Heuristics.hs

-- | Some OpenCL platforms have a SIMD/warp/wavefront-based execution
-- model that execute groups of threads in lockstep, permitting us to
-- perform cross-thread synchronisation within each such group 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 group sizes and group
-- counts.
module Futhark.CodeGen.OpenCL.Heuristics
  ( SizeHeuristic (..),
    DeviceType (..),
    WhichSize (..),
    DeviceInfo (..),
    sizeHeuristicsTable,
  )
where

import Futhark.Analysis.PrimExp
import Futhark.Util.Pretty

-- | The type of OpenCL device that this heuristic applies to.
data DeviceType = DeviceCPU | DeviceGPU

-- | 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.
newtype DeviceInfo = DeviceInfo String

instance Pretty DeviceInfo where
  pretty (DeviceInfo s) = "device_info" <> parens (pretty s)

-- | A size that can be assigned a default.
data WhichSize = LockstepWidth | NumGroups | GroupSize | TileSize | RegTileSize | Threshold

-- | A heuristic for setting the default value for something.
data SizeHeuristic = SizeHeuristic
  { platformName :: String,
    deviceType :: DeviceType,
    heuristicSize :: WhichSize,
    heuristicValue :: TPrimExp Int32 DeviceInfo
  }

-- | All of our heuristics.
sizeHeuristicsTable :: [SizeHeuristic]
sizeHeuristicsTable =
  [ SizeHeuristic "NVIDIA CUDA" DeviceGPU LockstepWidth 32,
    SizeHeuristic "AMD Accelerated Parallel Processing" DeviceGPU LockstepWidth 32,
    SizeHeuristic "" DeviceGPU LockstepWidth 1,
    -- We calculate the number of groups to aim for 1024 threads per
    -- compute unit if we also use the default group size.  This seems
    -- to perform well in practice.
    SizeHeuristic "" DeviceGPU NumGroups $ 4 * max_compute_units,
    SizeHeuristic "" DeviceGPU GroupSize 256,
    SizeHeuristic "" DeviceGPU TileSize 16,
    SizeHeuristic "" DeviceGPU RegTileSize 4,
    SizeHeuristic "" DeviceGPU Threshold $ 32 * 1024,
    SizeHeuristic "" DeviceCPU LockstepWidth 1,
    SizeHeuristic "" DeviceCPU NumGroups max_compute_units,
    SizeHeuristic "" DeviceCPU GroupSize 32,
    SizeHeuristic "" DeviceCPU TileSize 4,
    SizeHeuristic "" DeviceCPU RegTileSize 1,
    SizeHeuristic "" DeviceCPU Threshold max_compute_units
  ]
  where
    max_compute_units =
      TPrimExp $ LeafExp (DeviceInfo "MAX_COMPUTE_UNITS") $ IntType Int32