derek/gem5
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

gpu-compute, arch-gcn3: Change how waitcnts are implemented

Browse Source 
Use single counters per memory operation type and increment
them upon issue, not execute.

Change-Id: I6afc0b66b21882538ef90a14a57a3ab3cc7bd6f3
Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/29973
Reviewed-by: Anthony Gutierrez <anthony.gutierrez@amd.com>
Maintainer: Anthony Gutierrez <anthony.gutierrez@amd.com>
Tested-by: kokoro <noreply+kokoro@google.com>
This commit is contained in:
Tony Gutierrez
2018-08-10 15:01:12 -04:00
committed by Anthony Gutierrez
parent 63c76448eb
commit 4d737462c2
8 changed files with 106 additions and 18 deletions
Download Patch File Download Diff File Expand all files Collapse all files

36
src/arch/gcn3/insts/instructions.cc
View File

@@ -32565,6 +32565,7 @@ namespace Gcn3ISA
vdst.write();
wf->decLGKMInstsIssued();
wf->rdLmReqsInPipe--;
wf->validateRequestCounters();
} // execute
@@ -32635,6 +32636,7 @@ namespace Gcn3ISA
vdst.write();
wf->decLGKMInstsIssued();
wf->rdLmReqsInPipe--;
wf->validateRequestCounters();
} // execute
@@ -39400,6 +39402,8 @@ namespace Gcn3ISA
Wavefront *wf = gpuDynInst->wavefront();
if (wf->execMask().none()) {
wf->decVMemInstsIssued();
wf->decLGKMInstsIssued();
wf->rdGmReqsInPipe--;
wf->rdLmReqsInPipe--;
return;
@@ -39496,6 +39500,8 @@ namespace Gcn3ISA
Wavefront *wf = gpuDynInst->wavefront();
if (wf->execMask().none()) {
wf->decVMemInstsIssued();
wf->decLGKMInstsIssued();
wf->rdGmReqsInPipe--;
wf->rdLmReqsInPipe--;
return;
@@ -39592,6 +39598,8 @@ namespace Gcn3ISA
Wavefront *wf = gpuDynInst->wavefront();
if (wf->execMask().none()) {
wf->decVMemInstsIssued();
wf->decLGKMInstsIssued();
wf->rdGmReqsInPipe--;
wf->rdLmReqsInPipe--;
return;
@@ -39660,6 +39668,8 @@ namespace Gcn3ISA
Wavefront *wf = gpuDynInst->wavefront();
if (wf->execMask().none()) {
wf->decVMemInstsIssued();
wf->decLGKMInstsIssued();
wf->rdGmReqsInPipe--;
wf->rdLmReqsInPipe--;
return;
@@ -39728,6 +39738,8 @@ namespace Gcn3ISA
Wavefront *wf = gpuDynInst->wavefront();
if (wf->execMask().none()) {
wf->decVMemInstsIssued();
wf->decLGKMInstsIssued();
wf->rdGmReqsInPipe--;
wf->rdLmReqsInPipe--;
return;
@@ -39805,6 +39817,8 @@ namespace Gcn3ISA
Wavefront *wf = gpuDynInst->wavefront();
if (wf->execMask().none()) {
wf->decVMemInstsIssued();
wf->decLGKMInstsIssued();
wf->rdGmReqsInPipe--;
wf->rdLmReqsInPipe--;
}
@@ -39884,6 +39898,8 @@ namespace Gcn3ISA
Wavefront *wf = gpuDynInst->wavefront();
if (wf->execMask().none()) {
wf->decVMemInstsIssued();
wf->decLGKMInstsIssued();
wf->wrGmReqsInPipe--;
wf->wrLmReqsInPipe--;
return;
@@ -39952,6 +39968,8 @@ namespace Gcn3ISA
Wavefront *wf = gpuDynInst->wavefront();
if (wf->execMask().none()) {
wf->decVMemInstsIssued();
wf->decLGKMInstsIssued();
wf->wrGmReqsInPipe--;
wf->wrLmReqsInPipe--;
return;
@@ -40021,6 +40039,8 @@ namespace Gcn3ISA
Wavefront *wf = gpuDynInst->wavefront();
if (wf->execMask().none()) {
wf->decVMemInstsIssued();
wf->decLGKMInstsIssued();
wf->wrGmReqsInPipe--;
wf->wrLmReqsInPipe--;
return;
@@ -40090,6 +40110,8 @@ namespace Gcn3ISA
Wavefront *wf = gpuDynInst->wavefront();
if (wf->execMask().none()) {
wf->decVMemInstsIssued();
wf->decLGKMInstsIssued();
wf->wrGmReqsInPipe--;
wf->wrLmReqsInPipe--;
return;
@@ -40159,6 +40181,8 @@ namespace Gcn3ISA
Wavefront *wf = gpuDynInst->wavefront();
if (wf->execMask().none()) {
wf->decVMemInstsIssued();
wf->decLGKMInstsIssued();
wf->wrGmReqsInPipe--;
wf->wrLmReqsInPipe--;
return;
@@ -40237,6 +40261,8 @@ namespace Gcn3ISA
Wavefront *wf = gpuDynInst->wavefront();
if (wf->execMask().none()) {
wf->decVMemInstsIssued();
wf->decLGKMInstsIssued();
wf->wrGmReqsInPipe--;
wf->wrLmReqsInPipe--;
return;
@@ -40325,6 +40351,8 @@ namespace Gcn3ISA
Wavefront *wf = gpuDynInst->wavefront();
if (wf->execMask().none()) {
wf->decVMemInstsIssued();
wf->decLGKMInstsIssued();
wf->wrGmReqsInPipe--;
wf->rdGmReqsInPipe--;
return;
@@ -40425,6 +40453,8 @@ namespace Gcn3ISA
Wavefront *wf = gpuDynInst->wavefront();
if (wf->execMask().none()) {
wf->decVMemInstsIssued();
wf->decLGKMInstsIssued();
wf->wrGmReqsInPipe--;
wf->rdGmReqsInPipe--;
return;
@@ -40526,6 +40556,8 @@ namespace Gcn3ISA
Wavefront *wf = gpuDynInst->wavefront();
if (wf->execMask().none()) {
wf->decVMemInstsIssued();
wf->decLGKMInstsIssued();
wf->wrGmReqsInPipe--;
wf->rdGmReqsInPipe--;
return;
@@ -40893,6 +40925,8 @@ namespace Gcn3ISA
Wavefront *wf = gpuDynInst->wavefront();
if (wf->execMask().none()) {
wf->decVMemInstsIssued();
wf->decLGKMInstsIssued();
wf->wrGmReqsInPipe--;
wf->rdGmReqsInPipe--;
return;
@@ -40995,6 +41029,8 @@ namespace Gcn3ISA
Wavefront *wf = gpuDynInst->wavefront();
if (wf->execMask().none()) {
wf->decVMemInstsIssued();
wf->decLGKMInstsIssued();
wf->wrGmReqsInPipe--;
wf->rdGmReqsInPipe--;
return;

1
src/gpu-compute/global_memory_pipeline.cc
View File

@@ -130,6 +130,7 @@ GlobalMemPipeline::exec()
DPRINTF(GPUMem, "CU%d: WF[%d][%d]: Completing global mem instr %s\n",
m->cu_id, m->simdId, m->wfSlotId, m->disassemble());
m->completeAcc(m);
w->decVMemInstsIssued();
if (m->isLoad() || m->isAtomicRet()) {
w->computeUnit->vrf[w->simdId]->

3
src/gpu-compute/gpu_dyn_inst.cc
View File

@@ -819,6 +819,7 @@ GPUDynInst::resolveFlatSegment(const VectorMask &mask)
if (executedAs() == Enums::SC_GLOBAL) {
// no transormation for global segment
wavefront()->execUnitId = wavefront()->flatGmUnitId;
wavefront()->decLGKMInstsIssued();
if (isLoad()) {
wavefront()->rdLmReqsInPipe--;
} else if (isStore()) {
@@ -838,6 +839,7 @@ GPUDynInst::resolveFlatSegment(const VectorMask &mask)
}
}
wavefront()->execUnitId = wavefront()->flatLmUnitId;
wavefront()->decVMemInstsIssued();
if (isLoad()) {
wavefront()->rdGmReqsInPipe--;
} else if (isStore()) {
@@ -897,6 +899,7 @@ GPUDynInst::resolveFlatSegment(const VectorMask &mask)
}
}
wavefront()->execUnitId = wavefront()->flatLmUnitId;
wavefront()->decLGKMInstsIssued();
if (isLoad()) {
wavefront()->rdGmReqsInPipe--;
} else if (isStore()) {

1
src/gpu-compute/local_memory_pipeline.cc
View File

@@ -76,6 +76,7 @@ LocalMemPipeline::exec()
DPRINTF(GPUMem, "CU%d: WF[%d][%d]: Completing local mem instr %s\n",
m->cu_id, m->simdId, m->wfSlotId, m->disassemble());
m->completeAcc(m);
w->decLGKMInstsIssued();
if (m->isLoad() || m->isAtomicRet()) {
w->computeUnit->vrf[w->simdId]->

1
src/gpu-compute/scalar_memory_pipeline.cc
View File

@@ -85,6 +85,7 @@ ScalarMemPipeline::exec()
}
m->completeAcc(m);
w->decLGKMInstsIssued();
if (m->isLoad() || m->isAtomic()) {
returnedLoads.pop();

9
src/gpu-compute/schedule_stage.cc
View File

@@ -135,6 +135,15 @@ ScheduleStage::exec()
// this wave spends in SCH stage.
wf->schCycles++;
addToSchListStalls[j]++;
} else {
if (gpu_dyn_inst->isScalar() || gpu_dyn_inst->isGroupSeg()) {
wf->incLGKMInstsIssued();
} else {
wf->incVMemInstsIssued();
if (gpu_dyn_inst->isFlat()) {
wf->incLGKMInstsIssued();
}
}
}
}

63
src/gpu-compute/wavefront.cc
View File

@@ -53,6 +53,7 @@ Wavefront::Wavefront(const Params *p)
: SimObject(p), wfSlotId(p->wf_slot_id), simdId(p->simdId),
maxIbSize(p->max_ib_size), _gpuISA(*this),
vmWaitCnt(-1), expWaitCnt(-1), lgkmWaitCnt(-1),
vmemInstsIssued(0), expInstsIssued(0), lgkmInstsIssued(0),
barId(WFBarrier::InvalidID)
{
lastTrace = 0;
@@ -1253,37 +1254,27 @@ Wavefront::waitCntsSatisfied()
return false;
}
// If we reach here, that means waitCnt instruction is executed and
// the waitcnts are set by the execute method. Check if waitcnts are
// satisfied.
// current number of vector memory ops in flight
int vm_cnt = outstandingReqsWrGm + outstandingReqsRdGm;
// current number of export insts or vector memory writes in flight
int exp_cnt = outstandingReqsWrGm;
// current number of scalar/LDS memory ops in flight
// we do not consider GDS/message ops
int lgkm_cnt = outstandingReqsWrLm + outstandingReqsRdLm +
scalarOutstandingReqsRdGm + scalarOutstandingReqsWrGm;
/**
* If we reach here, that means an s_waitcnt instruction was executed
* and the waitcnts are set by the execute method. Check if waitcnts
* are satisfied.
*/
if (vmWaitCnt != -1) {
if (vm_cnt > vmWaitCnt) {
if (vmemInstsIssued > vmWaitCnt) {
// vmWaitCnt not satisfied
return false;
}
}
if (expWaitCnt != -1) {
if (exp_cnt > expWaitCnt) {
if (expInstsIssued > expWaitCnt) {
// expWaitCnt not satisfied
return false;
}
}
if (lgkmWaitCnt != -1) {
if (lgkm_cnt > lgkmWaitCnt) {
if (lgkmInstsIssued > lgkmWaitCnt) {
// lgkmWaitCnt not satisfied
return false;
}
@@ -1355,6 +1346,42 @@ Wavefront::clearWaitCnts()
status = S_RUNNING;
}
void
Wavefront::incVMemInstsIssued()
{
++vmemInstsIssued;
}
void
Wavefront::incExpInstsIssued()
{
++expInstsIssued;
}
void
Wavefront::incLGKMInstsIssued()
{
++lgkmInstsIssued;
}
void
Wavefront::decVMemInstsIssued()
{
--vmemInstsIssued;
}
void
Wavefront::decExpInstsIssued()
{
--expInstsIssued;
}
void
Wavefront::decLGKMInstsIssued()
{
--lgkmInstsIssued;
}
Addr
Wavefront::pc() const
{

10
src/gpu-compute/wavefront.hh
View File

@@ -304,6 +304,13 @@ class Wavefront : public SimObject
void setWaitCnts(int vm_wait_cnt, int exp_wait_cnt, int lgkm_wait_cnt);
void clearWaitCnts();
void incVMemInstsIssued();
void incExpInstsIssued();
void incLGKMInstsIssued();
void decVMemInstsIssued();
void decExpInstsIssued();
void decLGKMInstsIssued();
/** Freeing VRF space */
void freeRegisterFile();
@@ -343,6 +350,9 @@ class Wavefront : public SimObject
int vmWaitCnt;
int expWaitCnt;
int lgkmWaitCnt;
int vmemInstsIssued;
int expInstsIssued;
int lgkmInstsIssued;
status_e status;
Addr _pc;
VectorMask _execMask;