GPU RTC: Fix launch bounds, cut 3rd runtime parameter to force number of blocks started

Author: davidrohr

GPU/GPUTracking/Definitions/GPUDefParametersDefault.h

@@ -24,6 +24,8 @@
 #include "GPUCommonDef.h"
 #include "GPUDefMacros.h"
 
+// Launch bound definition, 3 optional parameters: maxThreads per block, minBlocks per multiprocessor, force number of blocks (not passed to compiler as launch bounds)
+
 // GPU Run Configuration
 #ifdef GPUCA_GPUCODE
 #if defined(GPUCA_GPUTYPE_MI2xx)

GPU/GPUTracking/Definitions/GPUDefParametersLoad.template.inc

@@ -37,7 +37,7 @@ static GPUDefParameters GPUDefParametersLoad()
     if (par.par_LB_minBlocks[i] > 0) {                                                                             \
       o << ", " << par.par_LB_minBlocks[i];                                                                        \
     }                                                                                                              \
-    if (par.par_LB_forceBlocks[i] > 0) {                                                                           \
+    if (!forRTC && par.par_LB_forceBlocks[i] > 0) {                                                                \
       o << ", " << par.par_LB_forceBlocks[i];                                                                      \
     }                                                                                                              \
     o << "\n";                                                                                                     \