P4 to Git Change 1507569 by gandryey@gera-w8 on 2018/01/24 17:56:10

SWDEV-142271 - Performance drop is observed in Ocean Surface Simulation of Compubenchcl in 17.50 when compared to 17.Q4.1
	- Rewrite the adaptive mode for waveliimiter. Make sure the performance feedback corresponds to the right wave count. Add the new sampling logic to find the best number, based on average performance.

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/device.hpp#295 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuwavelimiter.cpp#14 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuwavelimiter.hpp#10 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palkernel.hpp#15 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palvirtual.cpp#71 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palvirtual.hpp#39 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palwavelimiter.cpp#4 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palwavelimiter.hpp#5 edit
... //depot/stg/opencl/drivers/opencl/runtime/platform/command.cpp#80 edit
... //depot/stg/opencl/drivers/opencl/runtime/platform/command.hpp#88 edit
... //depot/stg/opencl/drivers/opencl/runtime/utils/flags.hpp#282 edit

rocclr/runtime/device/gpu/gpuwavelimiter.cpp

@@ -121,7 +121,7 @@ void WLAlgorithmSmooth::outputTrace() {
 }
 
 
-void WLAlgorithmSmooth::callback(ulong duration) {
+void WLAlgorithmSmooth::callback(ulong duration, uint32_t waves) {
   dumper_.addData(duration, currWaves_, static_cast<char>(state_));
 
   if (!enable_ || (duration == 0)) {
@@ -205,65 +205,6 @@ 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) {
-  measure_.resize(MaxWave + 1);
-  clear(measure_);
-  countAll_ = 0;
-}
-
-WLAlgorithmAvrg::~WLAlgorithmAvrg() {}
-
-void WLAlgorithmAvrg::outputTrace() {
-  if (!traceStream_.is_open()) {
-    return;
-  }
-
-  traceStream_ << "[WaveLimiter] " << manager_->name() << " state=" << state_
-               << " currWaves=" << currWaves_ << " waves=" << waves_ << " bestWave=" << bestWave_
-               << '\n';
-  output(traceStream_, "\n measure = ", measure_);
-  traceStream_ << "\n\n";
-}
-
-
-void WLAlgorithmAvrg::callback(ulong duration) {
-  dumper_.addData(duration, currWaves_, static_cast<char>(state_));
-
-  if (!enable_) {
-    return;
-  }
-
-  countAll_++;
-
-  switch (state_) {
-    case WARMUP:
-      state_ = ADAPT;
-    case ADAPT:
-      measure_[waves_] += duration;
-      if (countAll_ <= MaxWave * 5) {
-        waves_--;
-        if (waves_ == 0) {
-          waves_ = MaxWave;
-        }
-      } else {
-        bestWave_ = MaxWave;
-        for (uint i = 1; i < MaxWave; i++) {
-          if (measure_[i] < measure_[bestWave_]) {
-            bestWave_ = i;
-          }
-        }
-        waves_ = bestWave_;
-        state_ = RUN;
-      }
-      break;
-    case RUN:
-    default:
-      break;
-  }
-}
-
 WaveLimiterManager::WaveLimiterManager(device::Kernel* kernel, const uint simdPerSH)
     : owner_(kernel), enable_(false), enableDump_(!flagIsDefault(GPU_WAVE_LIMIT_DUMP)) {
   setIfNotDefault(simdPerSH_, GPU_WAVE_LIMIT_CU_PER_SH, simdPerSH);