Task #2669: Use 2x2 correlator kernel for better performance, and cleaned up global_defines a bit

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

@@ -32,6 +32,10 @@
 
 #include <GPUProc/global_defines.h>
 
+// For Cobalt (= up to 80 antenna fields), the 2x2 kernel gives the best
+// performance.
+#define USE_2X2
+
 namespace LOFAR
 {
   namespace Cobalt
@@ -56,8 +60,7 @@ namespace LOFAR
       setArg(0, buffers.output);
       setArg(1, buffers.input);
 
-      size_t maxNrThreads, preferredMultiple;
-      maxNrThreads = getAttribute(CU_FUNC_ATTRIBUTE_MAX_THREADS_PER_BLOCK);
+      size_t preferredMultiple;
 
       gpu::Platform pf;
       if (pf.getName() == "AMD Accelerated Parallel Processing") {
@@ -80,7 +83,7 @@ namespace LOFAR
 # else
       unsigned nrBlocks = nrBaselines;
 # endif
-      unsigned nrPasses = (nrBlocks + maxNrThreads - 1) / maxNrThreads;
+      unsigned nrPasses = (nrBlocks + maxThreadsPerBlock - 1) / maxThreadsPerBlock;
       unsigned nrThreads = (nrBlocks + nrPasses - 1) / nrPasses;
       nrThreads = (nrThreads + preferredMultiple - 1) / preferredMultiple * preferredMultiple;
 

RTCP/Cobalt/GPUProc/src/global_defines.cc

@@ -34,9 +34,6 @@ namespace LOFAR
   {
     bool profiling = false;
     bool gpuProfiling = true;
-    const char *str = getenv("NR_GPUS");
-    unsigned nrGPUs = str ? atoi(str) : 1;
-
 
     inline void set_affinity(unsigned device)
     {

RTCP/Cobalt/GPUProc/src/global_defines.h

@@ -22,11 +22,8 @@
 #define LOFAR_GPUPROC_GLOBAL_DEFINES_H
 
 #define NR_STATION_FILTER_TAPS  16
-#undef USE_NEW_CORRELATOR
 #define NR_POLARIZATIONS         2 // TODO: get the nr of pol symbol from an LCS/Common header and/or from CoInterface/Config.h (if that isn't a dup too)
 #define NR_TAPS                 16
-#undef USE_2X2
-#undef USE_TEST_DATA
 #undef USE_B7015
 
 namespace LOFAR
@@ -35,7 +32,6 @@ namespace LOFAR
   {
     extern bool profiling;
     extern bool gpuProfiling;
-    extern unsigned nrGPUs;
 
     void set_affinity(unsigned device);
   }