Merge branch 'default-n' into 'master'

Use a heuristic for the best number of receivers per block.

See merge request !11

libtcc/Correlator.cc

@@ -41,12 +41,29 @@ Correlator::Correlator(unsigned nrBits,
 		       const std::string &customStoreVisibility
 		      )
 :
-  correlatorModule(compileModule(nrBits, nrReceivers, nrChannels, nrSamplesPerChannel, nrPolarizations, nrReceiversPerBlock, customStoreVisibility)),
-  correlatorKernel(correlatorModule, nrBits, nrReceivers, nrChannels, nrSamplesPerChannel, nrPolarizations, nrReceiversPerBlock)
+  capability([] { 
+    cu::Device device(0);
+    return 10 * device.getAttribute<CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR>() + device.getAttribute<CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR>();
+  } ()),
+  nrReceiversPerBlock(nrReceiversPerBlock != 0 ? nrReceiversPerBlock : defaultNrReceiversPerBlock(nrReceivers)),
+  correlatorModule(compileModule(nrBits, nrReceivers, nrChannels, nrSamplesPerChannel, nrPolarizations, this->nrReceiversPerBlock, customStoreVisibility)),
+  correlatorKernel(correlatorModule, nrBits, nrReceivers, nrChannels, nrSamplesPerChannel, nrPolarizations, this->nrReceiversPerBlock)
 {
 }
 
 
+unsigned Correlator::defaultNrReceiversPerBlock(unsigned nrReceivers) const
+{
+  return nrReceivers <=  32 ? 32 :
+	 nrReceivers <=  48 ? 48 :
+	 nrReceivers <=  64 || capability == 750 || capability == 870 ? 64 :
+	 nrReceivers <=  96 ? 32 :
+	 nrReceivers <= 128 ? 64 :
+	 nrReceivers <= 144 && capability == 900 ? 48 :
+	 nrReceivers <= 160 && capability == 900 ? 32 : 64;
+}
+
+
 cu::Module Correlator::compileModule(unsigned nrBits,
 				     unsigned nrReceivers,
 				     unsigned nrChannels,
@@ -56,9 +73,6 @@ cu::Module Correlator::compileModule(unsigned nrBits,
 				     const std::string &customStoreVisibility
 				    )
 {
-  cu::Device device(0);
-  int capability = 10 * device.getAttribute<CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR>() + device.getAttribute<CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR>();
-
   std::vector<std::string> options =
   {
     "-I" + findNVRTCincludePath(),

libtcc/Correlator.h

@@ -16,7 +16,7 @@ namespace tcc {
 		 unsigned nrChannels,
 		 unsigned nrSamplesPerChannel,
 		 unsigned nrPolarizations = 2,
-		 unsigned nrReceiversPerBlock = 64,
+		 unsigned nrReceiversPerBlock = 0, // 0: use a heuristic value that should work well
 		 const std::string &customStoreVisibility = ""
 		); // throw (cu::Error, nvrtc::Error)
 
@@ -27,6 +27,7 @@ namespace tcc {
 
     private:
       std::string      findNVRTCincludePath() const;
+      unsigned         defaultNrReceiversPerBlock(unsigned nrReceivers) const;
       cu::Module       compileModule(unsigned nrBits,
 				     unsigned nrReceivers,
 				     unsigned nrChannels,
@@ -36,6 +37,8 @@ namespace tcc {
 				     const std::string &customStoreVisibility
 				    );
 
+      unsigned         capability;
+      unsigned         nrReceiversPerBlock;
       cu::Module       correlatorModule;
       CorrelatorKernel correlatorKernel;
   };