GPU: Use launch_bounds also for HIP instead of attribute(num_vgpr)

Author: davidrohr

GPU/Common/GPUDefGPUParameters.h

@@ -27,7 +27,7 @@
   #define GPUCA_THREAD_COUNT_CONSTRUCTOR 128
   #define GPUCA_THREAD_COUNT_SELECTOR 128
   #define GPUCA_THREAD_COUNT_FINDER 128
-  #define GPUCA_NEIGHBORSFINDER_REGS REG, 64
+  #define GPUCA_NEIGHBORSFINDER_REGS REG, (GPUCA_THREAD_COUNT_FINDER, 1)
   #define GPUCA_NEIGHBOURS_FINDER_MAX_NNEIGHUP 0
   #define GPUCA_TRACKLET_SELECTOR_HITS_REG_SIZE 12
 #elif defined(GPUCA_GPUTYPE_TURING)

GPU/GPUTracking/Base/hip/GPUReconstructionHIP.hip.cxx

@@ -72,7 +72,7 @@ GPUg() void runKernelHIP(GPUCA_CONSMEM_PTR int iSlice, Args... args)
 */
 
 #undef GPUCA_KRNL_REG
-#define GPUCA_KRNL_REG(num) __attribute__((amdgpu_num_vgpr(num)))
+#define GPUCA_KRNL_REG(args) __launch_bounds__(GPUCA_M_STRIP(args))
 #define GPUCA_KRNL(x_class, x_attributes, x_arguments, x_forward) GPUCA_KRNL_WRAP(GPUCA_KRNL_, x_class, x_attributes, x_arguments, x_forward)
 #define GPUCA_KRNL_BACKEND_CLASS GPUReconstructionHIPBackend
 #define GPUCA_KRNL_CALL_single(x_class, x_attributes, x_arguments, x_forward) \