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

gpu-compute: Fix stat bucket sizes

Browse Source 
Change-Id: If30505515867a866c631cb117d3d22e19814a2f2
This commit is contained in:
Matthew Poremba
2024-04-13 15:49:53 -07:00
parent bdaeb082c3
commit 01f2df4b8a
3 changed files with 12 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
src/gpu-compute/compute_unit.cc
View File

@@ -2436,15 +2436,15 @@ ComputeUnit::ComputeUnitStats::ComputeUnitStats(statistics::Group *parent,
instCyclesLdsPerSimd.init(cu->numVectorALUs);
hitsPerTLBLevel.init(4);
execRateDist.init(0, 10, 2);
ldsBankConflictDist.init(0, cu->wfSize(), 2);
execRateDist.init(0, 10-1, 2);
ldsBankConflictDist.init(0, cu->wfSize()-1, 2);
pageDivergenceDist.init(1, cu->wfSize(), 4);
controlFlowDivergenceDist.init(1, cu->wfSize(), 4);
activeLanesPerGMemInstrDist.init(1, cu->wfSize(), 4);
activeLanesPerLMemInstrDist.init(1, cu->wfSize(), 4);
headTailLatency.init(0, 1000000, 10000).flags(statistics::pdf |
headTailLatency.init(0, 1000000-1, 10000).flags(statistics::pdf |
statistics::oneline);
waveLevelParallelism.init(0, n_wf * cu->numVectorALUs, 1);
instInterleave.init(cu->numVectorALUs, 0, 20, 1);

2
src/gpu-compute/exec_stage.cc
View File

@@ -216,7 +216,7 @@ ExecStage::ExecStageStats::ExecStageStats(statistics::Group *parent)
ComputeUnit *compute_unit = static_cast<ComputeUnit*>(parent);
spc.init(0, compute_unit->numExeUnits(), 1);
idleDur.init(0, 75, 5);
idleDur.init(0, 75-1, 5);
numCyclesWithInstrTypeIssued.init(compute_unit->numExeUnits());
numCyclesWithNoInstrTypeIssued.init(compute_unit->numExeUnits());

16
src/gpu-compute/shader.cc
View File

@@ -584,31 +584,31 @@ Shader::ShaderStats::ShaderStats(statistics::Group *parent, int wf_size)
"vector instruction destination operand distribution")
{
allLatencyDist
.init(0, 1600000, 10000)
.init(0, 1600000-1, 10000)
.flags(statistics::pdf | statistics::oneline);
loadLatencyDist
.init(0, 1600000, 10000)
.init(0, 1600000-1, 10000)
.flags(statistics::pdf | statistics::oneline);
storeLatencyDist
.init(0, 1600000, 10000)
.init(0, 1600000-1, 10000)
.flags(statistics::pdf | statistics::oneline);
initToCoalesceLatency
.init(0, 1600000, 10000)
.init(0, 1600000-1, 10000)
.flags(statistics::pdf | statistics::oneline);
rubyNetworkLatency
.init(0, 1600000, 10000)
.init(0, 1600000-1, 10000)
.flags(statistics::pdf | statistics::oneline);
gmEnqueueLatency
.init(0, 1600000, 10000)
.init(0, 1600000-1, 10000)
.flags(statistics::pdf | statistics::oneline);
gmToCompleteLatency
.init(0, 1600000, 10000)
.init(0, 1600000-1, 10000)
.flags(statistics::pdf | statistics::oneline);
coalsrLineAddresses
@@ -624,7 +624,7 @@ Shader::ShaderStats::ShaderStats(statistics::Group *parent, int wf_size)
ccprintf(namestr, "%s.cacheBlockRoundTrip%d",
static_cast<Shader*>(parent)->name(), idx);
cacheBlockRoundTrip[idx]
.init(0, 1600000, 10000)
.init(0, 1600000-1, 10000)
.name(namestr.str())
.desc("Coalsr-to-coalsr time for the Nth cache block in an inst")
.flags(statistics::pdf | statistics::oneline);