From 2a24e30a242be0bd30a6540d3f3cde4676bdfb21 Mon Sep 17 00:00:00 2001 From: Joe Groff Date: Sat, 10 Oct 2009 17:45:15 -0500 Subject: [PATCH] store math.matrices.simd matrices in column-major order so that m4.v, frustum, and translation construction are faster --- extra/math/matrices/simd/simd-tests.factor | 68 +++++----- extra/math/matrices/simd/simd.factor | 142 +++++++++++---------- 2 files changed, 106 insertions(+), 104 deletions(-) diff --git a/extra/math/matrices/simd/simd-tests.factor b/extra/math/matrices/simd/simd-tests.factor index 3798c3e98e..965c2bddb5 100644 --- a/extra/math/matrices/simd/simd-tests.factor +++ b/extra/math/matrices/simd/simd-tests.factor @@ -52,10 +52,10 @@ IN: math.matrices.simd.tests [ S{ matrix4 f float-4-array{ - float-4{ 1.0 0.0 0.0 3.0 } - float-4{ 0.0 1.0 0.0 4.0 } - float-4{ 0.0 0.0 1.0 2.0 } - float-4{ 0.0 0.0 0.0 1.0 } + float-4{ 1.0 0.0 0.0 0.0 } + float-4{ 0.0 1.0 0.0 0.0 } + float-4{ 0.0 0.0 1.0 0.0 } + float-4{ 3.0 4.0 2.0 1.0 } } } ] [ float-4{ 3.0 4.0 2.0 0.0 } translation-matrix4 ] unit-test @@ -77,9 +77,9 @@ IN: math.matrices.simd.tests float-4{ 0.0 1.0 0.0 1.0 } pi 1/2. * rotation-matrix4 S{ matrix4 f float-4-array{ - float-4{ 0.0 0.0 1.0 0.0 } + float-4{ 0.0 0.0 -1.0 0.0 } float-4{ 0.0 1.0 0.0 0.0 } - float-4{ -1.0 0.0 0.0 0.0 } + float-4{ 1.0 0.0 0.0 0.0 } float-4{ 0.0 0.0 0.0 1.0 } } } @@ -89,10 +89,10 @@ IN: math.matrices.simd.tests [ S{ matrix4 f float-4-array{ - float-4{ 2.0 0.0 0.0 10.0 } - float-4{ 0.0 3.0 0.0 18.0 } - float-4{ 0.0 0.0 4.0 28.0 } - float-4{ 0.0 0.0 0.0 1.0 } + float-4{ 2.0 0.0 0.0 0.0 } + float-4{ 0.0 3.0 0.0 0.0 } + float-4{ 0.0 0.0 4.0 0.0 } + float-4{ 10.0 18.0 28.0 1.0 } } } ] [ @@ -106,10 +106,10 @@ IN: math.matrices.simd.tests } S{ matrix4 f float-4-array{ - float-4{ 1.0 0.0 0.0 5.0 } - float-4{ 0.0 1.0 0.0 6.0 } - float-4{ 0.0 0.0 1.0 7.0 } - float-4{ 0.0 0.0 0.0 1.0 } + float-4{ 1.0 0.0 0.0 0.0 } + float-4{ 0.0 1.0 0.0 0.0 } + float-4{ 0.0 0.0 1.0 0.0 } + float-4{ 5.0 6.0 7.0 1.0 } } } m4. @@ -118,10 +118,10 @@ IN: math.matrices.simd.tests [ S{ matrix4 f float-4-array{ - float-4{ 3.0 0.0 0.0 5.0 } - float-4{ 0.0 4.0 0.0 6.0 } - float-4{ 0.0 0.0 5.0 7.0 } - float-4{ 0.0 0.0 0.0 2.0 } + float-4{ 3.0 0.0 0.0 0.0 } + float-4{ 0.0 4.0 0.0 0.0 } + float-4{ 0.0 0.0 5.0 0.0 } + float-4{ 5.0 6.0 7.0 2.0 } } } ] [ @@ -135,10 +135,10 @@ IN: math.matrices.simd.tests } S{ matrix4 f float-4-array{ - float-4{ 1.0 0.0 0.0 5.0 } - float-4{ 0.0 1.0 0.0 6.0 } - float-4{ 0.0 0.0 1.0 7.0 } - float-4{ 0.0 0.0 0.0 1.0 } + float-4{ 1.0 0.0 0.0 0.0 } + float-4{ 0.0 1.0 0.0 0.0 } + float-4{ 0.0 0.0 1.0 0.0 } + float-4{ 5.0 6.0 7.0 1.0 } } } m4+ @@ -147,10 +147,10 @@ IN: math.matrices.simd.tests [ S{ matrix4 f float-4-array{ - float-4{ 1.0 0.0 0.0 -5.0 } - float-4{ 0.0 2.0 0.0 -6.0 } - float-4{ 0.0 0.0 3.0 -7.0 } - float-4{ 0.0 0.0 0.0 0.0 } + float-4{ 1.0 0.0 0.0 0.0 } + float-4{ 0.0 2.0 0.0 0.0 } + float-4{ 0.0 0.0 3.0 0.0 } + float-4{ -5.0 -6.0 -7.0 0.0 } } } ] [ @@ -164,10 +164,10 @@ IN: math.matrices.simd.tests } S{ matrix4 f float-4-array{ - float-4{ 1.0 0.0 0.0 5.0 } - float-4{ 0.0 1.0 0.0 6.0 } - float-4{ 0.0 0.0 1.0 7.0 } - float-4{ 0.0 0.0 0.0 1.0 } + float-4{ 1.0 0.0 0.0 0.0 } + float-4{ 0.0 1.0 0.0 0.0 } + float-4{ 0.0 0.0 1.0 0.0 } + float-4{ 5.0 6.0 7.0 1.0 } } } m4- @@ -219,10 +219,10 @@ IN: math.matrices.simd.tests [ S{ matrix4 f float-4-array{ - float-4{ 1/2. 0.0 0.0 0.0 } - float-4{ 0.0 1/2. 0.0 0.0 } - float-4{ 0.0 0.0 -6/4. -10/4. } - float-4{ 0.0 0.0 -1.0 0.0 } + float-4{ 1/2. 0.0 0.0 0.0 } + float-4{ 0.0 1/2. 0.0 0.0 } + float-4{ 0.0 0.0 -6/4. -1.0 } + float-4{ 0.0 0.0 -10/4. 0.0 } } } ] [ diff --git a/extra/math/matrices/simd/simd.factor b/extra/math/matrices/simd/simd.factor index da1149dfec..edbe77781f 100644 --- a/extra/math/matrices/simd/simd.factor +++ b/extra/math/matrices/simd/simd.factor @@ -9,34 +9,34 @@ SPECIALIZED-ARRAY: float-4 IN: math.matrices.simd STRUCT: matrix4 - { rows float-4[4] } ; + { columns float-4[4] } ; INSTANCE: matrix4 immutable-sequence M: matrix4 length drop 4 ; inline -M: matrix4 nth-unsafe rows>> nth-unsafe ; inline +M: matrix4 nth-unsafe columns>> nth-unsafe ; inline M: matrix4 new-sequence 2drop matrix4 (struct) ; inline > 4 firstn ; inline +: columns ( a -- a1 a2 a3 a4 ) + columns>> 4 firstn ; inline -:: set-rows ( c1 c2 c3 c4 c -- c ) - c rows>> :> rows - c1 rows set-first - c2 rows set-second - c3 rows set-third - c4 rows set-fourth +:: set-columns ( c1 c2 c3 c4 c -- c ) + c columns>> :> columns + c1 columns set-first + c2 columns set-second + c3 columns set-third + c4 columns set-fourth c ; inline : make-matrix4 ( quot: ( -- c1 c2 c3 c4 ) -- c ) - matrix4 (struct) swap dip set-rows ; inline + matrix4 (struct) swap dip set-columns ; inline -:: 2map-rows ( a b quot -- c ) +:: 2map-columns ( a b quot -- c ) [ - a rows :> a4 :> a3 :> a2 :> a1 - b rows :> b4 :> b3 :> b2 :> b1 + a columns :> a4 :> a3 :> a2 :> a1 + b columns :> b4 :> b3 :> b2 :> b1 a1 b1 quot call a2 b2 quot call @@ -44,57 +44,57 @@ M: matrix4 new-sequence 2drop matrix4 (struct) ; inline a4 b4 quot call ] make-matrix4 ; inline -: map-rows ( a quot -- c ) - '[ rows _ 4 napply ] make-matrix4 ; inline +: map-columns ( a quot -- c ) + '[ columns _ 4 napply ] make-matrix4 ; inline PRIVATE> -TYPED: m4+ ( a: matrix4 b: matrix4 -- c: matrix4 ) [ v+ ] 2map-rows ; -TYPED: m4- ( a: matrix4 b: matrix4 -- c: matrix4 ) [ v- ] 2map-rows ; -TYPED: m4* ( a: matrix4 b: matrix4 -- c: matrix4 ) [ v* ] 2map-rows ; -TYPED: m4/ ( a: matrix4 b: matrix4 -- c: matrix4 ) [ v/ ] 2map-rows ; +TYPED: m4+ ( a: matrix4 b: matrix4 -- c: matrix4 ) [ v+ ] 2map-columns ; +TYPED: m4- ( a: matrix4 b: matrix4 -- c: matrix4 ) [ v- ] 2map-columns ; +TYPED: m4* ( a: matrix4 b: matrix4 -- c: matrix4 ) [ v* ] 2map-columns ; +TYPED: m4/ ( a: matrix4 b: matrix4 -- c: matrix4 ) [ v/ ] 2map-columns ; -TYPED: m4*n ( a: matrix4 b: float -- c: matrix4 ) [ v*n ] curry map-rows ; -TYPED: m4/n ( a: matrix4 b: float -- c: matrix4 ) [ v/n ] curry map-rows ; -TYPED: n*m4 ( a: float b: matrix4 -- c: matrix4 ) [ n*v ] with map-rows ; -TYPED: n/m4 ( a: float b: matrix4 -- c: matrix4 ) [ n/v ] with map-rows ; +TYPED: m4*n ( a: matrix4 b: float -- c: matrix4 ) [ v*n ] curry map-columns ; +TYPED: m4/n ( a: matrix4 b: float -- c: matrix4 ) [ v/n ] curry map-columns ; +TYPED: n*m4 ( a: float b: matrix4 -- c: matrix4 ) [ n*v ] with map-columns ; +TYPED: n/m4 ( a: float b: matrix4 -- c: matrix4 ) [ n/v ] with map-columns ; TYPED:: m4. ( a: matrix4 b: matrix4 -- c: matrix4 ) [ - a rows :> a4 :> a3 :> a2 :> a1 - b rows :> b4 :> b3 :> b2 :> b1 + a columns :> a4 :> a3 :> a2 :> a1 + b columns :> b4 :> b3 :> b2 :> b1 - a1 first b1 n*v :> c1a - a2 first b1 n*v :> c2a - a3 first b1 n*v :> c3a - a4 first b1 n*v :> c4a + b1 first a1 n*v :> c1a + b2 first a1 n*v :> c2a + b3 first a1 n*v :> c3a + b4 first a1 n*v :> c4a - a1 second b2 n*v c1a v+ :> c1b - a2 second b2 n*v c2a v+ :> c2b - a3 second b2 n*v c3a v+ :> c3b - a4 second b2 n*v c4a v+ :> c4b + b1 second a2 n*v c1a v+ :> c1b + b2 second a2 n*v c2a v+ :> c2b + b3 second a2 n*v c3a v+ :> c3b + b4 second a2 n*v c4a v+ :> c4b - a1 third b3 n*v c1b v+ :> c1c - a2 third b3 n*v c2b v+ :> c2c - a3 third b3 n*v c3b v+ :> c3c - a4 third b3 n*v c4b v+ :> c4c + b1 third a3 n*v c1b v+ :> c1c + b2 third a3 n*v c2b v+ :> c2c + b3 third a3 n*v c3b v+ :> c3c + b4 third a3 n*v c4b v+ :> c4c - a1 fourth b4 n*v c1c v+ - a2 fourth b4 n*v c2c v+ - a3 fourth b4 n*v c3c v+ - a4 fourth b4 n*v c4c v+ + b1 fourth a4 n*v c1c v+ + b2 fourth a4 n*v c2c v+ + b3 fourth a4 n*v c3c v+ + b4 fourth a4 n*v c4c v+ ] make-matrix4 ; -TYPED:: v.m4 ( a: float-4 b: matrix4 -- c: float-4 ) - b rows :> b4 :> b3 :> b2 :> b1 +TYPED:: m4.v ( m: matrix4 v: float-4 -- v': float-4 ) + m columns :> m4 :> m3 :> m2 :> m1 - a first b1 n*v - a second b2 n*v v+ - a third b3 n*v v+ - a fourth b4 n*v v+ ; + v first m1 n*v + v second m2 n*v v+ + v third m3 n*v v+ + v fourth m4 n*v v+ ; -TYPED:: m4.v ( a: matrix4 b: float-4 -- c: float-4 ) - a rows [ b v. ] 4 napply float-4-boa ; +TYPED:: v.m4 ( v: float-4 m: matrix4 -- c: float-4 ) + m columns [ v v. ] 4 napply float-4-boa ; CONSTANT: identity-matrix4 S{ matrix4 f @@ -131,37 +131,37 @@ TYPED: diagonal-matrix4 ( diagonal: float-4 -- matrix: matrix4 ) [ (vmerge) ] bi-curry@ bi* ; inline TYPED: transpose-matrix4 ( matrix: matrix4 -- matrix: matrix4 ) - [ rows vmerge-transpose vmerge-transpose ] make-matrix4 ; + [ columns vmerge-transpose vmerge-transpose ] make-matrix4 ; + +: linear>homogeneous ( v -- v' ) + [ float-4{ t t t f } ] dip float-4{ 0.0 0.0 0.0 1.0 } v? ; inline : scale-matrix4 ( factors -- matrix ) - [ float-4{ t t t f } ] dip float-4{ 0.0 0.0 0.0 1.0 } v? - diagonal-matrix4 ; inline + linear>homogeneous diagonal-matrix4 ; inline : ortho-matrix4 ( factors -- matrix ) float-4{ 1.0 1.0 1.0 1.0 } swap v/ scale-matrix4 ; inline -TYPED:: translation-matrix4 ( offset: float-4 -- matrix: matrix4 ) +TYPED: translation-matrix4 ( offset: float-4 -- matrix: matrix4 ) [ - float-4{ 1.0 1.0 1.0 1.0 } :> diagonal - - offset 0 float-4-with (vmerge) - [ 0 float-4-with swap (vmerge) ] bi@ drop :> z :> y :> x - - diagonal y vmerge-diagonal* - [ x vmerge-diagonal* ] - [ z vmerge-diagonal* ] bi* + linear>homogeneous + [ + float-4{ 1.0 0.0 0.0 0.0 } + float-4{ 0.0 1.0 0.0 0.0 } + float-4{ 0.0 0.0 1.0 0.0 } + ] dip ] make-matrix4 ; TYPED:: rotation-matrix4 ( axis: float-4 theta: float -- matrix: matrix4 ) - ! x*x + c*(1.0 - x*x) x*y*(1.0 - c) - s*z x*z*(1.0 - c) + s*y 0 - ! x*y*(1.0 - c) + s*z y*y + c*(1.0 - y*y) y*z*(1.0 - c) - s*x 0 - ! x*z*(1.0 - c) - s*y y*z*(1.0 - c) + s*x z*z + c*(1.0 - z*z) 0 + ! x*x + c*(1.0 - x*x) x*y*(1.0 - c) + s*z x*z*(1.0 - c) - s*y 0 + ! x*y*(1.0 - c) - s*z y*y + c*(1.0 - y*y) y*z*(1.0 - c) + s*x 0 + ! x*z*(1.0 - c) + s*y y*z*(1.0 - c) - s*x z*z + c*(1.0 - z*z) 0 ! 0 0 0 1 matrix4 (struct) :> triangle-m theta cos :> c theta sin :> s - float-4{ 1.0 -1.0 1.0 0.0 } :> triangle-sign + float-4{ -1.0 1.0 -1.0 0.0 } :> triangle-sign c float-4-with :> cc s float-4-with :> ss @@ -184,7 +184,7 @@ TYPED:: rotation-matrix4 ( axis: float-4 theta: float -- matrix: matrix4 ) triangle-lo { 1 0 3 3 } vshuffle float-4 new - triangle-m set-rows drop + triangle-m set-columns drop diagonal-m triangle-m m4+ ; @@ -194,8 +194,10 @@ TYPED:: frustum-matrix4 ( xy: float-4 near: float far: float -- matrix: matrix4 float-4{ t t f f } xy near far - float-4-with v? ! denom v/ :> fov - fov 0.0 float-4-with (vmerge-head) vmerge-diagonal - fov float-4{ f f t t } vand - float-4{ 0.0 0.0 -1.0 0.0 } + float-4{ 0.0 -1.0 0.0 0.0 } :> negone + + fov vmerge-diagonal + [ vmerge-diagonal ] + [ negone (vmerge) ] bi* ] make-matrix4 ;