1 USING: accessors alien arrays byte-arrays classes combinators
2 cpu.architecture effects functors generalizations kernel lexer
3 literals math math.bitwise math.vectors
4 math.vectors.simd.intrinsics parser prettyprint.custom
5 quotations sequences sequences.generalizations sequences.private
7 QUALIFIED-WITH: alien.c-types c
10 ERROR: bad-simd-length got expected ;
11 ERROR: bad-simd-vector obj ;
15 ! Primitive SIMD constructors
17 GENERIC: new-underlying ( underlying seq -- seq' )
19 : make-underlying ( seq quot -- seq' )
20 dip new-underlying ; inline
21 : change-underlying ( seq quot -- seq' )
22 '[ underlying>> @ ] keep new-underlying ; inline
28 ! Helper for boolean vector literals
30 : vector-true-value ( class -- value )
31 { c:float c:double } member? [ -1 bits>double ] [ -1 ] if ; foldable
33 : vector-false-value ( type -- value )
34 { c:float c:double } member? [ 0.0 ] [ 0 ] if ; foldable
36 : boolean>element ( bool/elt type -- elt )
38 { t [ vector-true-value ] }
39 { f [ vector-false-value ] }
48 { underlying byte-array read-only initial: $[ 16 <byte-array> ] } ;
50 GENERIC: simd-element-type ( obj -- c-type )
51 GENERIC: simd-rep ( simd -- rep )
52 GENERIC: simd-with ( n exemplar -- v )
54 M: object simd-element-type drop f ;
55 M: object simd-rep drop f ;
60 DEFER: simd-construct-op
62 ! Unboxers for SIMD operations
63 : if-both-vectors ( a b rep t f -- )
64 [ 2over [ simd-128? ] both? ] 2dip if ; inline
66 : if-both-vectors-match ( a b rep t f -- )
67 [ 3dup [ drop [ simd-128? ] both? ] [ '[ simd-rep _ eq? ] both? ] 3bi and ]
70 : simd-unbox ( a -- a (a) )
71 [ ] [ underlying>> ] bi ; inline
73 : v->v-op ( a rep quot: ( (a) rep -- (c) ) fallback-quot -- c )
74 drop [ simd-unbox ] 2dip 2curry make-underlying ; inline
76 : vx->v-op ( a obj rep quot: ( (a) obj rep -- (c) ) fallback-quot -- c )
77 drop [ simd-unbox ] 3dip 3curry make-underlying ; inline
79 : vn->v-op ( a n rep quot: ( (a) n rep -- (c) ) fallback-quot -- c )
80 drop [ [ simd-unbox ] [ >fixnum ] bi* ] 2dip 3curry make-underlying ; inline
82 : vx->x-op ( a obj rep quot: ( (a) obj rep -- obj ) fallback-quot -- obj )
83 drop [ underlying>> ] 3dip call ; inline
85 : v->x-op ( a rep quot: ( (a) rep -- obj ) fallback-quot -- obj )
86 drop [ underlying>> ] 2dip call ; inline
88 : (vv->v-op) ( a b rep quot: ( (a) (b) rep -- (c) ) -- c )
89 [ [ simd-unbox ] [ underlying>> ] bi* ] 2dip 3curry make-underlying ; inline
91 : (vv->x-op) ( a b rep quot: ( (a) (b) rep -- n ) -- n )
92 [ [ underlying>> ] bi@ ] 2dip 3curry call ; inline
94 : (vvx->v-op) ( a b obj rep quot: ( (a) (b) obj rep -- (c) ) -- c )
95 [ [ simd-unbox ] [ underlying>> ] bi* ] 3dip 2curry 2curry make-underlying ; inline
97 : vv->v-op ( a b rep quot: ( (a) (b) rep -- (c) ) fallback-quot -- c )
98 [ '[ _ (vv->v-op) ] ] [ '[ drop @ ] ] bi* if-both-vectors-match ; inline
100 :: vvx->v-op ( a b obj rep quot: ( (a) (b) obj rep -- (c) ) fallback-quot -- c )
102 [ obj swap quot (vvx->v-op) ]
103 [ drop obj fallback-quot call ] if-both-vectors-match ; inline
105 : vv'->v-op ( a b rep quot: ( (a) (b) rep -- (c) ) fallback-quot -- c )
106 [ '[ _ (vv->v-op) ] ] [ '[ drop @ ] ] bi* if-both-vectors ; inline
108 : vv->x-op ( a b rep quot: ( (a) (b) rep -- obj ) fallback-quot -- obj )
109 [ '[ _ (vv->x-op) ] ] [ '[ drop @ ] ] bi* if-both-vectors-match ; inline
111 : mask>count ( n rep -- n' )
114 { double-2-rep [ -1 shift ] }
117 { ushort-8-rep [ -1 shift ] }
118 { short-8-rep [ -1 shift ] }
119 { ushort-8-rep [ -1 shift ] }
120 { int-4-rep [ -2 shift ] }
121 { uint-4-rep [ -2 shift ] }
122 { longlong-2-rep [ -3 shift ] }
123 { ulonglong-2-rep [ -3 shift ] }
131 ! SIMD vectors as sequences
133 M: simd-128 hashcode* underlying>> hashcode* ; inline
134 M: simd-128 clone [ clone ] change-underlying ; inline
135 M: simd-128 byte-length drop 16 ; inline
137 M: simd-128 new-sequence
139 [ nip [ 16 (byte-array) ] make-underlying ]
140 [ length bad-simd-length ] if ; inline
143 dup simd-rep [ drop v= vall? ] [ 3drop f ] if-both-vectors-match ; inline
145 ! SIMD primitive operations
148 dup simd-rep [ (simd-v+) ] [ call-next-method ] vv->v-op ; inline
150 dup simd-rep [ (simd-v-) ] [ call-next-method ] vv->v-op ; inline
152 dup simd-rep [ (simd-vneg) ] [ call-next-method ] v->v-op ; inline
154 dup simd-rep [ (simd-v+-) ] [ call-next-method ] vv->v-op ; inline
156 dup simd-rep [ (simd-vs+) ] [ call-next-method ] vv->v-op ; inline
158 dup simd-rep [ (simd-vs-) ] [ call-next-method ] vv->v-op ; inline
160 dup simd-rep [ (simd-vs*) ] [ call-next-method ] vv->v-op ; inline
162 dup simd-rep [ (simd-v*) ] [ call-next-method ] vv->v-op ; inline
164 dup simd-rep [ (simd-v*high) ] [ call-next-method ] vv->v-op ; inline
166 dup simd-rep [ (simd-v/) ] [ call-next-method ] vv->v-op ; inline
168 dup simd-rep [ (simd-vavg) ] [ call-next-method ] vv->v-op ; inline
170 dup simd-rep [ (simd-vmin) ] [ call-next-method ] vv->v-op ; inline
172 dup simd-rep [ (simd-vmax) ] [ call-next-method ] vv->v-op ; inline
174 dup simd-rep [ (simd-vdot) ] [ call-next-method ] vv->x-op ; inline
176 dup simd-rep [ (simd-vsad) ] [ call-next-method ] vv->x-op ; inline
178 dup simd-rep [ (simd-vsqrt) ] [ call-next-method ] v->v-op ; inline
180 dup simd-rep [ (simd-sum) ] [ call-next-method ] v->x-op ; inline
182 dup simd-rep [ (simd-vabs) ] [ call-next-method ] v->v-op ; inline
184 dup simd-rep [ (simd-vbitand) ] [ call-next-method ] vv->v-op ; inline
186 dup simd-rep [ (simd-vbitandn) ] [ call-next-method ] vv->v-op ; inline
188 dup simd-rep [ (simd-vbitor) ] [ call-next-method ] vv->v-op ; inline
190 dup simd-rep [ (simd-vbitxor) ] [ call-next-method ] vv->v-op ; inline
192 dup simd-rep [ (simd-vbitnot) ] [ call-next-method ] v->v-op ; inline
194 dup simd-rep [ (simd-vand) ] [ call-next-method ] vv->v-op ; inline
196 dup simd-rep [ (simd-vandn) ] [ call-next-method ] vv->v-op ; inline
198 dup simd-rep [ (simd-vor) ] [ call-next-method ] vv->v-op ; inline
200 dup simd-rep [ (simd-vxor) ] [ call-next-method ] vv->v-op ; inline
202 dup simd-rep [ (simd-vnot) ] [ call-next-method ] v->v-op ; inline
204 over simd-rep [ (simd-vlshift) ] [ call-next-method ] vn->v-op ; inline
206 over simd-rep [ (simd-vrshift) ] [ call-next-method ] vn->v-op ; inline
208 over simd-rep [ (simd-hlshift) ] [ call-next-method ] vn->v-op ; inline
210 over simd-rep [ (simd-hrshift) ] [ call-next-method ] vn->v-op ; inline
211 M: simd-128 vshuffle-elements
212 over simd-rep [ (simd-vshuffle-elements) ] [ call-next-method ] vx->v-op ; inline
213 M: simd-128 vshuffle2-elements
214 over simd-rep [ (simd-vshuffle2-elements) ] [ call-next-method ] vvx->v-op ; inline
215 M: simd-128 vshuffle-bytes
216 dup simd-rep [ (simd-vshuffle-bytes) ] [ call-next-method ] vv'->v-op ; inline
217 M: simd-128 (vmerge-head)
218 dup simd-rep [ (simd-vmerge-head) ] [ call-next-method ] vv->v-op ; inline
219 M: simd-128 (vmerge-tail)
220 dup simd-rep [ (simd-vmerge-tail) ] [ call-next-method ] vv->v-op ; inline
222 dup simd-rep [ (simd-v<=) ] [ call-next-method ] vv->v-op ; inline
224 dup simd-rep [ (simd-v<) ] [ call-next-method ] vv->v-op ; inline
226 dup simd-rep [ (simd-v=) ] [ call-next-method ] vv->v-op ; inline
228 dup simd-rep [ (simd-v>) ] [ call-next-method ] vv->v-op ; inline
230 dup simd-rep [ (simd-v>=) ] [ call-next-method ] vv->v-op ; inline
231 M: simd-128 vunordered?
232 dup simd-rep [ (simd-vunordered?) ] [ call-next-method ] vv->v-op ; inline
234 dup simd-rep [ (simd-vany?) ] [ call-next-method ] v->x-op ; inline
236 dup simd-rep [ (simd-vall?) ] [ call-next-method ] v->x-op ; inline
238 dup simd-rep [ (simd-vnone?) ] [ call-next-method ] v->x-op ; inline
241 [ [ (simd-vgetmask) assert-positive ] [ call-next-method ] v->x-op ]
242 [ mask>count ] bi ; inline
244 ! SIMD high-level specializations
246 M: simd-128 vbroadcast swap [ nth ] [ simd-with ] bi ; inline
247 M: simd-128 n+v [ simd-with ] keep v+ ; inline
248 M: simd-128 n-v [ simd-with ] keep v- ; inline
249 M: simd-128 n*v [ simd-with ] keep v* ; inline
250 M: simd-128 n/v [ simd-with ] keep v/ ; inline
251 M: simd-128 v+n over simd-with v+ ; inline
252 M: simd-128 v-n over simd-with v- ; inline
253 M: simd-128 v*n over simd-with v* ; inline
254 M: simd-128 v/n over simd-with v/ ; inline
255 M: simd-128 norm-sq dup vdot assert-positive ; inline
256 M: simd-128 distance v- norm ; inline
258 M: simd-128 >pprint-sequence ;
259 M: simd-128 pprint* pprint-object ;
263 ! SIMD concrete type functor
265 <FUNCTOR: define-simd-128 ( T -- )
270 A-boa DEFINES ${T}-boa
271 A-with DEFINES ${T}-with
272 A-cast DEFINES ${T}-cast
275 ELT [ A-rep rep-component-type ]
276 N [ A-rep rep-length ]
277 COERCER [ ELT c:c-type-class "coercer" word-prop [ ] or ]
279 BOA-EFFECT [ N "n" <array> { "v" } <effect> ]
283 TUPLE: A < simd-128 ; final
285 M: A new-underlying drop \ A boa ; inline
286 M: A simd-rep drop A-rep ; inline
287 M: A simd-element-type drop ELT ; inline
288 M: A simd-with drop A-with ; inline
291 swap \ A-rep [ (simd-select) ] [ call-next-method ] vx->x-op ; inline
293 [ ELT boolean>element ] 2dip
294 underlying>> ELT c:set-alien-element ; inline
296 : >A ( seq -- simd ) \ A new clone-like ; inline
298 M: A like drop dup \ A instance? [ >A ] unless ; inline
300 : A-with ( n -- v ) COERCER call \ A-rep (simd-with) \ A boa ; inline
301 : A-cast ( v -- v' ) underlying>> \ A boa ; inline
303 M: A length drop N ; inline
306 [ COERCER N napply ] N {
307 { 2 [ [ A-rep (simd-gather-2) A boa ] ] }
308 { 4 [ [ A-rep (simd-gather-4) A boa ] ] }
309 [ \ A new '[ _ _ nsequence ] ]
311 BOA-EFFECT define-inline
313 M: A pprint-delims drop \ A{ \ } ;
314 SYNTAX: A{ \ } [ >A ] parse-literal ;
321 { A-rep alien-vector A boa } >quotation >>getter
323 [ dup simd-128? [ bad-simd-vector ] unless underlying>> ] 2dip
324 A-rep set-alien-vector
325 } >quotation >>setter
334 scan-token define-simd-128 ;
349 SIMD-128: ulonglong-2
356 vshuffle-bytes ; inline
359 uchar-16-rep [ (simd-v*hs+) ] [ call-next-method ] vv->v-op short-8-cast ; inline
361 ushort-8-rep [ (simd-v*hs+) ] [ call-next-method ] vv->v-op uint-4-cast ; inline
363 uint-4-rep [ (simd-v*hs+) ] [ call-next-method ] vv->v-op ulonglong-2-cast ; inline
365 char-16-rep [ (simd-v*hs+) ] [ call-next-method ] vv->v-op short-8-cast ; inline
367 short-8-rep [ (simd-v*hs+) ] [ call-next-method ] vv->v-op int-4-cast ; inline
369 int-4-rep [ (simd-v*hs+) ] [ call-next-method ] vv->v-op longlong-2-cast ; inline
371 { "math.vectors.simd" "mirrors" } "math.vectors.simd.mirrors" require-when