Switch to using faster assembly routine (instead of C) for 640-mode pixel decoding loop.
This commit is contained in:
Stephen Heumann
parent
12999358b5
commit
9e401f4a83
4
make
4
make
|
|
@ -26,7 +26,7 @@ for file in vncview vncsession vncdisplay colortables \
|
|||
end
|
||||
end
|
||||
|
||||
for file in tables
|
||||
for file in tables rawdec
|
||||
unset exit
|
||||
newer {file}.a {file}.asm
|
||||
if {status} != 0
|
||||
|
|
@ -45,7 +45,7 @@ if {status} != 0
|
|||
end
|
||||
|
||||
if {link} == true
|
||||
link vncview vncsession vncdisplay colortables tables \
|
||||
link vncview vncsession vncdisplay colortables tables rawdec \
|
||||
desktopsize mouse keyboard copyrect raw hextile clipboard \
|
||||
keep=VNCview.GS
|
||||
filetype VNCview.GS S16 $DB03
|
||||
|
|
|
|||
54
raw.cc
54
raw.cc
|
|
@ -47,6 +47,9 @@ extern unsigned char destBuf[];
|
|||
|
||||
static unsigned char *destPtr;
|
||||
|
||||
unsigned char * rawDecode640(unsigned startOffset, unsigned endOffset,
|
||||
unsigned char *lineDataPtr);
|
||||
|
||||
/* Ends drawing of a raw rectangle when it is complete or aborted
|
||||
* because the rectangle is not visible.
|
||||
*/
|
||||
|
|
@ -116,55 +119,8 @@ void RawDraw (void) {
|
|||
finalDestPtr = destPtr + lineBytes - 1;
|
||||
if (hRez == 640) {
|
||||
initialLineDataPtr = lineDataPtr;
|
||||
while (destPtr + 7 < finalDestPtr) { /* Unrolled loop */
|
||||
*(unsigned*)destPtr =
|
||||
outPixels = (*(unsigned*)(bigcoltab640a + (inPixelsA = *(unsigned*)lineDataPtr))
|
||||
^ *(unsigned*)(bigcoltab640b + (inPixelsB = ((unsigned*)lineDataPtr)[1])));
|
||||
*(unsigned*)(destPtr+1) =
|
||||
((inPixelsA ^ (inPixelsA2 = ((unsigned*)lineDataPtr)[2])) |
|
||||
(inPixelsB ^ (inPixelsB2 = ((unsigned*)lineDataPtr)[3]))) == 0 ? outPixels :
|
||||
(outPixels = (*(unsigned*)(bigcoltab640a + inPixelsA2)
|
||||
^ *(unsigned*)(bigcoltab640b + inPixelsB2)));
|
||||
*(unsigned*)(destPtr+2) =
|
||||
((inPixelsA2 ^ (inPixelsA = ((unsigned*)lineDataPtr)[4])) |
|
||||
(inPixelsB2 ^ (inPixelsB = ((unsigned*)lineDataPtr)[5]))) == 0 ? outPixels :
|
||||
(outPixels = (*(unsigned*)(bigcoltab640a + inPixelsA)
|
||||
^ *(unsigned*)(bigcoltab640b + inPixelsB)));
|
||||
*(unsigned*)(destPtr+3) =
|
||||
((inPixelsA ^ (inPixelsA2 = ((unsigned*)lineDataPtr)[6])) |
|
||||
(inPixelsB ^ (inPixelsB2 = ((unsigned*)lineDataPtr)[7]))) == 0 ? outPixels :
|
||||
(outPixels = (*(unsigned*)(bigcoltab640a + inPixelsA2)
|
||||
^ *(unsigned*)(bigcoltab640b + inPixelsB2)));
|
||||
*(unsigned*)(destPtr+4) =
|
||||
((inPixelsA2 ^ (inPixelsA = ((unsigned*)lineDataPtr)[8])) |
|
||||
(inPixelsB2 ^ (inPixelsB = ((unsigned*)lineDataPtr)[9]))) == 0 ? outPixels :
|
||||
(outPixels = (*(unsigned*)(bigcoltab640a + inPixelsA)
|
||||
^ *(unsigned*)(bigcoltab640b + inPixelsB)));
|
||||
*(unsigned*)(destPtr+5) =
|
||||
((inPixelsA ^ (inPixelsA2 = ((unsigned*)lineDataPtr)[10])) |
|
||||
(inPixelsB ^ (inPixelsB2 = ((unsigned*)lineDataPtr)[11]))) == 0 ? outPixels :
|
||||
(outPixels = (*(unsigned*)(bigcoltab640a + inPixelsA2)
|
||||
^ *(unsigned*)(bigcoltab640b + inPixelsB2)));
|
||||
*(unsigned*)(destPtr+6) =
|
||||
((inPixelsA2 ^ (inPixelsA = ((unsigned*)lineDataPtr)[12])) |
|
||||
(inPixelsB2 ^ (inPixelsB = ((unsigned*)lineDataPtr)[13]))) == 0 ? outPixels :
|
||||
(outPixels = (*(unsigned*)(bigcoltab640a + inPixelsA)
|
||||
^ *(unsigned*)(bigcoltab640b + inPixelsB)));
|
||||
*(destPtr+7) =
|
||||
((inPixelsA ^ (inPixelsA2 = ((unsigned*)lineDataPtr)[14])) |
|
||||
(inPixelsB ^ (inPixelsB2 = ((unsigned*)lineDataPtr)[15]))) == 0 ? outPixels :
|
||||
((*(unsigned*)(bigcoltab640a + inPixelsA2)
|
||||
^ *(unsigned*)(bigcoltab640b + inPixelsB2)));
|
||||
destPtr += 8;
|
||||
lineDataPtr += 32;
|
||||
}
|
||||
while (destPtr < finalDestPtr) {
|
||||
*(destPtr++) =
|
||||
*(unsigned*)(bigcoltab640a + *(unsigned*)lineDataPtr)
|
||||
+ *(unsigned*)(bigcoltab640b + ((unsigned*)lineDataPtr)[1]);
|
||||
lineDataPtr += 4;
|
||||
}
|
||||
/* Final byte to produce */
|
||||
lineDataPtr = rawDecode640(destPtr-destBuf, finalDestPtr-destBuf, lineDataPtr);
|
||||
destPtr = finalDestPtr;
|
||||
*destPtr = pixTransTbl[*(lineDataPtr++)] & 0xC0;
|
||||
for (i = lineDataPtr - initialLineDataPtr; i < rectWidth; i++)
|
||||
switch (i & 0x03) {
|
||||
|
|
|
|||
|
|
@ -0,0 +1,118 @@
|
|||
mcopy rawdec.macros
|
||||
gen on
|
||||
case on
|
||||
|
||||
* Pointer to direct page space for use by these routines
|
||||
dpPtr data
|
||||
ds 4
|
||||
end
|
||||
|
||||
* DP locations
|
||||
destOfst gequ 0 offset into destBuf to start at
|
||||
endOfst gequ 2 offset into destBuf to end before
|
||||
loop1End gequ 4 offset into destBuf to end loop 1
|
||||
oldDB gequ 6 data bank on entry
|
||||
oldDP gequ 8 direct page on entry
|
||||
|
||||
* Generates 640-mode SHR pixels [destBuf+startOffset, destBuf+endOffset)
|
||||
* from raw pixels starting at lineDataPtr. Returns next lineDataPtr value.
|
||||
*
|
||||
* First loop is unrolled to minimize index calculation overhead,
|
||||
* and runs until less that 8 iterations (output bytes) remain.
|
||||
* Second loop generates the remaining bytes.
|
||||
*
|
||||
* unsigned char * rawDecode640(unsigned startOffset, unsigned endOffset,
|
||||
* unsigned char *lineDataPtr);
|
||||
rawDecode640 start rawDec640
|
||||
unroll equ 8 loop unrolling factor
|
||||
|
||||
tdc
|
||||
tax
|
||||
lda |dpPtr
|
||||
tcd set new direct page
|
||||
stx oldDP save direct page on entry
|
||||
|
||||
phb
|
||||
phb
|
||||
pla
|
||||
sta oldDB save data bank on entry
|
||||
|
||||
lda 10,S
|
||||
pha leaves extra byte: clean up later
|
||||
plb initialize data bank=bank of lineDataPtr
|
||||
lda 8+1,S initialize y = lineDataPtr (low 16 bits)
|
||||
tay
|
||||
|
||||
pla move return address to proper position
|
||||
sta 8-1,S
|
||||
pla
|
||||
sta 10-3,S
|
||||
|
||||
plx
|
||||
stx destOfst initialize x = destOfst = startOffset
|
||||
|
||||
pla
|
||||
sta endOfst initialize endOfst = endOffset
|
||||
|
||||
sec
|
||||
sbc #unroll-1
|
||||
bcs doLoop1 if endOffset-7 did not underflow...
|
||||
jmp test2
|
||||
|
||||
doLoop1 sta loop1End initialize loop1End = endOffset - 7
|
||||
txa a = startOffset
|
||||
jmp test1
|
||||
|
||||
loop1 sep #$20
|
||||
longa off
|
||||
loopBody640 unroll
|
||||
rep #$20
|
||||
longa on
|
||||
tya
|
||||
clc
|
||||
adc #unroll*4 carry must be clear
|
||||
tay
|
||||
bcs incDB1
|
||||
cont1 txa
|
||||
clc
|
||||
adc #unroll carry must be clear
|
||||
sta destOfst
|
||||
test1 cmp loop1End
|
||||
bge check2
|
||||
jmp loop1
|
||||
|
||||
check2 cmp endOfst
|
||||
bge end
|
||||
loop2 anop
|
||||
loopBody640 1
|
||||
inx
|
||||
stx destOfst
|
||||
tya
|
||||
clc
|
||||
adc #4 carry must be clear
|
||||
tay
|
||||
bcs incDB2
|
||||
test2 cpx endOfst
|
||||
blt loop2
|
||||
|
||||
end phb
|
||||
plx x = old DB (high byte of lineDataPtr)
|
||||
pei (oldDB)
|
||||
plb restore data bank
|
||||
plb
|
||||
lda oldDP
|
||||
tcd restore direct page
|
||||
tya a = lineDataPtr (low 16 bits)
|
||||
rtl
|
||||
|
||||
incDB1 pea cont1-1
|
||||
bra incDB
|
||||
incDB2 pea test2-1
|
||||
incDB phb
|
||||
pla
|
||||
inc A
|
||||
pha
|
||||
plb
|
||||
clc
|
||||
rts
|
||||
end
|
||||
|
|
@ -0,0 +1,13 @@
|
|||
macro
|
||||
loopBody640 &iters
|
||||
lcla &n
|
||||
.top
|
||||
ldx |4*&n,Y
|
||||
lda >BCT640A,X
|
||||
ldx |4*&n+2,Y
|
||||
ora >BCT640B,X
|
||||
LDX destOfst
|
||||
STA >destBuf+&n,X
|
||||
&n seta &n+1
|
||||
aif &n<&iters,.top
|
||||
mend
|
||||
12
vncview.cc
12
vncview.cc
|
|
@ -76,7 +76,8 @@ BOOLEAN vncConnected = FALSE; /* are we connected to a VNC host */
|
|||
int menuOffset; /* Indicates which menu bar is active */
|
||||
static Ref startStopParm; /* tool start/shutdown parameter */
|
||||
BOOLEAN colorTablesComplete = FALSE; /* Are the big color tables complete */
|
||||
|
||||
static Handle dpHndl; /* direct page space for assembly routines */
|
||||
extern void *dpPtr; /* pointer to DP for assembly routines */
|
||||
|
||||
/* Connection options */
|
||||
int hRez = 320;
|
||||
|
|
@ -357,6 +358,8 @@ static void Quit (void) {
|
|||
|
||||
if (readBufferHndl)
|
||||
DisposeHandle(readBufferHndl); /* Get rid of TCPIP read buffer hndl */
|
||||
if (dpHndl)
|
||||
DisposeHandle(dpHndl);
|
||||
|
||||
if (cursor)
|
||||
free(cursor);
|
||||
|
|
@ -400,6 +403,12 @@ int main (void) {
|
|||
}
|
||||
|
||||
readBufferHndl = NewHandle(1, userid(), 0, NULL);
|
||||
if (toolerror()) goto oomQuit;
|
||||
|
||||
dpHndl = NewHandle(0x0100, userid(),
|
||||
attrLocked|attrFixed|attrPage|attrNoCross|attrBank, 0x00000000);
|
||||
if (toolerror()) goto oomQuit;
|
||||
dpPtr = *dpHndl;
|
||||
|
||||
LoadOneTool(54, 0x200); /* load Marinetti 2.0+ */
|
||||
if (toolerror()) { /* Check that Marinetti is available */
|
||||
|
|
@ -412,6 +421,7 @@ int main (void) {
|
|||
TCPIPStartUp(); /* ... so activate it now */
|
||||
|
||||
if (toolerror()) { /* Get handle for TCPIP read buffer */
|
||||
oomQuit:
|
||||
SysBeep();
|
||||
InitCursor();
|
||||
AlertWindow(awResource, NULL, outOfMemoryError);
|
||||
|
|
|
|||
Loading…
Reference in New Issue