Futhark.IR.Prop.Reshape

Facilities for creating, inspecting, and simplifying reshape and coercion operations.

Basic tools

newDim :: DimChange d -> d

The new dimension.

newDims :: ShapeChange d -> [d]

The new dimensions resulting from a reshape operation.

newShape :: ShapeChange SubExp -> Shape

The new shape resulting from a reshape operation.

Construction

shapeCoerce :: [SubExp] -> VName -> Exp lore

Construct a Reshape where all dimension changes are DimCoercions.

Execution

reshapeOuter :: ShapeChange SubExp -> Int -> Shape -> ShapeChange SubExp

reshapeOuter newshape n oldshape returns a Reshape expression that replaces the outer n dimensions of oldshape with newshape.

reshapeInner :: ShapeChange SubExp -> Int -> Shape -> ShapeChange SubExp

reshapeInner newshape n oldshape returns a Reshape expression that replaces the inner m-n dimensions (where m is the rank of oldshape) of src with newshape.

Inspection

shapeCoercion :: ShapeChange d -> Maybe [d]

If the shape change is nothing but shape coercions, return the new dimensions. Otherwise, return Nothing.

Simplification

fuseReshape :: Eq d => ShapeChange d -> ShapeChange d -> ShapeChange d

fuseReshape s1 s2 creates a new ShapeChange that is semantically the same as first applying s1 and then s2. This may take advantage of properties of DimCoercion versus DimNew to preserve information.

informReshape :: Eq d => [d] -> ShapeChange d -> ShapeChange d

Given concrete information about the shape of the source array, convert some DimNews into DimCoercions.

Shape calculations

reshapeIndex :: IntegralExp num => [num] -> [num] -> [num] -> [num]

reshapeIndex to_dims from_dims is transforms the index list is (which is into an array of shape from_dims) into an index list is', which is into an array of shape to_dims. is must have the same length as from_dims, and is' will have the same length as to_dims.

flattenIndex :: IntegralExp num => [num] -> [num] -> num

flattenIndex dims is computes the flat index of is into an array with dimensions dims. The length of dims and is must be the same.

unflattenIndex :: IntegralExp num => [num] -> num -> [num]

unflattenIndex dims i computes a list of indices into an array with dimension dims given the flat index i. The resulting list will have the same size as dims.

sliceSizes :: IntegralExp num => [num] -> [num]

Given a length n list of dimensions dims, sizeSizes dims will compute a length n+1 list of the size of each possible array slice. The first element of this list will be the product of dims, and the last element will be 1.