294 lines
9.3 KiB
C++
294 lines
9.3 KiB
C++
/* -*- C++ -*- ------------------------------------------------------------
|
|
|
|
Copyright (c) 2007 Jesse Anders and Demian Nave http://cmldev.net/
|
|
|
|
The Configurable Math Library (CML) is distributed under the terms of the
|
|
Boost Software License, v1.0 (see cml/LICENSE for details).
|
|
|
|
*-----------------------------------------------------------------------*/
|
|
/** @file
|
|
* @brief
|
|
*
|
|
* @todo Need to implement unrolling for efficient col-major array access.
|
|
*
|
|
* @todo Does it make sense to unroll an assignment if either side of the
|
|
* assignment has a fixed size, or just when the target matrix is fixed
|
|
* size?
|
|
*/
|
|
|
|
#ifndef matrix_unroller_h
|
|
#define matrix_unroller_h
|
|
|
|
#include <cml/et/traits.h>
|
|
#include <cml/et/size_checking.h>
|
|
#include <cml/et/scalar_ops.h>
|
|
|
|
#if !defined(CML_2D_UNROLLER) && !defined(CML_NO_2D_UNROLLER)
|
|
#error "The matrix unroller has not been defined."
|
|
#endif
|
|
|
|
#if defined(CML_2D_UNROLLER) && !defined(CML_MATRIX_UNROLL_LIMIT)
|
|
#error "CML_MATRIX_UNROLL_LIMIT is undefined."
|
|
#endif
|
|
|
|
namespace cml {
|
|
namespace et {
|
|
namespace detail {
|
|
|
|
/** Unroll a binary assignment operator on a fixed-size matrix.
|
|
*
|
|
* This uses a forward iteration to make better use of the cache.
|
|
*
|
|
* @sa cml::matrix
|
|
* @sa cml::et::OpAssign
|
|
*
|
|
* @bug Need to verify that OpT is actually an assignment operator.
|
|
* @bug The 2D unroller needs to be specified for efficient col-major
|
|
* access.
|
|
*/
|
|
template<class OpT, typename E, class AT, typename BO, typename L, class SrcT>
|
|
class MatrixAssignmentUnroller
|
|
{
|
|
protected:
|
|
|
|
/* The matrix type being assigned to: */
|
|
typedef cml::matrix<E,AT,BO,L> matrix_type;
|
|
|
|
/* Record traits for the arguments: */
|
|
typedef ExprTraits<matrix_type> dest_traits;
|
|
typedef ExprTraits<SrcT> src_traits;
|
|
|
|
#if defined(CML_2D_UNROLLER)
|
|
|
|
/* Forward declare: */
|
|
template<int R, int C, int LastRow, int LastCol, bool can_unroll>
|
|
struct Eval;
|
|
|
|
/* XXX This needs to be specified for efficient col-major access also! */
|
|
|
|
/** Evaluate the binary operator at element R,C. */
|
|
template<int R, int C, int LastRow, int LastCol>
|
|
struct Eval<R,C,LastRow,LastCol,true> {
|
|
void operator()(matrix_type& dest, const SrcT& src) const {
|
|
|
|
/* Apply to current R,C: */
|
|
OpT().apply(dest(R,C), src_traits().get(src,R,C));
|
|
|
|
/* Evaluate at R,C+1: */
|
|
Eval<R,C+1,LastRow,LastCol,true>()(dest,src);
|
|
}
|
|
};
|
|
|
|
/** Evaluate the binary operator at element R,LastCol. */
|
|
template<int R, int LastRow, int LastCol>
|
|
struct Eval<R,LastCol,LastRow,LastCol,true> {
|
|
void operator()(matrix_type& dest, const SrcT& src) const {
|
|
|
|
/* Apply to R,LastCol: */
|
|
OpT().apply(dest(R,LastCol), src_traits().get(src,R,LastCol));
|
|
|
|
/* Evaluate at R+1,0; i.e. move to next row and start the
|
|
* col iteration from 0:
|
|
*/
|
|
Eval<R+1,0,LastRow,LastCol,true>()(dest,src);
|
|
}
|
|
};
|
|
|
|
/** Evaluate the binary operator at element LastRow,C. */
|
|
template<int C, int LastRow, int LastCol>
|
|
struct Eval<LastRow,C,LastRow,LastCol,true> {
|
|
void operator()(matrix_type& dest, const SrcT& src) const {
|
|
|
|
/* Apply to LastRow,C: */
|
|
OpT().apply(dest(LastRow,C), src_traits().get(src,LastRow,C));
|
|
|
|
/* Evaluate at LastRow,C+1: */
|
|
Eval<LastRow,C+1,LastRow,LastCol,true>()(dest,src);
|
|
}
|
|
};
|
|
|
|
/** Evaluate the binary operator at element LastRow,LastCol. */
|
|
template<int LastRow, int LastCol>
|
|
struct Eval<LastRow,LastCol,LastRow,LastCol,true> {
|
|
void operator()(matrix_type& dest, const SrcT& src) const {
|
|
|
|
/* Apply to LastRow,LastCol: */
|
|
OpT().apply(
|
|
dest(LastRow,LastCol),
|
|
src_traits().get(src,LastRow,LastCol));
|
|
}
|
|
};
|
|
|
|
|
|
/** Evaluate operators on large matrices using a loop. */
|
|
template<int R, int C, int LastRow, int LastCol>
|
|
struct Eval<R,C,LastRow,LastCol,false> {
|
|
void operator()(matrix_type& dest, const SrcT& src) const {
|
|
for(size_t i = 0; i <= LastRow; ++i) {
|
|
for(size_t j = 0; j <= LastCol; ++j) {
|
|
OpT().apply(dest(i,j), src_traits().get(src,i,j));
|
|
}
|
|
}
|
|
}
|
|
};
|
|
|
|
#endif // CML_2D_UNROLLER
|
|
|
|
#if defined(CML_NO_2D_UNROLLER)
|
|
|
|
/** Evaluate the binary operator using a loop. */
|
|
template<int R, int C, int LastRow, int LastCol> struct Eval {
|
|
void operator()(matrix_type& dest, const SrcT& src) const {
|
|
for(size_t i = 0; i <= LastRow; ++i) {
|
|
for(size_t j = 0; j <= LastCol; ++j) {
|
|
OpT().apply(dest(i,j), src_traits().get(src,i,j));
|
|
}
|
|
}
|
|
}
|
|
};
|
|
|
|
#endif // CML_NO_2D_UNROLLER
|
|
|
|
|
|
public:
|
|
|
|
/** Unroll assignment for a fixed-sized matrix. */
|
|
void operator()(
|
|
cml::matrix<E,AT,BO,L>& dest, const SrcT& src, cml::fixed_size_tag)
|
|
{
|
|
typedef cml::matrix<E,AT,BO,L> matrix_type;
|
|
enum {
|
|
LastRow = matrix_type::array_rows-1,
|
|
LastCol = matrix_type::array_cols-1,
|
|
Max = (LastRow+1)*(LastCol+1)
|
|
};
|
|
|
|
#if defined(CML_2D_UNROLLER)
|
|
typedef typename MatrixAssignmentUnroller<OpT,E,AT,BO,L,SrcT>
|
|
::template Eval<0, 0, LastRow, LastCol,
|
|
(Max <= CML_MATRIX_UNROLL_LIMIT)> Unroller;
|
|
#endif
|
|
|
|
#if defined(CML_NO_2D_UNROLLER)
|
|
/* Use a loop: */
|
|
typedef typename MatrixAssignmentUnroller<OpT,E,AT,BO,L,SrcT>
|
|
::template Eval<0, 0, LastRow, LastCol> Unroller;
|
|
#endif
|
|
|
|
/* Use a run-time check if src is a run-time sized expression: */
|
|
typedef typename ExprTraits<SrcT>::size_tag src_size;
|
|
typedef typename select_if<
|
|
same_type<src_size,dynamic_size_tag>::is_true,
|
|
dynamic_size_tag, fixed_size_tag>::result size_tag;
|
|
|
|
/* Check the expression size (the returned size isn't needed): */
|
|
CheckedSize(dest,src,size_tag());
|
|
/* Note: for two fixed-size expressions, the if-statements and
|
|
* comparisons should be completely eliminated as dead code. If
|
|
* src is a dynamic-sized expression, the check will still happen.
|
|
*/
|
|
|
|
Unroller()(dest,src);
|
|
}
|
|
|
|
|
|
private:
|
|
/* XXX Blah, a temp. hack to fix the auto-resizing stuff below. */
|
|
matrix_size hack_actual_size(
|
|
matrix_type& dest, const SrcT& /*src*/, scalar_result_tag
|
|
)
|
|
{
|
|
typedef ExprTraits<matrix_type> dest_traits;
|
|
return dest_traits().size(dest);
|
|
}
|
|
|
|
matrix_size hack_actual_size(
|
|
matrix_type& /*dest*/, const SrcT& src, matrix_result_tag
|
|
)
|
|
{
|
|
typedef ExprTraits<SrcT> src_traits;
|
|
return src_traits().size(src);
|
|
}
|
|
|
|
matrix_size CheckOrResize(
|
|
matrix_type& dest, const SrcT& src, cml::resizable_tag)
|
|
{
|
|
#if defined(CML_AUTOMATIC_MATRIX_RESIZE_ON_ASSIGNMENT)
|
|
/* Get the size of src. This also causes src to check its size: */
|
|
matrix_size N = hack_actual_size(
|
|
dest, src, typename src_traits::result_tag());
|
|
|
|
/* Set the destination matrix's size: */
|
|
dest.resize(N.first,N.second);
|
|
#else
|
|
matrix_size N = CheckedSize(dest,src,dynamic_size_tag());
|
|
#endif
|
|
return N;
|
|
}
|
|
|
|
matrix_size CheckOrResize(
|
|
matrix_type& dest, const SrcT& src, cml::not_resizable_tag)
|
|
{
|
|
return CheckedSize(dest,src,dynamic_size_tag());
|
|
}
|
|
|
|
|
|
public:
|
|
|
|
|
|
/** Use a loop for dynamic-sized matrix assignment.
|
|
*
|
|
* @note The target matrix must already have the correct size.
|
|
*
|
|
* @todo This needs to be specialized for efficient row-major or col-major
|
|
* layout access.
|
|
*/
|
|
void operator()(matrix_type& dest, const SrcT& src, cml::dynamic_size_tag)
|
|
{
|
|
typedef ExprTraits<SrcT> src_traits;
|
|
matrix_size N = this->CheckOrResize(
|
|
dest,src,typename matrix_type::resizing_tag());
|
|
for(size_t i = 0; i < N.first; ++i) {
|
|
for(size_t j = 0; j < N.second; ++j) {
|
|
OpT().apply(dest(i,j), src_traits().get(src,i,j));
|
|
/* Note: we don't need get(), since dest is a matrix. */
|
|
}
|
|
}
|
|
}
|
|
};
|
|
|
|
}
|
|
|
|
/** This constructs an assignment unroller for fixed-size arrays.
|
|
*
|
|
* The operator must be an assignment op (otherwise, this doesn't make any
|
|
* sense). Also, automatic unrolling is only performed for fixed-size
|
|
* matrices; a loop is used for dynamic-sized matrices.
|
|
*
|
|
* @sa cml::matrix
|
|
* @sa cml::et::OpAssign
|
|
*
|
|
* @bug Need to verify that OpT is actually an assignment operator.
|
|
*/
|
|
template<class OpT, class SrcT, typename E, class AT, typename BO, typename L>
|
|
inline void UnrollAssignment(cml::matrix<E,AT,BO,L>& dest, const SrcT& src)
|
|
{
|
|
/* Record the destination matrix type, and the expression traits: */
|
|
typedef cml::matrix<E,AT,BO,L> matrix_type;
|
|
|
|
/* Record the type of the unroller: */
|
|
typedef detail::MatrixAssignmentUnroller<OpT,E,AT,BO,L,SrcT> unroller;
|
|
|
|
/* Finally, do the unroll call: */
|
|
unroller()(dest, src, typename matrix_type::size_tag());
|
|
/* XXX It may make sense to unroll if either side is a fixed size. */
|
|
}
|
|
|
|
} // namespace et
|
|
} // namespace cml
|
|
|
|
#endif
|
|
|
|
// -------------------------------------------------------------------------
|
|
// vim:ft=cpp
|