd_decode_latency;
list_remove(simd->issue_buffer, uop);
list_enqueue(simd->decode_buffer, uop);
if (si_spatial_report_active)
si_alu_report_new_inst(simd->compute_unit);
si_trace("si.inst id=%lld cu=%d wf=%d uop_id=%lld "
"stg=\"simd-d\"\n", uop->id_in_compute_unit,
simd->compute_unit->id, uop->wavefront->id,
uop->id_in_wavefront);
}
}
static int X86ThreadIssueLQ(X86Thread *self, int quant)
{
X86Core *core = self->core;
X86Cpu *cpu = self->cpu;
struct linked_list_t *lq = self->lq;
struct x86_uop_t *load;
struct mod_client_info_t *client_info;
/* Process lq */
linked_list_head(lq);
while (!linked_list_is_end(lq) && quant)
{
/* Get element from load queue. If it is not ready, go to the next one */
load = linked_list_get(lq);
if (!load->ready && !X86ThreadIsUopReady(self, load))
{
linked_list_next(lq);
continue;
}
load->ready = 1;
/* Check that memory system is accessible */
if (!mod_can_access(self->data_mod, load->phy_addr))
{
linked_list_next(lq);
continue;
}
/* Remove from load queue */
assert(load->uinst->opcode == x86_uinst_load);
X86ThreadRemoveFromLQ(self);
/* create and fill the mod_client_info_t object */
client_info = mod_client_info_create(self->data_mod);
client_info->prefetcher_eip = load->eip;
/* Access memory system */
mod_access(self->data_mod, mod_access_load,
load->phy_addr, NULL, core->event_queue, load, client_info);
/* The cache system will place the load at the head of the
* event queue when it is ready. For now, mark "in_event_queue" to
* prevent the uop from being freed. */
load->in_event_queue = 1;
load->issued = 1;
load->issue_when = asTiming(cpu)->cycle;
/* Statistics */
core->num_issued_uinst_array[load->uinst->opcode]++;
core->lsq_reads++;
core->reg_file_int_reads += load->ph_int_idep_count;
core->reg_file_fp_reads += load->ph_fp_idep_count;
self->num_issued_uinst_array[load->uinst->opcode]++;
self->lsq_reads++;
self->reg_file_int_reads += load->ph_int_idep_count;
self->reg_file_fp_reads += load->ph_fp_idep_count;
cpu->num_issued_uinst_array[load->uinst->opcode]++;
if (load->trace_cache)
self->trace_cache->num_issued_uinst++;
/* One more instruction issued, update quantum. */
quant--;
/* MMU statistics */
MMUAccessPage(cpu->mmu, load->phy_addr, mmu_access_read);
/* Trace */
x86_trace("x86.inst id=%lld core=%d stg=\"i\"\n",
load->id_in_core, core->id);
}
return quant;
}
static int X86ThreadIssuePreQ(X86Thread *self, int quantum)
{
X86Core *core = self->core;
X86Cpu *cpu = self->cpu;
struct linked_list_t *preq = self->preq;
struct x86_uop_t *prefetch;
/* Process preq */
linked_list_head(preq);
while (!linked_list_is_end(preq) && quantum)
{
/* Get element from prefetch queue. If it is not ready, go to the next one */
prefetch = linked_list_get(preq);
if (!prefetch->ready && !X86ThreadIsUopReady(self, prefetch))
{
linked_list_next(preq);
continue;
}
/*
* Make sure its not been prefetched recently. This is just to avoid unnecessary
* memory traffic. Even though the cache will realise a "hit" on redundant
* prefetches, its still helpful to avoid going to the memory (cache).
*/
if (prefetch_history_is_redundant(core->prefetch_history,
self->data_mod, prefetch->phy_addr))
{
/* remove from queue. do not prefetch. */
assert(prefetch->uinst->opcode == x86_uinst_prefetch);
X86ThreadRemovePreQ(self);
prefetch->completed = 1;
x86_uop_free_if_not_queued(prefetch);
continue;
}
prefetch->ready = 1;
/* Check that memory system is accessible */
if (!mod_can_access(self->data_mod, prefetch->phy_addr))
{
linked_list_next(preq);
continue;
}
/* Remove from prefetch queue */
assert(prefetch->uinst->opcode == x86_uinst_prefetch);
X86ThreadRemovePreQ(self);
/* Access memory system */
mod_access(self->data_mod, mod_access_prefetch,
prefetch->phy_addr, NULL, core->event_queue, prefetch, NULL);
/* Record prefetched address */
prefetch_history_record(core->prefetch_history, prefetch->phy_addr);
/* The cache system will place the prefetch at the head of the
* event queue when it is ready. For now, mark "in_event_queue" to
* prevent the uop from being freed. */
prefetch->in_event_queue = 1;
prefetch->issued = 1;
prefetch->issue_when = asTiming(cpu)->cycle;
/* Statistics */
core->num_issued_uinst_array[prefetch->uinst->opcode]++;
core->lsq_reads++;
core->reg_file_int_reads += prefetch->ph_int_idep_count;
core->reg_file_fp_reads += prefetch->ph_fp_idep_count;
self->num_issued_uinst_array[prefetch->uinst->opcode]++;
self->lsq_reads++;
self->reg_file_int_reads += prefetch->ph_int_idep_count;
self->reg_file_fp_reads += prefetch->ph_fp_idep_count;
cpu->num_issued_uinst_array[prefetch->uinst->opcode]++;
if (prefetch->trace_cache)
self->trace_cache->num_issued_uinst++;
/* One more instruction issued, update quantum. */
quantum--;
/* MMU statistics */
MMUAccessPage(cpu->mmu, prefetch->phy_addr, mmu_access_read);
/* Trace */
x86_trace("x86.inst id=%lld core=%d stg=\"i\"\n",
prefetch->id_in_core, core->id);
}
return quantum;
}
static int X86ThreadIssueIQ(X86Thread *self, int quant)
{
X86Cpu *cpu = self->cpu;
X86Core *core = self->core;
struct linked_list_t *iq = self->iq;
struct x86_uop_t *uop;
int lat;
/* Find instruction to issue */
linked_list_head(iq);
while (!linked_list_is_end(iq) && quant)
{
/* Get element from IQ */
uop = linked_list_get(iq);
assert(x86_uop_exists(uop));
assert(!(uop->flags & X86_UINST_MEM));
if (!uop->ready && !X86ThreadIsUopReady(self, uop))
{
linked_list_next(iq);
continue;
}
uop->ready = 1; /* avoid next call to 'X86ThreadIsUopReady' */
/* Run the instruction in its corresponding functional unit.
* If the instruction does not require a functional unit, 'X86CoreReserveFunctionalUnit'
* returns 1 cycle latency. If there is no functional unit available,
* 'X86CoreReserveFunctionalUnit' returns 0. */
lat = X86CoreReserveFunctionalUnit(core, uop);
if (!lat)
{
linked_list_next(iq);
continue;
}
/* Instruction was issued to the corresponding fu.
* Remove it from IQ */
X86ThreadRemoveFromIQ(self);
/* Schedule inst in Event Queue */
assert(!uop->in_event_queue);
assert(lat > 0);
uop->issued = 1;
uop->issue_when = asTiming(cpu)->cycle;
uop->when = asTiming(cpu)->cycle + lat;
X86CoreInsertInEventQueue(core, uop);
/* Statistics */
core->num_issued_uinst_array[uop->uinst->opcode]++;
core->iq_reads++;
core->reg_file_int_reads += uop->ph_int_idep_count;
core->reg_file_fp_reads += uop->ph_fp_idep_count;
self->num_issued_uinst_array[uop->uinst->opcode]++;
self->iq_reads++;
self->reg_file_int_reads += uop->ph_int_idep_count;
self->reg_file_fp_reads += uop->ph_fp_idep_count;
cpu->num_issued_uinst_array[uop->uinst->opcode]++;
if (uop->trace_cache)
self->trace_cache->num_issued_uinst++;
/* One more instruction issued, update quantum. */
quant--;
/* Trace */
x86_trace("x86.inst id=%lld core=%d stg=\"i\"\n",
uop->id_in_core, core->id);
}
return quant;
}
