GpuWideLoad¶

Reads consecutive elements of an array as one vector.

Declaration¶

Syntax¶

GpuWideLoad(expr* targets, expr source, expr array, expr? aligned)

Arguments¶

Argument	Description
`targets`	Variables of the type of the elements; there are as many as fill 16 bytes (2, 4 or 8)
`source`	The first element read, an element of `array`
`array`	The array `source` is an element of, one the kernel is given
`aligned`	A condition on values the launch fixes (the extents and lower bound of `array`) that holds when the elements start at a multiple of 16 bytes if `array` does; absent when they always do

Return values¶

None.

Description¶

Assigns targets(j) the element that follows source by j - 1 places in the first dimension of array. When aligned holds and the data of array starts at a multiple of 16 bytes, the device reads the 16 bytes at source at once; otherwise it reads the elements one by one. Either way each target holds what the element holds, and every element is one that the statement after the node reads, so none is read that the program does not read. The node is made by the gpu_wide_loads pass in device code on an NVIDIA GPU and is only allowed there.

Examples¶

(GpuWideLoad
    [(Var 3 __gpu_wide) (Var 3 __gpu_wide1) (Var 3 __gpu_wide2) (Var 3 __gpu_wide3)]
    (ArrayItem (Var 3 w) [(() (IntegerBinOp (IntegerBinOp (IntegerConstant 4 (Integer 4)) Mul (Var 3 c) (Integer 4) ()) Sub (IntegerConstant 3 (Integer 4)) (Integer 4) ()) ()) (() (Var 3 n) ())] (Integer 4) ColMajor ())
    (Var 3 w)
    ())