GPU: Rewrite virtual kernel call to a single virtual function, should enable further simplifications in the future

Author: davidrohr

GPU/GPUTracking/Base/GPUReconstructionCPU.h

@@ -75,10 +75,12 @@ class GPUReconstructionCPU : public GPUReconstructionKernels<GPUReconstructionCP
 #define GPUCA_KRNL(x_class, x_attributes, x_arguments, x_forward, x_types, ...)                                                                                                              \
   inline void runKernelImplWrapper(gpu_reconstruction_kernels::classArgument<GPUCA_M_KRNL_TEMPLATE(x_class)>, bool cpuFallback, double& timer, krnlSetup&& setup GPUCA_M_STRIP(x_arguments)) \
   {                                                                                                                                                                                          \
+    krnlSetupArgs<GPUCA_M_KRNL_TEMPLATE(x_class) GPUCA_M_STRIP(x_types)> args(setup.x, setup.y, setup.z, timer GPUCA_M_STRIP(x_forward));                                                    \
+    const uint32_t num = GetKernelNum<GPUCA_M_KRNL_TEMPLATE(x_class)>();                                                                                                                     \
     if (cpuFallback) {                                                                                                                                                                       \
-      GPUReconstructionCPU::runKernelImpl(krnlSetupArgs<GPUCA_M_KRNL_TEMPLATE(x_class) GPUCA_M_STRIP(x_types)>(setup.x, setup.y, setup.z, timer GPUCA_M_STRIP(x_forward)));                  \
+      GPUReconstructionCPU::runKernelImpl(num, &args);                                                                                                                                       \
     } else {                                                                                                                                                                                 \
-      runKernelImpl(krnlSetupArgs<GPUCA_M_KRNL_TEMPLATE(x_class) GPUCA_M_STRIP(x_types)>(setup.x, setup.y, setup.z, timer GPUCA_M_STRIP(x_forward)));                                        \
+      runKernelImpl(num, &args);                                                                                                                                                             \
     }                                                                                                                                                                                        \
   }
 #include "GPUReconstructionKernelList.h"

GPU/GPUTracking/Base/GPUReconstructionKernels.h

@@ -95,13 +95,19 @@ class GPUReconstructionKernels : public T
   template <class S, int32_t I = 0, typename... Args>
   using krnlSetupArgs = gpu_reconstruction_kernels::krnlSetupArgs<S, I, Args...>;
 
-#define GPUCA_KRNL(x_class, x_attributes, x_arguments, x_forward, x_types)                                     \
-  virtual void runKernelImpl(const krnlSetupArgs<GPUCA_M_KRNL_TEMPLATE(x_class) GPUCA_M_STRIP(x_types)>& args) \
-  {                                                                                                            \
-    T::template runKernelBackend<GPUCA_M_KRNL_TEMPLATE(x_class)>(args);                                        \
-  }
+  virtual void runKernelImpl(const int num, const void* args)
+  {
+    switch (num) { // clang-format off
+#define GPUCA_KRNL(x_class, x_attributes, x_arguments, x_forward, x_types, x_num)                               \
+      case x_num: {                                                                                             \
+        const auto& args2 = *(const krnlSetupArgs<GPUCA_M_KRNL_TEMPLATE(x_class) GPUCA_M_STRIP(x_types)>*)args; \
+        T::template runKernelBackend<GPUCA_M_KRNL_TEMPLATE(x_class)>(args2);                                    \
+        break; \
+      }
 #include "GPUReconstructionKernelList.h"
 #undef GPUCA_KRNL
+    } // clang-format on
+  }
 };
 
 } // namespace o2::gpu