Jonas Hahnfeld authored Aug 25, 2021 and Andrei Gheata committed Sep 11, 2021

All accesses and the related variables are already unsigned, make
the underlying data type match for consistency.

Jonas Hahnfeld authored Aug 25, 2021 and Andrei Gheata committed Sep 11, 2021

They were doing exactly the same, only the returned data was slightly
different.

The default constructor of Planes was marked for the device, but
the constructor of Quadrilaterals was also marked for the host
(despite having a different implementation there). Get rid of this
problem by properly initializing the arrays in Quadrilaterals.

See previous commit; the compiler warning was a bit misleading, which
is why I missed this before.

CUDA support in Clang requires that all definitions and declarations
specify a consistent set of attributes.

On CUDA, this macro expands to "static __constant__ const" and Clang
(correctly) complains if kTolerance is used to compute the value of
a constexpr.

Clang confuses the host function and the kernel if they have the
same name. I don't want to argue with the compiler if it can be
easily avoided.

UnplacedTrd exposes now the normal interface on GPU. Normal computed even if point not exactly on surface..

The normal computation was done previously only if the point was kHalfTolerance from surface. Now further points will still compute the normal but return valid=false.

When copying host instances with bulk transfers, fewer kernels are
invoked, which reduces the overhead, and construction on the GPU runs in
parallel.
With the trackML geometry, bulk copying placed volumes and transformations
is 10x faster on a TeslaV100 GPU.
The total transfer time reduces 5x.

Diff of AdePT't example11 synchronising trackML.gdml:
```
New:                                                            Old:
INFO: using default trackML.gdml for option -gdml_name		INFO: using default trackML.gdml for option -gdml_name
INFO: using default 0 for option -cache_depth			INFO: using default 0 for option -cache_depth
INFO: using default 1 for option -particles			INFO: using default 1 for option -particles
INFO: using default 100 for option -energy			INFO: using default 100 for option -energy
(II) vgdml::Frontend::Load: VecGeom millimeter is 1		(II) vgdml::Frontend::Load: VecGeom millimeter is 1
Starting synchronization to GPU.				Starting synchronization to GPU.
Allocating geometry on GPU...Allocating logical volumes... OK	Allocating geometry on GPU...Allocating logical volumes... OK
Allocating unplaced volumes... OK: #elems in alloc_mem=2, mem	Allocating unplaced volumes... OK: #elems in alloc_mem=2, mem
Allocating placed volumes... OK					Allocating placed volumes... OK
Allocating navigation index table... OK				Allocating navigation index table... OK
Allocating transformations... OK: #elems in alloc_mem=5, mem_	Allocating transformations... OK: #elems in alloc_mem=5, mem_
Allocating daughter lists... OK					Allocating daughter lists... OK
 geometry OK: #elems in alloc_mem=7, mem_map=38013, dau_gpu_c	 geometry OK: #elems in alloc_mem=7, mem_map=38013, dau_gpu_c
NUMBER OF PLACED VOLUMES 18789					NUMBER OF PLACED VOLUMES 18789
NUMBER OF UNPLACED VOLUMES 145					NUMBER OF UNPLACED VOLUMES 145
Copying geometry to GPU...					Copying geometry to GPU...

Copying logical volumes... OK;	TIME NEEDED 0.000615441s      |	Copying logical volumes... OK;	TIME NEEDED 0.000695619s
Copying unplaced volumes... OK;	TIME NEEDED 0.000503695s      |	Copying unplaced volumes... OK;	TIME NEEDED 0.000411785s
Copying transformations_... OK;	TIME NEEDED 0.00657207s       |	Copying transformations_... OK;	TIME NEEDED 0.0558993s
Copying placed volumes... OK;	TIME NEEDED 0.00866507s       |	Copying placed volumes... OK;	TIME NEEDED 0.0913828s
Copying daughter arrays... OK;	TIME NEEDED 0.00384211s       |	Copying daughter arrays... OK;	TIME NEEDED 0.00892739s
Geometry synchronized to GPU in 0.036591 s.		      |	Geometry synchronized to GPU in 0.173805 s.
     ---  InitElectronData ... 					     ---  InitElectronData ...
     ---  BuildELossTables ...					     ---  BuildELossTables ...
...								...
iter  221 -- tracks in flight:     2 energy deposition:    77	iter  221 -- tracks in flight:     2 energy deposition:    77
iter  222 -- tracks in flight:     0 energy deposition:    77	iter  222 -- tracks in flight:     0 energy deposition:    77
Run time: 0.0552					      |	Run time: 0.0532
```

Instead of starting one kernel to construct one placed volume on the GPU, one can
collect all instances of the same type, and construct these in a single kernel call.
This drastically reduces the number of kernel calls for larger geometries.

This required defining template functions that
- Collect all constructor arguments in arrays
- Copy those to the GPU
- Run all constructors in parallel
- Free the memory occupied by the constructor arguments.

For each type of placed volumes, the helper ConstructManyOnGPU<Type> must be instantiated
explicitly in the cxx namespace, as implicit instantiation doesn't reach it automatically.
Most instantiations happen via the macros in PlacedVolume.h, but PlacedAssembly, UnplacedExtruded,
UnplacedMultiUnion and UnplacedTesselated needed explicit dummy instantiations to fix linker
problems.

CudaManager can print details about synchronising a geometry to the GPU.
For some reason, parts of this go to cerr while the rest goes to cout.
Here, all is directed to cout to allow for better logging.

Improve formatting of the timing printouts while synchronising geometry.