static void
tree_output()
{
Symbol **symbols, *main_sym;
size_t i, num;
cflow_depmap_t depmap;
/* Collect functions and assign them ordinal numbers */
num = collect_functions(&symbols);
for (i = 0; i < num; i++)
symbols[i]->ord = i;
/* Create a dependency matrix */
depmap = depmap_alloc(num);
for (i = 0; i < num; i++) {
if (symbols[i]->callee) {
struct linked_list_entry *p;
for (p = linked_list_head(symbols[i]->callee); p;
p = p->next) {
Symbol *s = (Symbol*) p->data;
if (symbol_is_function(s))
depmap_set(depmap, i, ((Symbol*)p->data)->ord);
}
}
}
depmap_tc(depmap);
/* Mark recursive calls */
for (i = 0; i < num; i++)
if (depmap_isset(depmap, i, i))
symbols[i]->recursive = 1;
free(depmap);
free(symbols);
/* Collect and sort all symbols */
num = collect_symbols(&symbols, is_var, 0);
qsort(symbols, num, sizeof(*symbols), compare);
/* Produce output */
begin();
if (reverse_tree) {
for (i = 0; i < num; i++) {
inverted_tree(0, 0, symbols[i]);
separator();
}
} else {
main_sym = lookup(start_name);
if (main_sym) {
direct_tree(0, 0, main_sym);
separator();
} else {
for (i = 0; i < num; i++) {
if (symbols[i]->callee == NULL)
continue;
direct_tree(0, 0, symbols[i]);
separator();
}
}
}
end();
free(symbols);
}
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 */
if (*mmu_report_file_name)
mmu_access_page(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;
}
static int x86_cpu_issue_lq(int core, int thread, int quant)
{
struct linked_list_t *lq = X86_THREAD.lq;
struct x86_uop_t *load;
/* 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 && !x86_reg_file_ready(load))
{
linked_list_next(lq);
continue;
}
load->ready = 1;
/* Check that memory system is accessible */
if (!mod_can_access(X86_THREAD.data_mod, load->phy_addr))
{
linked_list_next(lq);
continue;
}
/* Remove from load queue */
assert(load->uinst->opcode == x86_uinst_load);
x86_lq_remove(core, thread);
/* Access memory system */
mod_access(X86_THREAD.data_mod, mod_access_load,
load->phy_addr, NULL, X86_CORE.event_queue, load);
/* 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 = x86_cpu->cycle;
/* Instruction issued */
X86_CORE.issued[load->uinst->opcode]++;
X86_CORE.lsq_reads++;
X86_CORE.reg_file_int_reads += load->ph_int_idep_count;
X86_CORE.reg_file_fp_reads += load->ph_fp_idep_count;
X86_THREAD.issued[load->uinst->opcode]++;
X86_THREAD.lsq_reads++;
X86_THREAD.reg_file_int_reads += load->ph_int_idep_count;
X86_THREAD.reg_file_fp_reads += load->ph_fp_idep_count;
x86_cpu->issued[load->uinst->opcode]++;
quant--;
/* MMU statistics */
if (*mmu_report_file_name)
mmu_access_page(load->phy_addr, mmu_access_read);
/* Trace */
x86_trace("x86.inst id=%lld core=%d stg=\"i\"\n",
load->id_in_core, load->core);
}
return quant;
}
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 */
if (*mmu_report_file_name) mmu_access_page(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 x86_cpu_issue_iq(int core, int thread, int quant)
{
struct linked_list_t *iq = X86_THREAD.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 && !x86_reg_file_ready(uop))
{
linked_list_next(iq);
continue;
}
uop->ready = 1; /* avoid next call to 'x86_reg_file_ready' */
/* Run the instruction in its corresponding functional unit.
* If the instruction does not require a functional unit, 'x86_fu_reserve'
* returns 1 cycle latency. If there is no functional unit available,
* 'x86_fu_reserve' returns 0. */
lat = x86_fu_reserve(uop);
if (!lat)
{
linked_list_next(iq);
continue;
}
/* Instruction was issued to the corresponding fu.
* Remove it from IQ */
x86_iq_remove(core, thread);
/* Schedule inst in Event Queue */
assert(!uop->in_event_queue);
assert(lat > 0);
uop->issued = 1;
uop->issue_when = x86_cpu->cycle;
uop->when = x86_cpu->cycle + lat;
x86_event_queue_insert(X86_CORE.event_queue, uop);
/* Instruction issued */
X86_CORE.issued[uop->uinst->opcode]++;
X86_CORE.iq_reads++;
X86_CORE.reg_file_int_reads += uop->ph_int_idep_count;
X86_CORE.reg_file_fp_reads += uop->ph_fp_idep_count;
X86_THREAD.issued[uop->uinst->opcode]++;
X86_THREAD.iq_reads++;
X86_THREAD.reg_file_int_reads += uop->ph_int_idep_count;
X86_THREAD.reg_file_fp_reads += uop->ph_fp_idep_count;
x86_cpu->issued[uop->uinst->opcode]++;
quant--;
/* Trace */
x86_trace("x86.inst id=%lld core=%d stg=\"i\"\n",
uop->id_in_core, uop->core);
}
return quant;
}
void evg_isa_write_task_commit(struct evg_work_item_t *work_item)
{
struct linked_list_t *task_list = work_item->write_task_list;
struct evg_wavefront_t *wavefront = work_item->wavefront;
struct evg_work_group_t *work_group = work_item->work_group;
struct evg_isa_write_task_t *wt;
struct evg_inst_t *inst;
/* Process first tasks of type:
* - EVG_ISA_WRITE_TASK_WRITE_DEST
* - EVG_ISA_WRITE_TASK_WRITE_LDS
*/
for (linked_list_head(task_list); !linked_list_is_end(task_list); )
{
/* Get task */
wt = linked_list_get(task_list);
assert(wt->work_item == work_item);
inst = wt->inst;
switch (wt->kind)
{
case EVG_ISA_WRITE_TASK_WRITE_DEST:
{
if (wt->write_mask)
evg_isa_write_gpr(work_item, wt->gpr, wt->rel, wt->chan, wt->value);
work_item->pv.elem[wt->inst->alu] = wt->value;
/* Debug */
if (evg_isa_debugging())
{
evg_isa_debug(" i%d:%s", work_item->id,
map_value(&evg_pv_map, wt->inst->alu));
if (wt->write_mask)
{
evg_isa_debug(",");
evg_inst_dump_gpr(wt->gpr, wt->rel, wt->chan, 0,
debug_file(evg_isa_debug_category));
}
evg_isa_debug("<=");
gpu_isa_dest_value_dump(inst, &wt->value,
debug_file(evg_isa_debug_category));
}
break;
}
case EVG_ISA_WRITE_TASK_WRITE_LDS:
{
struct mem_t *local_mem;
union evg_reg_t lds_value;
local_mem = work_group->local_mem;
assert(local_mem);
assert(wt->lds_value_size);
mem_write(local_mem, wt->lds_addr, wt->lds_value_size, &wt->lds_value);
/* Debug */
lds_value.as_uint = wt->lds_value;
evg_isa_debug(" i%d:LDS[0x%x]<=(%u,%gf) (%d bytes)", work_item->id, wt->lds_addr,
lds_value.as_uint, lds_value.as_float, (int) wt->lds_value_size);
break;
}
default:
linked_list_next(task_list);
continue;
}
/* Done with this task */
repos_free_object(evg_isa_write_task_repos, wt);
linked_list_remove(task_list);
}
/* Process PUSH_BEFORE, PRED_SET */
for (linked_list_head(task_list); !linked_list_is_end(task_list); )
{
/* Get task */
wt = linked_list_get(task_list);
inst = wt->inst;
/* Process */
switch (wt->kind)
{
case EVG_ISA_WRITE_TASK_PUSH_BEFORE:
{
if (!wavefront->push_before_done)
evg_wavefront_stack_push(wavefront);
wavefront->push_before_done = 1;
break;
}
case EVG_ISA_WRITE_TASK_SET_PRED:
{
int update_pred = EVG_ALU_WORD1_OP2.update_pred;
int update_exec_mask = EVG_ALU_WORD1_OP2.update_exec_mask;
assert(inst->info->fmt[1] == EVG_FMT_ALU_WORD1_OP2);
if (update_pred)
evg_work_item_set_pred(work_item, wt->cond);
if (update_exec_mask)
evg_work_item_set_active(work_item, wt->cond);
/* Debug */
if (debug_status(evg_isa_debug_category))
{
if (update_pred && update_exec_mask)
evg_isa_debug(" i%d:act/pred<=%d", work_item->id, wt->cond);
else if (update_pred)
evg_isa_debug(" i%d:pred=%d", work_item->id, wt->cond);
else if (update_exec_mask)
evg_isa_debug(" i%d:pred=%d", work_item->id, wt->cond);
}
break;
}
default:
abort();
}
/* Done with task */
repos_free_object(evg_isa_write_task_repos, wt);
linked_list_remove(task_list);
}
/* List should be empty */
assert(!linked_list_count(task_list));
}
static void evg_tex_engine_fetch(EvgComputeUnit *compute_unit)
{
EvgGpu *gpu = compute_unit->gpu;
struct linked_list_t *pending_queue = compute_unit->tex_engine.pending_queue;
struct linked_list_t *finished_queue = compute_unit->tex_engine.finished_queue;
EvgWavefront *wavefront;
struct evg_uop_t *cf_uop, *uop;
struct evg_work_item_uop_t *work_item_uop;
EvgInst *inst;
int inst_num;
EvgWorkItem *work_item;
int work_item_id;
char str[MAX_LONG_STRING_SIZE];
char str_trimmed[MAX_LONG_STRING_SIZE];
/* Get wavefront to fetch from */
linked_list_head(pending_queue);
cf_uop = linked_list_get(pending_queue);
if (!cf_uop)
return;
wavefront = cf_uop->wavefront;
assert(wavefront->clause_kind == EvgInstClauseTEX);
/* If fetch queue is full, cannot fetch until space is made */
if (compute_unit->tex_engine.fetch_queue_length >= evg_gpu_tex_engine_fetch_queue_size)
return;
/* Emulate instruction and create uop */
inst_num = (wavefront->clause_buf - wavefront->clause_buf_start) / 16;
EvgWavefrontExecute(wavefront);
inst = wavefront->tex_inst;
uop = evg_uop_create(compute_unit);
uop->wavefront = wavefront;
uop->work_group = wavefront->work_group;
uop->cf_uop = cf_uop;
uop->id_in_compute_unit = compute_unit->gpu_uop_id_counter++;
uop->last = wavefront->clause_kind != EvgInstClauseTEX;
uop->global_mem_read = wavefront->global_mem_read;
uop->global_mem_write = wavefront->global_mem_write;
uop->vliw_slots = 1;
/* If TEX clause finished, extract CF uop from 'pending_queue' and
* insert it into 'finished_queue'. */
if (uop->last)
{
linked_list_remove(pending_queue);
linked_list_add(finished_queue, cf_uop);
}
/* If instruction is a global memory read (should be), record addresses */
if (uop->global_mem_read)
{
assert((inst->info->flags & EvgInstFlagMemRead));
EVG_FOREACH_WORK_ITEM_IN_WAVEFRONT(wavefront, work_item_id)
{
work_item = gpu->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;
}
void evg_faults_init(void)
{
FILE *f;
char line[MAX_STRING_SIZE];
char *line_ptr;
struct evg_fault_t *fault;
int line_num;
long long last_cycle;
evg_fault_list = linked_list_create();
if (!*evg_faults_file_name)
return;
f = fopen(evg_faults_file_name, "rt");
if (!f)
fatal("%s: cannot open file", evg_faults_file_name);
line_num = 0;
last_cycle = 0;
while (!feof(f))
{
const char *delim = " ";
/* Read a line */
line_num++;
line_ptr = fgets(line, MAX_STRING_SIZE, f);
if (!line_ptr)
break;
/* Allocate new fault */
fault = calloc(1, sizeof(struct evg_fault_t));
if (!fault)
fatal("%s: out of memory", __FUNCTION__);
/* Read <cycle> field */
line_ptr = strtok(line_ptr, delim);
if (!line_ptr)
goto wrong_format;
fault->cycle = atoll(line_ptr);
if (fault->cycle < 1)
fatal("%s: line %d: lowest possible cycle is 1",
evg_faults_file_name, line_num);
if (fault->cycle < last_cycle)
fatal("%s: line %d: cycles must be ordered",
evg_faults_file_name, line_num);
/* <fault> - Type of fault */
line_ptr = strtok(NULL, delim);
if (!line_ptr)
goto wrong_format;
if (!strcmp(line_ptr, "ams"))
fault->type = evg_fault_ams;
else if (!strcmp(line_ptr, "reg"))
fault->type = evg_fault_reg;
else if (!strcmp(line_ptr, "mem"))
fault->type = evg_fault_mem;
else
fatal("%s: line %d: invalid value for <fault> ('%s')",
evg_faults_file_name, line_num, line_ptr);
/* <cu_id> - Compute unit */
line_ptr = strtok(NULL, delim);
if (!line_ptr)
goto wrong_format;
fault->compute_unit_id = atoi(line_ptr);
if (fault->compute_unit_id >= evg_gpu_num_compute_units || fault->compute_unit_id < 0)
fatal("%s: line %d: invalid compute unit ID",
evg_faults_file_name, line_num);
/* Analyze rest of the line depending on fault type */
switch (fault->type)
{
case evg_fault_ams:
/* <stack_id> - Stack ID */
line_ptr = strtok(NULL, delim);
if (!line_ptr)
goto wrong_format;
fault->stack_id = atoi(line_ptr);
if (fault->stack_id >= evg_gpu_max_wavefronts_per_compute_unit)
fatal("%s: line %d: invalid stack ID",
evg_faults_file_name, line_num);
/* <am_id> - Active mask ID */
line_ptr = strtok(NULL, delim);
if (!line_ptr)
goto wrong_format;
fault->active_mask_id = atoi(line_ptr);
if (fault->active_mask_id >= EVG_MAX_STACK_SIZE)
fatal("%s: line %d: invalid active mask ID",
evg_faults_file_name, line_num);
/* <bit> */
line_ptr = strtok(NULL, delim);
if (!line_ptr)
goto wrong_format;
fault->bit = atoi(line_ptr);
if (fault->bit >= evg_emu_wavefront_size)
fatal("%s: line %d: invalid bit index",
evg_faults_file_name, line_num);
/* No more tokens */
if (strtok(NULL, delim))
fatal("%s: line %d: too many arguments",
evg_faults_file_name, line_num);
break;
case evg_fault_reg:
/* <reg_id> - Register ID */
line_ptr = strtok(NULL, delim);
if (!line_ptr)
goto wrong_format;
fault->reg_id = atoi(line_ptr);
if (fault->reg_id >= evg_gpu_num_registers || fault->reg_id < 0)
fatal("%s: line %d: invalid compute unit ID",
evg_faults_file_name, line_num);
/* <bit> */
line_ptr = strtok(NULL, delim);
if (!line_ptr)
goto wrong_format;
fault->bit = atoi(line_ptr);
if (fault->bit < 0 || fault->bit >= 128)
fatal("%s: line %d: invalid bit index",
evg_faults_file_name, line_num);
break;
case evg_fault_mem:
/* <byte> - Byte position in local memory */
line_ptr = strtok(NULL, delim);
if (!line_ptr)
goto wrong_format;
fault->byte = atoi(line_ptr);
if (fault->byte >= evg_gpu_local_mem_size || fault->byte < 0)
fatal("%s: line %d: invalid byte position",
evg_faults_file_name, line_num);
/* <bit> - Bit position */
line_ptr = strtok(NULL, delim);
if (!line_ptr)
goto wrong_format;
fault->bit = atoi(line_ptr);
if (fault->bit > 7 || fault->bit < 0)
fatal("%s: line %d: invalid bit position",
evg_faults_file_name, line_num);
break;
}
/* Insert fault in fault list */
linked_list_out(evg_fault_list);
linked_list_insert(evg_fault_list, fault);
last_cycle = fault->cycle;
continue;
wrong_format:
fatal("%s: line %d: not enough arguments",
evg_faults_file_name, line_num);
}
linked_list_head(evg_fault_list);
}
void evg_faults_insert(void)
{
struct evg_fault_t *fault;
struct evg_compute_unit_t *compute_unit;
for (;;)
{
linked_list_head(evg_fault_list);
fault = linked_list_get(evg_fault_list);
if (!fault || fault->cycle > evg_gpu->cycle)
break;
/* Insert fault depending on fault type */
switch (fault->type)
{
case evg_fault_ams:
{
struct evg_work_group_t *work_group;
struct evg_wavefront_t *wavefront;
struct evg_work_item_t *work_item;
int work_group_id; /* in compute unit */
int wavefront_id; /* in compute unit */
int value;
/* Initial debug */
evg_faults_debug("fault clk=%lld cu=%d type=\"ams\" stack=%d am=%d bit=%d ",
evg_gpu->cycle,
fault->compute_unit_id, fault->stack_id,
fault->active_mask_id, fault->bit);
assert(fault->cycle == evg_gpu->cycle);
compute_unit = evg_gpu->compute_units[fault->compute_unit_id];
/* If compute unit is idle, dismiss */
if (!compute_unit->work_group_count)
{
evg_faults_debug("effect=\"cu_idle\"");
goto end_loop;
}
/* Get work-group and wavefront. If wavefront ID exceeds current number, dismiss */
work_group_id = fault->stack_id / evg_gpu->ndrange->wavefronts_per_work_group;
wavefront_id = fault->stack_id % evg_gpu->ndrange->wavefronts_per_work_group;
if (work_group_id >= evg_gpu_max_work_groups_per_compute_unit
|| !compute_unit->work_groups[work_group_id])
{
evg_faults_debug("effect=\"wf_idle\"");
goto end_loop;
}
work_group = compute_unit->work_groups[work_group_id];
wavefront = work_group->wavefronts[wavefront_id];
/* If active_mask_id exceeds stack top, dismiss */
if (fault->active_mask_id > wavefront->stack_top)
{
evg_faults_debug("effect=\"am_idle\"");
goto end_loop;
}
/* If 'bit' exceeds number of work-items in wavefront, dismiss */
if (fault->bit >= wavefront->work_item_count)
{
evg_faults_debug("effect=\"wi_idle\"");
goto end_loop;
}
/* Fault caused an error, show affected software entities */
work_item = wavefront->work_items[fault->bit];
evg_faults_debug("effect=\"error\" wg=%d wf=%d wi=%d",
work_group->id,
wavefront->id,
work_item->id);
/* Inject fault */
value = bit_map_get(wavefront->active_stack,
fault->active_mask_id * wavefront->work_item_count
+ fault->bit, 1);
bit_map_set(wavefront->active_stack,
fault->active_mask_id * wavefront->work_item_count
+ fault->bit, 1, !value);
evg_fault_errors++;
break;
}
case evg_fault_reg:
{
struct evg_opencl_kernel_t *kernel = evg_gpu->ndrange->kernel;
int work_group_id_in_compute_unit;
struct evg_work_group_t *work_group;
struct evg_wavefront_t *wavefront;
int num_registers_per_work_group;
int work_item_id_in_compute_unit;
int work_item_id_in_work_group;
struct evg_work_item_t *work_item;
struct linked_list_t *fetch_queue;
struct evg_uop_t *inst_buffer;
struct evg_uop_t *exec_buffer;
struct heap_t *event_queue;
struct evg_uop_t *uop;
int lo_reg;
/* Initial debug */
evg_faults_debug("fault clk=%lld cu=%d type=\"reg\" reg=%d bit=%d ",
evg_gpu->cycle,
fault->compute_unit_id,
fault->reg_id,
fault->bit);
assert(fault->cycle == evg_gpu->cycle);
compute_unit = evg_gpu->compute_units[fault->compute_unit_id];
/* If compute unit is idle, dismiss */
if (!compute_unit->work_group_count)
{
evg_faults_debug("effect=\"cu_idle\"");
goto end_loop;
}
/* Get work-group */
num_registers_per_work_group = kernel->bin_file->enc_dict_entry_evergreen->num_gpr_used
* kernel->local_size;
work_group_id_in_compute_unit = fault->reg_id / num_registers_per_work_group;
if (work_group_id_in_compute_unit >= evg_gpu_max_work_groups_per_compute_unit)
{
evg_faults_debug("effect=\"reg_idle\"");
goto end_loop;
}
/* Get work-group (again) */
work_group = compute_unit->work_groups[work_group_id_in_compute_unit];
if (!work_group)
{
evg_faults_debug("effect=\"reg_idle\"");
goto end_loop;
}
/* Get affected entities */
work_item_id_in_compute_unit = fault->reg_id
/ kernel->bin_file->enc_dict_entry_evergreen->num_gpr_used;
work_item_id_in_work_group = work_item_id_in_compute_unit % kernel->local_size;
work_item = work_group->work_items[work_item_id_in_work_group];
wavefront = work_item->wavefront;
lo_reg = fault->reg_id % kernel->bin_file->enc_dict_entry_evergreen->num_gpr_used;
/* Fault falling between Fetch and Read stage of an instruction
* consuming register. This case cannot be modeled due to functional
* simulation skew. */
fetch_queue = compute_unit->alu_engine.fetch_queue;
inst_buffer = compute_unit->alu_engine.inst_buffer;
for (linked_list_head(fetch_queue); !linked_list_is_end(fetch_queue);
linked_list_next(fetch_queue))
{
uop = linked_list_get(fetch_queue);
if (evg_stack_faults_is_idep(uop, wavefront, lo_reg))
{
evg_faults_debug("effect=\"reg_read\"");
goto end_loop;
}
}
uop = inst_buffer;
if (uop && evg_stack_faults_is_idep(uop, wavefront, lo_reg))
{
evg_faults_debug("effect=\"reg_read\"");
goto end_loop;
}
/* Fault falling between Fetch and Write stage of an instruction
* writing on the register. The instruction will overwrite the fault,
* so this shouldn't cause its injection. */
exec_buffer = compute_unit->alu_engine.exec_buffer;
for (linked_list_head(fetch_queue); !linked_list_is_end(fetch_queue);
linked_list_next(fetch_queue))
{
uop = linked_list_get(fetch_queue);
if (evg_stack_faults_is_odep(uop, wavefront, lo_reg))
{
evg_faults_debug("effect=\"reg_write\"");
goto end_loop;
}
}
uop = inst_buffer;
if (uop && evg_stack_faults_is_odep(uop, wavefront, lo_reg))
{
evg_faults_debug("effect=\"reg_write\"");
goto end_loop;
}
uop = exec_buffer;
if (uop && evg_stack_faults_is_odep(uop, wavefront, lo_reg))
{
evg_faults_debug("effect=\"reg_write\"");
goto end_loop;
}
event_queue = compute_unit->alu_engine.event_queue;
for (heap_first(event_queue, (void **) &uop); uop;
heap_next(event_queue, (void **) &uop))
{
if (evg_stack_faults_is_odep(uop, wavefront, lo_reg))
{
evg_faults_debug("effect=\"reg_write\"");
goto end_loop;
}
}
/* Fault caused error */
evg_faults_debug("effect=\"error\" ");
evg_faults_debug("wg=%d wf=%d wi=%d lo_reg=%d ",
work_group->id, work_item->wavefront->id, work_item->id, lo_reg);
/* Insert the fault */
if (fault->bit < 32)
work_item->gpr[lo_reg].elem[0] ^= 1 << fault->bit;
else if (fault->bit < 64)
work_item->gpr[lo_reg].elem[1] ^= 1 << (fault->bit - 32);
else if (fault->bit < 96)
work_item->gpr[lo_reg].elem[2] ^= 1 << (fault->bit - 64);
else
work_item->gpr[lo_reg].elem[3] ^= 1 << (fault->bit - 96);
evg_fault_errors++;
break;
}
case evg_fault_mem:
{
struct evg_work_group_t *work_group;
int work_group_id_in_compute_unit;
unsigned char value;
/* Initial debug */
evg_faults_debug("fault clk=%lld cu=%d type=\"mem\" byte=%d bit=%d ",
evg_gpu->cycle,
fault->compute_unit_id,
fault->byte,
fault->bit);
assert(fault->cycle == evg_gpu->cycle);
compute_unit = evg_gpu->compute_units[fault->compute_unit_id];
/* If compute unit is idle, dismiss */
if (!compute_unit->work_group_count)
{
evg_faults_debug("effect=\"cu_idle\"");
goto end_loop;
}
/* Check if there is any local memory used at all */
if (!evg_gpu->ndrange->local_mem_top)
{
evg_faults_debug("effect=\"mem_idle\"");
goto end_loop;
}
/* Get work-group */
work_group_id_in_compute_unit = fault->byte / evg_gpu->ndrange->local_mem_top;
if (work_group_id_in_compute_unit >= evg_gpu_max_work_groups_per_compute_unit)
{
evg_faults_debug("effect=\"mem_idle\"");
goto end_loop;
}
/* Get work-group (again) */
work_group = compute_unit->work_groups[work_group_id_in_compute_unit];
if (!work_group)
{
evg_faults_debug("effect=\"mem_idle\"");
goto end_loop;
}
/* Inject fault */
evg_faults_debug("effect=\"error\" wg=%d ",
work_group->id);
mem_read(work_group->local_mem, fault->byte, 1, &value);
value ^= 1 << fault->bit;
mem_write(work_group->local_mem, fault->byte, 1, &value);
evg_fault_errors++;
break;
}
default:
panic("invalid fault type");
}
end_loop:
/* Extract and free */
free(fault);
linked_list_remove(evg_fault_list);
evg_faults_debug("\n");
/* If all faults were inserted and no error was caused, end simulation */
if (!linked_list_count(evg_fault_list) && !evg_fault_errors)
esim_finish = esim_finish_evg_no_faults;
}
}
