ParallelAccelerator.API.Stencil
Exported
runStencil(inputs...) ¶
"runStencil" takes arguments in the form of "(kernel_function, A, B, C, ..., iteration, border_style)" where "kernel_function" is a lambda that represent stencil kernel, and "A", "B", "C", ... are arrays (of same dimension and size) that will be traversed by the stencil, and they must match the number of arguments of the stencil kernel. "iteration" and "border_style" are optional, and if present, "iteration" is the number of steps to repeat the stencil computation, and "border_style" is a symbol of the following:
:oob_src_zero, returns zero from a read when it is out-of-bound.
:oob_dst_zero, writes zero to destination array in an assigment should any read in its right-hand-side become out-of-bound.
:oob_wraparound, wraps around the index (with respect to source array dimension and sizes) of a read operation when it is out-of-bound.
:oob_skip, skips write to destination array in an assignment should any read in its right-hand-side become out-of-bound.
The "kernel_function" should take a set of arrays as input, and in the function body only index them with relative indices as if there is a cursor (at index 0) traversing them. It may contain more than one assignment to any input arrays, but such writes must always be indexed at 0 to guarantee write operation never goes out-of-bound. Also care must be taken when the same array is both read from and written into in the kernel function, as they'll result in non-deterministic behavior when the stencil is parallelized by ParallelAccelerator.
The "kernel_function" may optionally access scalar variables from outerscope, but no write is permitted. It may optinally contain a return statement, which acts as a specification for buffer swapping when it is an interative stencil loop (ISL). The number of arrays returned must match the input arrays. If it is not an ISL, it should always return nothing.
This function is a reference implementation of stencil in native Julia for two purposes: to verify expected return result, and to make sure user code type checks. It runs very very slow, so any real usage should go through ParallelAccelerator optimizations.
"runStencil" always returns nothing.
source: ParallelAccelerator/src/api-stencil.jl:142
Internal
process_node(node, state, top_level_number, is_top_level, read) ¶
This function is a AstWalker callback.
source: ParallelAccelerator/src/api-stencil.jl:377
@comprehend(ast) ¶
Translate all comprehension in an AST into equivalent code that uses cartesianarray call.