Merge branch 'master' of git://factorcode.org/git/factor

db4
Doug Coleman 2009-10-21 12:11:16 -05:00
commit c5dd8d0b20
8 changed files with 275 additions and 26 deletions

View File

@ -163,8 +163,8 @@ IN: compiler.cfg.intrinsics
{ math.vectors.simd.intrinsics:(simd-v*) [ [ ^^mul-vector ] emit-binary-vector-op ] }
{ math.vectors.simd.intrinsics:(simd-vs*) [ [ ^^saturated-mul-vector ] emit-binary-vector-op ] }
{ math.vectors.simd.intrinsics:(simd-v/) [ [ ^^div-vector ] emit-binary-vector-op ] }
{ math.vectors.simd.intrinsics:(simd-vmin) [ [ ^^min-vector ] emit-binary-vector-op ] }
{ math.vectors.simd.intrinsics:(simd-vmax) [ [ ^^max-vector ] emit-binary-vector-op ] }
{ math.vectors.simd.intrinsics:(simd-vmin) [ [ generate-min-vector ] emit-binary-vector-op ] }
{ math.vectors.simd.intrinsics:(simd-vmax) [ [ generate-max-vector ] emit-binary-vector-op ] }
{ math.vectors.simd.intrinsics:(simd-v.) [ [ ^^dot-vector ] emit-binary-vector-op ] }
{ math.vectors.simd.intrinsics:(simd-vabs) [ [ generate-abs-vector ] emit-unary-vector-op ] }
{ math.vectors.simd.intrinsics:(simd-vsqrt) [ [ ^^sqrt-vector ] emit-unary-vector-op ] }

View File

@ -10,8 +10,8 @@ compiler.cfg.stacks compiler.cfg.stacks.local compiler.cfg.hats
compiler.cfg.instructions compiler.cfg.registers
compiler.cfg.intrinsics.alien
specialized-arrays ;
FROM: alien.c-types => heap-size char uchar float double ;
SPECIALIZED-ARRAYS: float double ;
FROM: alien.c-types => heap-size uchar ushort uint ulonglong float double ;
SPECIALIZED-ARRAYS: uchar ushort uint ulonglong float double ;
IN: compiler.cfg.intrinsics.simd
MACRO: check-elements ( quots -- )
@ -155,28 +155,79 @@ MACRO: if-literals-match ( quots -- )
[ ^^not-vector ]
[ [ ^^fill-vector ] [ ^^xor-vector ] bi ] if ;
:: (generate-compare-vector) ( src1 src2 rep {cc,swap} -- dst )
:: ((generate-compare-vector)) ( src1 src2 rep {cc,swap} -- dst )
{cc,swap} first2 :> swap? :> cc
swap?
[ src2 src1 rep cc ^^compare-vector ]
[ src1 src2 rep cc ^^compare-vector ] if ;
:: generate-compare-vector ( src1 src2 rep orig-cc -- dst )
:: (generate-compare-vector) ( src1 src2 rep orig-cc -- dst )
rep orig-cc %compare-vector-ccs :> not? :> ccs
ccs empty?
[ rep not? [ ^^fill-vector ] [ ^^zero-vector ] if ]
[
ccs unclip :> first-cc :> rest-ccs
src1 src2 rep first-cc (generate-compare-vector) :> first-dst
src1 src2 rep first-cc ((generate-compare-vector)) :> first-dst
rest-ccs first-dst
[ [ src1 src2 rep ] dip (generate-compare-vector) rep ^^or-vector ]
[ [ src1 src2 rep ] dip ((generate-compare-vector)) rep ^^or-vector ]
reduce
not? [ rep generate-not-vector ] when
] if ;
: sign-bit-mask ( rep -- byte-array )
unsign-rep {
{ char-16-rep [ uchar-array{
HEX: 80 HEX: 80 HEX: 80 HEX: 80
HEX: 80 HEX: 80 HEX: 80 HEX: 80
HEX: 80 HEX: 80 HEX: 80 HEX: 80
HEX: 80 HEX: 80 HEX: 80 HEX: 80
} underlying>> ] }
{ short-8-rep [ ushort-array{
HEX: 8000 HEX: 8000 HEX: 8000 HEX: 8000
HEX: 8000 HEX: 8000 HEX: 8000 HEX: 8000
} underlying>> ] }
{ int-4-rep [ uint-array{
HEX: 8000,0000 HEX: 8000,0000
HEX: 8000,0000 HEX: 8000,0000
} underlying>> ] }
{ longlong-2-rep [ ulonglong-array{
HEX: 8000,0000,0000,0000
HEX: 8000,0000,0000,0000
} underlying>> ] }
} case ;
:: (generate-minmax-compare-vector) ( src1 src2 rep orig-cc -- dst )
orig-cc order-cc {
{ cc< [ src1 src2 rep ^^max-vector src1 rep cc/= (generate-compare-vector) ] }
{ cc<= [ src1 src2 rep ^^min-vector src1 rep cc= (generate-compare-vector) ] }
{ cc> [ src1 src2 rep ^^min-vector src1 rep cc/= (generate-compare-vector) ] }
{ cc>= [ src1 src2 rep ^^max-vector src1 rep cc= (generate-compare-vector) ] }
} case ;
:: generate-compare-vector ( src1 src2 rep orig-cc -- dst )
{
{
[ rep orig-cc %compare-vector-reps member? ]
[ src1 src2 rep orig-cc (generate-compare-vector) ]
}
{
[ rep %min-vector-reps member? ]
[ src1 src2 rep orig-cc (generate-minmax-compare-vector) ]
}
{
[ rep unsign-rep orig-cc %compare-vector-reps member? ]
[
rep sign-bit-mask ^^load-constant :> sign-bits
src1 sign-bits rep ^^xor-vector
src2 sign-bits rep ^^xor-vector
rep unsign-rep orig-cc (generate-compare-vector)
]
}
} cond ;
:: generate-unpack-vector-head ( src rep -- dst )
{
{
@ -265,3 +316,17 @@ MACRO: if-literals-match ( quots -- )
]
} cond ;
: generate-min-vector ( src1 src2 rep -- dst )
dup %min-vector-reps member?
[ ^^min-vector ] [
[ cc< generate-compare-vector ]
[ generate-blend-vector ] 3bi
] if ;
: generate-max-vector ( src1 src2 rep -- dst )
dup %max-vector-reps member?
[ ^^max-vector ] [
[ cc> generate-compare-vector ]
[ generate-blend-vector ] 3bi
] if ;

View File

@ -515,3 +515,48 @@ M: ##scalar>vector rewrite
M: ##xor-vector rewrite
dup [ src1>> vreg>vn ] [ src2>> vreg>vn ] bi eq?
[ [ dst>> ] [ rep>> ] bi \ ##zero-vector new-insn ] [ drop f ] if ;
: vector-not? ( expr -- ? )
{
[ not-vector-expr? ]
[ {
[ xor-vector-expr? ]
[ [ src1>> ] [ src2>> ] bi [ vn>expr fill-vector-expr? ] either? ]
} 1&& ]
} 1|| ;
GENERIC: vector-not-src ( expr -- vreg )
M: not-vector-expr vector-not-src src>> vn>vreg ;
M: xor-vector-expr vector-not-src
dup src1>> vn>expr fill-vector-expr? [ src2>> ] [ src1>> ] if vn>vreg ;
M: ##and-vector rewrite
{
{ [ dup src1>> vreg>expr vector-not? ] [
{
[ dst>> ]
[ src1>> vreg>expr vector-not-src ]
[ src2>> ]
[ rep>> ]
} cleave \ ##andn-vector new-insn
] }
{ [ dup src2>> vreg>expr vector-not? ] [
{
[ dst>> ]
[ src2>> vreg>expr vector-not-src ]
[ src1>> ]
[ rep>> ]
} cleave \ ##andn-vector new-insn
] }
[ drop f ]
} cond ;
M: ##andn-vector rewrite
dup src1>> vreg>expr vector-not? [
{
[ dst>> ]
[ src1>> vreg>expr vector-not-src ]
[ src2>> ]
[ rep>> ]
} cleave \ ##and-vector new-insn
] [ drop f ] if ;

View File

@ -1281,6 +1281,128 @@ cell 8 = [
} value-numbering-step
] unit-test
! NOT x AND y => x ANDN y
[
{
T{ ##fill-vector f 3 float-4-rep }
T{ ##xor-vector f 4 0 3 float-4-rep }
T{ ##andn-vector f 5 0 1 float-4-rep }
}
] [
{
T{ ##fill-vector f 3 float-4-rep }
T{ ##xor-vector f 4 0 3 float-4-rep }
T{ ##and-vector f 5 4 1 float-4-rep }
} value-numbering-step
] unit-test
[
{
T{ ##not-vector f 4 0 float-4-rep }
T{ ##andn-vector f 5 0 1 float-4-rep }
}
] [
{
T{ ##not-vector f 4 0 float-4-rep }
T{ ##and-vector f 5 4 1 float-4-rep }
} value-numbering-step
] unit-test
! x AND NOT y => y ANDN x
[
{
T{ ##fill-vector f 3 float-4-rep }
T{ ##xor-vector f 4 0 3 float-4-rep }
T{ ##andn-vector f 5 0 1 float-4-rep }
}
] [
{
T{ ##fill-vector f 3 float-4-rep }
T{ ##xor-vector f 4 0 3 float-4-rep }
T{ ##and-vector f 5 1 4 float-4-rep }
} value-numbering-step
] unit-test
[
{
T{ ##not-vector f 4 0 float-4-rep }
T{ ##andn-vector f 5 0 1 float-4-rep }
}
] [
{
T{ ##not-vector f 4 0 float-4-rep }
T{ ##and-vector f 5 1 4 float-4-rep }
} value-numbering-step
] unit-test
! NOT x ANDN y => x AND y
[
{
T{ ##fill-vector f 3 float-4-rep }
T{ ##xor-vector f 4 0 3 float-4-rep }
T{ ##and-vector f 5 0 1 float-4-rep }
}
] [
{
T{ ##fill-vector f 3 float-4-rep }
T{ ##xor-vector f 4 0 3 float-4-rep }
T{ ##andn-vector f 5 4 1 float-4-rep }
} value-numbering-step
] unit-test
[
{
T{ ##not-vector f 4 0 float-4-rep }
T{ ##and-vector f 5 0 1 float-4-rep }
}
] [
{
T{ ##not-vector f 4 0 float-4-rep }
T{ ##andn-vector f 5 4 1 float-4-rep }
} value-numbering-step
] unit-test
! AND <=> ANDN
[
{
T{ ##fill-vector f 3 float-4-rep }
T{ ##xor-vector f 4 0 3 float-4-rep }
T{ ##andn-vector f 5 0 1 float-4-rep }
T{ ##and-vector f 6 0 2 float-4-rep }
T{ ##or-vector f 7 5 6 float-4-rep }
}
] [
{
T{ ##fill-vector f 3 float-4-rep }
T{ ##xor-vector f 4 0 3 float-4-rep }
T{ ##and-vector f 5 4 1 float-4-rep }
T{ ##andn-vector f 6 4 2 float-4-rep }
T{ ##or-vector f 7 5 6 float-4-rep }
} value-numbering-step
] unit-test
[
{
T{ ##not-vector f 4 0 float-4-rep }
T{ ##andn-vector f 5 0 1 float-4-rep }
T{ ##and-vector f 6 0 2 float-4-rep }
T{ ##or-vector f 7 5 6 float-4-rep }
}
] [
{
T{ ##not-vector f 4 0 float-4-rep }
T{ ##and-vector f 5 4 1 float-4-rep }
T{ ##andn-vector f 6 4 2 float-4-rep }
T{ ##or-vector f 7 5 6 float-4-rep }
} value-numbering-step
] unit-test
! branch folding
: test-branch-folding ( insns -- insns' n )
<basic-block>
[ V{ 0 1 } clone >>successors basic-block set value-numbering-step ] keep

View File

@ -893,7 +893,7 @@ M: x86 %compare-vector ( dst src1 src2 rep cc -- )
M: x86 %compare-vector-reps
{
{ [ dup { cc= cc/= } memq? ] [ drop %compare-vector-eq-reps ] }
{ [ dup { cc= cc/= cc/<>= cc<>= } memq? ] [ drop %compare-vector-eq-reps ] }
[ drop %compare-vector-ord-reps ]
} cond ;
@ -1098,7 +1098,7 @@ M: x86 %min-vector ( dst src1 src2 rep -- )
M: x86 %min-vector-reps
{
{ sse? { float-4-rep } }
{ sse2? { uchar-16-rep short-8-rep double-2-rep short-8-rep uchar-16-rep } }
{ sse2? { uchar-16-rep short-8-rep double-2-rep } }
{ sse4.1? { char-16-rep ushort-8-rep int-4-rep uint-4-rep } }
} available-reps ;
@ -1118,7 +1118,7 @@ M: x86 %max-vector ( dst src1 src2 rep -- )
M: x86 %max-vector-reps
{
{ sse? { float-4-rep } }
{ sse2? { uchar-16-rep short-8-rep double-2-rep short-8-rep uchar-16-rep } }
{ sse2? { uchar-16-rep short-8-rep double-2-rep } }
{ sse4.1? { char-16-rep ushort-8-rep int-4-rep uint-4-rep } }
} available-reps ;

View File

@ -163,8 +163,8 @@ M: vector-rep supported-simd-op?
{ \ (simd-v*) [ %mul-vector-reps ] }
{ \ (simd-vs*) [ %saturated-mul-vector-reps ] }
{ \ (simd-v/) [ %div-vector-reps ] }
{ \ (simd-vmin) [ %min-vector-reps ] }
{ \ (simd-vmax) [ %max-vector-reps ] }
{ \ (simd-vmin) [ %min-vector-reps cc< %compare-vector-reps union ] }
{ \ (simd-vmax) [ %max-vector-reps cc> %compare-vector-reps union ] }
{ \ (simd-v.) [ %dot-vector-reps ] }
{ \ (simd-vsqrt) [ %sqrt-vector-reps ] }
{ \ (simd-sum) [ %horizontal-add-vector-reps ] }
@ -193,12 +193,12 @@ M: vector-rep supported-simd-op?
{ \ (simd-(vpack-unsigned)) [ %unsigned-pack-vector-reps ] }
{ \ (simd-(vunpack-head)) [ (%unpack-reps) ] }
{ \ (simd-(vunpack-tail)) [ (%unpack-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<=) [ unsign-rep cc<= %compare-vector-reps ] }
{ \ (simd-v<) [ unsign-rep cc< %compare-vector-reps ] }
{ \ (simd-v=) [ unsign-rep cc= %compare-vector-reps ] }
{ \ (simd-v>) [ unsign-rep cc> %compare-vector-reps ] }
{ \ (simd-v>=) [ unsign-rep cc>= %compare-vector-reps ] }
{ \ (simd-vunordered?) [ unsign-rep cc/<>= %compare-vector-reps ] }
{ \ (simd-gather-2) [ %gather-vector-2-reps ] }
{ \ (simd-gather-4) [ %gather-vector-4-reps ] }
{ \ (simd-vany?) [ %test-vector-reps ] }

View File

@ -101,6 +101,7 @@ $nl
vxor
vnot
v?
vif
}
"Entire vector tests:"
{ $subsections
@ -534,10 +535,19 @@ HELP: vnot
{ $notes "See " { $link "math-vectors-simd-logic" } " for notes on dealing with vector boolean inputs and results when using SIMD types." } ;
HELP: v?
{ $values { "mask" "a sequence of booleans" } { "true" "a sequence of numbers" } { "false" "a sequence of numbers" } { "w" "a sequence of numbers" } }
{ $values { "mask" "a sequence of booleans" } { "true" "a sequence of numbers" } { "false" "a sequence of numbers" } { "result" "a sequence of numbers" } }
{ $description "Creates a new sequence by selecting elements from the " { $snippet "true" } " and " { $snippet "false" } " sequences based on whether the corresponding bits of the " { $snippet "mask" } " sequence are set or not." }
{ $notes "See " { $link "math-vectors-simd-logic" } " for notes on dealing with vector boolean inputs and results when using SIMD types." } ;
HELP: vif
{ $values { "mask" "a sequence of booleans" } { "true-quot" { $quotation "( -- vector )" } } { "false-quot" { $quotation "( -- vector )" } } { "result" "a sequence" } }
{ $description "If all of the elements of " { $snippet "mask" } " are true, " { $snippet "true-quot" } " is called and its output value returned. If all of the elements of " { $snippet "mask" } " are false, " { $snippet "false-quot" } " is called and its output value returned. Otherwise, both quotations are called and " { $snippet "mask" } " is used to select elements from each output as with " { $link v? } "." }
{ $notes "See " { $link "math-vectors-simd-logic" } " for notes on dealing with vector boolean inputs and results when using SIMD types."
$nl
"For most conditional SIMD code, unless a case is exceptionally expensive to compute, it is usually most efficient to just compute all cases and blend them with " { $link v? } " instead of using " { $snippet "vif" } "." } ;
{ v? vif } related-words
HELP: vany?
{ $values { "v" "a sequence of booleans" } { "?" "a boolean" } }
{ $description "Returns true if any element of " { $snippet "v" } " is true." }

View File

@ -142,9 +142,16 @@ M: simd-128 vshuffle ( u perm -- v )
: vunordered? ( u v -- w ) [ unordered? ] 2map ;
: v= ( u v -- w ) [ = ] 2map ;
: v? ( mask true false -- w )
: v? ( mask true false -- result )
[ vand ] [ vandn ] bi-curry* bi vor ; inline
:: vif ( mask true-quot false-quot -- result )
{
{ [ mask vall? ] [ true-quot call ] }
{ [ mask vnone? ] [ false-quot call ] }
[ mask true-quot call false-quot call v? ]
} cond ; inline
: vfloor ( u -- v ) [ floor ] map ;
: vceiling ( u -- v ) [ ceiling ] map ;
: vtruncate ( u -- v ) [ truncate ] map ;
@ -175,20 +182,20 @@ PRIVATE>
: bilerp ( aa ba ab bb {t,u} -- a_tu )
[ first lerp ] [ second lerp ] bi-curry
[ 2bi@ ] [ call ] bi* ;
[ 2bi@ ] [ call ] bi* ; inline
: vlerp ( a b t -- a_t )
[ lerp ] 3map ;
[ over v- ] dip v* v+ ; inline
: vnlerp ( a b t -- a_t )
[ lerp ] curry 2map ;
[ over v- ] dip v*n v+ ; inline
: vbilerp ( aa ba ab bb {t,u} -- a_tu )
[ first vnlerp ] [ second vnlerp ] bi-curry
[ 2bi@ ] [ call ] bi* ;
[ 2bi@ ] [ call ] bi* ; inline
: v~ ( a b epsilon -- ? )
[ ~ ] curry 2all? ;
[ ~ ] curry 2all? ; inline
HINTS: vneg { array } ;
HINTS: norm-sq { array } ;