mirror of
https://github.com/Rockbox/rockbox.git
synced 2025-11-13 07:02:31 -05:00
Port Tom Meyer's lcd_yuv_blit() optimisations for the ipod 5g to the ipod Photo/Color and Nano - gives a similar 10-15% speedup in mpegplayer.
git-svn-id: svn://svn.rockbox.org/rockbox/trunk@10629 a1c6a512-1295-4272-9138-f99709370657
This commit is contained in:
Dave Chapman
parent
30f237303b
commit
5b71470185
1 changed files with 103 additions and 83 deletions
|
|
@ -404,14 +404,15 @@ void lcd_blit(const fb_data* data, int x, int by, int width,
|
||||||
#define CSUB_Y 2
|
#define CSUB_Y 2
|
||||||
|
|
||||||
#define RYFAC (31*257)
|
#define RYFAC (31*257)
|
||||||
#define GYFAC (63*257)
|
#define GYFAC (31*257)
|
||||||
#define BYFAC (31*257)
|
#define BYFAC (31*257)
|
||||||
#define RVFAC 11170 /* 31 * 257 * 1.402 */
|
#define RVFAC 11170 /* 31 * 257 * 1.402 */
|
||||||
#define GVFAC (-11563) /* 63 * 257 * -0.714136 */
|
#define GVFAC (-5690) /* 31 * 257 * -0.714136 */
|
||||||
#define GUFAC (-5572) /* 63 * 257 * -0.344136 */
|
#define GUFAC (-2742) /* 31 * 257 * -0.344136 */
|
||||||
#define BUFAC 14118 /* 31 * 257 * 1.772 */
|
#define BUFAC 14118 /* 31 * 257 * 1.772 */
|
||||||
|
|
||||||
#define ROUNDOFFS (127*257)
|
#define ROUNDOFFS (127*257)
|
||||||
|
#define ROUNDOFFSG (63*257)
|
||||||
|
|
||||||
/* Performance function to blit a YUV bitmap directly to the LCD */
|
/* Performance function to blit a YUV bitmap directly to the LCD */
|
||||||
void lcd_yuv_blit(unsigned char * const src[3],
|
void lcd_yuv_blit(unsigned char * const src[3],
|
||||||
|
|
@ -478,14 +479,27 @@ void lcd_yuv_blit(unsigned char * const src[3],
|
||||||
lcd_send_lo(LCD_CNTL_WRITE_TO_GRAM);
|
lcd_send_lo(LCD_CNTL_WRITE_TO_GRAM);
|
||||||
}
|
}
|
||||||
|
|
||||||
|
const int stride_div_csub_x = stride/CSUB_X;
|
||||||
|
|
||||||
h=0;
|
h=0;
|
||||||
while (1) {
|
while (1) {
|
||||||
int pixels_to_write;
|
/* upsampling, YUV->RGB conversion and reduction to RGB565 in one go */
|
||||||
const unsigned char *ysrc = src[0] + stride * src_y + src_x;
|
const unsigned char *ysrc = src[0] + stride * src_y + src_x;
|
||||||
|
|
||||||
|
const int uvoffset = stride_div_csub_x * (src_y/CSUB_Y) +
|
||||||
|
(src_x/CSUB_X);
|
||||||
|
|
||||||
|
const unsigned char *usrc = src[1] + uvoffset;
|
||||||
|
const unsigned char *vsrc = src[2] + uvoffset;
|
||||||
const unsigned char *row_end = ysrc + width;
|
const unsigned char *row_end = ysrc + width;
|
||||||
|
|
||||||
int y, u, v;
|
int y, u, v;
|
||||||
int red, green, blue;
|
int red1, green1, blue1;
|
||||||
|
int red2, green2, blue2;
|
||||||
unsigned rbits, gbits, bbits;
|
unsigned rbits, gbits, bbits;
|
||||||
|
|
||||||
|
int rc, gc, bc;
|
||||||
|
int pixels_to_write;
|
||||||
fb_data pixel1,pixel2;
|
fb_data pixel1,pixel2;
|
||||||
|
|
||||||
if (h==0) {
|
if (h==0) {
|
||||||
|
|
@ -509,91 +523,97 @@ void lcd_yuv_blit(unsigned char * const src[3],
|
||||||
outl(0x34000000, 0x70008a20);
|
outl(0x34000000, 0x70008a20);
|
||||||
}
|
}
|
||||||
|
|
||||||
/* upsampling, YUV->RGB conversion and reduction to RGB565 in one go */
|
do
|
||||||
const unsigned char *usrc = src[1] + (stride/CSUB_X) * (src_y/CSUB_Y)
|
{
|
||||||
+ (src_x/CSUB_X);
|
|
||||||
const unsigned char *vsrc = src[2] + (stride/CSUB_X) * (src_y/CSUB_Y)
|
|
||||||
+ (src_x/CSUB_X);
|
|
||||||
int rc, gc, bc;
|
|
||||||
|
|
||||||
u = *usrc++ - 128;
|
u = *usrc++ - 128;
|
||||||
v = *vsrc++ - 128;
|
v = *vsrc++ - 128;
|
||||||
rc = RVFAC * v + ROUNDOFFS;
|
rc = RVFAC * v + ROUNDOFFS;
|
||||||
gc = GVFAC * v + GUFAC * u + ROUNDOFFS;
|
gc = GVFAC * v + GUFAC * u + ROUNDOFFSG;
|
||||||
bc = BUFAC * u + ROUNDOFFS;
|
bc = BUFAC * u + ROUNDOFFS;
|
||||||
|
|
||||||
do
|
/* Pixel 1 */
|
||||||
{
|
|
||||||
y = *ysrc++;
|
y = *ysrc++;
|
||||||
red = RYFAC * y + rc;
|
|
||||||
green = GYFAC * y + gc;
|
|
||||||
blue = BYFAC * y + bc;
|
|
||||||
|
|
||||||
if ((unsigned)red > (RYFAC*255+ROUNDOFFS))
|
red1 = RYFAC * y + rc;
|
||||||
|
green1 = GYFAC * y + gc;
|
||||||
|
blue1 = BYFAC * y + bc;
|
||||||
|
|
||||||
|
/* Pixel 2 */
|
||||||
|
y = *ysrc++;
|
||||||
|
red2 = RYFAC * y + rc;
|
||||||
|
green2 = GYFAC * y + gc;
|
||||||
|
blue2 = BYFAC * y + bc;
|
||||||
|
|
||||||
|
/* Since out of bounds errors are relatively rare, we check two
|
||||||
|
pixels at once to see if any components are out of bounds, and
|
||||||
|
then fix whichever is broken. This works due to high values and
|
||||||
|
negative values both becoming larger than the cutoff when
|
||||||
|
casted to unsigned. And ORing them together checks all of them
|
||||||
|
simultaneously. */
|
||||||
|
if (((unsigned)(red1 | green1 | blue1 |
|
||||||
|
red2 | green2 | blue2)) > (RYFAC*255+ROUNDOFFS)) {
|
||||||
|
if (((unsigned)(red1 | green1 | blue1)) >
|
||||||
|
(RYFAC*255+ROUNDOFFS)) {
|
||||||
|
if ((unsigned)red1 > (RYFAC*255+ROUNDOFFS))
|
||||||
{
|
{
|
||||||
if (red < 0)
|
if (red1 < 0)
|
||||||
red = 0;
|
red1 = 0;
|
||||||
else
|
else
|
||||||
red = (RYFAC*255+ROUNDOFFS);
|
red1 = (RYFAC*255+ROUNDOFFS);
|
||||||
}
|
}
|
||||||
if ((unsigned)green > (GYFAC*255+ROUNDOFFS))
|
if ((unsigned)green1 > (GYFAC*255+ROUNDOFFSG))
|
||||||
{
|
{
|
||||||
if (green < 0)
|
if (green1 < 0)
|
||||||
green = 0;
|
green1 = 0;
|
||||||
else
|
else
|
||||||
green = (GYFAC*255+ROUNDOFFS);
|
green1 = (GYFAC*255+ROUNDOFFSG);
|
||||||
}
|
}
|
||||||
if ((unsigned)blue > (BYFAC*255+ROUNDOFFS))
|
if ((unsigned)blue1 > (BYFAC*255+ROUNDOFFS))
|
||||||
{
|
{
|
||||||
if (blue < 0)
|
if (blue1 < 0)
|
||||||
blue = 0;
|
blue1 = 0;
|
||||||
else
|
else
|
||||||
blue = (BYFAC*255+ROUNDOFFS);
|
blue1 = (BYFAC*255+ROUNDOFFS);
|
||||||
}
|
}
|
||||||
rbits = ((unsigned)red) >> 16 ;
|
}
|
||||||
gbits = ((unsigned)green) >> 16 ;
|
|
||||||
bbits = ((unsigned)blue) >> 16 ;
|
if (((unsigned)(red2 | green2 | blue2)) >
|
||||||
|
(RYFAC*255+ROUNDOFFS)) {
|
||||||
|
if ((unsigned)red2 > (RYFAC*255+ROUNDOFFS))
|
||||||
|
{
|
||||||
|
if (red2 < 0)
|
||||||
|
red2 = 0;
|
||||||
|
else
|
||||||
|
red2 = (RYFAC*255+ROUNDOFFS);
|
||||||
|
}
|
||||||
|
if ((unsigned)green2 > (GYFAC*255+ROUNDOFFSG))
|
||||||
|
{
|
||||||
|
if (green2 < 0)
|
||||||
|
green2 = 0;
|
||||||
|
else
|
||||||
|
green2 = (GYFAC*255+ROUNDOFFSG);
|
||||||
|
}
|
||||||
|
if ((unsigned)blue2 > (BYFAC*255+ROUNDOFFS))
|
||||||
|
{
|
||||||
|
if (blue2 < 0)
|
||||||
|
blue2 = 0;
|
||||||
|
else
|
||||||
|
blue2 = (BYFAC*255+ROUNDOFFS);
|
||||||
|
}
|
||||||
|
}
|
||||||
|
}
|
||||||
|
|
||||||
|
rbits = red1 >> 16 ;
|
||||||
|
gbits = green1 >> 15 ;
|
||||||
|
bbits = blue1 >> 16 ;
|
||||||
|
|
||||||
pixel1 = swap16((rbits << 11) | (gbits << 5) | bbits);
|
pixel1 = swap16((rbits << 11) | (gbits << 5) | bbits);
|
||||||
|
|
||||||
y = *ysrc++;
|
rbits = red2 >> 16 ;
|
||||||
red = RYFAC * y + rc;
|
gbits = green2 >> 15 ;
|
||||||
green = GYFAC * y + gc;
|
bbits = blue2 >> 16 ;
|
||||||
blue = BYFAC * y + bc;
|
|
||||||
|
|
||||||
if ((unsigned)red > (RYFAC*255+ROUNDOFFS))
|
|
||||||
{
|
|
||||||
if (red < 0)
|
|
||||||
red = 0;
|
|
||||||
else
|
|
||||||
red = (RYFAC*255+ROUNDOFFS);
|
|
||||||
}
|
|
||||||
if ((unsigned)green > (GYFAC*255+ROUNDOFFS))
|
|
||||||
{
|
|
||||||
if (green < 0)
|
|
||||||
green = 0;
|
|
||||||
else
|
|
||||||
green = (GYFAC*255+ROUNDOFFS);
|
|
||||||
}
|
|
||||||
if ((unsigned)blue > (BYFAC*255+ROUNDOFFS))
|
|
||||||
{
|
|
||||||
if (blue < 0)
|
|
||||||
blue = 0;
|
|
||||||
else
|
|
||||||
blue = (BYFAC*255+ROUNDOFFS);
|
|
||||||
}
|
|
||||||
rbits = ((unsigned)red) >> 16 ;
|
|
||||||
gbits = ((unsigned)green) >> 16 ;
|
|
||||||
bbits = ((unsigned)blue) >> 16 ;
|
|
||||||
|
|
||||||
pixel2 = swap16((rbits << 11) | (gbits << 5) | bbits);
|
pixel2 = swap16((rbits << 11) | (gbits << 5) | bbits);
|
||||||
|
|
||||||
u = *usrc++ - 128;
|
|
||||||
v = *vsrc++ - 128;
|
|
||||||
rc = RVFAC * v + ROUNDOFFS;
|
|
||||||
gc = GVFAC * v + GUFAC * u + ROUNDOFFS;
|
|
||||||
bc = BUFAC * u + ROUNDOFFS;
|
|
||||||
|
|
||||||
while ((inl(0x70008a20) & 0x1000000) == 0);
|
while ((inl(0x70008a20) & 0x1000000) == 0);
|
||||||
|
|
||||||
/* output 2 pixels */
|
/* output 2 pixels */
|
||||||
|
|
|
||||||
Loading…
Add table
Add a link
Reference in a new issue