Task #4520: Moved defaultDefinitions() to Kernel::compileDefinitions(). All...

Task #4520: Moved defaultDefinitions() to Kernel::compileDefinitions(). All kernel still get the whole she-bang of compile options; this can later be made more kernel-specific.

RTCP/Cobalt/GPUProc/src/cuda/CudaRuntimeCompiler.cc

@@ -68,54 +68,14 @@ namespace LOFAR
     return flags;    
   };
 
-  // Return empty set of definitions for the nvcc compilation of a cuda kernel
-  definitions_type defaultDefinitions()
-  {
-    definitions_type defs;
-
-    return defs;
-  }
-
-  // Return the set of default definitions for the nvcc compilation of a cuda kernel in Cobalt
-  definitions_type defaultDefinitions(const Parset &ps)
-  {
-    definitions_type defs;
+  // // Return empty set of definitions for the nvcc compilation of a cuda kernel
+  // definitions_type defaultDefinitions()
+  // {
+  //   definitions_type defs;
 
-    using boost::format;
+  //   return defs;
+  // }
 
-    defs["NVIDIA_CUDA"] = ""; // left-over from OpenCL for Correlator.cl/.cu 
-
-    // TODO: support device specific defs somehow (createPTX() knows about targets, but may be kernel and target specific)
-    //if (devices[0].getInfo<CL_DEVICE_NAME>() == "GeForce GTX 680")
-    //  defs["USE_FLOAT4_IN_CORRELATOR"] = "";
-
-    // TODO: kernel-specific defs should be specified in the XXXKernel class
-    defs["COMPLEX"] = "2";
-
-    defs["NR_BITS_PER_SAMPLE"] = str(format("%u") % ps.nrBitsPerSample());
-    defs["SUBBAND_BANDWIDTH"]  = str(format("%.7ff") % ps.subbandBandwidth()); // returns double, so rounding issue?
-    defs["NR_SUBBANDS"]        = str(format("%u") % ps.nrSubbands()); // size_t, but %zu not supp
-    defs["NR_CHANNELS"]        = str(format("%u") % ps.nrChannelsPerSubband());
-    defs["NR_STATIONS"]        = str(format("%u") % ps.nrStations());
-    defs["NR_SAMPLES_PER_CHANNEL"] = str(format("%u") % ps.nrSamplesPerChannel());
-    defs["NR_SAMPLES_PER_SUBBAND"] = str(format("%u") % ps.nrSamplesPerSubband());
-    defs["NR_BEAMS"]           = str(format("%u") % ps.nrBeams());
-    defs["NR_TABS"]            = str(format("%u") % ps.nrTABs(0)); // TODO: 0 should be dep on #beams
-    defs["NR_COHERENT_STOKES"] = str(format("%u") % ps.nrCoherentStokes()); // size_t
-    defs["NR_INCOHERENT_STOKES"] = str(format("%u") % ps.nrIncoherentStokes()); // size_t
-    defs["COHERENT_STOKES_TIME_INTEGRATION_FACTOR"]   = str(format("%u") % ps.coherentStokesTimeIntegrationFactor());
-    defs["INCOHERENT_STOKES_TIME_INTEGRATION_FACTOR"] = str(format("%u") % ps.incoherentStokesTimeIntegrationFactor());
-    defs["NR_POLARIZATIONS"]   = str(format("%u") % NR_POLARIZATIONS);
-    defs["NR_TAPS"]            = str(format("%u") % NR_TAPS);
-    defs["NR_STATION_FILTER_TAPS"] = str(format("%u") % NR_STATION_FILTER_TAPS);
-    if (ps.delayCompensation())
-      defs["DELAY_COMPENSATION"] = "";
-    if (ps.correctBandPass())
-      defs["BANDPASS_CORRECTION"] = "";
-    defs["DEDISPERSION_FFT_SIZE"] = str(format("%u") % ps.dedispersionFFTsize()); // size_t
-
-    return defs;  
-  }
   
   // Performs a 'system' call of nvcc. Return the stdout of the command
   // on error no stdout is created and an exception is thrown
@@ -162,7 +122,9 @@ namespace LOFAR
   // Create a nvcc command line string based on the input path, a set of flags and a map
   // of definitions. Use this command to call nvcc and compile the file at input path to a ptx file
   // which content is returned as a string
-  std::string compileToPtx(const std::string& pathToCuFile, const flags_type& flags, const definitions_type& definitions)
+  std::string compileToPtx(const std::string& pathToCuFile,
+                           const flags_type& flags,
+                           const definitions_type& definitions)
   {
     const string cudaCompiler = "nvcc"; 
     stringstream cmd("");
@@ -189,13 +151,6 @@ namespace LOFAR
     return runNVCC(cmd.str());
   };
 
-  // overloaded function. Use the path and default flags and definitions to call nvcc
-  std::string compileToPtx(const std::string& pathToCuFile)
-  {
-    // compile with the default flags and definitions
-    return compileToPtx(pathToCuFile, defaultFlags(), defaultDefinitions());
-  };
-
   }
 }
 

RTCP/Cobalt/GPUProc/src/cuda/CudaRuntimeCompiler.h

@@ -48,10 +48,9 @@ namespace LOFAR
   // Return the set of default flags for the nvcc compilation of a cuda kernel in Cobalt
   flags_type defaultFlags();
   
-  // Return the set of default definitions for the nvcc compilation of a cuda kernel in Cobalt
-  // DEFINITION=0 results in a definition and cannot for used for undef. Do not insert parameters in this case
-  definitions_type defaultDefinitions();
-  definitions_type defaultDefinitions(const Parset &ps);
+  // // Return the set of default definitions for the nvcc compilation of a cuda kernel in Cobalt
+  // // DEFINITION=0 results in a definition and cannot for used for undef. Do not insert parameters in this case
+  // definitions_type defaultDefinitions();
   
   // Performs a 'system' call of nvcc. Return the stdout of the command
   // on error no stdout is created and an exception is thrown
@@ -60,11 +59,12 @@ namespace LOFAR
   // Create a nvcc command line string based on the input path, a set of flags and a map
   // of definitions. Use this command to call nvcc and compile the file at input path to a ptx file
   // which content is returned as a string
-  std::string compileToPtx(const std::string& pathToCuFile, const flags_type& flags, const definitions_type& definitions);
+  std::string 
+  compileToPtx(const std::string& pathToCuFile, 
+               const flags_type& flags = defaultFlags(),
+               const definitions_type& definitions = definitions_type());
   
   // overloaded function. Use the path and default flags and definitions to call nvcc
-  std::string compileToPtx(const std::string& pathToCuFile);
-
   }
 }
 

RTCP/Cobalt/GPUProc/src/cuda/Kernels/Kernel.cc

@@ -20,6 +20,9 @@
 
 #include <lofar_config.h>
 
+#include <boost/format.hpp>
+
+#include <GPUProc/global_defines.h>
 #include "Kernel.h"
 
 namespace LOFAR
@@ -53,6 +56,46 @@ namespace LOFAR
 //      counter.doOperation(event, nrOperations, nrBytesRead, nrBytesWritten);
     }
 
+    Kernel::definitions_type Kernel::compileDefinitions(const Parset& ps)
+    {
+      definitions_type defs;
+
+      using boost::format;
+
+      defs["NVIDIA_CUDA"] = ""; // left-over from OpenCL for Correlator.cl/.cu 
+
+      // TODO: support device specific defs somehow (createPTX() knows about targets, but may be kernel and target specific)
+      //if (devices[0].getInfo<CL_DEVICE_NAME>() == "GeForce GTX 680")
+      //  defs["USE_FLOAT4_IN_CORRELATOR"] = "";
+
+      // TODO: kernel-specific defs should be specified in the XXXKernel class
+      defs["COMPLEX"] = "2";
+
+      defs["NR_BITS_PER_SAMPLE"] = str(format("%u") % ps.nrBitsPerSample());
+      defs["SUBBAND_BANDWIDTH"]  = str(format("%.7ff") % ps.subbandBandwidth()); // returns double, so rounding issue?
+      defs["NR_SUBBANDS"]        = str(format("%u") % ps.nrSubbands()); // size_t, but %zu not supp
+      defs["NR_CHANNELS"]        = str(format("%u") % ps.nrChannelsPerSubband());
+      defs["NR_STATIONS"]        = str(format("%u") % ps.nrStations());
+      defs["NR_SAMPLES_PER_CHANNEL"] = str(format("%u") % ps.nrSamplesPerChannel());
+      defs["NR_SAMPLES_PER_SUBBAND"] = str(format("%u") % ps.nrSamplesPerSubband());
+      defs["NR_BEAMS"]           = str(format("%u") % ps.nrBeams());
+      defs["NR_TABS"]            = str(format("%u") % ps.nrTABs(0)); // TODO: 0 should be dep on #beams
+      defs["NR_COHERENT_STOKES"] = str(format("%u") % ps.nrCoherentStokes()); // size_t
+      defs["NR_INCOHERENT_STOKES"] = str(format("%u") % ps.nrIncoherentStokes()); // size_t
+      defs["COHERENT_STOKES_TIME_INTEGRATION_FACTOR"]   = str(format("%u") % ps.coherentStokesTimeIntegrationFactor());
+      defs["INCOHERENT_STOKES_TIME_INTEGRATION_FACTOR"] = str(format("%u") % ps.incoherentStokesTimeIntegrationFactor());
+      defs["NR_POLARIZATIONS"]   = str(format("%u") % NR_POLARIZATIONS);
+      defs["NR_TAPS"]            = str(format("%u") % NR_TAPS);
+      defs["NR_STATION_FILTER_TAPS"] = str(format("%u") % NR_STATION_FILTER_TAPS);
+      if (ps.delayCompensation())
+        defs["DELAY_COMPENSATION"] = "";
+      if (ps.correctBandPass())
+        defs["BANDPASS_CORRECTION"] = "";
+      defs["DEDISPERSION_FFT_SIZE"] = str(format("%u") % ps.dedispersionFFTsize()); // size_t
+
+      return defs;  
+    }
+
   }
 }
 

RTCP/Cobalt/GPUProc/src/cuda/Kernels/Kernel.h

@@ -40,6 +40,12 @@ namespace LOFAR
 
       void enqueue(gpu::Stream &queue/*, PerformanceCounter &counter*/);
 
+      // Map used for storing compile defintions as key/value pairs.
+      typedef std::map<std::string, std::string> definitions_type;
+
+      // Return required compile definitions given the Parset \a ps.
+      static definitions_type compileDefinitions(const Parset& ps);
+
     protected:
       gpu::Event event;
       const Parset &ps;

RTCP/Cobalt/GPUProc/src/cuda/Pipelines/Pipeline.cc

@@ -26,6 +26,7 @@
 #include <Common/lofar_iomanip.h>
 
 #include <GPUProc/gpu_utils.h>
+#include <GPUProc/Kernels/Kernel.h>
 
 namespace LOFAR
 {
@@ -46,7 +47,7 @@ namespace LOFAR
     std::string Pipeline::createPTX(const string &srcFilename)
     {
       flags_type flags(defaultFlags());
-      definitions_type definitions(defaultDefinitions(ps));
+      Kernel::definitions_type definitions(Kernel::compileDefinitions(ps));
 
       return LOFAR::Cobalt::createPTX(devices, srcFilename, flags, definitions);
     }

RTCP/Cobalt/GPUProc/test/cuda/tCorrelator.cc

@@ -31,6 +31,7 @@
 
 #include <GPUProc/gpu_wrapper.h>
 #include <GPUProc/gpu_utils.h>
+#include <GPUProc/Kernels/Kernel.h>
 #include <GPUProc/cuda/CudaRuntimeCompiler.h>
 
 #include "TestUtil.h"
@@ -59,7 +60,7 @@ HostMemory runTest(gpu::Context ctx,
   cout << "\n==== runTest: function = " << function << " ====\n" << endl;
 
   // Get an instantiation of the default parameters
-  definitions_type definitions = defaultDefinitions();
+  definitions_type definitions = Kernel::definitions_type();
   flags_type flags = defaultFlags();
 
   // ****************************************

RTCP/Cobalt/GPUProc/test/cuda/tCorrelatorWorkQueueProcessSb.cc

@@ -25,6 +25,7 @@
 #include <Common/LofarLogger.h>
 #include <CoInterface/Parset.h>
 #include <GPUProc/gpu_utils.h>
+#include <GPUProc/Kernels/Kernel.h>
 #include <GPUProc/WorkQueues/CorrelatorWorkQueue.h>
 
 using namespace std;
@@ -61,7 +62,7 @@ int main() {
 
   map<string, string> ptx;
   flags_type flags(defaultFlags());
-  definitions_type definitions(defaultDefinitions(ps));
+  Kernel::definitions_type definitions(Kernel::compileDefinitions(ps));
   ptx[kfilenameFIR] = createPTX(devices, kfilenameFIR, flags, definitions);
   ptx[kfilenameDBP] = createPTX(devices, kfilenameDBP, flags, definitions);
   ptx[kfilenameCor] = createPTX(devices, kfilenameCor, flags, definitions);

RTCP/Cobalt/GPUProc/test/cuda/tCudaRuntimeCompiler.cc

@@ -21,6 +21,7 @@
 #include <lofar_config.h>
 
 #include <string>
+#include <GPUProc/Kernels/Kernel.h>
 #include <GPUProc/cuda/CudaRuntimeCompiler.h>
 #include <Common/LofarLogger.h>
 
@@ -35,7 +36,7 @@ int main()
   string kernelPath = "tCudaRuntimeCompiler.in_.cu";
   
   // Get an instantiation of the default parameters
-  definitions_type definitions = defaultDefinitions();
+  definitions_type definitions = Kernel::definitions_type();
 
   // override the default with a magic number
   definitions["NVIDIA_CUDA"] = "123456";

RTCP/Cobalt/GPUProc/test/cuda/tDelayAndBandPass.cc

@@ -32,6 +32,7 @@
 
 #include <GPUProc/gpu_wrapper.h>
 #include <GPUProc/gpu_utils.h>
+#include <GPUProc/Kernels/Kernel.h>
 #include <GPUProc/cuda/CudaRuntimeCompiler.h>
 #include <UnitTest++.h>
 
@@ -67,7 +68,7 @@ float * runTest(float bandPassFactor,
   string kernelPath = "DelayAndBandPass.cu";  //The test copies the kernel to the current dir (also the complex header, needed for compilation)
  
   // Get an instantiation of the default parameters
-  definitions_type definitions = defaultDefinitions();
+  definitions_type definitions = Kernel::definitions_type();
   flags_type flags = defaultFlags();
 
   // ****************************************

RTCP/Cobalt/GPUProc/test/cuda/tDelayAndBandPassKernel.cc

@@ -52,7 +52,7 @@ int main() {
 
   // Get default parameters for the compiler
   flags_type flags = defaultFlags();
-  definitions_type definitions = defaultDefinitions(ps);
+  Kernel::definitions_type definitions(Kernel::compileDefinitions(ps));
 
   string ptx = createPTX(devices, srcFilename, flags, definitions);
   gpu::Module module(createModule(ctx, srcFilename, ptx));

RTCP/Cobalt/GPUProc/test/cuda/tIntToFloat.cc

@@ -32,6 +32,7 @@
 
 #include <GPUProc/gpu_wrapper.h>
 #include <GPUProc/gpu_utils.h>
+#include <GPUProc/Kernels/Kernel.h>
 #include <GPUProc/cuda/CudaRuntimeCompiler.h>
 #include <UnitTest++.h>
 
@@ -65,7 +66,7 @@ float * runTest(unsigned NR_BITS_PER_SAMPLE = 16,
   // ****************************************
   // Compile to ptx  
   // Get an instantiation of the default parameters
-  definitions_type definitions = defaultDefinitions();
+  definitions_type definitions = Kernel::definitions_type();
   flags_type flags = defaultFlags();
 
   // Set op string string pairs to be provided to the compiler as defines

RTCP/Cobalt/GPUProc/test/cuda/tIntToFloatKernel.cc

@@ -53,7 +53,7 @@ int main() {
 
   // Get default parameters for the compiler
   flags_type flags = defaultFlags();
-  definitions_type definitions = defaultDefinitions(ps);
+  Kernel::definitions_type definitions(Kernel::compileDefinitions(ps));
 
   string ptx = createPTX(devices, srcFilename, flags, definitions);
   gpu::Module module(createModule(ctx, srcFilename, ptx));

RTCP/Cobalt/GPUProc/test/cuda/tKernel.cc

@@ -54,7 +54,7 @@ int main() {
 
   // Get default parameters for the compiler
   flags_type flags = defaultFlags();
-  definitions_type definitions = defaultDefinitions(ps);
+  Kernel::definitions_type definitions(Kernel::compileDefinitions(ps));
 
   string ptx = createPTX(devices, srcFilename, flags, definitions);
   gpu::Module module(createModule(ctx, srcFilename, ptx));

RTCP/Cobalt/GPUProc/test/cuda/tcreateProgram.cc

@@ -27,6 +27,7 @@
 #include <Common/LofarLogger.h>
 #include <CoInterface/Parset.h>
 #include <GPUProc/gpu_utils.h>
+#include <GPUProc/Kernels/Kernel.h>
 #include <GPUProc/cuda/CudaRuntimeCompiler.h>
 #include <GPUProc/global_defines.h>
 
@@ -59,7 +60,7 @@ int main(int argc, char *argv[]) {
   // Collect inputs from the parste and assign them to CudaRuntimeCompiler
   // input_types.
   flags_type flags = defaultFlags();
-  definitions_type definitions = defaultDefinitions(ps);
+  Kernel::definitions_type definitions(Kernel::compileDefinitions(ps));
 
   string ptx = createPTX(devices, srcFilename, flags, definitions);
   gpu::Module module(createModule(ctx, srcFilename, ptx));