steve
/
factor
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

add %merge-vector-head and %merge-vector-tail instructions to back vmerge

db4
Joe Groff 2009-10-03 21:48:53 -05:00
parent 05c722ea0c
commit 0c9c3d4859
10 changed files with 112 additions and 40 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
basis/compiler/cfg/instructions/instructions.factor
View File

@ -280,6 +280,16 @@ def: dst
use: src use: src
literal: shuffle rep ; literal: shuffle rep ;
PURE-INSN: ##merge-vector-head
def: dst
use: src1 src2
literal: rep ;
PURE-INSN: ##merge-vector-tail
def: dst
use: src1 src2
literal: rep ;
PURE-INSN: ##compare-vector PURE-INSN: ##compare-vector
def: dst def: dst
use: src1 src2 use: src1 src2

2
basis/compiler/cfg/intrinsics/intrinsics.factor
View File

@ -194,6 +194,8 @@ IN: compiler.cfg.intrinsics
{ math.vectors.simd.intrinsics:(simd-gather-2) [ emit-gather-vector-2 ] } { math.vectors.simd.intrinsics:(simd-gather-2) [ emit-gather-vector-2 ] }
{ math.vectors.simd.intrinsics:(simd-gather-4) [ emit-gather-vector-4 ] } { math.vectors.simd.intrinsics:(simd-gather-4) [ emit-gather-vector-4 ] }
{ math.vectors.simd.intrinsics:(simd-vshuffle) [ emit-shuffle-vector ] } { math.vectors.simd.intrinsics:(simd-vshuffle) [ emit-shuffle-vector ] }
{ math.vectors.simd.intrinsics:(simd-vmerge-head) [ [ ^^merge-vector-head ] emit-binary-vector-op ] }
{ math.vectors.simd.intrinsics:(simd-vmerge-tail) [ [ ^^merge-vector-tail ] emit-binary-vector-op ] }
{ math.vectors.simd.intrinsics:(simd-select) [ emit-select-vector ] } { math.vectors.simd.intrinsics:(simd-select) [ emit-select-vector ] }
{ math.vectors.simd.intrinsics:(simd-sum) [ [ ^^horizontal-add-vector ] emit-unary-vector-op ] } { math.vectors.simd.intrinsics:(simd-sum) [ [ ^^horizontal-add-vector ] emit-unary-vector-op ] }
{ math.vectors.simd.intrinsics:alien-vector [ emit-alien-vector ] } { math.vectors.simd.intrinsics:alien-vector [ emit-alien-vector ] }

2
basis/compiler/codegen/codegen.factor
View File

@ -163,6 +163,8 @@ CODEGEN: ##zero-vector %zero-vector
CODEGEN: ##gather-vector-2 %gather-vector-2 CODEGEN: ##gather-vector-2 %gather-vector-2
CODEGEN: ##gather-vector-4 %gather-vector-4 CODEGEN: ##gather-vector-4 %gather-vector-4
CODEGEN: ##shuffle-vector %shuffle-vector CODEGEN: ##shuffle-vector %shuffle-vector
CODEGEN: ##merge-vector-head %merge-vector-head
CODEGEN: ##merge-vector-tail %merge-vector-tail
CODEGEN: ##compare-vector %compare-vector CODEGEN: ##compare-vector %compare-vector
CODEGEN: ##test-vector %test-vector CODEGEN: ##test-vector %test-vector
CODEGEN: ##add-vector %add-vector CODEGEN: ##add-vector %add-vector

2
basis/compiler/tree/propagation/simd/simd.factor
View File

@ -31,6 +31,8 @@ IN: compiler.tree.propagation.simd
(simd-hlshift) (simd-hlshift)
(simd-hrshift) (simd-hrshift)
(simd-vshuffle) (simd-vshuffle)
(simd-vmerge-head)
(simd-vmerge-tail)
(simd-v<=) (simd-v<=)
(simd-v<) (simd-v<)
(simd-v=) (simd-v=)

3
basis/cpu/architecture/architecture.factor
View File

@ -218,6 +218,8 @@ HOOK: %fill-vector cpu ( dst rep -- )
HOOK: %gather-vector-2 cpu ( dst src1 src2 rep -- ) HOOK: %gather-vector-2 cpu ( dst src1 src2 rep -- )
HOOK: %gather-vector-4 cpu ( dst src1 src2 src3 src4 rep -- ) HOOK: %gather-vector-4 cpu ( dst src1 src2 src3 src4 rep -- )
HOOK: %shuffle-vector cpu ( dst src shuffle rep -- ) HOOK: %shuffle-vector cpu ( dst src shuffle rep -- )
HOOK: %merge-vector-head cpu ( dst src1 src2 rep -- )
HOOK: %merge-vector-tail cpu ( dst src1 src2 rep -- )
HOOK: %compare-vector cpu ( dst src1 src2 temp rep cc -- ) HOOK: %compare-vector cpu ( dst src1 src2 temp rep cc -- )
HOOK: %test-vector cpu ( dst src1 temp rep vcc -- ) HOOK: %test-vector cpu ( dst src1 temp rep vcc -- )
HOOK: %test-vector-branch cpu ( label src1 temp rep vcc -- ) HOOK: %test-vector-branch cpu ( label src1 temp rep vcc -- )
@ -256,6 +258,7 @@ HOOK: %fill-vector-reps cpu ( -- reps )
HOOK: %gather-vector-2-reps cpu ( -- reps ) HOOK: %gather-vector-2-reps cpu ( -- reps )
HOOK: %gather-vector-4-reps cpu ( -- reps ) HOOK: %gather-vector-4-reps cpu ( -- reps )
HOOK: %shuffle-vector-reps cpu ( -- reps ) HOOK: %shuffle-vector-reps cpu ( -- reps )
HOOK: %merge-vector-reps cpu ( -- reps )
HOOK: %compare-vector-reps cpu ( cc -- reps ) HOOK: %compare-vector-reps cpu ( cc -- reps )
HOOK: %test-vector-reps cpu ( -- reps ) HOOK: %test-vector-reps cpu ( -- reps )
HOOK: %add-vector-reps cpu ( -- reps ) HOOK: %add-vector-reps cpu ( -- reps )

1
basis/cpu/ppc/ppc.factor
View File

@ -262,6 +262,7 @@ M: ppc %fill-vector-reps { } ;
M: ppc %gather-vector-2-reps { } ; M: ppc %gather-vector-2-reps { } ;
M: ppc %gather-vector-4-reps { } ; M: ppc %gather-vector-4-reps { } ;
M: ppc %shuffle-vector-reps { } ; M: ppc %shuffle-vector-reps { } ;
M: ppc %merge-vector-reps { } ;
M: ppc %compare-vector-reps drop { } ; M: ppc %compare-vector-reps drop { } ;
M: ppc %test-vector-reps { } ; M: ppc %test-vector-reps { } ;
M: ppc %add-vector-reps { } ; M: ppc %add-vector-reps { } ;

28
basis/cpu/x86/x86.factor
View File

@ -721,6 +721,34 @@ M: x86 %shuffle-vector-reps
{ sse2? { double-2-rep int-4-rep uint-4-rep longlong-2-rep ulonglong-2-rep } } { sse2? { double-2-rep int-4-rep uint-4-rep longlong-2-rep ulonglong-2-rep } }
} available-reps ; } available-reps ;
M: x86 %merge-vector-head
[ two-operand ] keep
unsign-rep {
{ double-2-rep [ UNPCKLPD ] }
{ float-4-rep [ UNPCKLPS ] }
{ longlong-2-rep [ PUNPCKLQDQ ] }
{ int-4-rep [ PUNPCKLDQ ] }
{ short-8-rep [ PUNPCKLWD ] }
{ char-16-rep [ PUNPCKLBW ] }
} case ;
M: x86 %merge-vector-tail
[ two-operand ] keep
unsign-rep {
{ double-2-rep [ UNPCKHPD ] }
{ float-4-rep [ UNPCKHPS ] }
{ longlong-2-rep [ PUNPCKHQDQ ] }
{ int-4-rep [ PUNPCKHDQ ] }
{ short-8-rep [ PUNPCKHWD ] }
{ char-16-rep [ PUNPCKHBW ] }
} case ;
M: x86 %merge-vector-reps
{
{ sse? { float-4-rep } }
{ sse2? { double-2-rep char-16-rep uchar-16-rep short-8-rep ushort-8-rep int-4-rep uint-4-rep longlong-2-rep ulonglong-2-rep } }
} available-reps ;
:: compare-float-v-operands ( dst src1 src2 temp rep cc -- dst' src' rep cc' ) :: compare-float-v-operands ( dst src1 src2 temp rep cc -- dst' src' rep cc' )
cc { cc> cc>= cc/> cc/>= } member? cc { cc> cc>= cc/> cc/>= } member?
[ dst src2 src1 rep two-operand rep cc swap-cc ] [ dst src2 src1 rep two-operand rep cc swap-cc ]

18
basis/math/vectors/simd/functor/functor.factor
View File

@ -325,6 +325,8 @@ A-v.-op DEFINES-PRIVATE ${A}-v.-op
A-sum-op DEFINES-PRIVATE ${A}-sum-op A-sum-op DEFINES-PRIVATE ${A}-sum-op
A-vany-op DEFINES-PRIVATE ${A}-vany-op A-vany-op DEFINES-PRIVATE ${A}-vany-op
A-vall-op DEFINES-PRIVATE ${A}-vall-op A-vall-op DEFINES-PRIVATE ${A}-vall-op
A-vmerge-head-op DEFINES-PRIVATE ${A}-vmerge-head-op
A-vmerge-tail-op DEFINES-PRIVATE ${A}-vmerge-tail-op
WHERE WHERE
@ -419,6 +421,20 @@ INSTANCE: A sequence
: A-vall-op ( v1 quot -- n ) : A-vall-op ( v1 quot -- n )
[ (simd-vbitand) ] (A-v->n-op) ; inline [ (simd-vbitand) ] (A-v->n-op) ; inline
: A-vmerge-head-op ( v1 v2 quot -- v )
drop
[ underlying1>> ] bi@
[ A-rep (simd-vmerge-head) ]
[ A-rep (simd-vmerge-tail) ] 2bi
\ A boa ;
: A-vmerge-tail-op ( v1 v2 quot -- v )
drop
[ underlying2>> ] bi@
[ A-rep (simd-vmerge-head) ]
[ A-rep (simd-vmerge-tail) ] 2bi
\ A boa ;
simd new simd new
\ A >>class \ A >>class
\ A-with >>ctor \ A-with >>ctor
@ -429,6 +445,8 @@ simd new
{ vnone? A-vany-op } { vnone? A-vany-op }
{ vany? A-vany-op } { vany? A-vany-op }
{ vall? A-vall-op } { vall? A-vall-op }
{ vmerge-head A-vmerge-head-op }
{ vmerge-tail A-vmerge-tail-op }
} >>special-wrappers } >>special-wrappers
{ {
{ { +vector+ +vector+ -> +vector+ } A-vv->v-op } { { +vector+ +vector+ -> +vector+ } A-vv->v-op }

84
basis/math/vectors/simd/intrinsics/intrinsics.factor
View File

@ -55,6 +55,8 @@ SIMD-OP: vrshift
SIMD-OP: hlshift SIMD-OP: hlshift
SIMD-OP: hrshift SIMD-OP: hrshift
SIMD-OP: vshuffle SIMD-OP: vshuffle
SIMD-OP: vmerge-head
SIMD-OP: vmerge-tail
SIMD-OP: v<= SIMD-OP: v<=
SIMD-OP: v< SIMD-OP: v<
SIMD-OP: v= SIMD-OP: v=
@ -118,44 +120,46 @@ GENERIC# supported-simd-op? 1 ( rep intrinsic -- ? )
M: vector-rep supported-simd-op? M: vector-rep supported-simd-op?
{ {
{ \ (simd-v+) [ %add-vector-reps ] } { \ (simd-v+) [ %add-vector-reps ] }
{ \ (simd-vs+) [ %saturated-add-vector-reps ] } { \ (simd-vs+) [ %saturated-add-vector-reps ] }
{ \ (simd-v+-) [ %add-sub-vector-reps ] } { \ (simd-v+-) [ %add-sub-vector-reps ] }
{ \ (simd-v-) [ %sub-vector-reps ] } { \ (simd-v-) [ %sub-vector-reps ] }
{ \ (simd-vs-) [ %saturated-sub-vector-reps ] } { \ (simd-vs-) [ %saturated-sub-vector-reps ] }
{ \ (simd-v*) [ %mul-vector-reps ] } { \ (simd-v*) [ %mul-vector-reps ] }
{ \ (simd-vs*) [ %saturated-mul-vector-reps ] } { \ (simd-vs*) [ %saturated-mul-vector-reps ] }
{ \ (simd-v/) [ %div-vector-reps ] } { \ (simd-v/) [ %div-vector-reps ] }
{ \ (simd-vmin) [ %min-vector-reps ] } { \ (simd-vmin) [ %min-vector-reps ] }
{ \ (simd-vmax) [ %max-vector-reps ] } { \ (simd-vmax) [ %max-vector-reps ] }
{ \ (simd-v.) [ %dot-vector-reps ] } { \ (simd-v.) [ %dot-vector-reps ] }
{ \ (simd-vsqrt) [ %sqrt-vector-reps ] } { \ (simd-vsqrt) [ %sqrt-vector-reps ] }
{ \ (simd-sum) [ %horizontal-add-vector-reps ] } { \ (simd-sum) [ %horizontal-add-vector-reps ] }
{ \ (simd-vabs) [ %abs-vector-reps ] } { \ (simd-vabs) [ %abs-vector-reps ] }
{ \ (simd-vbitand) [ %and-vector-reps ] } { \ (simd-vbitand) [ %and-vector-reps ] }
{ \ (simd-vbitandn) [ %andn-vector-reps ] } { \ (simd-vbitandn) [ %andn-vector-reps ] }
{ \ (simd-vbitor) [ %or-vector-reps ] } { \ (simd-vbitor) [ %or-vector-reps ] }
{ \ (simd-vbitxor) [ %xor-vector-reps ] } { \ (simd-vbitxor) [ %xor-vector-reps ] }
{ \ (simd-vbitnot) [ %not-vector-reps ] } { \ (simd-vbitnot) [ %not-vector-reps ] }
{ \ (simd-vand) [ %and-vector-reps ] } { \ (simd-vand) [ %and-vector-reps ] }
{ \ (simd-vandn) [ %andn-vector-reps ] } { \ (simd-vandn) [ %andn-vector-reps ] }
{ \ (simd-vor) [ %or-vector-reps ] } { \ (simd-vor) [ %or-vector-reps ] }
{ \ (simd-vxor) [ %xor-vector-reps ] } { \ (simd-vxor) [ %xor-vector-reps ] }
{ \ (simd-vnot) [ %not-vector-reps ] } { \ (simd-vnot) [ %not-vector-reps ] }
{ \ (simd-vlshift) [ %shl-vector-reps ] } { \ (simd-vlshift) [ %shl-vector-reps ] }
{ \ (simd-vrshift) [ %shr-vector-reps ] } { \ (simd-vrshift) [ %shr-vector-reps ] }
{ \ (simd-hlshift) [ %horizontal-shl-vector-reps ] } { \ (simd-hlshift) [ %horizontal-shl-vector-reps ] }
{ \ (simd-hrshift) [ %horizontal-shr-vector-reps ] } { \ (simd-hrshift) [ %horizontal-shr-vector-reps ] }
{ \ (simd-vshuffle) [ %shuffle-vector-reps ] } { \ (simd-vshuffle) [ %shuffle-vector-reps ] }
{ \ (simd-v<=) [ cc<= %compare-vector-reps ] } { \ (simd-vmerge-head) [ %merge-vector-reps ] }
{ \ (simd-v<) [ cc< %compare-vector-reps ] } { \ (simd-vmerge-tail) [ %merge-vector-reps ] }
{ \ (simd-v=) [ cc= %compare-vector-reps ] } { \ (simd-v<=) [ cc<= %compare-vector-reps ] }
{ \ (simd-v>) [ cc> %compare-vector-reps ] } { \ (simd-v<) [ cc< %compare-vector-reps ] }
{ \ (simd-v>=) [ cc>= %compare-vector-reps ] } { \ (simd-v=) [ cc= %compare-vector-reps ] }
{ \ (simd-vunordered?) [ cc/<>= %compare-vector-reps ] } { \ (simd-v>) [ cc> %compare-vector-reps ] }
{ \ (simd-gather-2) [ %gather-vector-2-reps ] } { \ (simd-v>=) [ cc>= %compare-vector-reps ] }
{ \ (simd-gather-4) [ %gather-vector-4-reps ] } { \ (simd-vunordered?) [ cc/<>= %compare-vector-reps ] }
{ \ (simd-vany?) [ %test-vector-reps ] } { \ (simd-gather-2) [ %gather-vector-2-reps ] }
{ \ (simd-vall?) [ %test-vector-reps ] } { \ (simd-gather-4) [ %gather-vector-4-reps ] }
{ \ (simd-vnone?) [ %test-vector-reps ] } { \ (simd-vany?) [ %test-vector-reps ] }
{ \ (simd-vall?) [ %test-vector-reps ] }
{ \ (simd-vnone?) [ %test-vector-reps ] }
} case member? ; } case member? ;

2
basis/math/vectors/specialization/specialization.factor
View File

@ -98,6 +98,8 @@ H{
{ hrshift { +vector+ +literal+ -> +vector+ } } { hrshift { +vector+ +literal+ -> +vector+ } }
{ vshuffle { +vector+ +literal+ -> +vector+ } } { vshuffle { +vector+ +literal+ -> +vector+ } }
{ vbroadcast { +vector+ +literal+ -> +vector+ } } { vbroadcast { +vector+ +literal+ -> +vector+ } }
{ vmerge-head { +vector+ +vector+ -> +vector+ } }
{ vmerge-tail { +vector+ +vector+ -> +vector+ } }
{ v<= { +vector+ +vector+ -> +vector+ } } { v<= { +vector+ +vector+ -> +vector+ } }
{ v< { +vector+ +vector+ -> +vector+ } } { v< { +vector+ +vector+ -> +vector+ } }
{ v= { +vector+ +vector+ -> +vector+ } } { v= { +vector+ +vector+ -> +vector+ } }