-! Copyright (C) 2009 Slava Pestov.
-! See http://factorcode.org/license.txt for BSD license.
-USING: alien.c-types combinators fry kernel parser math math.parser
-math.vectors.simd.functor sequences splitting vocabs.generated
-vocabs.loader vocabs.parser words accessors vocabs compiler.units
-definitions ;
+USING: accessors alien arrays byte-arrays classes combinators
+cpu.architecture effects functors generalizations kernel lexer
+literals math math.bitwise math.vectors
+math.vectors.simd.intrinsics parser prettyprint.custom
+quotations sequences sequences.generalizations sequences.private
+vocabs.loader words ;
QUALIFIED-WITH: alien.c-types c
IN: math.vectors.simd
-ERROR: bad-base-type type ;
+ERROR: bad-simd-length got expected ;
+ERROR: bad-simd-vector obj ;
+<<
<PRIVATE
+! Primitive SIMD constructors
-: simd-vocab ( base-type -- vocab )
- name>> "math.vectors.simd.instances." prepend ;
+GENERIC: new-underlying ( underlying seq -- seq' )
-: parse-base-type ( c-type -- c-type )
- dup { c:char c:uchar c:short c:ushort c:int c:uint c:longlong c:ulonglong c:float c:double } memq?
- [ bad-base-type ] unless ;
+: make-underlying ( seq quot -- seq' )
+ dip new-underlying ; inline
+: change-underlying ( seq quot -- seq' )
+ '[ underlying>> @ ] keep new-underlying ; inline
+PRIVATE>
+>>
+
+<PRIVATE
+
+! Helper for boolean vector literals
+
+: vector-true-value ( class -- value )
+ { c:float c:double } member? [ -1 bits>double ] [ -1 ] if ; foldable
+
+: vector-false-value ( type -- value )
+ { c:float c:double } member? [ 0.0 ] [ 0 ] if ; foldable
+
+: boolean>element ( bool/elt type -- elt )
+ swap {
+ { t [ vector-true-value ] }
+ { f [ vector-false-value ] }
+ [ nip ]
+ } case ; inline
+
+PRIVATE>
+
+! SIMD base type
+
+TUPLE: simd-128
+ { underlying byte-array read-only initial: $[ 16 <byte-array> ] } ;
+
+GENERIC: simd-element-type ( obj -- c-type )
+GENERIC: simd-rep ( simd -- rep )
+GENERIC: simd-with ( n exemplar -- v )
+
+M: object simd-element-type drop f ;
+M: object simd-rep drop f ;
+
+<<
+<PRIVATE
+
+DEFER: simd-construct-op
+
+! Unboxers for SIMD operations
+: if-both-vectors ( a b rep t f -- )
+ [ 2over [ simd-128? ] both? ] 2dip if ; inline
+
+: if-both-vectors-match ( a b rep t f -- )
+ [ 3dup [ drop [ simd-128? ] both? ] [ '[ simd-rep _ eq? ] both? ] 3bi and ]
+ 2dip if ; inline
+
+: simd-unbox ( a -- a (a) )
+ [ ] [ underlying>> ] bi ; inline
+
+: v->v-op ( a rep quot: ( (a) rep -- (c) ) fallback-quot -- c )
+ drop [ simd-unbox ] 2dip 2curry make-underlying ; inline
+
+: vx->v-op ( a obj rep quot: ( (a) obj rep -- (c) ) fallback-quot -- c )
+ drop [ simd-unbox ] 3dip 3curry make-underlying ; inline
+
+: vn->v-op ( a n rep quot: ( (a) n rep -- (c) ) fallback-quot -- c )
+ drop [ [ simd-unbox ] [ >fixnum ] bi* ] 2dip 3curry make-underlying ; inline
+
+: vx->x-op ( a obj rep quot: ( (a) obj rep -- obj ) fallback-quot -- obj )
+ drop [ underlying>> ] 3dip call ; inline
+
+: v->x-op ( a rep quot: ( (a) rep -- obj ) fallback-quot -- obj )
+ drop [ underlying>> ] 2dip call ; inline
+
+: (vv->v-op) ( a b rep quot: ( (a) (b) rep -- (c) ) -- c )
+ [ [ simd-unbox ] [ underlying>> ] bi* ] 2dip 3curry make-underlying ; inline
+
+: (vv->x-op) ( a b rep quot: ( (a) (b) rep -- n ) -- n )
+ [ [ underlying>> ] bi@ ] 2dip 3curry call ; inline
+
+: (vvx->v-op) ( a b obj rep quot: ( (a) (b) obj rep -- (c) ) -- c )
+ [ [ simd-unbox ] [ underlying>> ] bi* ] 3dip 2curry 2curry make-underlying ; inline
+
+: vv->v-op ( a b rep quot: ( (a) (b) rep -- (c) ) fallback-quot -- c )
+ [ '[ _ (vv->v-op) ] ] [ '[ drop @ ] ] bi* if-both-vectors-match ; inline
+
+:: vvx->v-op ( a b obj rep quot: ( (a) (b) obj rep -- (c) ) fallback-quot -- c )
+ a b rep
+ [ obj swap quot (vvx->v-op) ]
+ [ drop obj fallback-quot call ] if-both-vectors-match ; inline
+
+: vv'->v-op ( a b rep quot: ( (a) (b) rep -- (c) ) fallback-quot -- c )
+ [ '[ _ (vv->v-op) ] ] [ '[ drop @ ] ] bi* if-both-vectors ; inline
+
+: vv->x-op ( a b rep quot: ( (a) (b) rep -- obj ) fallback-quot -- obj )
+ [ '[ _ (vv->x-op) ] ] [ '[ drop @ ] ] bi* if-both-vectors-match ; inline
+
+: mask>count ( n rep -- n' )
+ [ bit-count ] dip {
+ { float-4-rep [ ] }
+ { double-2-rep [ -1 shift ] }
+ { uchar-16-rep [ ] }
+ { char-16-rep [ ] }
+ { ushort-8-rep [ -1 shift ] }
+ { short-8-rep [ -1 shift ] }
+ { ushort-8-rep [ -1 shift ] }
+ { int-4-rep [ -2 shift ] }
+ { uint-4-rep [ -2 shift ] }
+ { longlong-2-rep [ -3 shift ] }
+ { ulonglong-2-rep [ -3 shift ] }
+ } case ; inline
+
+PRIVATE>
+>>
-: forget-instances ( -- )
- [
- "math.vectors.simd.instances" child-vocabs
- [ forget-vocab ] each
- ] with-compilation-unit ;
+<<
+
+! SIMD vectors as sequences
+
+M: simd-128 hashcode* underlying>> hashcode* ; inline
+M: simd-128 clone [ clone ] change-underlying ; inline
+M: simd-128 byte-length drop 16 ; inline
+
+M: simd-128 new-sequence
+ 2dup length =
+ [ nip [ 16 (byte-array) ] make-underlying ]
+ [ length bad-simd-length ] if ; inline
+
+M: simd-128 equal?
+ dup simd-rep [ drop v= vall? ] [ 3drop f ] if-both-vectors-match ; inline
+
+! SIMD primitive operations
+
+M: simd-128 v+
+ dup simd-rep [ (simd-v+) ] [ call-next-method ] vv->v-op ; inline
+M: simd-128 v-
+ dup simd-rep [ (simd-v-) ] [ call-next-method ] vv->v-op ; inline
+M: simd-128 vneg
+ dup simd-rep [ (simd-vneg) ] [ call-next-method ] v->v-op ; inline
+M: simd-128 v+-
+ dup simd-rep [ (simd-v+-) ] [ call-next-method ] vv->v-op ; inline
+M: simd-128 vs+
+ dup simd-rep [ (simd-vs+) ] [ call-next-method ] vv->v-op ; inline
+M: simd-128 vs-
+ dup simd-rep [ (simd-vs-) ] [ call-next-method ] vv->v-op ; inline
+M: simd-128 vs*
+ dup simd-rep [ (simd-vs*) ] [ call-next-method ] vv->v-op ; inline
+M: simd-128 v*
+ dup simd-rep [ (simd-v*) ] [ call-next-method ] vv->v-op ; inline
+M: simd-128 v*high
+ dup simd-rep [ (simd-v*high) ] [ call-next-method ] vv->v-op ; inline
+M: simd-128 v/
+ dup simd-rep [ (simd-v/) ] [ call-next-method ] vv->v-op ; inline
+M: simd-128 vavg
+ dup simd-rep [ (simd-vavg) ] [ call-next-method ] vv->v-op ; inline
+M: simd-128 vmin
+ dup simd-rep [ (simd-vmin) ] [ call-next-method ] vv->v-op ; inline
+M: simd-128 vmax
+ dup simd-rep [ (simd-vmax) ] [ call-next-method ] vv->v-op ; inline
+M: simd-128 vdot
+ dup simd-rep [ (simd-vdot) ] [ call-next-method ] vv->x-op ; inline
+M: simd-128 vsad
+ dup simd-rep [ (simd-vsad) ] [ call-next-method ] vv->x-op ; inline
+M: simd-128 vsqrt
+ dup simd-rep [ (simd-vsqrt) ] [ call-next-method ] v->v-op ; inline
+M: simd-128 sum
+ dup simd-rep [ (simd-sum) ] [ call-next-method ] v->x-op ; inline
+M: simd-128 vabs
+ dup simd-rep [ (simd-vabs) ] [ call-next-method ] v->v-op ; inline
+M: simd-128 vbitand
+ dup simd-rep [ (simd-vbitand) ] [ call-next-method ] vv->v-op ; inline
+M: simd-128 vbitandn
+ dup simd-rep [ (simd-vbitandn) ] [ call-next-method ] vv->v-op ; inline
+M: simd-128 vbitor
+ dup simd-rep [ (simd-vbitor) ] [ call-next-method ] vv->v-op ; inline
+M: simd-128 vbitxor
+ dup simd-rep [ (simd-vbitxor) ] [ call-next-method ] vv->v-op ; inline
+M: simd-128 vbitnot
+ dup simd-rep [ (simd-vbitnot) ] [ call-next-method ] v->v-op ; inline
+M: simd-128 vand
+ dup simd-rep [ (simd-vand) ] [ call-next-method ] vv->v-op ; inline
+M: simd-128 vandn
+ dup simd-rep [ (simd-vandn) ] [ call-next-method ] vv->v-op ; inline
+M: simd-128 vor
+ dup simd-rep [ (simd-vor) ] [ call-next-method ] vv->v-op ; inline
+M: simd-128 vxor
+ dup simd-rep [ (simd-vxor) ] [ call-next-method ] vv->v-op ; inline
+M: simd-128 vnot
+ dup simd-rep [ (simd-vnot) ] [ call-next-method ] v->v-op ; inline
+M: simd-128 vlshift
+ over simd-rep [ (simd-vlshift) ] [ call-next-method ] vn->v-op ; inline
+M: simd-128 vrshift
+ over simd-rep [ (simd-vrshift) ] [ call-next-method ] vn->v-op ; inline
+M: simd-128 hlshift
+ over simd-rep [ (simd-hlshift) ] [ call-next-method ] vn->v-op ; inline
+M: simd-128 hrshift
+ over simd-rep [ (simd-hrshift) ] [ call-next-method ] vn->v-op ; inline
+M: simd-128 vshuffle-elements
+ over simd-rep [ (simd-vshuffle-elements) ] [ call-next-method ] vx->v-op ; inline
+M: simd-128 vshuffle2-elements
+ over simd-rep [ (simd-vshuffle2-elements) ] [ call-next-method ] vvx->v-op ; inline
+M: simd-128 vshuffle-bytes
+ dup simd-rep [ (simd-vshuffle-bytes) ] [ call-next-method ] vv'->v-op ; inline
+M: simd-128 (vmerge-head)
+ dup simd-rep [ (simd-vmerge-head) ] [ call-next-method ] vv->v-op ; inline
+M: simd-128 (vmerge-tail)
+ dup simd-rep [ (simd-vmerge-tail) ] [ call-next-method ] vv->v-op ; inline
+M: simd-128 v<=
+ dup simd-rep [ (simd-v<=) ] [ call-next-method ] vv->v-op ; inline
+M: simd-128 v<
+ dup simd-rep [ (simd-v<) ] [ call-next-method ] vv->v-op ; inline
+M: simd-128 v=
+ dup simd-rep [ (simd-v=) ] [ call-next-method ] vv->v-op ; inline
+M: simd-128 v>
+ dup simd-rep [ (simd-v>) ] [ call-next-method ] vv->v-op ; inline
+M: simd-128 v>=
+ dup simd-rep [ (simd-v>=) ] [ call-next-method ] vv->v-op ; inline
+M: simd-128 vunordered?
+ dup simd-rep [ (simd-vunordered?) ] [ call-next-method ] vv->v-op ; inline
+M: simd-128 vany?
+ dup simd-rep [ (simd-vany?) ] [ call-next-method ] v->x-op ; inline
+M: simd-128 vall?
+ dup simd-rep [ (simd-vall?) ] [ call-next-method ] v->x-op ; inline
+M: simd-128 vnone?
+ dup simd-rep [ (simd-vnone?) ] [ call-next-method ] v->x-op ; inline
+M: simd-128 vcount
+ dup simd-rep
+ [ [ (simd-vgetmask) assert-positive ] [ call-next-method ] v->x-op ]
+ [ mask>count ] bi ; inline
+
+! SIMD high-level specializations
+
+M: simd-128 vbroadcast swap [ nth ] [ simd-with ] bi ; inline
+M: simd-128 n+v [ simd-with ] keep v+ ; inline
+M: simd-128 n-v [ simd-with ] keep v- ; inline
+M: simd-128 n*v [ simd-with ] keep v* ; inline
+M: simd-128 n/v [ simd-with ] keep v/ ; inline
+M: simd-128 v+n over simd-with v+ ; inline
+M: simd-128 v-n over simd-with v- ; inline
+M: simd-128 v*n over simd-with v* ; inline
+M: simd-128 v/n over simd-with v/ ; inline
+M: simd-128 norm-sq dup vdot assert-positive ; inline
+M: simd-128 distance v- norm ; inline
+
+M: simd-128 >pprint-sequence ;
+M: simd-128 pprint* pprint-object ;
+
+<PRIVATE
+
+! SIMD concrete type functor
+
+<FUNCTOR: define-simd-128 ( T -- )
+
+A DEFINES-CLASS ${T}
+A-rep IS ${T}-rep
+>A DEFINES >${T}
+A-boa DEFINES ${T}-boa
+A-with DEFINES ${T}-with
+A-cast DEFINES ${T}-cast
+A{ DEFINES ${T}{
+
+ELT [ A-rep rep-component-type ]
+N [ A-rep rep-length ]
+COERCER [ ELT c:c-type-class "coercer" word-prop [ ] or ]
+
+BOA-EFFECT [ N "n" <array> { "v" } <effect> ]
+
+WHERE
+
+TUPLE: A < simd-128 ; final
+
+M: A new-underlying drop \ A boa ; inline
+M: A simd-rep drop A-rep ; inline
+M: A simd-element-type drop ELT ; inline
+M: A simd-with drop A-with ; inline
+
+M: A nth-unsafe
+ swap \ A-rep [ (simd-select) ] [ call-next-method ] vx->x-op ; inline
+M: A set-nth-unsafe
+ [ ELT boolean>element ] 2dip
+ underlying>> ELT c:set-alien-element ; inline
+
+: >A ( seq -- simd ) \ A new clone-like ; inline
+
+M: A like drop dup \ A instance? [ >A ] unless ; inline
+
+: A-with ( n -- v ) COERCER call \ A-rep (simd-with) \ A boa ; inline
+: A-cast ( v -- v' ) underlying>> \ A boa ; inline
+
+M: A length drop N ; inline
+
+\ A-boa
+[ COERCER N napply ] N {
+ { 2 [ [ A-rep (simd-gather-2) A boa ] ] }
+ { 4 [ [ A-rep (simd-gather-4) A boa ] ] }
+ [ \ A new '[ _ _ nsequence ] ]
+} case compose
+BOA-EFFECT define-inline
+
+M: A pprint-delims drop \ A{ \ } ;
+SYNTAX: A{ \ } [ >A ] parse-literal ;
+
+INSTANCE: A sequence
+
+c:<c-type>
+ byte-array >>class
+ A >>boxed-class
+ { A-rep alien-vector A boa } >quotation >>getter
+ {
+ [ dup simd-128? [ bad-simd-vector ] unless underlying>> ] 2dip
+ A-rep set-alien-vector
+ } >quotation >>setter
+ 16 >>size
+ 16 >>align
+ A-rep >>rep
+\ A c:typedef
+
+;FUNCTOR>
+
+SYNTAX: SIMD-128:
+ scan-token define-simd-128 ;
PRIVATE>
-: define-simd-vocab ( type -- vocab )
- parse-base-type
- [ simd-vocab ] keep '[
- _
- [ define-simd-128 ]
- [ define-simd-256 ] bi
- ] generate-vocab ;
+>>
+
+! SIMD instances
+
+SIMD-128: char-16
+SIMD-128: uchar-16
+SIMD-128: short-8
+SIMD-128: ushort-8
+SIMD-128: int-4
+SIMD-128: uint-4
+SIMD-128: longlong-2
+SIMD-128: ulonglong-2
+SIMD-128: float-4
+SIMD-128: double-2
+
+! misc
+
+M: simd-128 vshuffle
+ vshuffle-bytes ; inline
-SYNTAX: SIMD:
- scan-word define-simd-vocab use-vocab ;
+M: uchar-16 v*hs+
+ uchar-16-rep [ (simd-v*hs+) ] [ call-next-method ] vv->v-op short-8-cast ; inline
+M: ushort-8 v*hs+
+ ushort-8-rep [ (simd-v*hs+) ] [ call-next-method ] vv->v-op uint-4-cast ; inline
+M: uint-4 v*hs+
+ uint-4-rep [ (simd-v*hs+) ] [ call-next-method ] vv->v-op ulonglong-2-cast ; inline
+M: char-16 v*hs+
+ char-16-rep [ (simd-v*hs+) ] [ call-next-method ] vv->v-op short-8-cast ; inline
+M: short-8 v*hs+
+ short-8-rep [ (simd-v*hs+) ] [ call-next-method ] vv->v-op int-4-cast ; inline
+M: int-4 v*hs+
+ int-4-rep [ (simd-v*hs+) ] [ call-next-method ] vv->v-op longlong-2-cast ; inline
+{ "math.vectors.simd" "mirrors" } "math.vectors.simd.mirrors" require-when