Accelerate is an embedded language that distinguishes between vanilla arrays (e.g. in Haskell memory on the CPU) and embedded arrays (e.g. in device memory on a GPU), as well as the computations on both of these. Since Accelerate is an embedded language, programs written in Accelerate are not compiled by the Haskell compiler (GHC). Rather, each Accelerate backend is a runtime compiler which generates and executes parallel SIMD code of the target language at application runtime.

The type constructor Acc represents embedded collective array operations. A term of type Acc a is an Accelerate program which, once executed, will produce a value of type a (an Array or a tuple of Arrays). Collective operations of type Acc a comprise many scalar expressions, wrapped in type constructor Exp, which will be executed in parallel. Although collective operations comprise many scalar operations executed in parallel, scalar operations cannot initiate new collective operations: this stratification between scalar operations in Exp and array operations in Acc helps statically exclude nested data parallelism, which is difficult to execute efficiently on constrained hardware such as GPUs.

For example, to compute a vector dot product we could write:

dotp :: Num a => Vector a -> Vector a -> Acc (Scalar a)
dotp xs ys =
  let
      xs' = use xs
      ys' = use ys
  in
  fold (+) 0 ( zipWith (*) xs' ys' )

The function dotp consumes two one-dimensional arrays (Vectors) of values, and produces a single (Scalar) result as output. As the return type is wrapped in the type Acc, we see that it is an embedded Accelerate computation - it will be evaluated in the object language of dynamically generated parallel code, rather than the meta language of vanilla Haskell.

As the arguments to dotp are plain Haskell arrays, to make these available to Accelerate computations they must be embedded with the use function.

An Accelerate backend is used to evaluate the embedded computation and return the result back to vanilla Haskell. Calling the run function of a backend will generate code for the target architecture, compile, and execute it. For example, the following backends are available:

• accelerate-llvm-native: for execution on multicore CPUs
• accelerate-llvm-ptx: for execution on NVIDIA CUDA-capable GPUs

See also Exp, which encapsulates embedded scalar computations.

Fusion:

Array computations of type Acc will be subject to array fusion; Accelerate will combine individual Acc computations into a single computation, which reduces the number of traversals over the input data and thus improves performance. As such, it is often useful to have some intuition on when fusion should occur.

The main idea is to first partition array operations into two categories:

1. Element-wise operations, such as map, generate, and backpermute. Each element of these operations can be computed independently of all others.
2. Collective operations such as fold, scanl, and stencil. To compute each output element of these operations requires reading multiple elements from the input array(s).

Element-wise operations fuse together whenever the consumer operation uses a single element of the input array. Element-wise operations can both fuse their inputs into themselves, as well be fused into later operations. Both these examples should fuse into a single loop:

map -> reverse -> reshape -> map -> map

map -> backpermute ->
                      zipWith -> map
          generate ->

If the consumer operation uses more than one element of the input array (typically, via generate indexing an array multiple times), then the input array will be completely evaluated first; no fusion occurs in this case, because fusing the first operation into the second implies duplicating work.

On the other hand, collective operations can fuse their input arrays into themselves, but on output always evaluate to an array; collective operations will not be fused into a later step. For example:

     use ->
            zipWith -> fold |-> map
generate ->

Here the element-wise sequence (use + generate + zipWith) will fuse into a single operation, which then fuses into the collective fold operation. At this point in the program the fold must now be evaluated. In the final step the map reads in the array produced by fold. As there is no fusion between the fold and map steps, this program consists of two "loops"; one for the use + generate + zipWith + fold step, and one for the final map step.

You can see how many operations will be executed in the fused program by Show-ing the Acc program, or by using the debugging option -ddump-dot to save the program as a graphviz DOT file.

As a special note, the operations unzip and reshape, when applied to a real array, are executed in constant time, so in this situation these operations will not be fused.

Tips:

• Since Acc represents embedded computations that will only be executed when evaluated by a backend, we can programatically generate these computations using the meta language Haskell; for example, unrolling loops or embedding input values into the generated code.
• It is usually best to keep all intermediate computations in Acc, and only run the computation at the very end to produce the final result. This enables optimisations between intermediate results (e.g. array fusion) and, if the target architecture has a separate memory space as is the case of GPUs, to prevent excessive data transfers.

Arrays consists of nested tuples of individual Arrays, currently up to 15-elements wide. Accelerate computations can thereby return multiple results.

Compile and run a complete embedded array program.

The result is copied back to the host only once the arrays are demanded (or the result is forced to normal form). For results consisting of multiple components (a tuple of arrays or array of tuples) this applies per primitive array. Evaluating the result of run to WHNF will initiate the computation, but does not copy the results back from the device.

NOTE: it is recommended to use runN or runQ whenever possible.

As run, but execute using the specified target rather than using the default, automatically selected device.

Contexts passed to this function may all target to the same device, or to separate devices of differing compute capabilities.

This is runN, specialised to an array program of one argument.

As run1, but execute using the specified target rather than using the default, automatically selected device.

Prepare and execute an embedded array program.

This function can be used to improve performance in cases where the array program is constant between invocations, because it enables us to bypass front-end conversion stages and move directly to the execution phase. If you have a computation applied repeatedly to different input data, use this, specifying any changing aspects of the computation via the input parameters. If the function is only evaluated once, this is equivalent to run.

In order to use runN you must express your Accelerate program as a function of array terms:

f :: (Arrays a, Arrays b, ... Arrays c) => Acc a -> Acc b -> ... -> Acc c

This function then returns the compiled version of f:

runN f :: (Arrays a, Arrays b, ... Arrays c) => a -> b -> ... -> c

At an example, rather than:

step :: Acc (Vector a) -> Acc (Vector b)
step = ...

simulate :: Vector a -> Vector b
simulate xs = run $ step (use xs)

Instead write:

simulate = runN step

You can use the debugging options to check whether this is working successfully. For example, running with the -ddump-phases flag should show that the compilation steps only happen once, not on the second and subsequent invocations of simulate. Note that this typically relies on GHC knowing that it can lift out the function returned by runN and reuse it.

As with run, the resulting array(s) are only copied back to the host once they are actually demanded (forced to normal form). Thus, splitting a program into multiple runN steps does not imply transferring intermediate computations back and forth between host and device. However note that Accelerate is not able to optimise (fuse) across separate runN invocations.

See the programs in the 'accelerate-examples' package for examples.

See also runQ, which compiles the Accelerate program at _Haskell_ compile time, thus eliminating the runtime overhead altogether.

As runN, but execute using the specified target device.

Stream a lazily read list of input arrays through the given program, collecting results as we go.

As stream, but execute using the specified target.

Block the calling thread until the computation completes, then return the result.

Test whether the asynchronous computation has already completed. If so, return the result, else Nothing.

Cancel a running asynchronous computation.

As run, but run the computation asynchronously and return immediately without waiting for the result. The status of the computation can be queried using wait, poll, and cancel.

Note that a CUDA context can be active on only one host thread at a time. If you want to execute multiple computations in parallel, on the same or different devices, use runAsyncWith.

As runWith, but execute asynchronously. Be sure not to destroy the context, or attempt to attach it to a different host thread, before all outstanding operations have completed.

As run1, but the computation is executed asynchronously.

As run1With, but execute asynchronously.

As runN, but execute asynchronously.

As runNWith, but execute asynchronously.

Ahead-of-time compilation for an embedded array program.

This function will generate, compile, and link into the final executable, code to execute the given Accelerate computation at Haskell compile time. This eliminates any runtime overhead associated with the other run* operations. The generated code will be compiled for the current (default) GPU architecture.

Since the Accelerate program will be generated at Haskell compile time, construction of the Accelerate program, in particular via meta-programming, will be limited to operations available to that phase. Also note that any arrays which are embedded into the program via use will be stored as part of the final executable.

Usage of this function in your program is similar to that of runN. First, express your Accelerate program as a function of array terms:

f :: (Arrays a, Arrays b, ... Arrays c) => Acc a -> Acc b -> ... -> Acc c

This function then returns a compiled version of f as a Template Haskell splice, to be added into your program at Haskell compile time:

{-# LANGUAGE TemplateHaskell #-}

f' :: a -> b -> ... -> c
f' = $( runQ f )

Note that at the splice point the usage of f must monomorphic; i.e. the types a, b and c must be at some known concrete type.

See the lulesh-accelerate project for an example.

Note:

Due to GHC#13587, this currently must be as an untyped splice.

The correct type of this function is similar to that of runN:

runQ :: Afunction f => f -> Q (TExp (AfunctionR f))

Since: 1.1.0.0

Ahead-of-time analogue of runNWith. See runQ for more information.

NOTE: The supplied (at runtime) target must be compatible with the architecture that this function was compiled for (the defaultTarget of the compiling machine). Running on a device with the same compute capability is best, but this should also be forward compatible to newer architectures.

The correct type of this function is:

runQWith :: Afunction f => f -> Q (TExp (PTX -> AfunctionR f))

Since: 1.1.0.0

Ahead-of-time analogue of runNAsync. See runQ for more information.

The correct type of this function is:

runQAsync :: (Afunction f, RunAsync r, AfunctionR f ~ RunAsyncR r) => f -> Q (TExp r)

Since: 1.1.0.0

Ahead-of-time analogue of runNAsyncWith. See runQWith for more information.

The correct type of this function is:

runQAsyncWith :: (Afunction f, RunAsync r, AfunctionR f ~ RunAsyncR r) => f -> Q (TExp (PTX -> r))

Since: 1.1.0.0

The PTX execution target for NVIDIA GPUs.

The execution target carries state specific for the current execution context. The data here --- device memory and execution streams --- are implicitly tied to this CUDA execution context.

Don't store anything here that is independent of the context, for example state related to [persistent] kernel caching should _not_ go here.

Create a new PTX execution target for the given device

Create a PTX execute target for the given device context

Configure the default execution target to allocate all future host-side arrays using (CUDA) pinned memory. Any newly allocated arrays will be page-locked and directly accessible from the device, enabling high-speed (asynchronous) DMA.

Note that since the amount of available pageable memory will be reduced, overall system performance can suffer.

As with registerPinnedAllocator, but configure the given execution context.