Futhark.CodeGen.OpenCL.Heuristics

Some GPU platforms have a SIMDwarpwavefront-based execution model that execute blocks of threads in lockstep, permitting us to perform cross-thread synchronisation within each such block without the use of barriers. Unfortunately, there seems to be no reliable way to query these sizes at runtime. Instead, we use builtin tables to figure out which size we should use for a specific platform and device. If nothing matches here, the wave size should be set to one.

We also use this to select reasonable default block sizes and block counts.

Documentation

data SizeHeuristic

A heuristic for setting the default value for something.

Constructors

SizeHeuristic

Fields

• platformName :: String
• deviceType :: DeviceType
• heuristicSize :: WhichSize
• heuristicValue :: TPrimExp Int32 DeviceInfo

data DeviceType

The type of OpenCL device that this heuristic applies to.

Constructors

DeviceCPU

DeviceGPU

data WhichSize

A size that can be assigned a default.

Constructors

LockstepWidth

NumBlocks

BlockSize

TileSize

RegTileSize

Threshold

newtype DeviceInfo

The value supplies by a heuristic can depend on some device information. This will be translated into a call to clGetDeviceInfo(). Make sure to only request info that can be casted to a scalar type.

Constructors

DeviceInfo String

Instances

Pretty DeviceInfo

Defined in Futhark.CodeGen.OpenCL.Heuristics

Methods

pretty :: DeviceInfo -> Doc ann

prettyList :: [DeviceInfo] -> Doc ann

sizeHeuristicsTable :: [SizeHeuristic]

All of our heuristics.