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

SSE integer gather and broadcast

db4
Joe Groff 2009-09-28 11:24:08 -05:00
parent 97785c0c09
commit 965730efbe
1 changed files with 61 additions and 34 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

95
basis/cpu/x86/x86.factor
View File

@ -578,40 +578,66 @@ MACRO: available-reps ( alist -- )
reverse [ { } ] suffix reverse [ { } ] suffix
'[ _ cond ] ; '[ _ cond ] ;
M: x86 %broadcast-vector ( dst src rep -- ) : unsign-rep ( rep -- rep' )
{ dup {
{ float-4-rep [ [ float-4-rep %copy ] [ drop dup 0 SHUFPS ] 2bi ] } { uint-4-rep int-4-rep }
{ double-2-rep [ [ double-2-rep %copy ] [ drop dup UNPCKLPD ] 2bi ] } { ulonglong-2-rep longlong-2-rep }
{ ushort-8-rep short-8-rep }
{ uchar-16-rep char-16-rep }
} at* [ nip ] [ drop ] if ;
M:: x86 %broadcast-vector ( dst src rep -- )
rep unsign-rep {
{ float-4-rep [
dst src float-4-rep %copy
dst dst 0 SHUFPS
] }
{ double-2-rep [
dst src double-2-rep %copy
dst dst UNPCKLPD
] }
{ longlong-2-rep [ dst src BIN: 01000100 PSHUFD ] }
{ int-4-rep [ dst src 0 PSHUFD ] }
{ short-8-rep [
dst src 0 PSHUFLW
dst dst PUNPCKLQDQ
] }
{ char-16-rep [
dst src char-16-rep %copy
dst dst PUNPCKLBW
dst dst 0 PSHUFLW
dst dst PUNPCKLQDQ
] }
} case ; } case ;
M: x86 %broadcast-vector-reps M: x86 %broadcast-vector-reps
{ {
! Can't do this with sse1 since it will want to unbox ! Can't do this with sse1 since it will want to unbox
! a double-precision float and convert to single precision ! a double-precision float and convert to single precision
{ sse2? { float-4-rep double-2-rep } } { sse2? {
float-4-rep double-2-rep
longlong-2-rep ulonglong-2-rep
int-4-rep uint-4-rep
short-8-rep ushort-8-rep
char-16-rep uchar-16-rep
} }
} available-reps ; } available-reps ;
M:: x86 %gather-vector-4 ( dst src1 src2 src3 src4 rep -- ) M:: x86 %gather-vector-4 ( dst src1 src2 src3 src4 rep -- )
{ rep unsign-rep {
{ { float-4-rep [
[ rep float-4-rep eq? ] dst src1 float-4-rep %copy
[ dst src2 UNPCKLPS
dst src1 float-4-rep %copy src3 src4 UNPCKLPS
dst src2 UNPCKLPS dst src3 MOVLHPS
src3 src4 UNPCKLPS ] }
dst src3 MOVLHPS { int-4-rep [
] dst src1 int-4-rep %copy
} dst src2 PUNPCKLDQ
{ src3 src4 PUNPCKLDQ
[ rep { int-4-rep uint-4-rep } memq? ] dst src3 PUNPCKLQDQ
[ ] }
dst src1 int-4-rep %copy } case ;
dst src2 PUNPCKLDQ
src3 src4 PUNPCKLDQ
dst src3 PUNPCKLQDQ
]
}
} cond ;
M: x86 %gather-vector-4-reps M: x86 %gather-vector-4-reps
{ {
@ -621,19 +647,20 @@ M: x86 %gather-vector-4-reps
} available-reps ; } available-reps ;
M:: x86 %gather-vector-2 ( dst src1 src2 rep -- ) M:: x86 %gather-vector-2 ( dst src1 src2 rep -- )
rep { rep unsign-rep {
{ { double-2-rep [
double-2-rep dst src1 double-2-rep %copy
[ dst src2 UNPCKLPD
dst src1 double-2-rep %copy ] }
dst src2 UNPCKLPD { longlong-2-rep [
] dst src1 longlong-2-rep %copy
} dst src2 PUNPCKLQDQ
] }
} case ; } case ;
M: x86 %gather-vector-2-reps M: x86 %gather-vector-2-reps
{ {
{ sse2? { double-2-rep } } { sse2? { double-2-rep longlong-2-rep ulonglong-2-rep } }
} available-reps ; } available-reps ;
M: x86 %add-vector ( dst src1 src2 rep -- ) M: x86 %add-vector ( dst src1 src2 rep -- )