AI/rocm-systems
2
0 Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

P4 to Git Change 1200252 by jatang@jatang-opencl-hsa-stg1 on 2015/10/15 11:56:48

Browse Source 
SWDEV-56468 - Back out changelist 1200206

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpukernel.cpp#303 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpukernel.hpp#121 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuvirtual.cpp#387 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuwavelimiter.cpp#10 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuwavelimiter.hpp#7 edit
This commit is contained in:
foreman committed 2015-10-15 12:08:34 -04:00
1 parent 19ecdf3fd1
commit 4b286ad468
5 files changed
+42 -63
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

rocclr/runtime/device/gpu/gpuwavelimiter.cpp
+31 -20
View File
@@ -20,21 +20,24 @@ uint WLAlgorithmSmooth::AbandonThresh;
uint WLAlgorithmSmooth::DscThresh;
WaveLimiter::WaveLimiter(
WaveLimiterManager* manager,
Kernel* owner,
uint seqNum,
bool enable,
bool enableDump):
manager_(manager),
dumper_(manager_->name() + "_" + std::to_string(seqNum), enableDump) {
setIfNotDefault(SIMDPerSH_, GPU_WAVE_LIMIT_CU_PER_SH, manager->getSimdPerSH());
owner_(owner),
dumper_(owner_->name() + "_" + std::to_string(seqNum), enableDump) {
auto gpuDev = static_cast<const Device*>(&owner_->dev());
auto attrib = gpuDev->getAttribs();
auto hwInfo = gpuDev->hwInfo();
setIfNotDefault(SIMDPerSH_, GPU_WAVE_LIMIT_CU_PER_SH,
attrib.numberOfCUsperShaderArray * hwInfo->simdPerCU_);
MaxWave = GPU_WAVE_LIMIT_MAX_WAVE;
WarmUpCount = GPU_WAVE_LIMIT_WARMUP;
RunCount = GPU_WAVE_LIMIT_RUN * MaxWave;
state_ = WARMUP;
if (!flagIsDefault(GPU_WAVE_LIMIT_TRACE)) {
traceStream_.open(std::string(GPU_WAVE_LIMIT_TRACE) + manager_->name() +
traceStream_.open(std::string(GPU_WAVE_LIMIT_TRACE) + owner_->name() +
".txt");
}
@@ -55,8 +58,8 @@ uint WaveLimiter::getWavesPerSH(){
return waves_ * SIMDPerSH_;
}
WLAlgorithmSmooth::WLAlgorithmSmooth(WaveLimiterManager* manager, uint seqNum, bool enable, bool enableDump):
WaveLimiter(manager, seqNum, enable, enableDump) {
WLAlgorithmSmooth::WLAlgorithmSmooth(Kernel* owner, uint seqNum, bool enable, bool enableDump):
WaveLimiter(owner, seqNum, enable, enableDump) {
AdaptCount = 2 * MaxWave + 1;
AbandonThresh = GPU_WAVE_LIMIT_ABANDON;
DscThresh = GPU_WAVE_LIMIT_DSC_THRESH;
@@ -119,7 +122,7 @@ void WLAlgorithmSmooth::outputTrace() {
return;
}
traceStream_ << "[WaveLimiter] " << manager_->name() << " state=" << state_
traceStream_ << "[WaveLimiter] " << owner_->name() << " state=" << state_
<< " currWaves=" << currWaves_ << " waves=" << waves_
<< " bestWave=" << bestWave_ << '\n';
output(traceStream_, "\n measure = ", measure_);
@@ -214,8 +217,8 @@ void WaveLimiter::DataDumper::addData(ulong time, uint wave, char state) {
state_.push_back(state);
}
WLAlgorithmAvrg::WLAlgorithmAvrg(WaveLimiterManager* manager, uint seqNum, bool enable, bool enableDump):
WaveLimiter(manager, seqNum, enable, enableDump) {
WLAlgorithmAvrg::WLAlgorithmAvrg(Kernel* owner, uint seqNum, bool enable, bool enableDump):
WaveLimiter(owner, seqNum, enable, enableDump) {
measure_.resize(MaxWave + 1);
clear(measure_);
@@ -231,7 +234,7 @@ void WLAlgorithmAvrg::outputTrace() {
return;
}
traceStream_ << "[WaveLimiter] " << manager_->name() << " state=" << state_
traceStream_ << "[WaveLimiter] " << owner_->name() << " state=" << state_
<< " currWaves=" << currWaves_ << " waves=" << waves_
<< " bestWave=" << bestWave_ << '\n';
output(traceStream_, "\n measure = ", measure_);
@@ -276,12 +279,17 @@ void WLAlgorithmAvrg::callback(ulong duration) {
}
}
WaveLimiterManager::WaveLimiterManager(device::Kernel* kernel, const uint simdPerSH):
WaveLimiterManager::WaveLimiterManager(Kernel* kernel):
owner_(kernel),
enable_(false),
enableDump_(!flagIsDefault(GPU_WAVE_LIMIT_DUMP)) {
setIfNotDefault(simdPerSH_, GPU_WAVE_LIMIT_CU_PER_SH, simdPerSH);
fixed_ = GPU_WAVES_PER_SIMD * simdPerSH_;
auto gpuDev = static_cast<const Device*>(&owner_->dev());
auto attrib = gpuDev->getAttribs();
auto hwInfo = gpuDev->hwInfo();
unsigned simdPerSH = 0;
setIfNotDefault(simdPerSH, GPU_WAVE_LIMIT_CU_PER_SH,
attrib.numberOfCUsperShaderArray * hwInfo->simdPerCU_);
fixed_ = GPU_WAVES_PER_SIMD * simdPerSH;
}
WaveLimiterManager::~WaveLimiterManager() {
@@ -318,7 +326,7 @@ amd::ProfilingCallback* WaveLimiterManager::getProfilingCallback(
return loc->second;
}
auto limiter = new WLAlgorithmSmooth(this, limiters_.size(), enable_,
auto limiter = new WLAlgorithmSmooth(owner_, limiters_.size(), enable_,
enableDump_);
if (limiter == NULL) {
enable_ = false;
@@ -328,23 +336,26 @@ amd::ProfilingCallback* WaveLimiterManager::getProfilingCallback(
return limiter;
}
void WaveLimiterManager::enable(const bool isCiPlus) {
void WaveLimiterManager::enable() {
if (fixed_ > 0) {
return;
}
auto gpuDev = static_cast<const Device*>(&owner_->dev());
auto hwInfo = gpuDev->hwInfo();
// Enable it only for CI+, unless GPU_WAVE_LIMIT_ENABLE is set to 1
// Disabled for SI due to bug #10817
if (!flagIsDefault(GPU_WAVE_LIMIT_ENABLE)) {
enable_ = GPU_WAVE_LIMIT_ENABLE;
}
else {
if (isCiPlus) {
if (gpuDev->settings().ciPlus_) {
if (owner_->workGroupInfo()->wavesPerSimdHint_ == 0) {
enable_ = true;
}
else if (owner_->workGroupInfo()->wavesPerSimdHint_ <= GPU_WAVE_LIMIT_MAX_WAVE) {
fixed_ = owner_->workGroupInfo()->wavesPerSimdHint_ * getSimdPerSH();
//Todo:
//fixed_ = owner_->workGroupInfo()->wavesPerSimdHint_;
}
}
}