Cuda
Namespace: Tensor.Utils
Cuda support types functions
Nested types and modules
| Type | Description |
| BlockDim | CUDA block dimension (x, y, z) |
| GridDim | CUDA grid dimension (x, y, z) |
| LaunchDim | CUDA launch dimension |
| WorkDim | dimensionality of parallel work to perform (x, y, z) |
Functions and values
| Function or value | Description |
blas
Signature: CudaBlas
|
|
callback cuStream fn
Signature: cuStream:CUstream -> fn:(unit -> unit) -> unit
|
Places a callback function on a CUDA stream. The function is executed on a global callback thread and is allowed to make CUDA calls. The thread's CUDA context has been set to the libraries CUDA context. The CUDA stream continues execution while the callback function is being invoked. The callback can be blocked by waiting for other callbacks. |
callbackWithResult cuStream fn
Signature: cuStream:CUstream -> fn:(CUResult -> unit) -> unit
|
Places a callback function on a CUDA stream. The function is executed on a thread-pool thread and is allowed to make CUDA calls. This function is less efficient than Cuda.callback. |
checkContext ()
Signature: unit -> unit
|
Checks that the thread's current CUDA context is the CUDA context that was active or created while this module was initialized. |
computeLaunchDim (...)
Signature: (int64 * int64 * int64) -> maxBlockSize:int -> LaunchDim
|
Computes CUDA launch dimensions from work dimensions and maximum block size. It is possible that the calculated launch dimensions will be smaller than the specified work dimensions, since the maximum block and grid sizes are limited. |
context
Signature: CudaContext
|
CUDA context |
deviceInfo
Signature: CudaDeviceProperties
|
CUDA device info |
divCeil a b
Signature: a:int64 -> b:int64 -> int64
|
Is equivalent to int64 (ceil(float a / float b)). |
getIntPtr cuPtr
Signature: cuPtr:CUdeviceptr -> IntPtr
|
gets device pointer as IntPtr |
init ()
Signature: unit -> unit
|
Ensures that CUDA is initialized. Multiple calls are allowed and have no effect. |
keepAlive cuStream x
Signature: cuStream:CUstream -> x:obj -> unit
|
Keeps the given object alive (i.e. prevent it from being GCed) until all operations that were queued on the given CUDA stream up to now have been executed. |
keepAliveMany cuStream xs
Signature: cuStream:CUstream -> xs:obj list -> unit
|
Keeps the given objects alive (i.e. prevent them from being GCed) until all operations that were queued on the given CUDA stream up to now have been executed. |
maxBlockDim
Signature: BlockDim
|
CUDA maximum block dimension |
maxGridDim
Signature: GridDim
|
CUDA maximum grid dimension |
nvccArch
Signature: string
|
nvcc arch code |
nvccCode
Signature: string
|
nvcc sm code |
printDevice ()
Signature: unit -> unit
|
prints short CUDA device information |
printInfo ()
Signature: unit -> unit
|
prints CUDA info |
setContext ()
Signature: unit -> unit
|
Sets the thread's current CUDA context to the CUDA context that was active or created while this module was initialized. |
shutdown ()
Signature: unit -> unit
|
shutsdown CUDA (necessary for correct profiler results) |
toDim3 (arg1, arg2, arg3)
Signature: (int * int * int) -> dim3
|
convert block/grid dimension (x, y, z) to VectorTypes.dim3 |
