forked from len0rd/rockbox
Assembler optimised lcd_yuv_blit() for iPod Video (Fs #7951 by Andree Buschmann, adapted and separated into an .S file by me).
git-svn-id: svn://svn.rockbox.org/rockbox/trunk@15257 a1c6a512-1295-4272-9138-f99709370657
This commit is contained in:
Jens Arnold
parent
86a999c478
commit
3c1b9d9d70
3 changed files with 249 additions and 105 deletions
|
|
@ -730,6 +730,7 @@ target/arm/ipod/backlight-nano_video.c
|
||||||
target/arm/ipod/button-clickwheel.c
|
target/arm/ipod/button-clickwheel.c
|
||||||
target/arm/ipod/power-ipod.c
|
target/arm/ipod/power-ipod.c
|
||||||
target/arm/ipod/powermgmt-ipod-pcf.c
|
target/arm/ipod/powermgmt-ipod-pcf.c
|
||||||
|
target/arm/ipod/video/lcd-as-video.S
|
||||||
target/arm/ipod/video/lcd-video.c
|
target/arm/ipod/video/lcd-video.c
|
||||||
target/arm/usb-fw-pp502x.c
|
target/arm/usb-fw-pp502x.c
|
||||||
#endif /* SIMULATOR */
|
#endif /* SIMULATOR */
|
||||||
|
|
|
||||||
226
firmware/target/arm/ipod/video/lcd-as-video.S
Normal file
226
firmware/target/arm/ipod/video/lcd-as-video.S
Normal file
|
|
@ -0,0 +1,226 @@
|
||||||
|
/***************************************************************************
|
||||||
|
* __________ __ ___.
|
||||||
|
* Open \______ \ ____ ____ | | _\_ |__ _______ ___
|
||||||
|
* Source | _// _ \_/ ___\| |/ /| __ \ / _ \ \/ /
|
||||||
|
* Jukebox | | ( <_> ) \___| < | \_\ ( <_> > < <
|
||||||
|
* Firmware |____|_ /\____/ \___ >__|_ \|___ /\____/__/\_ \
|
||||||
|
* \/ \/ \/ \/ \/
|
||||||
|
* $Id$
|
||||||
|
*
|
||||||
|
* Copyright (C) 2007 by Andree Buschmann
|
||||||
|
*
|
||||||
|
* All files in this archive are subject to the GNU General Public License.
|
||||||
|
* See the file COPYING in the source tree root for full license agreement.
|
||||||
|
*
|
||||||
|
* This software is distributed on an "AS IS" basis, WITHOUT WARRANTY OF ANY
|
||||||
|
* KIND, either express or implied.
|
||||||
|
*
|
||||||
|
****************************************************************************/
|
||||||
|
|
||||||
|
/****************************************************************************
|
||||||
|
* void lcd_write_yuv_420_lines(unsigned char const * const src[3],
|
||||||
|
* int width,
|
||||||
|
* int stride);
|
||||||
|
*
|
||||||
|
* |R| |1.000000 -0.000001 1.402000| |Y'|
|
||||||
|
* |G| = |1.000000 -0.334136 -0.714136| |Pb|
|
||||||
|
* |B| |1.000000 1.772000 0.000000| |Pr|
|
||||||
|
* Scaled, normalized, rounded and tweaked to yield RGB 565:
|
||||||
|
* |R| |74 0 101| |Y' - 16| >> 9
|
||||||
|
* |G| = |74 -24 -51| |Cb - 128| >> 8
|
||||||
|
* |B| |74 128 0| |Cr - 128| >> 9
|
||||||
|
*/
|
||||||
|
.section .icode, "ax", %progbits
|
||||||
|
.align 2
|
||||||
|
.global lcd_write_yuv420_lines
|
||||||
|
.type lcd_write_yuv420_lines, %function
|
||||||
|
lcd_write_yuv420_lines:
|
||||||
|
/* r0 = yuv_src */
|
||||||
|
/* r1 = width */
|
||||||
|
/* r2 = stride */
|
||||||
|
stmfd sp!, { r4-r12 } /* save non-scratch */
|
||||||
|
ldmia r0, { r10, r11, r12 } /* r10 = yuv_src[0] = Y'_p */
|
||||||
|
/* r11 = yuv_src[1] = Cb_p */
|
||||||
|
/* r12 = yuv_src[2] = Cr_p */
|
||||||
|
add r2, r10, r2 /* r2 = &ysrc[stride] */
|
||||||
|
add r3, r1, r1, asl #1 /* number of bytes for chroma buffer */
|
||||||
|
add r3, r3, #15 /* plus room for 3 additional words, */
|
||||||
|
bic r3, r3, #3 /* rounded up to multiples of 4 byte */
|
||||||
|
sub sp, sp, r3 /* and allocate on stack */
|
||||||
|
stmia sp, {r1, r2, r3} /* width, &ysrc[stride], stack_alloc */
|
||||||
|
|
||||||
|
mov r7, r1 /* r7 = loop count */
|
||||||
|
add r8, sp, #12 /* chroma buffer */
|
||||||
|
mov r9, #0x30000000 /* LCD data port */
|
||||||
|
|
||||||
|
/* 1st loop start */
|
||||||
|
10: /* loop start */
|
||||||
|
|
||||||
|
ldrb r0, [r11], #1 /* r0 = *usrc++ = *Cb_p++ */
|
||||||
|
ldrb r1, [r12], #1 /* r1 = *vsrc++ = *Cr_p++ */
|
||||||
|
|
||||||
|
sub r0, r0, #128 /* r0 = Cb-128 */
|
||||||
|
sub r1, r1, #128 /* r1 = Cr-128 */
|
||||||
|
|
||||||
|
add r3, r1, r1, asl #1 /* r3 = Cr*51 + Cb*24 */
|
||||||
|
add r3, r3, r3, asl #4
|
||||||
|
add r3, r3, r0, asl #3
|
||||||
|
add r3, r3, r0, asl #4
|
||||||
|
|
||||||
|
add r4, r1, r1, asl #2 /* r1 = Cr*101 */
|
||||||
|
add r4, r4, r1, asl #5
|
||||||
|
add r1, r4, r1, asl #6
|
||||||
|
|
||||||
|
add r1, r1, #256 /* r1 = rv = (r1 + 256) >> 9 */
|
||||||
|
mov r1, r1, asr #9
|
||||||
|
strb r1, [r8], #1 /* store r1 to chroma_buf */
|
||||||
|
rsb r3, r3, #128 /* r3 = guv = (-r3 + 128) >> 8 */
|
||||||
|
mov r3, r3, asr #8
|
||||||
|
strb r3, [r8], #1 /* store r3 to chroma_buf */
|
||||||
|
add r0, r0, #2 /* r0 = bu = (Cb*128 + 256) >> 9 */
|
||||||
|
mov r0, r0, asr #2
|
||||||
|
strb r0, [r8], #1 /* store r0 to chroma_buf */
|
||||||
|
|
||||||
|
/* 1st loop, first pixel */
|
||||||
|
ldrb r5, [r10], #1 /* r5 = *ysrc++ = *Y'_p++ */
|
||||||
|
sub r5, r5, #16 /* r5 = (Y'-16) * 74 */
|
||||||
|
add r2, r5, r5, asl #2
|
||||||
|
add r5, r2, r5, asl #5
|
||||||
|
|
||||||
|
add r6, r1, r5, asr #8 /* r6 = r = (Y >> 9) + rv */
|
||||||
|
add r2, r3, r5, asr #7 /* r2 = g = (Y >> 8) + guv */
|
||||||
|
add r4, r0, r5, asr #8 /* r4 = b = (Y >> 9) + bu */
|
||||||
|
|
||||||
|
orr r5, r6, r4 /* check if clamping is needed... */
|
||||||
|
orr r5, r5, r2, asr #1 /* ...at all */
|
||||||
|
cmp r5, #31
|
||||||
|
bls 15f /* -> no clamp */
|
||||||
|
cmp r6, #31 /* clamp r */
|
||||||
|
mvnhi r6, r6, asr #31
|
||||||
|
andhi r6, r6, #31
|
||||||
|
cmp r2, #63 /* clamp g */
|
||||||
|
mvnhi r2, r2, asr #31
|
||||||
|
andhi r2, r2, #63
|
||||||
|
cmp r4, #31 /* clamp b */
|
||||||
|
mvnhi r4, r4, asr #31
|
||||||
|
andhi r4, r4, #31
|
||||||
|
15: /* no clamp */
|
||||||
|
|
||||||
|
orr r4, r4, r2, lsl #5 /* pixel = r<<11 | g<<5 | b */
|
||||||
|
orr r4, r4, r6, lsl #11
|
||||||
|
strh r4, [r9] /* write pixel */
|
||||||
|
|
||||||
|
/* 1st loop, second pixel */
|
||||||
|
ldrb r5, [r10], #1 /* r5 = *ysrc++ = *Y'_p++ */
|
||||||
|
sub r5, r5, #16 /* r5 = (Y'-16) * 74 */
|
||||||
|
add r2, r5, r5, asl #2
|
||||||
|
add r5, r2, r5, asl #5
|
||||||
|
|
||||||
|
add r6, r1, r5, asr #8 /* r6 = r = (Y >> 9) + rv */
|
||||||
|
add r2, r3, r5, asr #7 /* r2 = g = (Y >> 8) + guv */
|
||||||
|
add r4, r0, r5, asr #8 /* r4 = b = (Y >> 9) + bu */
|
||||||
|
|
||||||
|
orr r5, r6, r4 /* check if clamping is needed... */
|
||||||
|
orr r5, r5, r2, asr #1 /* ...at all */
|
||||||
|
cmp r5, #31
|
||||||
|
bls 15f /* -> no clamp */
|
||||||
|
cmp r6, #31 /* clamp r */
|
||||||
|
mvnhi r6, r6, asr #31
|
||||||
|
andhi r6, r6, #31
|
||||||
|
cmp r2, #63 /* clamp g */
|
||||||
|
mvnhi r2, r2, asr #31
|
||||||
|
andhi r2, r2, #63
|
||||||
|
cmp r4, #31 /* clamp b */
|
||||||
|
mvnhi r4, r4, asr #31
|
||||||
|
andhi r4, r4, #31
|
||||||
|
15: /* no clamp */
|
||||||
|
|
||||||
|
orr r4, r4, r2, lsl #5 /* pixel = r<<11 | g<<5 | b */
|
||||||
|
orr r4, r4, r6, lsl #11
|
||||||
|
strh r4, [r9] /* write pixel */
|
||||||
|
|
||||||
|
subs r7, r7, #2 /* check for loop end */
|
||||||
|
bgt 10b /* back to beginning */
|
||||||
|
/* 1st loop end */
|
||||||
|
|
||||||
|
add r8, sp, #12 /* chroma buffer */
|
||||||
|
ldmia sp, { r7, r10 } /* r7 = loop count */
|
||||||
|
/* r10 = &ysrc[stride] */
|
||||||
|
|
||||||
|
/* 2nd loop start */
|
||||||
|
20: /* loop start */
|
||||||
|
|
||||||
|
/* restore r1, r3 and r0 from chroma buffer */
|
||||||
|
ldrsb r1, [r8], #1
|
||||||
|
ldrsb r3, [r8], #1
|
||||||
|
ldrsb r0, [r8], #1
|
||||||
|
|
||||||
|
/* 2nd loop, first pixel */
|
||||||
|
ldrb r5, [r10], #1 /* r5 = *ysrc++ = *Y'_p++ */
|
||||||
|
sub r5, r5, #16 /* r5 = (Y'-16) * 74 */
|
||||||
|
add r2, r5, r5, asl #2
|
||||||
|
add r5, r2, r5, asl #5
|
||||||
|
|
||||||
|
add r6, r1, r5, asr #8 /* r6 = r = (Y >> 9) + rv */
|
||||||
|
add r2, r3, r5, asr #7 /* r2 = g = (Y >> 8) + guv */
|
||||||
|
add r4, r0, r5, asr #8 /* r4 = b = (Y >> 9) + bu */
|
||||||
|
|
||||||
|
orr r5, r6, r4 /* check if clamping is needed... */
|
||||||
|
orr r5, r5, r2, asr #1 /* ...at all */
|
||||||
|
cmp r5, #31
|
||||||
|
bls 15f /* -> no clamp */
|
||||||
|
cmp r6, #31 /* clamp r */
|
||||||
|
mvnhi r6, r6, asr #31
|
||||||
|
andhi r6, r6, #31
|
||||||
|
cmp r2, #63 /* clamp g */
|
||||||
|
mvnhi r2, r2, asr #31
|
||||||
|
andhi r2, r2, #63
|
||||||
|
cmp r4, #31 /* clamp b */
|
||||||
|
mvnhi r4, r4, asr #31
|
||||||
|
andhi r4, r4, #31
|
||||||
|
15: /* no clamp */
|
||||||
|
|
||||||
|
orr r4, r4, r2, lsl #5 /* pixel = r<<11 | g<<5 | b */
|
||||||
|
orr r4, r4, r6, lsl #11
|
||||||
|
strh r4, [r9] /* write pixel */
|
||||||
|
|
||||||
|
/* 2nd loop, second pixel */
|
||||||
|
ldrb r5, [r10], #1 /* r5 = *ysrc++ = *Y'_p++ */
|
||||||
|
sub r5, r5, #16 /* r5 = (Y'-16) * 74 */
|
||||||
|
add r2, r5, r5, asl #2
|
||||||
|
add r5, r2, r5, asl #5
|
||||||
|
|
||||||
|
add r6, r1, r5, asr #8 /* r6 = r = (Y >> 9) + rv */
|
||||||
|
add r2, r3, r5, asr #7 /* r2 = g = (Y >> 8) + guv */
|
||||||
|
add r4, r0, r5, asr #8 /* r4 = b = (Y >> 9) + bu */
|
||||||
|
|
||||||
|
orr r5, r6, r4 /* check if clamping is needed... */
|
||||||
|
orr r5, r5, r2, asr #1 /* ...at all */
|
||||||
|
cmp r5, #31
|
||||||
|
bls 15f /* -> no clamp */
|
||||||
|
cmp r6, #31 /* clamp r */
|
||||||
|
mvnhi r6, r6, asr #31
|
||||||
|
andhi r6, r6, #31
|
||||||
|
cmp r2, #63 /* clamp g */
|
||||||
|
mvnhi r2, r2, asr #31
|
||||||
|
andhi r2, r2, #63
|
||||||
|
cmp r4, #31 /* clamp b */
|
||||||
|
mvnhi r4, r4, asr #31
|
||||||
|
andhi r4, r4, #31
|
||||||
|
15: /* no clamp */
|
||||||
|
|
||||||
|
orr r4, r4, r2, lsl #5 /* pixel = r<<11 | g<<5 | b */
|
||||||
|
orr r4, r4, r6, lsl #11
|
||||||
|
strh r4, [r9] /* write pixel */
|
||||||
|
|
||||||
|
subs r7, r7, #2 /* check for loop end */
|
||||||
|
bgt 20b /* back to beginning */
|
||||||
|
/* 2nd loop end */
|
||||||
|
|
||||||
|
ldr r3, [sp, #8]
|
||||||
|
add sp, sp, r3 /* deallocate buffer */
|
||||||
|
ldmfd sp!, { r4-r12 } /* restore registers */
|
||||||
|
bx lr
|
||||||
|
|
||||||
|
.ltorg
|
||||||
|
.size lcd_write_yuv420_lines, .-lcd_write_yuv420_lines
|
||||||
|
|
@ -207,43 +207,26 @@ void lcd_update(void)
|
||||||
lcd_update_rect(0, 0, LCD_WIDTH, LCD_HEIGHT);
|
lcd_update_rect(0, 0, LCD_WIDTH, LCD_HEIGHT);
|
||||||
}
|
}
|
||||||
|
|
||||||
/* YUV- > RGB565 conversion
|
/* Line write helper function for lcd_yuv_blit. Write two lines of yuv420. */
|
||||||
* |R| |1.000000 -0.000001 1.402000| |Y'|
|
extern void lcd_write_yuv420_lines(unsigned char const * const src[3],
|
||||||
* |G| = |1.000000 -0.334136 -0.714136| |Pb|
|
int width,
|
||||||
* |B| |1.000000 1.772000 0.000000| |Pr|
|
int stride);
|
||||||
* Scaled, normalized, rounded and tweaked to yield RGB 565:
|
|
||||||
* |R| |74 0 101| |Y' - 16| >> 9
|
|
||||||
* |G| = |74 -24 -51| |Cb - 128| >> 8
|
|
||||||
* |B| |74 128 0| |Cr - 128| >> 9
|
|
||||||
*/
|
|
||||||
|
|
||||||
#define RGBYFAC 74 /* 1.0 */
|
|
||||||
#define RVFAC 101 /* 1.402 */
|
|
||||||
#define GVFAC (-51) /* -0.714136 */
|
|
||||||
#define GUFAC (-24) /* -0.334136 */
|
|
||||||
#define BUFAC 128 /* 1.772 */
|
|
||||||
|
|
||||||
/* ROUNDOFFS contain constant for correct round-offs as well as
|
|
||||||
constant parts of the conversion matrix (e.g. (Y'-16)*RGBYFAC
|
|
||||||
-> constant part = -16*RGBYFAC). Through extraction of these
|
|
||||||
constant parts we save at leat 4 substractions in the conversion
|
|
||||||
loop */
|
|
||||||
#define ROUNDOFFSR (256 - 16*RGBYFAC - 128*RVFAC)
|
|
||||||
#define ROUNDOFFSG (128 - 16*RGBYFAC - 128*GVFAC - 128*GUFAC)
|
|
||||||
#define ROUNDOFFSB (256 - 16*RGBYFAC - 128*BUFAC)
|
|
||||||
|
|
||||||
#define MAX_5BIT 0x1f
|
|
||||||
#define MAX_6BIT 0x3f
|
|
||||||
|
|
||||||
/* 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],
|
|
||||||
int src_x, int src_y, int stride,
|
|
||||||
int x, int y, int width, int height) ICODE_ATTR;
|
|
||||||
void lcd_yuv_blit(unsigned char * const src[3],
|
void lcd_yuv_blit(unsigned char * const src[3],
|
||||||
int src_x, int src_y, int stride,
|
int src_x, int src_y, int stride,
|
||||||
int x, int y, int width, int height)
|
int x, int y, int width, int height)
|
||||||
{
|
{
|
||||||
width = (width + 1) & ~1;
|
unsigned char const * yuv_src[3];
|
||||||
|
off_t z;
|
||||||
|
|
||||||
|
/* Sorry, but width and height must be >= 2 or else */
|
||||||
|
width &= ~1;
|
||||||
|
|
||||||
|
z = stride * src_y;
|
||||||
|
yuv_src[0] = src[0] + z + src_x;
|
||||||
|
yuv_src[1] = src[1] + (z >> 2) + (src_x >> 1);
|
||||||
|
yuv_src[2] = src[2] + (yuv_src[1] - src[1]);
|
||||||
|
|
||||||
if (finishup_needed)
|
if (finishup_needed)
|
||||||
{
|
{
|
||||||
|
|
@ -279,85 +262,19 @@ void lcd_yuv_blit(unsigned char * const src[3],
|
||||||
/* wait for it to be write ready */
|
/* wait for it to be write ready */
|
||||||
while ((inw(0x30030000) & 0x2) == 0);
|
while ((inw(0x30030000) & 0x2) == 0);
|
||||||
|
|
||||||
const int ymax = y + height - 1;
|
height >>= 1;
|
||||||
const int stride_div_sub_x = stride >> 1;
|
do
|
||||||
unsigned char *ysrc = 0;
|
|
||||||
unsigned char *usrc = 0;
|
|
||||||
unsigned char *vsrc = 0;
|
|
||||||
unsigned char *row_end = 0;
|
|
||||||
int uvoffset;
|
|
||||||
int yp, up, vp, rc, gc, bc; /* temporary variables */
|
|
||||||
int red1, green1, blue1; /* contain RGB of 1st pixel */
|
|
||||||
int red2, green2, blue2; /* contain RGB of 2nd pixel */
|
|
||||||
|
|
||||||
for (; y <= ymax ; y++)
|
|
||||||
{
|
{
|
||||||
/* upsampling, YUV->RGB conversion and reduction to RGB565 in one go */
|
lcd_write_yuv420_lines(yuv_src, width, stride);
|
||||||
uvoffset = stride_div_sub_x*(src_y >> 1) + (src_x >> 1);
|
|
||||||
ysrc = src[0] + stride * src_y + src_x;
|
|
||||||
usrc = src[1] + uvoffset;
|
|
||||||
vsrc = src[2] + uvoffset;
|
|
||||||
|
|
||||||
row_end = ysrc + width;
|
|
||||||
|
|
||||||
do
|
yuv_src[0] += stride << 1; /* Skip down two luma lines */
|
||||||
{
|
yuv_src[1] += stride >> 1; /* Skip down one chroma line */
|
||||||
up = *usrc++;
|
yuv_src[2] += stride >> 1;
|
||||||
vp = *vsrc++;
|
|
||||||
rc = RVFAC * vp + ROUNDOFFSR;
|
|
||||||
gc = GVFAC * vp + GUFAC * up + ROUNDOFFSG;
|
|
||||||
bc = BUFAC * up + ROUNDOFFSB;
|
|
||||||
|
|
||||||
/* Pixel 1 -> RGB565 */
|
|
||||||
yp = *ysrc++ * RGBYFAC;
|
|
||||||
red1 = (yp + rc) >> 9;
|
|
||||||
green1 = (yp + gc) >> 8;
|
|
||||||
blue1 = (yp + bc) >> 9;
|
|
||||||
|
|
||||||
/* Pixel 2 -> RGB565 */
|
|
||||||
yp = *ysrc++ * RGBYFAC;
|
|
||||||
red2 = (yp + rc) >> 9;
|
|
||||||
green2 = (yp + gc) >> 8;
|
|
||||||
blue2 = (yp + bc) >> 9;
|
|
||||||
|
|
||||||
/* 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 being !=0 when bitmasking them.
|
|
||||||
We first check for red and blue components (5bit range). */
|
|
||||||
if ((red1 | blue1 | red2 | blue2) & ~MAX_5BIT)
|
|
||||||
{
|
|
||||||
if (red1 & ~MAX_5BIT)
|
|
||||||
red1 = (red1 >> 31) ? 0 : MAX_5BIT;
|
|
||||||
if (blue1 & ~MAX_5BIT)
|
|
||||||
blue1 = (blue1 >> 31) ? 0 : MAX_5BIT;
|
|
||||||
if (red2 & ~MAX_5BIT)
|
|
||||||
red2 = (red2 >> 31) ? 0 : MAX_5BIT;
|
|
||||||
if (blue2 & ~MAX_5BIT)
|
|
||||||
blue2 = (blue2 >> 31) ? 0 : MAX_5BIT;
|
|
||||||
}
|
|
||||||
/* We second check for green component (6bit range) */
|
|
||||||
if ((green1 | green2) & ~MAX_6BIT)
|
|
||||||
{
|
|
||||||
if (green1 & ~MAX_6BIT)
|
|
||||||
green1 = (green1 >> 31) ? 0 : MAX_6BIT;
|
|
||||||
if (green2 & ~MAX_6BIT)
|
|
||||||
green2 = (green2 >> 31) ? 0 : MAX_6BIT;
|
|
||||||
}
|
|
||||||
|
|
||||||
/* pixel1 */
|
|
||||||
outw((red1 << 11) | (green1 << 5) | blue1, 0x30000000);
|
|
||||||
|
|
||||||
/* pixel2 */
|
|
||||||
outw((red2 << 11) | (green2 << 5) | blue2, 0x30000000);
|
|
||||||
}
|
|
||||||
while (ysrc < row_end);
|
|
||||||
|
|
||||||
src_y++;
|
|
||||||
}
|
}
|
||||||
|
while (--height > 0);
|
||||||
|
|
||||||
/* Top-half of original lcd_bcm_finishup() function */
|
/* Top-half of original lcd_bcm_finishup() function */
|
||||||
outw(0x31, 0x30030000);
|
outw(0x31, 0x30030000);
|
||||||
|
|
||||||
lcd_bcm_read32(0x1FC);
|
lcd_bcm_read32(0x1FC);
|
||||||
|
|
||||||
|
|
|
||||||
Loading…
Add table
Add a link
Reference in a new issue