/************************************************************************* * * * YAP Prolog * * * * Yap Prolog was developed at NCCUP - Universidade do Porto * * * * Copyright L.Damas, V.S.Costa and Universidade do Porto 1985-2006 * * * ************************************************************************** * * * File: matrix.yap * * Last rev: * * mods: * * comments: Have some fun with blobs * * * *************************************************************************/ /** * @file matrix.yap * @author VITOR SANTOS COSTA * @date Tue Nov 17 22:53:40 2015 * * @brief Vector, Array and Matrix library * * */ :- module( matrix, [(<==)/2, op(800, xfx, <==), (+=)/2, op(800, xfx, +=), (-=)/2, op(800, xfx, -=), op(700, xfx, in), op(700, xfx, ins), op(450, xfx, ..), % should bind more tightly than \/ op(710, xfx, of), of/2, matrix_new/3, matrix_new/4, matrix_new_set/4, matrix_dims/2, matrix_ndims/2, matrix_size/2, matrix_type/2, matrix_to_list/2, matrix_to_lists/2, matrix_get/3, matrix_set/3, matrix_set_all/2, matrix_add/3, matrix_inc/2, matrix_dec/2, matrix_mult/2, matrix_inc/3, matrix_dec/3, matrix_arg_to_offset/3, matrix_offset_to_arg/3, matrix_max/2, matrix_maxarg/2, matrix_min/2, matrix_minarg/2, matrix_sum/2, matrix_sum_out/3, matrix_sum_out_several/3, matrix_sum_logs_out/3, matrix_sum_logs_out_several/3, matrix_add_to_all/2, matrix_agg_lines/3, matrix_agg_cols/3, matrix_to_logs/1, matrix_to_exps/1, matrix_to_exps2/1, matrix_to_logs/2, matrix_to_exps/2, matrix_op/4, matrix_op_to_all/4, matrix_op_to_lines/4, matrix_op_to_cols/4, matrix_shuffle/3, matrix_transpose/2, matrix_set_all_that_disagree/5, matrix_expand/3, matrix_select/4, matrix_column/3, matrix_get/2, matrix_set/2, foreach/2, foreach/4, op(50, yf, []), op(50, yf, '()'), op(100, xfy, '.'), op(100, fy, '.') ]). /** @defgroup matrix Matrix Library @ingroup library @{ This package provides a fast implementation of multi-dimensional matrices of integers and floats. In contrast to dynamic arrays, these matrices are multi-dimensional and compact. In contrast to static arrays. these arrays are allocated in the stack, and disppear in backtracking. Matrices are available by loading the library `library(matrix)`. They are multimensional objects of type: + terms: Prolog terms + ints: bounded integers, represented as an opaque term. The maximum integer depends on hardware, but should be obtained from the natural size of the machine. + floats: floating-point numbers, represented as an opaque term. Matrix elements can be accessed through the `matrix_get/2` predicate or through an R-inspired access notation (that uses the ciao style extension to `[]`). Examples include: + Access the second row, third column of matrix X. Indices start from `0`, ~~~~ _E_ <== _X_[2,3] ~~~~ + Access all the second row, the output is a list ofe elements. ~~~~ _L_ <== _X_[2,_] ~~~~ + Access all the second, thrd and fourth rows, the output is a list of elements. ~~~~ _L_ <== _X_[2..4,_] ~~~~ + Access all the fifth, sixth and eight rows, the output is a list of elements. ~~~~ _L_ <== _X_[2..4+3,_] ~~~~ The matrix library also supports a B-Prolog/ECliPSe inspired `foreach`iterator to iterate over elements of a matrix: + Copy a vector, element by element. ~~~~ foreach(I in 0..N1, X[I] <== Y[I]) ~~~~ + The lower-triangular matrix _Z_ is the difference between the lower-triangular and upper-triangular parts of _X_. ~~~~ foreach([I in 0..N1, J in I..N1], Z[I,J] <== X[I,J] - X[I,J]) ~~~~ + Add all elements of a matrix by using _Sum_ as an accumulator. ~~~~ foreach([I in 0..N1, J in 0..N1], plus(X[I,J]), 0, Sum) ~~~~ Notice that the library does not support all known matrix operations. Please contact the YAP maintainers if you require extra functionality. + _X_ <== array[ _Dim1_,..., _Dimn_] of _Objects_ The of/2 operator can be used to create a new array of _Objects_. The objects supported are: + `Unbound Variable` create an array of free variables + `ints ` create an array of integers + `floats ` create an array of floating-point numbers + `_I_: _J_` create an array with integers from _I_ to _J_ + `[..]` create an array from the values in a list The dimensions can be given as an integer, and the matrix will be indexed `C`-style from `0..( _Max_-1)`, or can be given as an interval ` _Base_.. _Limit_`. In the latter case, matrices of integers and of floating-point numbers should have the same _Base_ on every dimension. */ /* A matrix is an object with integer or floating point numbers. A matrix may have a number of dimensions. These routines implement a number of routine manipulation procedures. '$matrix'(Type,D1,D2,...,Dn,data(......)) Type = int, float Operations: typedef enum { MAT_SUM=0, MAT_SUB=1, MAT_TIMES=2, MAT_DIV=3, MAT_IDIV=4, MAT_ZDIV=5 } op_type; */ /** @pred ?_LHS_ <== ?_RHS_ is semidet General matrix assignment operation. It evaluates the right-hand side according to the left-hand side and to the matrix: + if _LHS_ is part of an integer or floating-point matrix, perform non-backtrackable assignment. + other unify left-hand side and right-hand size. The right-hand side supports the following operators: + `[]/2` written as _M_[ _Offset_]: obtain an element or list of elements of matrix _M_ at offset _Offset_. + `matrix/1` create a vector from a list + `matrix/2` create a matrix from a list. Options are: + dim= a list of dimensions + type= integers, floating-point or terms + base= a list of base offsets per dimension (all must be the same for arrays of integers and floating-points + `matrix/3` create matrix giving two options + `dim/1` list with matrix dimensions + `nrow/1` number of rows in bi-dimensional matrix + `ncol/1` number of columns in bi-dimensional matrix + `length/1` size of a matrix + `size/1` size of a matrix + `max/1` maximum element of a numeric matrix + `maxarg/1` argument of maximum element of a numeric matrix + `min/1` minimum element of a numeric matrix + `minarg/1` argument of minimum element of a numeric matrix + `list/1` represent matrix as a list + `lists/2` represent matrix as list of embedded lists + `../2` _I_.. _J_ generates a list with all integers from _I_ to _J_, included. + `+/2` add two numbers, add two matrices element-by-element, or add a number to all elements of a matrix or list. + `-/2 ` subtract two numbers, subtract two matrices or lists element-by-element, or subtract a number from all elements of a matrix or list + `* /2` multiply two numbers, multiply two matrices or lists element-by-element, or multiply a number from all elements of a matrix or list + `log/1` natural logarithm of a number, matrix or list + `exp/1 ` natural exponentiation of a number, matrix or list */ /** @pred matrix_add(+ _Matrix_,+ _Position_,+ _Operand_) Add _Operand_ to the element of _Matrix_ at position _Position_. */ /** @pred matrix_agg_cols(+ _Matrix_,+Operator,+ _Aggregate_) If _Matrix_ is a n-dimensional matrix, unify _Aggregate_ with the one dimensional matrix where each element is obtained by adding all Matrix elements with same first index. Currently, only addition is supported. */ /** @pred matrix_agg_lines(+ _Matrix_,+Operator,+ _Aggregate_) If _Matrix_ is a n-dimensional matrix, unify _Aggregate_ with the n-1 dimensional matrix where each element is obtained by adding all _Matrix_ elements with same last n-1 index. Currently, only addition is supported. */ /** @pred matrix_arg_to_offset(+ _Matrix_,+ _Position_,- _Offset_) Given matrix _Matrix_ return what is the numerical _Offset_ of the element at _Position_. */ /** @pred matrix_column(+ _Matrix_,+ _Column_,- _NewMatrix_) Select from _Matrix_ the column matching _Column_ as new matrix _NewMatrix_. _Column_ must have one less dimension than the original matrix. */ /** @pred matrix_dec(+ _Matrix_,+ _Position_) Decrement the element of _Matrix_ at position _Position_. */ /** @pred matrix_dec(+ _Matrix_,+ _Position_,- _Element_) Decrement the element of _Matrix_ at position _Position_ and unify with _Element_. */ /** @pred matrix_dims(+ _Matrix_,- _Dims_) Unify _Dims_ with a list of dimensions for _Matrix_. */ /** @pred matrix_expand(+ _Matrix_,+ _NewDimensions_,- _New_) Expand _Matrix_ to occupy new dimensions. The elements in _NewDimensions_ are either 0, for an existing dimension, or a positive integer with the size of the new dimension. */ /** @pred matrix_get(+ _Matrix_,+ _Position_,- _Elem_) Unify _Elem_ with the element of _Matrix_ at position _Position_. */ /** @pred matrix_get(+ _Matrix_[+ _Position_],- _Elem_) Unify _Elem_ with the element _Matrix_[ _Position_]. */ /** @pred matrix_inc(+ _Matrix_,+ _Position_) Increment the element of _Matrix_ at position _Position_. */ /** @pred matrix_inc(+ _Matrix_,+ _Position_,- _Element_) Increment the element of _Matrix_ at position _Position_ and unify with _Element_. */ /** @pred matrix_max(+ _Matrix_,+ _Max_) Unify _Max_ with the maximum in matrix _Matrix_. */ /** @pred matrix_maxarg(+ _Matrix_,+ _Maxarg_) Unify _Max_ with the position of the maximum in matrix _Matrix_. */ /** @pred matrix_min(+ _Matrix_,+ _Min_) Unify _Min_ with the minimum in matrix _Matrix_. */ /** @pred matrix_minarg(+ _Matrix_,+ _Minarg_) Unify _Min_ with the position of the minimum in matrix _Matrix_. */ /** @pred matrix_ndims(+ _Matrix_,- _Dims_) Unify _NDims_ with the number of dimensions for _Matrix_. */ /** @pred matrix_new(+ _Type_,+ _Dims_,+ _List_,- _Matrix_) Create a new matrix _Matrix_ of type _Type_, which may be one of `ints` or `floats`, with dimensions _Dims_, and initialized from list _List_. */ /** @pred matrix_new(+ _Type_,+ _Dims_,- _Matrix_) Create a new matrix _Matrix_ of type _Type_, which may be one of `ints` or `floats`, and with a list of dimensions _Dims_. The matrix will be initialized to zeros. ~~~~~ ?- matrix_new(ints,[2,3],Matrix). Matrix = {..} ~~~~~ Notice that currently YAP will always write a matrix of numbers as `{..}`. */ /** @pred matrix_new_set(? _Dims_,+ _OldMatrix_,+ _Value_,- _NewMatrix_) Create a new matrix _NewMatrix_ of type _Type_, with dimensions _Dims_. The elements of _NewMatrix_ are set to _Value_. */ /** @pred matrix_offset_to_arg(+ _Matrix_,- _Offset_,+ _Position_) Given a position _Position _ for matrix _Matrix_ return the corresponding numerical _Offset_ from the beginning of the matrix. */ /** @pred matrix_op(+ _Matrix1_,+ _Matrix2_,+ _Op_,- _Result_) _Result_ is the result of applying _Op_ to matrix _Matrix1_ and _Matrix2_. Currently, only addition (`+`) is supported. */ /** @pred matrix_op_to_all(+ _Matrix1_,+ _Op_,+ _Operand_,- _Result_) _Result_ is the result of applying _Op_ to all elements of _Matrix1_, with _Operand_ as the second argument. Currently, only addition (`+`), multiplication (`\*`), and division (`/`) are supported. */ /** @pred matrix_op_to_cols(+ _Matrix1_,+ _Cols_,+ _Op_,- _Result_) _Result_ is the result of applying _Op_ to all elements of _Matrix1_, with the corresponding element in _Cols_ as the second argument. Currently, only addition (`+`) is supported. Notice that _Cols_ will have n-1 dimensions. */ /** @pred matrix_op_to_lines(+ _Matrix1_,+ _Lines_,+ _Op_,- _Result_) _Result_ is the result of applying _Op_ to all elements of _Matrix1_, with the corresponding element in _Lines_ as the second argument. Currently, only division (`/`) is supported. */ /** @pred matrix_select(+ _Matrix_,+ _Dimension_,+ _Index_,- _New_) Select from _Matrix_ the elements who have _Index_ at _Dimension_. */ /** @pred matrix_set(+ _Matrix_,+ _Position_,+ _Elem_) Set the element of _Matrix_ at position _Position_ to _Elem_. */ /** @pred matrix_set(+ _Matrix_[+ _Position_],+ _Elem_) Set the element of _Matrix_[ _Position_] to _Elem_. */ /** @pred matrix_set_all(+ _Matrix_,+ _Elem_) Set all element of _Matrix_ to _Elem_. */ /** @pred matrix_shuffle(+ _Matrix_,+ _NewOrder_,- _Shuffle_) Shuffle the dimensions of matrix _Matrix_ according to _NewOrder_. The list _NewOrder_ must have all the dimensions of _Matrix_, starting from 0. */ /** @pred matrix_size(+ _Matrix_,- _NElems_) Unify _NElems_ with the number of elements for _Matrix_. */ /** @pred matrix_sum(+ _Matrix_,+ _Sum_) Unify _Sum_ with the sum of all elements in matrix _Matrix_. */ /** @pred matrix_to_list(+ _Matrix_,- _Elems_) Unify _Elems_ with the list including all the elements in _Matrix_. */ /** @pred matrix_transpose(+ _Matrix_,- _Transpose_) Transpose matrix _Matrix_ to _Transpose_. Equivalent to: ~~~~~ matrix_transpose(Matrix,Transpose) :- matrix_shuffle(Matrix,[1,0],Transpose). ~~~~~ */ /** @pred matrix_type(+ _Matrix_,- _Type_) Unify _NElems_ with the type of the elements in _Matrix_. */ :- load_foreign_files([matrix], [], init_matrix). :- multifile rhs_opaque/1, array_extension/2. :- meta_predicate foreach(+,0), foreach(+,2, +, -). :- use_module(library(maplist)). :- use_module(library(mapargs)). :- use_module(library(lists)). ( X <== '[]'(Dims0, array) of V ) :- var(V), !, foldl( norm_dim, Dims0, Dims, Bases, 1, Size ), length( L, Size ), X <== matrix( L, [dim=Dims,base=Bases] ). ( X <== '[]'(Dims0, array) of ints ) :- !, foldl( norm_dim, Dims0, Dims, Bases, 1, _Size ), matrix_new( ints , Dims, X ), matrix_base(X, Bases). ( X <== '[]'(Dims0, array) of floats ) :- !, foldl( norm_dim, Dims0, Dims, Bases, 1, _Size ), matrix_new( floats , Dims, X ), matrix_base(X, Bases). ( X <== '[]'(Dims0, static.array) of floats ) :- atom(X), !, foldl( norm_dim, Dims0, Dims, Bases, 1, _Size ), static_array( Size, floats, X ). ( X <== '[]'(Dims0, array) of (I:J) ) :- !, foldl( norm_dim, Dims0, Dims, Bases, 1, Size ), matrix_seq(I, J, Dims, X), matrixn_size(X, Size), matrix_base(X, Bases). ( X <== '[]'(Dims0, array) of L ) :- length( L, Size ), !, foldl( norm_dim, Dims0, Dims, Bases, 1, Size ), X <== matrix( L, [dim=Dims,base=Bases] ). ( X <== '[]'(Dims0, array) of Pattern ) :- !, array_extension(Pattern, Goal), foldl( norm_dim, Dims0, Dims, Bases, 1, Size ), call(Goal, Pattern, Dims, Size, L), X <== matrix( L, [dim=Dims,base=Bases] ). ( LHS <== RHS ) :- rhs(RHS, R), set_lhs( LHS, R). norm_dim( I..J, D, I, P0, P) :- !, D is J+1-I, P is P0*D. norm_dim( I, I, 0, P0, P ) :- P is P0*I. rhs(RHS, RHS) :- var(RHS), !. % base case rhs(A, A) :- atom(A), !. rhs(RHS, RHS) :- number(RHS), !. rhs(RHS, RHS) :- opaque(RHS), !. rhs(RHS, RHS) :- RHS = '$matrix'(_, _, _, _, _), !. rhs(matrix(List), RHS) :- !, rhs( List, A1), new_matrix(A1, [], RHS). rhs(matrix(List, Opt1), RHS) :- !, rhs( List, A1), new_matrix(A1, Opt1, RHS). rhs(matrix(List, Opt1, Opt2), RHS) :- !, rhs( List, A1), new_matrix(A1, [Opt1, Opt2], RHS). rhs(dim(RHS), Dims) :- !, rhs(RHS, X1), matrix_dims( X1, Dims ). rhs(dims(RHS), Dims) :- !, rhs(RHS, X1), matrix_dims( X1, Dims ). rhs(nrow(RHS), NRow) :- !, rhs(RHS, X1), matrix_dims( X1, [NRow,_] ). rhs(ncol(RHS), NCol) :- !, rhs(RHS, X1), matrix_dims( X1, [_,NCol] ). rhs(length(RHS), Size) :- !, rhs(RHS, X1), matrix_size( X1, Size ). rhs(size(RHS), Size) :- !, rhs(RHS, X1), matrix_size( X1, Size ). rhs(max(RHS), Size) :- !, rhs(RHS, X1), matrix_max( X1, Size ). rhs(min(RHS), Size) :- !, rhs(RHS, X1), matrix_min( X1, Size ). rhs(maxarg(RHS), Size) :- !, rhs(RHS, X1), matrix_maxarg( X1, Size ). rhs(minarg(RHS), Size) :- !, rhs(RHS, X1), matrix_minarg( X1, Size ). rhs(list(RHS), List) :- !, rhs(RHS, X1), matrix_to_list( X1, List ). rhs(lists(RHS), List) :- !, rhs(RHS, X1), matrix_to_lists( X1, List ). rhs('[]'([Args], floats(RHS)), Val) :- integer(RHS), integer(Args), !, get_float_from_address(RHS,Args,Val). rhs('[]'(Args, RHS), Val) :- !, rhs(RHS, X1), matrix_dims( X1, Dims, Bases), maplist( index(Range), Args, Dims, Bases, NArgs), ( var(Range) -> matrix_get( X1, NArgs, Val ) ; matrix_get_range( X1, NArgs, Val ) ). rhs('[]'([Args], floats(RHS)), Val) :- atom(RHS), integer(Args), !, array_element(RHS,Args,Val). rhs('[]'(Args, RHS), Val) :- !, rhs(RHS, X1), matrix_dims( X1, Dims, Bases), maplist( index(Range), Args, Dims, Bases, NArgs), ( var(Range) -> array_element( X1, NArgs, Val ) ; matrix_get_range( X1, NArgs, Val ) ). rhs('..'(I, J), [I1|Is]) :- !, rhs(I, I1), rhs(J, J1), once( foldl(inc, Is, I1, J1) ). rhs([H|T], [NH|NT]) :- !, rhs(H, NH), rhs(T, NT). rhs(log(RHS), Logs ) :- !, rhs(RHS, X1), matrix_to_logs( X1, Logs ). rhs(exp(RHS), Logs ) :- !, rhs(RHS, X1), matrix_to_exps( X1, Logs ). rhs(sum(RHS), Logs ) :- !, rhs(RHS, X1), matrix_sum( X1, Logs ). rhs(S, NS) :- rhs_opaque( S ), !, S = NS. rhs(E1+E2, V) :- !, rhs(E1, R1), rhs(E2, R2), mplus(R1, R2, V). rhs(E1-E2, V) :- !, rhs(E1, R1), rhs(E2, R2), msub(R1, R2, V). rhs(S, NS) :- S =.. [N|As], maplist(rhs, As, Bs), NS =.. [N|Bs]. set_lhs(V, R) :- var(V), !, V = R. set_lhs(V, R) :- number(V), !, V = R. set_lhs('[]'([Args], floats(RHS)), Val) :- !, integer(RHS), integer(Args), set_float_from_address(RHS,Args,Val). set_lhs('[]'(Args, M), Val) :- matrix_dims( M, Dims, Bases), maplist( index(Range), Args, Dims, Bases, NArgs), ( var(Range) -> matrix_set( M, NArgs, Val ) ; matrix_set_range( M, NArgs, Val ) ). % % ranges of arguments % index(Range, V, M, Base, Indx) :- var(V), !, Max is (M-1)+Base, index(Range, Base..Max, M, Base, Indx). index(Range, '*', M, Base, Indx) :- !, Max is (M-1)+Base, index(Range, Base..Max, M, Base, Indx). index(Range, Exp, M, _Base, Indx) :- !, index(Exp, M, Indx0), ( integer(Indx0) -> Indx = Indx0 ; Indx0 = [Indx] -> true ; Indx0 = Indx, Range = range ). index(I, _M, I ) :- integer(I), !. index(I..J, _M, [I|O] ) :- !, I1 is I, J1 is J, once( foldl(inc, O, I1, J1) ). index(I:J, _M, [I|O] ) :- !, I1 is I, J1 is J, once( foldl(inc, O, I1, J1) ). index(I+J, M, O ) :- !, index(I, M, I1), index(J, M, J1), add_index(I1, J1, O). index(I-J, M, O ) :- !, index(I, M, I1), index(J, M, J1), sub_index(I1, J1, O). index(I*J, M, O ) :- !, index(I, M, I1), index(J, M, J1), O is I1*J1. index(I div J, M, O ) :- !, index(I, M, I1), index(J, M, J1), O is I1 div J1. index(I rem J, M, O ) :- !, index(I, M, I1), index(J, M, J1), O is I1 rem J1. index(I, M, NI ) :- maplist(indx(M), I, NI). indx(M, I, NI) :- index(I, M, NI). add_index(I1, J1, O) :- integer(I1), integer(J1), !, O is I1+J1. add_index(I1, J1, O) :- integer(I1), !, maplist(plus(I1), J1, O). add_index(I1, J1, O) :- integer(J1), !, maplist(plus(J1), I1, O). add_index(I1, J1, O) :- ord_union(I1, J1, O). sub_index(I1, J1, O) :- integer(I1), integer(J1), !, O is I1-J1. sub_index(I1, J1, O) :- integer(I1), !, maplist(rminus(I1), J1, O). sub_index(I1, J1, O) :- integer(J1), !, maplist(minus(J1), I1, O). sub_index(I1, J1, O) :- ord_subtract(I1, J1, O). minus(X, Y, Z) :- Z is X-Y. rminus(X, Y, Z) :- Z is Y-X. times(X, Y, Z) :- Z is Y*X. div(X, Y, Z) :- Z is X/Y. rdiv(X, Y, Z) :- Z is Y/X. zdiv(X, Y, Z) :- (X == 0 -> Z = 0 ; X == 0.0 -> Z = 0.0 ; Z is X / Y ). mplus(I1, I2, V) :- number(I1) -> ( number(I2) -> V is I1+I2 ; matrix(I2) -> matrix_op_to_all(I1, +, I2, V) ; is_list(I2) -> maplist(plus(I1), I2, V) ; V = I1+I2 ) ; matrix(I1) -> ( number(I2) -> matrix_op_to_all(I1, +, I2, V) ; matrix(I2) -> matrix_op(I1, I2, +, V) ; V = I1+I2 ) ; is_list(I1) -> ( number(I2) -> maplist(plus(I2), I1, V) ; is_list(I2) -> maplist(plus, I1, I2, V) ; V = I1+I2 ) ; V = I1 +I2. msub(I1, I2, V) :- number(I1) -> ( number(I2) -> V is I1-I2 ; matrix(I2) -> matrix_op_to_all(I1, -, NI2, V) ; is_list(I2) -> maplist(minus(I1), I2, V) ; V = I1-I2 ) ; matrix(I1) -> ( number(I2) -> NI2 is -I2, matrix_op_to_all(I1, +, NI2, V) ; matrix(I2) -> matrix_op(I1, I2, -, V) ; V = I1-I2 ) ; is_list(I1) -> ( number(I2) -> NI2 is -I2, maplist(plus(NI2), I1, V) ; is_list(I2) -> maplist(minus, I1, I2, V) ; V = I1-I2 ) ; V = I1-I2. mtimes(I1, I2, V) :- number(I1) -> ( number(I2) -> V is I1*I2 ; matrix(I2) -> matrix_op_to_all(I1, *, I2, V) ; is_list(I2) -> maplist(times(I1), I2, V) ; V = I1*I2 ) ; matrix(I1) -> ( number(I2) -> matrix_op_to_all(I1, *, I2, V) ; matrix(I2) -> matrix_op(I1, I2, *, V) ; V = I1*I2 ) ; is_list(I1) -> ( number(I2) -> maplist(times(I2), I1, V) ; is_list(I2) -> maplist(times, I1, I2, V) ; V = I1*I2 ) ; V = I1 *I2. % % three types of matrix: integers, floats and general terms. % matrix_new(terms.terms,Dims, '$matrix'(Dims, NDims, Size, Offsets, Matrix) ) :- length(Dims,NDims), foldl(size, Dims, 1, Size), maplist(zero, Dims, Offsets), functor( Matrix, c, Size). matrix_new(opaque.ints,Dims,Matrix) :- length(Dims,NDims), new_ints_matrix_set(NDims, Dims, 0, Matrix). matrix_new(opaque.floats,Dims,Matrix) :- length(Dims,NDims), new_floats_matrix_set(NDims, Dims, 0.0, Matrix). matrix_new(array.Type(Size), Dims, Data, '$array'(Id) ) :- length(Dims,NDims), foldl(size, Dims, 1, Size), maplist(zero, Dims, Offsets), functor( Matrix, c, Size), new_array(Size,Type,Dims,Data), matrix_new(terms, Dims, Data, '$matrix'(Dims, NDims, Size, Offsets, Matrix) ) :- length(Dims,NDims), foldl(size, Dims, 1, Size), maplist(zero, Dims, Offsets), functor( Matrix, c, Size), Matrix =.. [c|Data]. matrix_new(ints,Dims,Data,Matrix) :- length(Dims,NDims), new_ints_matrix(NDims, Dims, Data, Matrix). matrix_new(floats,Dims,Data,Matrix) :- length(Dims,NDims), new_floats_matrix(NDims, Dims, Data, Matrix). matrix_dims( Mat, Dims) :- ( opaque(Mat) -> matrixn_dims( Mat, Dims ) ; Mat = '$matrix'( Dims, _, _, _, _) ). matrix_dims( Mat, Dims, Bases) :- ( opaque(Mat) -> matrixn_dims( Mat, Dims, Bases ) ; Mat = '$matrix'( Dims, _, _, Bases, _) ). matrix_ndims( Mat, NDims) :- ( opaque(Mat) -> matrixn_ndims( Mat, NDims ) ; Mat = '$matrix'( _, NDims, _, _, _) ). matrix_size( Mat, Size) :- ( opaque(Mat) -> matrixn_size( Mat, Size ) ; Mat = '$matrix'( _, _, Size, _, _) ). matrix_to_list( Mat, ToList) :- ( opaque(Mat) -> matrixn_to_list( Mat, ToList ) ; Mat = '$matrix'( _, _, _, _, M), M=.. [_|ToList] ). matrix_to_lists( Mat, ToList) :- matrix_dims( Mat, [D|Dims] ), D1 is D-1, foreach( I in 0..D1, matrix_slicer( Dims, Mat, [I|L]-L), ToList, [] ). matrix_slicer( [_], M, Pos-[_], [O|L0], L0) :- !, O <== '[]'(Pos,M). matrix_slicer( [D|Dims], M, Pos-[I|L], [O|L0], L0) :- D1 is D-1, foreach( I in 0..D1 , L^matrix_slicer( Dims, M, Pos-L), O, [] ). matrix_get( Mat, Pos, El) :- ( opaque(Mat) -> matrixn_get( Mat, Pos, El ) ; m_get(Mat, Pos, El) ). matrix_get_range( Mat, Pos, Els) :- slice(Pos, Keys), maplist( matrix_get(Mat), Keys, Els). slice([], [[]]). slice([[H|T]|Extra], Els) :- !, slice(Extra, Els0), foldl(add_index_prefix( Els0 ), [H|T], Els, [] ). slice([H|Extra], Els) :- !, slice(Extra, Els0), add_index_prefix( Els0 , H, Els, [] ). add_index_prefix( [] , _H ) --> []. add_index_prefix( [L|Els0] , H ) --> [[H|L]], add_index_prefix( Els0 , H ). matrix_set_range( Mat, Pos, Els) :- slice(Pos, Keys), maplist( matrix_set(Mat), Keys, Els). matrix_set( Mat, Pos, El) :- ( opaque(Mat) -> matrixn_set( Mat, Pos, El ) ; m_set(Mat, Pos, El) ). matrix_new_set(ints,Dims,Elem,Matrix) :- length(Dims,NDims), new_ints_matrix_set(NDims, Dims, Elem, Matrix). matrix_new_set(floats,Dims,Elem,Matrix) :- length(Dims,NDims), new_floats_matrix_set(NDims, Dims, Elem, Matrix). matrix_type(Matrix,Type) :- ( matrix_type_as_number(Matrix, 0) -> Type = ints ; opaque( Matrix ) -> Type = floats ; Type = terms ). matrix_base(Matrix, Bases) :- ( opaque( Matrix ) -> maplist('='(Base), Bases), matrixn_set_base( Matrix, Base ) ; nb_setarg(4, Matrix, Bases ) ). matrix_arg_to_offset(M, Index, Offset) :- ( opaque(M) -> matrixn_arg_to_offset( M, Index, Offset ) ; M = '$matrix'(Dims, _, Size, Bases, _) -> foldl2(indx, Index, Dims, Bases, Size, _, 0, Offset) ). matrix_offset_to_arg(M, Offset, Index) :- ( opaque(M) -> matrixn_offset_to_arg( M, Offset, Index ) ; M = '$matrix'(Dims, _, Size, Bases, _) -> foldl2(offset, Index, Dims, Bases, Size, _, Offset, _) ). matrix_max(M, Max) :- ( opaque(M) -> matrixn_max( M, Max ) ; M = '$matrix'(_, _, _, _, C) -> arg(1,C,V0), foldargs(max, M, V0, Max) ; M = [V0|L], foldl(max, L, V0, Max) ). max(New, Old, Max) :- ( New >= Old -> New = Max ; Old = Max ). matrix_maxarg(M, MaxArg) :- ( opaque(M) -> matrixn_maxarg( M, MaxArg ); M = '$matrix'(_, _, _, _, C) -> arg(1,C,V0), foldargs(maxarg, M, V0-0-0, _-Offset-_), matrix_offset_to_arg(M, Offset, MaxArg) ; M = [V0|L], foldl(maxarg, L, V0-0-1, _Max-Off-_ ), MaxArg = [Off] ). maxarg(New, Old-OPos-I0, Max-MPos-I) :- I is I0+1, ( New > Old -> New = Max, MPos = I0 ; Old = Max, MPos = OPos ). matrix_min(M, Min) :- ( opaque(M) -> matrixn_min( M, Min ) ; M = '$matrix'(_, _, _, _, C) -> arg(1,C,V0), foldargs(min, M, V0, Max) ; M = [V0|L], foldl(min, L, V0, Max) ). min(New, Old, Max) :- ( New =< Old -> New = Max ; Old = Max ). matrix_minarg(M, MinArg) :- ( opaque(M) -> matrixn_minarg( M, MinArg ); M = '$matrix'(_, _, _, _, C) -> arg(1,C,V0), foldargs(minarg, M, V0-0-0, _-Offset-_), matrix_offset_to_arg(M, Offset, MinArg) ; M = [V0|L], foldl(minarg, L, V0-0-1, _Min-Off-_ ), MinArg = [Off] ). minarg(New, Old-OPos-I0, Min-MPos-I) :- I is I0+1, ( New < Old -> New = Min, MPos = I0 ; Old = Min, MPos = OPos ). matrix_to_logs(M, LogM) :- ( opaque(M) -> matrixn_to_logs( M, LogM ) ; M = '$matrix'(A, B, D, E, C) -> LogM = '$matrix'(A, B, D, E, LogC), mapargs(log, C, LogC) ; M = [V0|L] -> maplist(log, [V0|L], LogM ) ; LogM is log(M) ). log(X, Y) :- Y is log(X). matrix_to_exps(M, ExpM) :- ( opaque(M) -> matrixn_to_exps( M, ExpM ) ; M = '$matrix'(A, B, D, E, C) -> ExpM = '$matrix'(A, B, D, E, ExpC), mapargs(exp, C, ExpC) ; M = [V0|L] -> maplist(exp, [V0|L], ExpM ) ; ExpM is exp(M) ). exp(X, Y) :- Y is exp(X). matrix_agg_lines(M1,+,NM) :- do_matrix_agg_lines(M1,0,NM). /* other operations: *, logprod */ matrix_agg_cols(M1,+,NM) :- do_matrix_agg_cols(M1,0,NM). /* other operations: *, logprod */ matrix_op(M1,M2,+,NM) :- ( opaque(M1), opaque(M2) -> do_matrix_op(M1,M2,0,NM) ; matrix_m(M1, '$matrix'(A,B,D,E,C1)), matrix_m(M2, '$matrix'(A,B,D,E,C2)), mapargs(plus, C1, C2, C), NM = '$matrix'(A,B,D,E,C) ). matrix_op(M1,M2,-,NM) :- ( opaque(M1), opaque(M2) -> do_matrix_op(M1,M2,1,NM) ; matrix_m(M1, '$matrix'(A,B,D,E,C1)), matrix_m(M2, '$matrix'(A,B,D,E,C2)), mapargs(minus, C1, C2, C), NM = '$matrix'(A,B,D,E,C) ). matrix_op(M1,M2,*,NM) :- ( opaque(M1), opaque(M2) -> do_matrix_op(M1,M2,2,NM) ; matrix_m(M1, '$matrix'(A,B,D,E,C1)), matrix_m(M2, '$matrix'(A,B,D,E,C2)), mapargs(times, C1, C2, C), NM = '$matrix'(A,B,D,E,C) ). matrix_op(M1,M2,/,NM) :- ( opaque(M1), opaque(M2) -> do_matrix_op(M1,M2,3,NM) ; matrix_m(M1, '$matrix'(A,B,D,E,C1)), matrix_m(M2, '$matrix'(A,B,D,E,C2)), mapargs(div, C1, C2, C), NM = '$matrix'(A,B,D,E,C) ). matrix_op(M1,M2,zdiv,NM) :- ( opaque(M1), opaque(M2) -> do_matrix_op(M1,M2,5,NM) ; matrix_m(M1, '$matrix'(A,B,D,E,C1)), matrix_m(M2, '$matrix'(A,B,D,E,C2)), mapargs(zdiv, C1, C2, C), NM = '$matrix'(A,B,D,E,C) ). matrix_op_to_all(M1,+,Num,NM) :- ( opaque(M1) -> do_matrix_op_to_all(M1,0,Num,NM) ; M1 = '$matrix'(A,B,D,E,C), mapargs(plus(Num), C, NC), NM = '$matrix'(A,B,D,E,NC) ). matrix_op_to_all(M1,-,Num,NM) :- ( opaque(M1) -> do_matrix_op_to_all(M1,1,Num,NM) ; M1 = '$matrix'(A,B,D,E,C), mapargs(minus(Num), C, NC), NM = '$matrix'(A,B,D,E,NC) ). matrix_op_to_all(M1,*,Num,NM) :- ( opaque(M1) -> do_matrix_op_to_all(M1,2,Num,NM) ; M1 = '$matrix'(A,B,D,E,C), mapargs(times(Num), C, NC), NM = '$matrix'(A,B,D,E,NC) ). matrix_op_to_all(M1,/,Num,NM) :- % can only use floats. FNum is float(Num), ( opaque(M1) -> do_matrix_op_to_all(M1,3,FNum,NM) ; M1 = '$matrix'(A,B,D,E,C), mapargs(div(Num), C, NC), NM = '$matrix'(A,B,D,E,NC) ). /* other operations: *, logprod */ matrix_op_to_lines(M1,M2,/,NM) :- do_matrix_op_to_lines(M1,M2,3,NM). /* other operations: *, logprod */ matrix_op_to_cols(M1,M2,+,NM) :- do_matrix_op_to_cols(M1,M2,0,NM). /* other operations: *, logprod */ matrix_transpose(M1,M2) :- matrix_shuffle(M1,[1,0],M2). size(N0, N1, N2) :- N2 is N0*N1. % use 1 to get access to matrix m_get('$matrix'(Dims, _, Sz, Bases, M), Indx, V) :- foldl2(indx, Indx, Dims, Bases, Sz, _, 1, Offset), arg(Offset, M, V). m_set('$matrix'(Dims, _, Sz, Bases, M), Indx, V) :- foldl2(indx, Indx, Dims, Bases, Sz, _, 1, Offset), arg(Offset, M, V). indx( I, Dim, Base, BlkSz, NBlkSz, I0, IF) :- NBlkSz is BlkSz div Dim , IF is (I-Base)*NBlkSz + I0. offset( I, Dim, BlkSz, NBlkSz, Base, I0, IF) :- NBlkSz is BlkSz div Dim, I is I0 div NBlkSz + Base, IF is I0 rem NBlkSz. inc(I1, I, I1) :- I1 is I+1. new_matrix(M0, Opts0, M) :- opaque(M), !, matrix_to_list(M0, L), new_matrix(L, Opts0, M). new_matrix('$matrix'(_,_,_,_,C), Opts0, M) :- !, C =..[_|L], new_matrix(L, Opts0, M). new_matrix(C, Opts0, M) :- functor(C, c, _), !, C =..[_|L], new_matrix(L, Opts0, M). new_matrix(List, Opts0, M) :- foldl2(el_list(MDims), List, Flat, [], 0, Dim), !, fix_opts(Opts0, Opts), foldl2(process_new_opt, Opts, Type, TypeF, [Dim|MDims], Dims, Base), ( var(TypeF) -> guess_type( Flat, Type ) ; true ), matrix_new( Type, Dims, Flat, M), ( nonvar(Base) -> matrix_base(M, Base); true ). new_matrix([H|List], Opts0, M) :- length( [H|List], Size), fix_opts(Opts0, Opts), foldl2(process_new_opt(Base), Opts, Type, TypeF, [Size], Dims), ( var(TypeF) -> guess_type( [H|List], Type ) ; true ), matrix_new( Type, Dims, [H|List], M), ( nonvar(Base) -> matrix_base(M, Base); true ). fix_opts(V, _) :- var(V), !, throw(error(instantiation_error, V)). fix_opts(A=B, [A=B]). fix_opts(A, A) :- is_list(A), !. fix_opts(V, _) :- var(V), !, throw(error(domain_error(options=V), new_matrix)). guess_type( List, Type ) :- maplist( integer, List), !, Type = ints. guess_type( List, Type ) :- maplist( number, List), !, Type = floats. guess_type( _List, terms ). process_new_opt(_Base, dim=Dim, Type, Type, _, Dim) :- !. process_new_opt(_Base, type=Type, _, Type, Dim, Dim) :- !. process_new_opt( Base, base=Base, Type, Type, Dim, Dim) :- !. process_new_opt(_Base, Opt, Type, Type, Dim, Dim) :- throw(error(domain_error(opt=Opt), new_matrix)). el_list(_, V, _Els, _NEls, _I0, _I1) :- var(V), !, fail. el_list([N|Extra], El, Els, NEls, I0, I1) :- foldl2(el_list(Extra), El, Els, NEls, 0, N), !, I1 is I0+1. el_list([N], El, Els, NEls, I0, I1) :- El = [_|_], length(El, N), append(El, NEls, Els), I1 is I0+1. foreach( Domain, Goal) :- strip_module(Goal, M, Locals^NG), !, term_variables(Domain+Locals, LocalVarsL), LocalVars =.. [vs|LocalVarsL], iterate( Domain, [], LocalVars, M:NG, [], [] ), terms:reset_variables( LocalVars ). foreach( Domain, Goal ) :- strip_module(Goal, M, NG), term_variables(Domain, LocalVarsL), LocalVars =.. [vs|LocalVarsL], iterate( Domain, [], LocalVars, M:NG, [], [] ), terms:reset_variables( LocalVars ). foreach( Domain, Goal, Inp, Out) :- strip_module(Goal, M, Locals^NG), !, term_variables(Domain+Locals, LocalVarsL), LocalVars =.. [vs|LocalVarsL], iterate( Domain, [], LocalVars, M:NG, [], [], Inp, Out). foreach( Domain, Goal, Inp, Out ) :- strip_module(Goal, M, NG), term_variables(Domain, LocalVarsL), LocalVars =.. [vs|LocalVarsL], iterate( Domain, [], LocalVars, M:NG, [], [], Inp, Out ). iterate( [], [], LocalVars, Goal, Vs, Bs ) :- terms:freshen_variables(LocalVars), Vs = Bs, MG <== Goal, once( MG ), terms:reset_variables(LocalVars). iterate( [], [H|Cont], LocalVars, Goal, Vs, Bs ) :- iterate(H, Cont, LocalVars, Goal, Vs, Bs ). iterate( [H|L], [], LocalVars, Goal, Vs, Bs ) :- !, iterate(H, L, LocalVars, Goal, Vs, Bs ). iterate( [H|L], Cont, LocalVars, Goal, Vs, Bs ) :- !, append(L, Cont, LCont), iterate(H, LCont, LocalVars, Goal, Vs, Bs ). iterate( [] ins _A .. _B, [H|L], LocalVars, Goal, Vs, Bs ) :- !, iterate(H, L, LocalVars, Goal, Vs, Bs ). iterate( [] ins _A .. _B, [], LocalVars, Goal, Vs, Bs ) :- !, iterate([], [], LocalVars, Goal, Vs, Bs ). iterate( [V|Ps] ins A..B, Cont, LocalVars, Goal, Vs, Bs ) :- eval(A, Vs, Bs, NA), eval(B, Vs, Bs, NB), ( NA > NB -> true ; A1 is NA+1, iterate( Ps ins NA..NB, Cont, LocalVars, Goal, [V|Vs], [NA|Bs] ), iterate( [V|Ps] ins A1..NB, Cont, LocalVars, Goal, Vs, Bs ) ). iterate( V in A..B, Cont, LocalVars, Goal, Vs, Bs) :- var(V), eval(A, Vs, Bs, NA), eval(B, Vs, Bs, NB), ( NA > NB -> true ; A1 is NA+1, (Cont = [H|L] -> iterate( H, L, LocalVars, Goal, [V|Vs], [NA|Bs] ) ; iterate( [], [], LocalVars, Goal, [V|Vs], [NA|Bs] ) ), iterate( V in A1..NB, Cont, LocalVars, Goal, Vs, Bs ) ). iterate( [], [], LocalVars, Goal, Vs, Bs, Inp, Out ) :- terms:freshen_variables(LocalVars), Vs = Bs, MG <== Goal, once( call(MG, Inp, Out) ), terms:reset_variables(LocalVars). iterate( [], [H|Cont], LocalVars, Goal, Vs, Bs, Inp, Out ) :- iterate(H, Cont, LocalVars, Goal, Vs, Bs, Inp, Out ). iterate( [H|L], [], LocalVars, Goal, Vs, Bs, Inp, Out ) :- !, iterate(H, L, LocalVars, Goal, Vs, Bs, Inp, Out ). iterate( [H|L], Cont, LocalVars, Goal, Vs, Bs, Inp, Out ) :- !, append(L, Cont, LCont), iterate(H, LCont, LocalVars, Goal, Vs, Bs, Inp, Out ). iterate( [] ins _A .. _B, [], LocalVars, Goal, Vs, Bs, Inp, Out ) :- !, iterate([], [], LocalVars, Goal, Vs, Bs, Inp, Out ). iterate( [] ins _A .. _B, [H|L], LocalVars, Goal, Vs, Bs, Inp, Out ) :- !, iterate(H, L, LocalVars, Goal, Vs, Bs, Inp, Out ). iterate( [V|Ps] ins A..B, Cont, LocalVars, Goal, Vs, Bs, Inp, Out ) :- eval(A, Vs, Bs, NA), eval(B, Vs, Bs, NB), ( NA > NB -> Inp = Out ; A1 is NA+1, iterate( Ps ins A..B, Cont, LocalVars, Goal, [V|Vs], [NA|Bs], Inp, Mid ), iterate( [V|Ps] ins A1..NB, Cont, LocalVars, Goal, Vs, Bs, Mid, Out ) ). iterate( V in A..B, Cont, LocalVars, Goal, Vs, Bs, Inp, Out) :- var(V), eval(A, Vs, Bs, NA), eval(B, Vs, Bs, NB), ( NA > NB -> Inp = Out ; A1 is NA+1, (Cont = [H|L] -> iterate( H, L, LocalVars, Goal, [V|Vs], [NA|Bs], Inp, Mid ) ; iterate( [], [], LocalVars, Goal, [V|Vs], [NA|Bs], Inp, Mid ) ), iterate( V in A1..NB, Cont, LocalVars, Goal, Vs, Bs, Mid, Out ) ). eval(I, _Vs, _Bs, I) :- integer(I), !. eval(I, Vs, Bs, NI) :- copy_term(I+Vs, IA+Bs), NI <== IA. matrix_seq(A, B, Dims, M) :- ints(A, B, L), matrix_new(ints, Dims, L, M). ints(A,B,O) :- ( A > B -> O = [] ; O = [A|L], A1 is A+1, ints(A1,B,L) ). zero(_, 0). /** @} */