Merge branch 'master' of git://factorcode.org/git/factor
commit
9b64d206ff
|
@ -235,8 +235,13 @@ M:: ppc %box-float ( dst src temp -- )
|
|||
dst 16 float temp %allot
|
||||
src dst float-offset STFD ;
|
||||
|
||||
: float-function-param ( i spill-slot -- )
|
||||
[ float-regs param-regs nth 1 ] [ n>> spill@ ] bi* LFD ;
|
||||
GENERIC: float-function-param* ( dst src -- )
|
||||
|
||||
M: spill-slot float-function-param* [ 1 ] dip n>> spill@ LFD ;
|
||||
M: integer float-function-param* FMR ;
|
||||
|
||||
: float-function-param ( i src -- )
|
||||
[ float-regs param-regs nth ] dip float-function-param* ;
|
||||
|
||||
: float-function-return ( reg -- )
|
||||
float-regs return-reg double-rep %copy ;
|
||||
|
|
|
@ -618,11 +618,15 @@ ALIAS: PINSRQ PINSRD
|
|||
: MOVDQA ( dest src -- ) { HEX: 6f HEX: 7f } HEX: 66 2-operand-rm-mr-sse ;
|
||||
: MOVDQU ( dest src -- ) { HEX: 6f HEX: 7f } HEX: f3 2-operand-rm-mr-sse ;
|
||||
|
||||
<PRIVATE
|
||||
|
||||
: 2shuffler ( indexes/mask -- mask )
|
||||
dup integer? [ first2 { 1 0 } bitfield ] unless ;
|
||||
: 4shuffler ( indexes/mask -- mask )
|
||||
dup integer? [ first4 { 6 4 2 0 } bitfield ] unless ;
|
||||
|
||||
PRIVATE>
|
||||
|
||||
: PSHUFD ( dest src imm -- ) 4shuffler HEX: 70 HEX: 66 3-operand-rm-sse ;
|
||||
: PSHUFLW ( dest src imm -- ) 4shuffler HEX: 70 HEX: f2 3-operand-rm-sse ;
|
||||
: PSHUFHW ( dest src imm -- ) 4shuffler HEX: 70 HEX: f3 3-operand-rm-sse ;
|
||||
|
|
|
@ -579,12 +579,12 @@ MACRO: available-reps ( alist -- )
|
|||
'[ _ cond ] ;
|
||||
|
||||
: unsign-rep ( rep -- rep' )
|
||||
dup {
|
||||
{
|
||||
{ uint-4-rep int-4-rep }
|
||||
{ ulonglong-2-rep longlong-2-rep }
|
||||
{ ushort-8-rep short-8-rep }
|
||||
{ uchar-16-rep char-16-rep }
|
||||
} at* [ nip ] [ drop ] if ;
|
||||
} ?at drop ;
|
||||
|
||||
M:: x86 %broadcast-vector ( dst src rep -- )
|
||||
rep unsign-rep {
|
||||
|
@ -592,22 +592,23 @@ M:: x86 %broadcast-vector ( dst src rep -- )
|
|||
dst src float-4-rep %copy
|
||||
dst dst { 0 0 0 0 } SHUFPS
|
||||
] }
|
||||
{ double-2-rep [
|
||||
{ double-2-rep [
|
||||
dst src MOVDDUP
|
||||
] }
|
||||
{ longlong-2-rep [
|
||||
dst src = [
|
||||
dst dst PUNPCKLQDQ
|
||||
] [
|
||||
dst src { 0 1 0 1 } PSHUFD
|
||||
] if
|
||||
dst src =
|
||||
[ dst dst PUNPCKLQDQ ]
|
||||
[ dst src { 0 1 0 1 } PSHUFD ]
|
||||
if
|
||||
] }
|
||||
{ int-4-rep [ dst src { 0 0 0 0 } PSHUFD ] }
|
||||
{ short-8-rep [
|
||||
{ int-4-rep [
|
||||
dst src { 0 0 0 0 } PSHUFD
|
||||
] }
|
||||
{ short-8-rep [
|
||||
dst src { 0 0 0 0 } PSHUFLW
|
||||
dst dst PUNPCKLQDQ
|
||||
] }
|
||||
{ char-16-rep [
|
||||
{ char-16-rep [
|
||||
dst src char-16-rep %copy
|
||||
dst dst PUNPCKLBW
|
||||
dst dst { 0 0 0 0 } PSHUFLW
|
||||
|
@ -619,13 +620,7 @@ M: x86 %broadcast-vector-reps
|
|||
{
|
||||
! Can't do this with sse1 since it will want to unbox
|
||||
! a double-precision float and convert to single precision
|
||||
{ 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
|
||||
} }
|
||||
{ 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 ;
|
||||
|
||||
M:: x86 %gather-vector-4 ( dst src1 src2 src3 src4 rep -- )
|
||||
|
|
|
@ -20,8 +20,9 @@ ERROR: bad-base-type type ;
|
|||
PRIVATE>
|
||||
|
||||
: define-simd-vocab ( type -- vocab )
|
||||
parse-base-type
|
||||
[ simd-vocab ] keep '[
|
||||
_ parse-base-type
|
||||
_
|
||||
[ define-simd-128 ]
|
||||
[ define-simd-256 ] bi
|
||||
] generate-vocab ;
|
||||
|
|
|
@ -1,6 +1,7 @@
|
|||
USING: specialized-arrays sequences.complex
|
||||
kernel sequences tools.test arrays accessors ;
|
||||
SPECIALIZED-ARRAY: float
|
||||
QUALIFIED-WITH: alien.c-types c
|
||||
SPECIALIZED-ARRAY: c:float
|
||||
IN: sequences.complex.tests
|
||||
|
||||
: test-array ( -- x )
|
||||
|
|
|
@ -5,7 +5,8 @@ USING: arrays accessors io io.files io.files.temp
|
|||
io.encodings.binary kernel math math.constants math.functions
|
||||
math.vectors math.vectors.simd math.parser make sequences
|
||||
sequences.private words hints classes.struct ;
|
||||
SIMD: double
|
||||
QUALIFIED-WITH: alien.c-types c
|
||||
SIMD: c:double
|
||||
IN: benchmark.raytracer-simd
|
||||
|
||||
! parameters
|
||||
|
|
|
@ -2,7 +2,8 @@
|
|||
! See http://factorcode.org/license.txt for BSD license.
|
||||
USING: kernel io math math.functions math.parser math.vectors
|
||||
math.vectors.simd sequences specialized-arrays ;
|
||||
SIMD: float
|
||||
QUALIFIED-WITH: alien.c-types c
|
||||
SIMD: c:float
|
||||
SPECIALIZED-ARRAY: float-4
|
||||
IN: benchmark.simd-1
|
||||
|
||||
|
|
Loading…
Reference in New Issue