** NOTE on API changes and versioning **
Cutorch provides a CUDA backend for torch7.
Cutorch provides the following:
- a new tensor type:
torch.CudaTensor
that acts liketorch.FloatTensor
, but all it's operations are on the GPU. Most of the tensor operations are supported by cutorch. There are a few missing ones, which are being implemented. The missing list can be found here: torch#70 - several other GPU tensor types, with limited functionality. Currently limited to copying/conversion, and several indexing and shaping operations.
cutorch.*
- Functions to set/get GPU, get device properties, memory usage, set/get low-level streams, set/get random number generator's seed, synchronization etc. They are described in more detail below.
This new tensor type behaves exactly like a torch.FloatTensor
, but has a couple of extra functions of note:
t:getDevice()
- Given a CudaTensort
, you can call :getDevice on it to find out the GPU ID on which the tensor memory is allocated.
Most other (besides float) CPU torch tensor types now have a cutorch equivalent, with similar names:
torch.CudaDoubleTensor
torch.CudaByteTensor
torch.CudaCharTensor
torch.CudaIntTensor
torch.CudaShortTensor
torch.CudaLongTensor
- and
torch.CudaHalfTensor
when supported as indicated bycutorch.hasHalf
; these are half-precision (16-bit) floats.
Note: these are currently limited to copying/conversion, and several indexing and shaping operations (e.g. narrow
, select
, unfold
, transpose
).
###cutorch.*
API
cutorch.synchronize()
: All of the CUDA API is asynchronous (barring a few functions), which means that you can queue up operations. To wait for the operations to finish, you can issuecutorch.synchronize()
in your code, when the code waits for all GPU operations on the current GPU to finish. WARNING: synchronizes the CPU host with respect to the current device (as percutorch.getDevice()
) only.cutorch.synchronizeAll()
: Same ascutorch.synchronize()
except synchronizes the CPU host with all visible GPU devices in the system. Equivalent to callingcutorch.synchronize()
once per each device.cutorch.setDevice(i)
: If one has multiple-GPUs, you can switch the default GPU (to allocate CUDA tensors and do operations). The GPU IDs are 1-indexed, so having 4 GPUs means, you can setDevice(1), setDevice(2), setDevice(3), setDevice(4).idx = cutorch.getDevice()
: Returns the currently set GPU device index.count = cutorch.getDeviceCount()
: Gets the number of available GPUs.freeMemory, totalMemory = cutorch.getMemoryUsage(devID)
: Gets the total and free memory in bytes for the given device ID.cutorch.seed([devID])
- Sets and returns a random seed for the current or specified device.cutorch.seedAll()
- Sets and returns a random seed for all available GPU devices.cutorch.initialSeed([devID])
- Returns the seed for the current or specified devicecutorch.manualSeed(seed [, device])
- Sets a manually specified RNG seed for the current or specified devicecutorch.manualSeedAll(seed)
- Sets a manually specified RNG seed for all available GPUscutorch.getRNGState([device])
- returns the current RNG state in the form of a byte tensor, for the current or specified device.cutorch.setRNGState(state [, device])
- Sets the RNG state from a previously saved state, on the current or specified device.cutorch.getState()
- Returns the global state of the cutorch package. This state is not for users, it stores the raw RNG states, cublas handles and other thread and device-specific stuff.cutorch.withDevice(devID, f)
- This is a convenience for multi-GPU code, that takes in a device ID as well as a function f. It switches cutorch to the new device, executes the function f, and switches back cutorch to the original device.cutorch.createCudaHostTensor([...])
- Allocates atorch.FloatTensor
of host-pinned memory, where dimensions can be given as an argument list of sizes or atorch.LongStorage
.
cutorch.reserveStreams(n [, nonblocking])
: creates n user streams for use on every device. NOTE: stream indexs
on device 1 is a different cudaStream_t than streams
on device 2. Takes an optional non-blocking flag; by default, this is assumed to be false. If true, then the stream is created with cudaStreamNonBlocking.n = cutorch.getNumStreams()
: returns the number of user streams available on every device. Bydefault
, this is0
, meaning only the default stream (stream 0) is available.cutorch.setStream(n)
: specifies that the current stream active for the current device (or any other device) isn
. This is preserved across device switches. 1-N are user streams,0
is the default stream.n = cutorch.getStream()
: returns the current stream active. By default, returns0
.cutorch.setDefaultStream()
: an alias forcutorch.setStream(0)
cutorch.streamWaitFor(streamWaiting, {streamsToWaitOn...})
: A 1-to-N-way barrier.streamWaiting
will wait for the list of streams specified to finish executing all kernels/events/barriers. Does not block any of the streamsToWaitOn. Current device only.cutorch.streamWaitForMultiDevice(deviceWaiting, streamWaiting, {[device]={streamsToWaitOn...}...})
: (deviceWaiting, streamWaiting) will wait on the list of (device
,streams
...) pairs; handles single or multiple device.cutorch.streamWaitForMultiDevice, a, b, {[a]={streams...}})
is equivalent tocutorch.setDevice(a); cutorch.streamWaitFor(b, {streams...})
.cutorch.streamBarrier({streams...})
: an N-to-N-way barrier between all the streams; all streams will wait for the completion of all other streams on the current device only. More efficient than creating the same N-to-N-way dependency viastreamWaitFor
.cutorch.streamBarrierMultiDevice({[device]={streamsToWaitOn...}...})
: As with streamBarrier but allows barriers between streams on arbitrary devices. Creates a cross-device N-to-N-way barrier between all (device, stream) values listed.cutorch.streamSynchronize(stream)
: equivalent tocudaStreamSynchronize(stream)
for the current device. Blocks the CPU until stream completes its queued kernels/events.cutorch.setPeerToPeerAccess(dev, devToAccess, f)
: explicitly enable (f
true) or disable p2p access (f
false) fromdev
accessing memory ondevToAccess
. Affects copy efficiency (if disabled, copies will be d2d rather than p2p; i.e., the CPU intermediates), and affects kernel p2p access as well. Can only be enabled if the underlying hardware supports p2p access. p2p access is enabled by default for all pairs of devices if the underlying hardware supports it.cutorch.getPeerToPeerAccess(dev, devToAccess)
: returns whether or not p2p access is currently enabled or disabled, for reasons of a prior call ofsetPeerToPeerAccess
or underlying hardware support.cutorch.setKernelPeerToPeerAccess(f)
: by default, kernels running on one device cannot directly access memory on another device. This is a check imposed by cutorch, to prevent synchronization and performance issues. To disable the check, call this withf
true. Kernel p2p access is actually only allowed for a pair of devices if both this is true and the underlyinggetPeerToPeerAccess
for the pair involved is true.cutorch.getKernelPeerToPeerAccess()
: returns whether or not kernel p2p checks are enabled or disabled.
Transfering a FloatTensor src
to the GPU:
dest = src:cuda() -- dest is on the current GPU
Allocating a tensor on a given GPU:
Allocate src
on GPU 3
cutorch.setDevice(3)
src = torch.CudaTensor(100)
Copying a CUDA tensor from one GPU to another:
Given a tensor called src
on GPU 1, if you want to create it's clone on GPU 2, then:
cutorch.setDevice(2)
local dest = src:clone()
OR
local dest
cutorch.withDevice(2, function() dest = src:clone() end)
Version 1.0 can be installed via: luarocks install cutorch 1.0-0
Compared to version 1.0, these are the following API changes:
operators | 1.0 | master |
---|---|---|
lt , le , gt , ge , eq , ne return type |
torch.CudaTensor | torch.CudaByteTensor |
min ,max (2nd return value) |
torch.CudaTensor | torch.CudaLongTensor |
maskedFill , maskedCopy (mask input) |
torch.CudaTensor | torch.CudaByteTensor |
topk , sort (2nd return value) |
torch.CudaTensor | torch.CudaLongTensor |
operators | CPU | CUDA |
---|