diff --git a/basis/compiler/cfg/hats/hats.factor b/basis/compiler/cfg/hats/hats.factor index 36fa631050..4bfcb3dac8 100644 --- a/basis/compiler/cfg/hats/hats.factor +++ b/basis/compiler/cfg/hats/hats.factor @@ -51,9 +51,6 @@ insn-classes get [ : ^^unbox-c-ptr ( src class -- dst ) [ next-vreg dup ] 2dip next-vreg ##unbox-c-ptr ; -: ^^neg ( src -- dst ) - [ 0 ^^load-literal ] dip ^^sub ; - : ^^allot-tuple ( n -- dst ) 2 + cells tuple ^^allot ; diff --git a/basis/compiler/cfg/instructions/instructions.factor b/basis/compiler/cfg/instructions/instructions.factor index d560cfbab6..2e9d2863c3 100644 --- a/basis/compiler/cfg/instructions/instructions.factor +++ b/basis/compiler/cfg/instructions/instructions.factor @@ -186,6 +186,10 @@ PURE-INSN: ##not def: dst/int-rep use: src/int-rep ; +PURE-INSN: ##neg +def: dst/int-rep +use: src/int-rep ; + PURE-INSN: ##log2 def: dst/int-rep use: src/int-rep ; @@ -270,6 +274,10 @@ def: dst use: src/int-rep literal: rep ; +PURE-INSN: ##zero-vector +def: dst +literal: rep ; + PURE-INSN: ##broadcast-vector def: dst use: src/scalar-rep @@ -285,6 +293,16 @@ def: dst use: src1/scalar-rep src2/scalar-rep src3/scalar-rep src4/scalar-rep literal: rep ; +PURE-INSN: ##shuffle-vector +def: dst +use: src +literal: shuffle rep ; + +PURE-INSN: ##select-vector +def: dst +use: src +literal: n rep ; + PURE-INSN: ##add-vector def: dst use: src1 src2 @@ -335,6 +353,11 @@ def: dst use: src1 src2 literal: rep ; +PURE-INSN: ##dot-vector +def: dst/scalar-rep +use: src1 src2 +literal: rep ; + PURE-INSN: ##horizontal-add-vector def: dst/scalar-rep use: src diff --git a/basis/compiler/cfg/intrinsics/intrinsics.factor b/basis/compiler/cfg/intrinsics/intrinsics.factor index 7b8e0c5989..2f6529692d 100644 --- a/basis/compiler/cfg/intrinsics/intrinsics.factor +++ b/basis/compiler/cfg/intrinsics/intrinsics.factor @@ -164,6 +164,7 @@ IN: compiler.cfg.intrinsics { 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-v.) [ [ ^^dot-vector ] emit-binary-vector-op ] } { math.vectors.simd.intrinsics:(simd-vabs) [ [ ^^abs-vector ] emit-unary-vector-op ] } { math.vectors.simd.intrinsics:(simd-vsqrt) [ [ ^^sqrt-vector ] emit-unary-vector-op ] } { math.vectors.simd.intrinsics:(simd-vbitand) [ [ ^^and-vector ] emit-binary-vector-op ] } @@ -177,6 +178,8 @@ IN: compiler.cfg.intrinsics { math.vectors.simd.intrinsics:(simd-broadcast) [ [ ^^broadcast-vector ] emit-unary-vector-op ] } { 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-vshuffle) [ emit-shuffle-vector ] } + { math.vectors.simd.intrinsics:(simd-vselect) [ emit-select-vector ] } { 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:set-alien-vector [ emit-set-alien-vector ] } diff --git a/basis/compiler/cfg/intrinsics/simd/simd.factor b/basis/compiler/cfg/intrinsics/simd/simd.factor index f975c7b39f..68012cfeb5 100644 --- a/basis/compiler/cfg/intrinsics/simd/simd.factor +++ b/basis/compiler/cfg/intrinsics/simd/simd.factor @@ -45,6 +45,13 @@ IN: compiler.cfg.intrinsics.simd ds-push ] emit-vector-op ; +: emit-shuffle-vector ( node -- ) + ; + +: emit-select-vector ( node -- ) + + ; + : emit-alien-vector ( node -- ) dup [ '[ diff --git a/basis/compiler/codegen/codegen.factor b/basis/compiler/codegen/codegen.factor index 7a8253a35e..12cf303b4c 100755 --- a/basis/compiler/codegen/codegen.factor +++ b/basis/compiler/codegen/codegen.factor @@ -142,6 +142,7 @@ CODEGEN: ##sar-imm %sar-imm CODEGEN: ##min %min CODEGEN: ##max %max CODEGEN: ##not %not +CODEGEN: ##neg %neg CODEGEN: ##log2 %log2 CODEGEN: ##copy %copy CODEGEN: ##unbox-float %unbox-float @@ -160,9 +161,12 @@ CODEGEN: ##double>single-float %double>single-float CODEGEN: ##integer>float %integer>float CODEGEN: ##float>integer %float>integer CODEGEN: ##unbox-vector %unbox-vector +CODEGEN: ##zero-vector %zero-vector CODEGEN: ##broadcast-vector %broadcast-vector CODEGEN: ##gather-vector-2 %gather-vector-2 CODEGEN: ##gather-vector-4 %gather-vector-4 +CODEGEN: ##shuffle-vector %shuffle-vector +CODEGEN: ##select-vector %select-vector CODEGEN: ##box-vector %box-vector CODEGEN: ##add-vector %add-vector CODEGEN: ##saturated-add-vector %saturated-add-vector @@ -174,6 +178,7 @@ CODEGEN: ##saturated-mul-vector %saturated-mul-vector CODEGEN: ##div-vector %div-vector CODEGEN: ##min-vector %min-vector CODEGEN: ##max-vector %max-vector +CODEGEN: ##dot-vector %dot-vector CODEGEN: ##sqrt-vector %sqrt-vector CODEGEN: ##horizontal-add-vector %horizontal-add-vector CODEGEN: ##horizontal-sub-vector %horizontal-sub-vector diff --git a/basis/cpu/architecture/architecture.factor b/basis/cpu/architecture/architecture.factor index 1e4e27b6e4..9222fcc17c 100644 --- a/basis/cpu/architecture/architecture.factor +++ b/basis/cpu/architecture/architecture.factor @@ -180,6 +180,7 @@ HOOK: %sar-imm cpu ( dst src1 src2 -- ) HOOK: %min cpu ( dst src1 src2 -- ) HOOK: %max cpu ( dst src1 src2 -- ) HOOK: %not cpu ( dst src -- ) +HOOK: %neg cpu ( dst src -- ) HOOK: %log2 cpu ( dst src -- ) HOOK: %copy cpu ( dst src rep -- ) @@ -210,9 +211,12 @@ HOOK: %float>integer cpu ( dst src -- ) HOOK: %box-vector cpu ( dst src temp rep -- ) HOOK: %unbox-vector cpu ( dst src rep -- ) +HOOK: %zero-vector cpu ( dst rep -- ) HOOK: %broadcast-vector cpu ( dst src rep -- ) HOOK: %gather-vector-2 cpu ( dst src1 src2 rep -- ) HOOK: %gather-vector-4 cpu ( dst src1 src2 src3 src4 rep -- ) +HOOK: %shuffle-vector cpu ( dst src shuffle rep -- ) +HOOK: %select-vector cpu ( dst src n rep -- ) HOOK: %add-vector cpu ( dst src1 src2 rep -- ) HOOK: %saturated-add-vector cpu ( dst src1 src2 rep -- ) HOOK: %add-sub-vector cpu ( dst src1 src2 rep -- ) @@ -223,6 +227,7 @@ HOOK: %saturated-mul-vector cpu ( dst src1 src2 rep -- ) HOOK: %div-vector cpu ( dst src1 src2 rep -- ) HOOK: %min-vector cpu ( dst src1 src2 rep -- ) HOOK: %max-vector cpu ( dst src1 src2 rep -- ) +HOOK: %dot-vector cpu ( dst src1 src2 rep -- ) HOOK: %sqrt-vector cpu ( dst src rep -- ) HOOK: %horizontal-add-vector cpu ( dst src rep -- ) HOOK: %horizontal-sub-vector cpu ( dst src rep -- ) @@ -239,9 +244,12 @@ HOOK: %horizontal-shr-vector cpu ( dst src1 src2 rep -- ) HOOK: %integer>scalar cpu ( dst src rep -- ) HOOK: %scalar>integer cpu ( dst src rep -- ) +HOOK: %zero-vector-reps cpu ( -- reps ) HOOK: %broadcast-vector-reps cpu ( -- reps ) HOOK: %gather-vector-2-reps cpu ( -- reps ) HOOK: %gather-vector-4-reps cpu ( -- reps ) +HOOK: %shuffle-vector-reps cpu ( -- reps ) +HOOK: %select-vector-reps cpu ( -- reps ) HOOK: %add-vector-reps cpu ( -- reps ) HOOK: %saturated-add-vector-reps cpu ( -- reps ) HOOK: %add-sub-vector-reps cpu ( -- reps ) @@ -252,6 +260,7 @@ HOOK: %saturated-mul-vector-reps cpu ( -- reps ) HOOK: %div-vector-reps cpu ( -- reps ) HOOK: %min-vector-reps cpu ( -- reps ) HOOK: %max-vector-reps cpu ( -- reps ) +HOOK: %dot-vector-reps cpu ( -- reps ) HOOK: %sqrt-vector-reps cpu ( -- reps ) HOOK: %horizontal-add-vector-reps cpu ( -- reps ) HOOK: %horizontal-sub-vector-reps cpu ( -- reps ) diff --git a/basis/cpu/x86/x86.factor b/basis/cpu/x86/x86.factor index 857e1ede6b..4d9c53d774 100644 --- a/basis/cpu/x86/x86.factor +++ b/basis/cpu/x86/x86.factor @@ -129,6 +129,7 @@ M: x86 %min int-rep two-operand [ CMP ] [ CMOVG ] 2bi ; M: x86 %max int-rep two-operand [ CMP ] [ CMOVL ] 2bi ; M: x86 %not int-rep one-operand NOT ; +M: x86 %neg int-rep one-operand NEG ; M: x86 %log2 BSR ; GENERIC: copy-register* ( dst src rep -- ) @@ -578,6 +579,19 @@ MACRO: available-reps ( alist -- ) reverse [ { } ] suffix '[ _ cond ] ; +M: x86 %zero-vector + { + { double-2-rep [ dup XORPD ] } + { float-4-rep [ dup XORPS ] } + [ drop dup PXOR ] + } case ; + +M: x86 %zero-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 ; + : unsign-rep ( rep -- rep' ) { { uint-4-rep int-4-rep } @@ -663,6 +677,10 @@ M: x86 %gather-vector-2-reps { sse2? { double-2-rep longlong-2-rep ulonglong-2-rep } } } available-reps ; +M: x86 %shuffle-vector-reps { } ; + +M: x86 %select-vector-reps { } ; + M: x86 %add-vector ( dst src1 src2 rep -- ) [ two-operand ] keep { @@ -820,6 +838,28 @@ M: x86 %max-vector-reps { sse4.1? { char-16-rep ushort-8-rep int-4-rep uint-4-rep } } } available-reps ; +M: x86 %dot-vector + [ two-operand ] keep + { + { float-4-rep [ + sse4.1? + [ HEX: ff DPPS ] + [ [ MULPS ] [ drop dup float-4-rep %horizontal-add-vector ] 2bi ] + if + ] } + { double-2-rep [ + sse4.1? + [ HEX: ff DPPD ] + [ [ MULPD ] [ drop dup double-2-rep %horizontal-add-vector ] 2bi ] + if + ] } + } case ; + +M: x86 %dot-vector-reps + { + { sse3? { float-4-rep double-2-rep } } + } available-reps ; + M: x86 %horizontal-add-vector ( dst src rep -- ) { { float-4-rep [ [ float-4-rep %copy ] [ HADDPS ] [ HADDPS ] 2tri ] } diff --git a/basis/math/vectors/simd/functor/functor.factor b/basis/math/vectors/simd/functor/functor.factor index bc42bddf02..c2eb85e3a6 100644 --- a/basis/math/vectors/simd/functor/functor.factor +++ b/basis/math/vectors/simd/functor/functor.factor @@ -55,7 +55,7 @@ ERROR: bad-schema schema ; :: high-level-ops ( ctor elt-class -- assoc ) ! Some SIMD operations are defined in terms of others. { - { vneg [ [ dup v- ] keep v- ] } + { vneg [ [ dup vbitxor ] keep v- ] } { n+v [ [ ctor execute ] dip v+ ] } { v+n [ ctor execute v+ ] } { n-v [ [ ctor execute ] dip v- ] } @@ -71,12 +71,7 @@ ERROR: bad-schema schema ; ! To compute dot product and distance with integer vectors, we ! have to do things less efficiently, with integer overflow checks, ! in the general case. - elt-class m:float = [ - { - { distance [ v- norm ] } - { v. [ v* sum ] } - } append - ] when ; + elt-class m:float = [ { distance [ v- norm ] } suffix ] when ; :: simd-vector-words ( class ctor rep vv->v vn->v v->v v->n -- ) rep rep-component-type c-type-boxed-class :> elt-class diff --git a/basis/math/vectors/simd/intrinsics/intrinsics.factor b/basis/math/vectors/simd/intrinsics/intrinsics.factor index 9d31dfa813..78301ae3a5 100644 --- a/basis/math/vectors/simd/intrinsics/intrinsics.factor +++ b/basis/math/vectors/simd/intrinsics/intrinsics.factor @@ -36,6 +36,7 @@ SIMD-OP: v* SIMD-OP: v/ SIMD-OP: vmin SIMD-OP: vmax +SIMD-OP: v. SIMD-OP: vsqrt SIMD-OP: sum SIMD-OP: vabs @@ -47,10 +48,12 @@ SIMD-OP: vlshift SIMD-OP: vrshift SIMD-OP: hlshift SIMD-OP: hrshift +SIMD-OP: vshuffle : (simd-broadcast) ( x rep -- v ) bad-simd-call ; : (simd-gather-2) ( a b rep -- v ) bad-simd-call ; : (simd-gather-4) ( a b c d rep -- v ) bad-simd-call ; +: (simd-select) ( v n rep -- x ) bad-simd-call ; : assert-positive ( x -- y ) ; @@ -110,6 +113,7 @@ M: vector-rep supported-simd-op? { \ (simd-v/) [ %div-vector-reps ] } { \ (simd-vmin) [ %min-vector-reps ] } { \ (simd-vmax) [ %max-vector-reps ] } + { \ (simd-v.) [ %dot-vector-reps ] } { \ (simd-vsqrt) [ %sqrt-vector-reps ] } { \ (simd-sum) [ %horizontal-add-vector-reps ] } { \ (simd-vabs) [ %abs-vector-reps ] } @@ -121,7 +125,9 @@ M: vector-rep supported-simd-op? { \ (simd-vrshift) [ %shr-vector-reps ] } { \ (simd-hlshift) [ %horizontal-shl-vector-reps ] } { \ (simd-hrshift) [ %horizontal-shr-vector-reps ] } + { \ (simd-vshuffle) [ %shuffle-vector-reps ] } { \ (simd-broadcast) [ %broadcast-vector-reps ] } { \ (simd-gather-2) [ %gather-vector-2-reps ] } { \ (simd-gather-4) [ %gather-vector-4-reps ] } + { \ (simd-select) [ %select-vector-reps ] } } case member? ; diff --git a/basis/math/vectors/vectors.factor b/basis/math/vectors/vectors.factor index de9ba51aec..5f9b7e395b 100644 --- a/basis/math/vectors/vectors.factor +++ b/basis/math/vectors/vectors.factor @@ -66,6 +66,9 @@ PRIVATE> GENERIC: new-underlying ( underlying seq -- seq' ) +: change-underlying ( seq quot -- seq' ) + '[ underlying>> @ ] keep new-underlying ; inline + PRIVATE> : vbitand ( u v -- w ) over '[ _ [ bitand ] fp-bitwise-op ] 2map ; @@ -74,6 +77,14 @@ PRIVATE> : vbitxor ( u v -- w ) over '[ _ [ bitxor ] fp-bitwise-op ] 2map ; : vbitnot ( u -- w ) dup '[ _ [ bitnot ] fp-bitwise-unary ] map ; +: vshuffle ( u perm -- v ) swap nths ; + +: vlshift ( u n -- w ) '[ _ shift ] map ; +: vrshift ( u n -- w ) neg '[ _ shift ] map ; + +: hlshift ( u n -- w ) '[ _ prepend 16 head ] change-underlying ; +: hrshift ( u n -- w ) '[ _ append 16 tail* ] change-underlying ; + : vand ( u v -- w ) [ and ] 2map ; : vor ( u v -- w ) [ or ] 2map ; : vxor ( u v -- w ) [ xor ] 2map ; @@ -88,15 +99,6 @@ PRIVATE> : v? ( ? u v -- w ) [ ? ] pick 3map-as ; -: vlshift ( u n -- w ) '[ _ shift ] map ; -: vrshift ( u n -- w ) neg '[ _ shift ] map ; - -: hlshift ( u n -- w ) - [ [ underlying>> ] dip prepend 16 head ] [ drop ] 2bi new-underlying ; - -: hrshift ( u n -- w ) - [ [ underlying>> ] dip append 16 tail* ] [ drop ] 2bi new-underlying ; - : vfloor ( u -- v ) [ floor ] map ; : vceiling ( u -- v ) [ ceiling ] map ; : vtruncate ( u -- v ) [ truncate ] map ;