void mips_sys_call(struct mips_ctx_t *ctx)
{
struct mips_regs_t *regs = ctx->regs;
int code;
int err;
/* System call code */
code = regs->regs_R[2] - __NR_Linux;
if (code < 1 || code >= mips_sys_code_count)
fatal("%s: invalid system call code (%d)", __FUNCTION__, code);
/* Statistics */
mips_sys_call_freq[code]++;
/* Debug */
mips_sys_debug("'%s' (code %d, inst %lld, pid %d)\n",
mips_sys_call_name[code], code, asEmu(mips_emu)->instructions, ctx->pid);
mips_isa_call_debug("system call '%s' (code %d, inst %lld, pid %d)\n",
mips_sys_call_name[code], code, asEmu(mips_emu)->instructions, ctx->pid);
/* Perform system call */
err = mips_sys_call_func[code](ctx);
/* Set return value in 'eax', except for 'sigreturn' system call. Also, if the
* context got suspended, the wake up routine will set the return value. */
if (code != mips_sys_code_sigreturn && !mips_ctx_get_status(ctx, mips_ctx_suspended))
regs->regs_R[2] = err;
/* Debug */
mips_sys_debug(" ret=(%d, 0x%x)", err, err);
if (err < 0 && err >= -SIM_ERRNO_MAX)
mips_sys_debug(", errno=%s)", str_map_value(&mips_sys_error_code_map, -err));
mips_sys_debug("\n");
}
void MIPSEmuCreate(MIPSEmu *self)
{
/* Parent */
EmuCreate(asEmu(self), "MIPS");
/* Initialize */
self->current_pid = 100;
pthread_mutex_init(&self->process_events_mutex, NULL);
/* Virtual functions */
asObject(self)->Dump = MIPSEmuDump;
asEmu(self)->DumpSummary = MIPSEmuDumpSummary;
asEmu(self)->Run = MIPSEmuRun;
}
/* Run fast-forward simulation */
void X86CpuFastForward(X86Cpu *self) {
X86Emu *emu = self->emu;
/* Fast-forward simulation. Run 'x86_cpu_fast_forward' iterations of the x86
* emulation loop until any simulation end reason is detected. */
while (asEmu(emu)->instructions < x86_cpu_fast_forward_count && !esim_finish)
X86EmuRun(asEmu(emu));
/* Record number of instructions in fast-forward execution. */
self->num_fast_forward_inst = asEmu(emu)->instructions;
/* Output warning if simulation finished during fast-forward execution. */
if (esim_finish)
warning("x86 fast-forwarding finished simulation.\n%s",
x86_cpu_err_fast_forward);
}
void KplWarpExecute(KplWarp *self)
{
KplEmu *emu;
KplGrid *grid;
KplThreadBlock *thread_block;
KplThread *thread;
struct KplInstWrap *inst;
KplInstBytes inst_bytes;
KplInstOpcode inst_op;
int thread_id;
/* Get current arch, grid, and thread-block */
thread_block = self->thread_block;
grid = thread_block->grid;
emu = grid->emu;
/* Get instruction */
inst_bytes.as_uint[0] = self->inst_buffer[self->pc / self->inst_size] >> 32;
inst_bytes.as_uint[1] = self->inst_buffer[self->pc / self->inst_size];
kpl_isa_debug("%s:%d: warp[%d] executes instruction [0x%x] 0x%016llx\n",
__FILE__, __LINE__, self->id, self->pc, inst_bytes.as_dword);
/* Decode instruction */
inst = self->inst;
KplInstWrapDecode(inst, self->pc, &inst_bytes);
/* Execute instruction */
inst_op = KplInstWrapGetOpcode(inst);
if (!inst_op)
fatal("%s:%d: unrecognized instruction (%08x %08x)",
__FILE__, __LINE__, inst_bytes.as_uint[0], inst_bytes.as_uint[1]);
for (thread_id = 0; thread_id < self->thread_count; ++thread_id)
{
thread = self->threads[thread_id];
emu->inst_func[inst_op](thread, inst);
}
/* Finish */
if (self->finished)
{
assert(list_index_of(thread_block->running_warps, self) != -1);
assert(list_index_of(thread_block->finished_warps, self) == -1);
list_remove(thread_block->running_warps, self);
list_add(thread_block->finished_warps, self);
return;
}
/* Update PC */
/* if (KplInstWrapGetCategory(inst) != KplInstCategoryCtrl)
self->pc += self->inst_size;
else
self->pc = self->target_pc;
*/
/* Stats */
asEmu(emu)->instructions++;
self->inst_count++;
}
int MIPSEmuRun(Emu *self)
{
MIPSEmu *emu = asMIPSEmu(self);
struct mips_ctx_t *ctx;
/* Stop if there is no context running */
if (emu->finished_list_count >= emu->context_list_count)
return FALSE;
/* Stop if maximum number of CPU instructions exceeded */
if (mips_emu_max_inst && asEmu(mips_emu)->instructions >= mips_emu_max_inst)
esim_finish = esim_finish_mips_max_inst;
/* Stop if any previous reason met */
if (esim_finish)
return TRUE;
/* Run an instruction from every running process */
for (ctx = emu->running_list_head; ctx; ctx = ctx->running_list_next)
mips_ctx_execute(ctx);
/* Free finished contexts */
while (emu->finished_list_head)
mips_ctx_free(emu->finished_list_head);
/* Still running */
return TRUE;
}
void SIEmuCreate(SIEmu *self)
{
/* Parent */
EmuCreate(asEmu(self), "SouthernIslands");
/* Initialize */
self->video_mem = mem_create();
self->video_mem->safe = 0;
self->video_mem_top = 0;
self->waiting_work_groups = list_create();
self->running_work_groups = list_create();
/* Set global memory to video memory by default */
self->global_mem = self->video_mem;
/* Virtual functions */
asObject(self)->Dump = SIEmuDump;
asEmu(self)->DumpSummary = SIEmuDumpSummary;
asEmu(self)->Run = SIEmuRun;
}
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 FrmEmuCreate(FrmEmu *self)
{
/* Parent */
EmuCreate(asEmu(self), "Fermi");
/* Initialize */
self->grids = list_create();
self->pending_grids = list_create();
self->running_grids = list_create();
self->finished_grids = list_create();
self->global_mem = mem_create();
self->global_mem->safe = 0;
self->global_mem_top = 0;
self->total_global_mem_size = 1 << 31; /* 2GB */
self->free_global_mem_size = 1 << 31; /* 2GB */
self->const_mem = mem_create();
self->const_mem->safe = 0;
/* Virtual functions */
asObject(self)->Dump = FrmEmuDump;
asEmu(self)->DumpSummary = FrmEmuDumpSummary;
asEmu(self)->Run = FrmEmuRun;
}
void X86ContextExecute(X86Context *self)
{
X86Emu *emu = self->emu;
struct x86_regs_t *regs = self->regs;
struct mem_t *mem = self->mem;
unsigned char buffer[20];
unsigned char *buffer_ptr;
int spec_mode;
/* Memory permissions should not be checked if the context is executing in
* speculative mode. This will prevent guest segmentation faults to occur. */
spec_mode = X86ContextGetState(self, X86ContextSpecMode);
mem->safe = spec_mode ? 0 : mem_safe_mode;
/* Read instruction from memory. Memory should be accessed here in unsafe mode
* (i.e., allowing segmentation faults) if executing speculatively. */
buffer_ptr = mem_get_buffer(mem, regs->eip, 20, mem_access_exec);
if (!buffer_ptr)
{
/* Disable safe mode. If a part of the 20 read bytes does not belong to the
* actual instruction, and they lie on a page with no permissions, this would
* generate an undesired protection fault. */
mem->safe = 0;
buffer_ptr = buffer;
mem_access(mem, regs->eip, 20, buffer_ptr, mem_access_exec);
}
mem->safe = mem_safe_mode;
/* Disassemble */
X86InstDecode(&self->inst, regs->eip, buffer_ptr);
if (self->inst.opcode == X86InstOpcodeInvalid && !spec_mode)
fatal("0x%x: not supported x86 instruction (%02x %02x %02x %02x...)",
regs->eip, buffer_ptr[0], buffer_ptr[1], buffer_ptr[2], buffer_ptr[3]);
/* Stop if instruction matches last instruction bytes */
if (x86_emu_last_inst_size &&
x86_emu_last_inst_size == self->inst.size &&
!memcmp(x86_emu_last_inst_bytes, buffer_ptr, x86_emu_last_inst_size))
esim_finish = esim_finish_x86_last_inst;
/* Execute instruction */
X86ContextExecuteInst(self);
/* Statistics */
asEmu(emu)->instructions++;
}
int X86CpuRun(Timing *self) {
X86Cpu *cpu = asX86Cpu(self);
X86Emu *emu = cpu->emu;
/* Stop if no context is running */
if (emu->finished_list_count >= emu->context_list_count) return FALSE;
/* Fast-forward simulation */
if (x86_cpu_fast_forward_count &&
asEmu(emu)->instructions < x86_cpu_fast_forward_count)
X86CpuFastForward(cpu);
/* Stop if maximum number of CPU instructions exceeded */
if (x86_emu_max_inst &&
cpu->num_committed_inst >= x86_emu_max_inst - x86_cpu_fast_forward_count)
esim_finish = esim_finish_x86_max_inst;
/* Stop if maximum number of cycles exceeded */
if (x86_emu_max_cycles && self->cycle >= x86_emu_max_cycles)
esim_finish = esim_finish_x86_max_cycles;
/* Stop if any previous reason met */
if (esim_finish) return TRUE;
/* One more cycle of x86 timing simulation */
self->cycle++;
/* Empty uop trace list. This dumps the last trace line for instructions
* that were freed in the previous simulation cycle. */
X86CpuEmptyTraceList(cpu);
/* Processor stages */
X86CpuRunStages(cpu);
/* Process host threads generating events */
X86EmuProcessEvents(emu);
/* Still simulating */
return TRUE;
}
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);
}
/* FIXME - merge with ctx_execute */
void mips_isa_execute_inst(struct mips_ctx_t *ctx) {
// struct mips_regs_t *regs = ctx->regs;
ctx->next_ip = ctx->n_next_ip;
ctx->n_next_ip += 4;
/* Debug */
if (debug_status(mips_isa_inst_debug_category)) {
mips_isa_inst_debug("%d %8lld %x: ", ctx->pid,
asEmu(mips_emu)->instructions, ctx->regs->pc);
mips_inst_debug_dump(&ctx->inst, debug_file(mips_isa_inst_debug_category));
}
/* Call instruction emulation function */
// regs->pc = regs->pc + ctx->inst.info->size;
if (ctx->inst.info->opcode) mips_isa_inst_func[ctx->inst.info->opcode](ctx);
/* Statistics */
mips_inst_freq[ctx->inst.info->opcode]++;
/* Debug */
mips_isa_inst_debug("\n");
// if (debug_status(mips_isa_call_debug_category))
// mips_isa_debug_call(ctx);
}
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);
}
void X86CpuDumpReport(X86Cpu *self, FILE *f) {
X86Emu *emu = self->emu;
X86Core *core;
X86Thread *thread;
long long now;
int i;
int j;
/* Get CPU timer value */
now = m2s_timer_get_value(asEmu(emu)->timer);
/* Dump CPU configuration */
fprintf(f, ";\n; CPU Configuration\n;\n\n");
X86DumpCpuConfig(f);
/* Report for the complete processor */
fprintf(f, ";\n; Simulation Statistics\n;\n\n");
fprintf(f, "; Global statistics\n");
fprintf(f, "[ Global ]\n\n");
fprintf(f, "Cycles = %lld\n", asTiming(self)->cycle);
fprintf(f, "Time = %.2f\n", (double)now / 1000000);
fprintf(f, "CyclesPerSecond = %.0f\n",
now ? (double)asTiming(self)->cycle / now * 1000000 : 0.0);
fprintf(f, "MemoryUsed = %lu\n", (long)mem_mapped_space);
fprintf(f, "MemoryUsedMax = %lu\n", (long)mem_max_mapped_space);
fprintf(f, "\n");
/* Dispatch stage */
fprintf(f, "; Dispatch stage\n");
X86CpuDumpUopReport(self, f, self->num_dispatched_uinst_array, "Dispatch",
x86_cpu_dispatch_width);
/* Issue stage */
fprintf(f, "; Issue stage\n");
X86CpuDumpUopReport(self, f, self->num_issued_uinst_array, "Issue",
x86_cpu_issue_width);
/* Commit stage */
fprintf(f, "; Commit stage\n");
X86CpuDumpUopReport(self, f, self->num_committed_uinst_array, "Commit",
x86_cpu_commit_width);
/* Committed branches */
fprintf(f, "; Committed branches\n");
fprintf(f, "; Branches - Number of committed control uops\n");
fprintf(
f, "; Squashed - Number of mispredicted uops squashed from the ROB\n");
fprintf(
f,
"; Mispred - Number of mispredicted branches in the correct path\n");
fprintf(f, "; PredAcc - Prediction accuracy\n");
fprintf(f, "Commit.Branches = %lld\n", self->num_branch_uinst);
fprintf(f, "Commit.Squashed = %lld\n", self->num_squashed_uinst);
fprintf(f, "Commit.Mispred = %lld\n", self->num_mispred_branch_uinst);
fprintf(
f, "Commit.PredAcc = %.4g\n",
self->num_branch_uinst
? (double)(self->num_branch_uinst - self->num_mispred_branch_uinst) /
self->num_branch_uinst
: 0.0);
fprintf(f, "\n");
/* Report for each core */
for (i = 0; i < x86_cpu_num_cores; i++) {
/* Core */
core = self->cores[i];
fprintf(f, "\n; Statistics for core %d\n", core->id);
fprintf(f, "[ c%d ]\n\n", core->id);
/* Functional units */
X86CoreDumpFunctionalUnitsReport(core, f);
/* Dispatch slots */
if (x86_cpu_dispatch_kind == x86_cpu_dispatch_kind_timeslice) {
fprintf(f, "; Dispatch slots usage (sum = cycles * dispatch width)\n");
fprintf(f, "; used - dispatch slot was used by a non-spec uop\n");
fprintf(f, "; spec - used by a mispeculated uop\n");
fprintf(f, "; ctx - no context allocated to thread\n");
fprintf(f, "; uopq,rob,iq,lsq,rename - no space in structure\n");
DUMP_DISPATCH_STAT(used);
DUMP_DISPATCH_STAT(spec);
DUMP_DISPATCH_STAT(uop_queue);
DUMP_DISPATCH_STAT(rob);
DUMP_DISPATCH_STAT(iq);
DUMP_DISPATCH_STAT(lsq);
DUMP_DISPATCH_STAT(rename);
DUMP_DISPATCH_STAT(ctx);
fprintf(f, "\n");
}
/* Dispatch stage */
fprintf(f, "; Dispatch stage\n");
X86CpuDumpUopReport(self, f, core->num_dispatched_uinst_array, "Dispatch",
x86_cpu_dispatch_width);
/* Issue stage */
fprintf(f, "; Issue stage\n");
X86CpuDumpUopReport(self, f, core->num_issued_uinst_array, "Issue",
x86_cpu_issue_width);
/* Commit stage */
fprintf(f, "; Commit stage\n");
X86CpuDumpUopReport(self, f, core->num_committed_uinst_array, "Commit",
x86_cpu_commit_width);
/* Committed branches */
fprintf(f, "; Committed branches\n");
fprintf(f, "Commit.Branches = %lld\n", core->num_branch_uinst);
fprintf(f, "Commit.Squashed = %lld\n", core->num_squashed_uinst);
fprintf(f, "Commit.Mispred = %lld\n", core->num_mispred_branch_uinst);
fprintf(f, "Commit.PredAcc = %.4g\n",
core->num_branch_uinst
? (double)(core->num_branch_uinst -
core->num_mispred_branch_uinst) /
core->num_branch_uinst
: 0.0);
fprintf(f, "\n");
/* Occupancy stats */
fprintf(f, "; Structure statistics (reorder buffer, instruction queue,\n");
fprintf(f,
"; load-store queue, and integer/floating-point register file)\n");
fprintf(f, "; Size - Available size\n");
fprintf(f, "; Occupancy - Average number of occupied entries\n");
fprintf(f, "; Full - Number of cycles when the structure was full\n");
fprintf(f, "; Reads, Writes - Accesses to the structure\n");
if (x86_rob_kind == x86_rob_kind_shared) DUMP_CORE_STRUCT_STATS(ROB, rob);
if (x86_iq_kind == x86_iq_kind_shared) {
DUMP_CORE_STRUCT_STATS(IQ, iq);
fprintf(f, "IQ.WakeupAccesses = %lld\n", core->iq_wakeup_accesses);
}
if (x86_lsq_kind == x86_lsq_kind_shared) DUMP_CORE_STRUCT_STATS(LSQ, lsq);
if (x86_reg_file_kind == x86_reg_file_kind_shared) {
DUMP_CORE_STRUCT_STATS(RF_Int, reg_file_int);
DUMP_CORE_STRUCT_STATS(RF_Fp, reg_file_fp);
}
fprintf(f, "\n");
/* Report for each thread */
for (j = 0; j < x86_cpu_num_threads; j++) {
thread = core->threads[j];
fprintf(f, "\n; Statistics for core %d - thread %d\n", core->id,
thread->id_in_core);
fprintf(f, "[ %s ]\n\n", thread->name);
/* Dispatch stage */
fprintf(f, "; Dispatch stage\n");
X86CpuDumpUopReport(self, f, thread->num_dispatched_uinst_array,
"Dispatch", x86_cpu_dispatch_width);
/* Issue stage */
fprintf(f, "; Issue stage\n");
X86CpuDumpUopReport(self, f, thread->num_issued_uinst_array, "Issue",
x86_cpu_issue_width);
/* Commit stage */
fprintf(f, "; Commit stage\n");
X86CpuDumpUopReport(self, f, thread->num_committed_uinst_array, "Commit",
x86_cpu_commit_width);
/* Committed branches */
fprintf(f, "; Committed branches\n");
fprintf(f, "Commit.Branches = %lld\n", thread->num_branch_uinst);
fprintf(f, "Commit.Squashed = %lld\n", thread->num_squashed_uinst);
fprintf(f, "Commit.Mispred = %lld\n", thread->num_mispred_branch_uinst);
fprintf(f, "Commit.PredAcc = %.4g\n",
thread->num_branch_uinst
? (double)(thread->num_branch_uinst -
thread->num_mispred_branch_uinst) /
thread->num_branch_uinst
: 0.0);
fprintf(f, "\n");
/* Occupancy stats */
fprintf(f,
"; Structure statistics (reorder buffer, instruction queue, "
"load-store queue,\n");
fprintf(f,
"; integer/floating-point register file, and renaming table)\n");
if (x86_rob_kind == x86_rob_kind_private)
DUMP_THREAD_STRUCT_STATS(ROB, rob);
if (x86_iq_kind == x86_iq_kind_private) {
DUMP_THREAD_STRUCT_STATS(IQ, iq);
fprintf(f, "IQ.WakeupAccesses = %lld\n", thread->iq_wakeup_accesses);
}
if (x86_lsq_kind == x86_lsq_kind_private)
DUMP_THREAD_STRUCT_STATS(LSQ, lsq);
if (x86_reg_file_kind == x86_reg_file_kind_private) {
DUMP_THREAD_STRUCT_STATS(RF_Int, reg_file_int);
DUMP_THREAD_STRUCT_STATS(RF_Fp, reg_file_fp);
}
fprintf(f, "RAT.IntReads = %lld\n", thread->rat_int_reads);
fprintf(f, "RAT.IntWrites = %lld\n", thread->rat_int_writes);
fprintf(f, "RAT.FpReads = %lld\n", thread->rat_fp_reads);
fprintf(f, "RAT.FpWrites = %lld\n", thread->rat_fp_writes);
fprintf(f, "BTB.Reads = %lld\n", thread->btb_reads);
fprintf(f, "BTB.Writes = %lld\n", thread->btb_writes);
fprintf(f, "\n");
/* Trace cache stats */
if (thread->trace_cache) X86ThreadDumpTraceCacheReport(thread, f);
}
}
}
int EvgGpuRun(Timing *self)
{
EvgGpu *gpu = asEvgGpu(self);
struct evg_ndrange_t *ndrange;
struct evg_compute_unit_t *compute_unit;
struct evg_compute_unit_t *compute_unit_next;
/* For efficiency when no Evergreen emulation is selected, exit here
* if the list of existing ND-Ranges is empty. */
if (!evg_emu->ndrange_list_count)
return FALSE;
/* Start one ND-Range in state 'pending' */
while ((ndrange = evg_emu->pending_ndrange_list_head))
{
/* Currently not supported for more than 1 ND-Range */
if (gpu->ndrange)
fatal("%s: Evergreen GPU timing simulation not supported for multiple ND-Ranges",
__FUNCTION__);
/* Set ND-Range status to 'running' */
evg_ndrange_clear_status(ndrange, evg_ndrange_pending);
evg_ndrange_set_status(ndrange, evg_ndrange_running);
/* Trace */
evg_trace("evg.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 */
evg_gpu_map_ndrange(ndrange);
evg_calc_plot();
}
/* Mapped ND-Range */
ndrange = gpu->ndrange;
assert(ndrange);
/* Allocate work-groups to compute units */
while (gpu->ready_list_head && ndrange->pending_list_head)
evg_compute_unit_map_work_group(gpu->ready_list_head,
ndrange->pending_list_head);
/* One more cycle */
asTiming(evg_gpu)->cycle++;
/* Stop if maximum number of GPU cycles exceeded */
if (evg_emu_max_cycles && asTiming(evg_gpu)->cycle >= evg_emu_max_cycles)
esim_finish = esim_finish_evg_max_cycles;
/* Stop if maximum number of GPU instructions exceeded */
if (evg_emu_max_inst && asEmu(evg_emu)->instructions >= evg_emu_max_inst)
esim_finish = esim_finish_evg_max_inst;
/* Stop if there was a simulation stall */
if (asTiming(evg_gpu)->cycle - gpu->last_complete_cycle > 1000000)
{
warning("Evergreen GPU simulation stalled.\n%s",
evg_err_stall);
esim_finish = esim_finish_stall;
}
/* Stop if any reason met */
if (esim_finish)
return TRUE;
/* Free instructions in trash */
evg_gpu_uop_trash_empty();
/* Run one loop iteration on each busy compute unit */
for (compute_unit = gpu->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->busy_list_next;
/* Run one cycle */
evg_compute_unit_run(compute_unit);
}
/* GPU-REL: insert stack faults */
evg_faults_insert();
/* If ND-Range finished execution in all compute units, free it. */
if (!gpu->busy_list_count)
{
/* Dump ND-Range report */
evg_ndrange_dump(ndrange, evg_emu_report_file);
/* Stop if maximum number of kernels reached */
if (evg_emu_max_kernels && evg_emu->ndrange_count >= evg_emu_max_kernels)
esim_finish = esim_finish_evg_max_kernels;
/* Finalize and free ND-Range */
assert(evg_ndrange_get_status(ndrange, evg_ndrange_finished));
evg_gpu_uop_trash_empty();
evg_gpu_unmap_ndrange();
evg_ndrange_free(ndrange);
}
/* Still simulating */
return TRUE;
}
void evg_gpu_dump_report(void)
{
struct evg_compute_unit_t *compute_unit;
struct mod_t *local_mod;
int compute_unit_id;
FILE *f;
double inst_per_cycle;
double cf_inst_per_cycle;
double alu_inst_per_cycle;
double tex_inst_per_cycle;
long long coalesced_reads;
long long coalesced_writes;
char vliw_occupancy[MAX_STRING_SIZE];
/* Open file */
f = file_open_for_write(evg_gpu_report_file_name);
if (!f)
return;
/* Dump GPU configuration */
fprintf(f, ";\n; GPU Configuration\n;\n\n");
evg_config_dump(f);
/* Report for device */
fprintf(f, ";\n; Simulation Statistics\n;\n\n");
inst_per_cycle = asTiming(evg_gpu)->cycle ? (double) asEmu(evg_emu)->instructions
/ asTiming(evg_gpu)->cycle : 0.0;
fprintf(f, "[ Device ]\n\n");
fprintf(f, "NDRangeCount = %d\n", evg_emu->ndrange_count);
fprintf(f, "Instructions = %lld\n", asEmu(evg_emu)->instructions);
fprintf(f, "Cycles = %lld\n", asTiming(evg_gpu)->cycle);
fprintf(f, "InstructionsPerCycle = %.4g\n", inst_per_cycle);
fprintf(f, "\n\n");
/* Report for compute units */
EVG_GPU_FOREACH_COMPUTE_UNIT(compute_unit_id)
{
compute_unit = evg_gpu->compute_units[compute_unit_id];
local_mod = compute_unit->local_memory;
inst_per_cycle = compute_unit->cycle ? (double) compute_unit->inst_count
/ compute_unit->cycle : 0.0;
cf_inst_per_cycle = compute_unit->cycle ? (double) compute_unit->cf_engine.inst_count
/ compute_unit->cycle : 0.0;
alu_inst_per_cycle = compute_unit->alu_engine.cycle ? (double) compute_unit->alu_engine.inst_count
/ compute_unit->alu_engine.cycle : 0.0;
tex_inst_per_cycle = compute_unit->tex_engine.cycle ? (double) compute_unit->tex_engine.inst_count
/ compute_unit->tex_engine.cycle : 0.0;
coalesced_reads = local_mod->reads - local_mod->effective_reads;
coalesced_writes = local_mod->writes - local_mod->effective_writes;
snprintf(vliw_occupancy, MAX_STRING_SIZE, "%lld %lld %lld %lld %lld",
compute_unit->alu_engine.vliw_slots[0],
compute_unit->alu_engine.vliw_slots[1],
compute_unit->alu_engine.vliw_slots[2],
compute_unit->alu_engine.vliw_slots[3],
compute_unit->alu_engine.vliw_slots[4]);
fprintf(f, "[ ComputeUnit %d ]\n\n", compute_unit_id);
fprintf(f, "WorkGroupCount = %lld\n", compute_unit->mapped_work_groups);
fprintf(f, "Instructions = %lld\n", compute_unit->inst_count);
fprintf(f, "Cycles = %lld\n", compute_unit->cycle);
fprintf(f, "InstructionsPerCycle = %.4g\n", inst_per_cycle);
fprintf(f, "\n");
fprintf(f, "CFEngine.Instructions = %lld\n", compute_unit->cf_engine.inst_count);
fprintf(f, "CFEngine.InstructionsPerCycle = %.4g\n", cf_inst_per_cycle);
fprintf(f, "CFEngine.ALUClauseTriggers = %lld\n", compute_unit->cf_engine.alu_clause_trigger_count);
fprintf(f, "CFEngine.TEXClauseTriggers = %lld\n", compute_unit->cf_engine.tex_clause_trigger_count);
fprintf(f, "CFEngine.GlobalMemWrites = %lld\n", compute_unit->cf_engine.global_mem_write_count);
fprintf(f, "\n");
fprintf(f, "ALUEngine.WavefrontCount = %lld\n", compute_unit->alu_engine.wavefront_count);
fprintf(f, "ALUEngine.Instructions = %lld\n", compute_unit->alu_engine.inst_count);
fprintf(f, "ALUEngine.InstructionSlots = %lld\n", compute_unit->alu_engine.inst_slot_count);
fprintf(f, "ALUEngine.LocalMemorySlots = %lld\n", compute_unit->alu_engine.local_mem_slot_count);
fprintf(f, "ALUEngine.VLIWOccupancy = %s\n", vliw_occupancy);
fprintf(f, "ALUEngine.Cycles = %lld\n", compute_unit->alu_engine.cycle);
fprintf(f, "ALUEngine.InstructionsPerCycle = %.4g\n", alu_inst_per_cycle);
fprintf(f, "\n");
fprintf(f, "TEXEngine.WavefrontCount = %lld\n", compute_unit->tex_engine.wavefront_count);
fprintf(f, "TEXEngine.Instructions = %lld\n", compute_unit->tex_engine.inst_count);
fprintf(f, "TEXEngine.Cycles = %lld\n", compute_unit->tex_engine.cycle);
fprintf(f, "TEXEngine.InstructionsPerCycle = %.4g\n", tex_inst_per_cycle);
fprintf(f, "\n");
fprintf(f, "LocalMemory.Accesses = %lld\n", local_mod->reads + local_mod->writes);
fprintf(f, "LocalMemory.Reads = %lld\n", local_mod->reads);
fprintf(f, "LocalMemory.EffectiveReads = %lld\n", local_mod->effective_reads);
fprintf(f, "LocalMemory.CoalescedReads = %lld\n", coalesced_reads);
fprintf(f, "LocalMemory.Writes = %lld\n", local_mod->writes);
fprintf(f, "LocalMemory.EffectiveWrites = %lld\n", local_mod->effective_writes);
fprintf(f, "LocalMemory.CoalescedWrites = %lld\n", coalesced_writes);
fprintf(f, "\n\n");
}
}