Project
Loading...
Searching...
No Matches
LaunchGeometry.h File Reference

Compile-time launch geometry of the ITS tracking kernels, per GPU family. Poor man's RTC to be removed/reworked entirely once we can use Gabriele's tuner. More...

Go to the source code of this file.

Classes

struct  o2::its::gpu::KernelOccupancy
 Minimum resident blocks per compute unit to request when no per-kernel measurement exists. More...
 

Namespaces

namespace  o2
 a couple of static helper functions to create timestamp values for CCDB queries or override obsolete objects
 
namespace  o2::its
 
namespace  o2::its::gpu
 

Functions

constexpr KernelOccupancy o2::its::gpu::uniformOccupancy (int minBlocks)
 Use the same occupancy floor for every kernel when no per-kernel measurements are available.
 
constexpr int o2::its::gpu::gridBlocks (int residentBlocksPerComputeUnit)
 Number of blocks in a grid whose depth is residentBlocksPerComputeUnit blocks per CU.
 
constexpr int o2::its::gpu::gridThreads (int residentBlocksPerComputeUnit)
 Number of threads covered by a grid whose depth is residentBlocksPerComputeUnit blocks per CU.
 
constexpr bool o2::its::gpu::residentCoversFloor ()
 The grid must provide at least as many blocks per CU as the corresponding occupancy floor.
 

Variables

constexpr int o2::its::gpu::ComputeUnits = 60
 
constexpr int o2::its::gpu::WarpSize = 64
 
constexpr int o2::its::gpu::GPUThreads = 256
 
constexpr int o2::its::gpu::DefaultBlocksPerComputeUnit = 4
 
constexpr int o2::its::gpu::MaxBlocksPerComputeUnit = 10
 
constexpr KernelOccupancy o2::its::gpu::MinBlocks = uniformOccupancy(1)
 NVIDIA: unmeasured.
 
constexpr KernelOccupancy o2::its::gpu::ResidentBlocks = uniformOccupancy(DefaultBlocksPerComputeUnit)
 

Detailed Description

Compile-time launch geometry of the ITS tracking kernels, per GPU family. Poor man's RTC to be removed/reworked entirely once we can use Gabriele's tuner.

Definition in file LaunchGeometry.h.