P4 to Git Change 1510104 by gandryey@gera-w8 on 2018/01/30 16:21:25

SWDEV-144231 - [CQE OCL][DTB][Perf][DTB-BLOCKER] 10% Performance drop observed while running Subtests of IndigoBench due to faulty CL#1507569
	-Wavelimiter has more negative impact than improvements in a few benchmarks.
	- Reduce the number of adaptive samples during the search to 4 per each wave count.
	- Add worstWave_ to skip sampling with low performance settings.

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palwavelimiter.cpp#7 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palwavelimiter.hpp#7 edit

rocclr/runtime/device/pal/palwavelimiter.cpp

@@ -30,6 +30,7 @@ WaveLimiter::WaveLimiter(WaveLimiterManager* manager, uint seqNum, bool enable,
   waves_ = MaxWave;
   enable_ = (SIMDPerSH_ == 0) ? false : enable;
   bestWave_ = (enable_) ? MaxWave : 0;
+  worstWave_ = 0;
   sampleCount_ = 0;
   resultCount_ = 0;
   numContinuousSamples_ = 0;
@@ -47,6 +48,13 @@ uint WaveLimiter::getWavesPerSH() {
     if (numContinuousSamples_ == 0) {
         ++waves_;
         waves_ %= MaxWave + 1;
+        // Don't execute the wave count with the worst performance
+        if (waves_ != 0) {
+          while (worstWave_ >= waves_) {
+            ++waves_;
+            waves_ %= MaxWave + 1;
+          }
+        }
     }
     ++numContinuousSamples_;
     numContinuousSamples_ %= MaxContinuousSamples;
@@ -89,8 +97,8 @@ void WLAlgorithmSmooth::outputTrace() {
     return;
   }
 
-  traceStream_ << "[WaveLimiter] " << manager_->name() << " state=" << state_
-               << " waves=" << waves_ << " bestWave=" << bestWave_ << '\n';
+  traceStream_ << "[WaveLimiter] " << manager_->name() << " state=" << state_ <<
+    " waves=" << waves_ << " bestWave=" << bestWave_ << " worstWave=" << worstWave_ << '\n';
   output(traceStream_, "\n adaptive measure = ", adpMeasure_);
   output(traceStream_, "\n adaptive smaple count = ", adpSampleCnt_);
   output(traceStream_, "\n run measure = ", runMeasure_);
@@ -135,6 +143,7 @@ void WLAlgorithmSmooth::callback(ulong duration, uint32_t waves) {
           // Reset the counters
           resultCount_ = sampleCount_ = 0;
           float min = std::numeric_limits<float>::max();
+          float max = std::numeric_limits<float>::min();
           uint32_t best = bestWave_;
           // Check performance for the previous run if it's available
           if (runSampleCnt_[bestWave_] > 0) {
@@ -154,10 +163,15 @@ void WLAlgorithmSmooth::callback(ulong duration, uint32_t waves) {
             else {
               average = 0.0f;
             }
-            if (average < min) {
+            // More waves have 5% advantage over the lower number
+            if (average * 1.05f < min) {
               min = average;
               bestWave_ = i;
             }
+            if (average > max) {
+              max = average;
+              worstWave_ = i;
+            }
           }
           // Check for 5% acceptance
           if ((min * 1.05f > reference) || (bestWave_ == best)) {
@@ -169,6 +183,12 @@ void WLAlgorithmSmooth::callback(ulong duration, uint32_t waves) {
           else {
             dynRunCount_ = RunCount;
           }
+          // Find the middle between the best and the worst
+          if (worstWave_ < bestWave_) {
+            worstWave_ += ((bestWave_ - worstWave_) >> 1);
+          } else {
+            worstWave_ = 0;
+          }
           state_ = RUN;
           outputTrace();
           // Start to collect the new data for the best wave

rocclr/runtime/device/pal/palwavelimiter.hpp

@@ -52,6 +52,7 @@ class WaveLimiter : public amd::ProfilingCallback {
   uint SIMDPerSH_;  // Number of SIMDs per SH
   uint waves_;      // Waves per SIMD to be set
   uint bestWave_;   // Optimal waves per SIMD
+  uint worstWave_;  // Wave number with the worst performance 
   uint countAll_;   // Number of kernel executions
   StateKind state_;
   WaveLimiterManager* manager_;
@@ -64,7 +65,7 @@ class WaveLimiter : public amd::ProfilingCallback {
   static uint MaxWave;      // Maximum number of waves per SIMD
   static uint RunCount;     // Number of kernel executions for normal run
   static uint AdaptCount;   // Number of kernel executions for adapting
-  const static uint MaxContinuousSamples = 8;
+  const static uint MaxContinuousSamples = 2;
 
   //! Call back from Event::recordProfilingInfo to get execution time.
   virtual void callback(ulong duration, uint32_t waves) = 0;
@@ -73,8 +74,12 @@ class WaveLimiter : public amd::ProfilingCallback {
   virtual void outputTrace() = 0;
 
   template <class T> void clear(T& A) {
+    uint idx = 0;
     for (auto& I : A) {
-      I = 0;
+      if (idx > worstWave_) {
+        I = 0;
+      }
+      ++idx;
     }
   }
   template <class T> void output(std::ofstream& ofs, const std::string& prompt, T& A) {
@@ -136,9 +141,8 @@ class WaveLimiterManager {
  private:
   device::Kernel* owner_;  // The kernel which owns this object
   uint simdPerSH_;         // Simd Per SH
-  std::unordered_map<const device::VirtualDevice*,
-                     WaveLimiter*>
-      limiters_;          // Maps virtual device to wave limiter
+  std::unordered_map<const device::VirtualDevice*, WaveLimiter*>
+    limiters_;            // Maps virtual device to wave limiter
   bool enable_;           // Whether the adaptation is enabled
   bool enableDump_;       // Whether the data dumper is enabled
   uint fixed_;            // The fixed waves/simd value if not zero