FS#11807 - Major speedup of iPod nano 2G. Part 4: Introduce asm for yuv blitting. Overall speedup of part1-4 is +50% for RGB and +93% for YUV.

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@28813 a1c6a512-1295-4272-9138-f99709370657

3 files changed, 276 insertions, 110 deletions

diff --git a/firmware/SOURCES b/firmware/SOURCES
index 7f21a89..72b0217 100644
--- a/firmware/SOURCES
+++ b/firmware/SOURCES
@@ -1504,6 +1504,7 @@ target/arm/s5l8700/kernel-s5l8700.c
 target/arm/s5l8700/dma-s5l8700.c
 target/arm/s5l8700/ipodnano2g/backlight-nano2g.c
 target/arm/s5l8700/ipodnano2g/lcd-nano2g.c
+target/arm/s5l8700/ipodnano2g/lcd-asm-nano2g.S
 target/arm/s5l8700/ipodnano2g/powermgmt-nano2g.c
 target/arm/s5l8700/ipodnano2g/power-nano2g.c
 target/arm/s5l8700/ipodnano2g/ftl-nano2g.c
diff --git a/firmware/target/arm/s5l8700/ipodnano2g/lcd-asm-nano2g.S b/firmware/target/arm/s5l8700/ipodnano2g/lcd-asm-nano2g.S
new file mode 100755
index 0000000..25cf662
--- /dev/null
+++ b/firmware/target/arm/s5l8700/ipodnano2g/lcd-asm-nano2g.S
@@ -0,0 +1,254 @@
+/***************************************************************************
+ *             __________               __   ___.
+ *   Open      \______   \ ____   ____ |  | _\_ |__   _______  ___
+ *   Source     |       _//  _ \_/ ___\|  |/ /| __ \ /  _ \  \/  /
+ *   Jukebox    |    |   (  <_> )  \___|    < | \_\ (  <_> > <  <
+ *   Firmware   |____|_  /\____/ \___  >__|_ \|___  /\____/__/\_ \
+ *                     \/            \/     \/    \/            \/
+ * $Id: lcd-as-video.S 26756 2010-06-11 04:41:36Z funman $
+ *
+ * Copyright (C) 2010 by Andree Buschmann
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; either version 2
+ * of the License, or (at your option) any later version.
+ *
+ * This software is distributed on an "AS IS" basis, WITHOUT WARRANTY OF ANY
+ * KIND, either express or implied.
+ *
+ ****************************************************************************/
+
+#include "config.h"
+
+    .section .icode, "ax", %progbits
+
+/****************************************************************************
+ * extern void lcd_write_yuv420_lines(unsigned char const * const src[3],
+ *                                    unsigned LCD_BASE,
+ *                                    int width,
+ *                                    int stride);
+ *
+ *   Conversion from Motion JPEG and MPEG Y'PbPr to RGB is:
+ *   |R|   |1.164  0.000  1.596| |Y' -  16|
+ *   |G| = |1.164 -0.391 -0.813| |Pb - 128|
+ *   |B|   |1.164  2.018  0.000| |Pr - 128|
+ *
+ *   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
+ *
+ * Converts two lines from YUV to RGB565 and writes to LCD at once. First loop
+ * loads Cb/Cr, calculates the chroma offset and saves them to buffer. Within
+ * the second loop these chroma offset are reloaded from buffer. Within each 
+ * loop two pixels are calculated and written to LCD. 
+ */
+    .align      2
+    .global     lcd_write_yuv420_lines
+    .type       lcd_write_yuv420_lines, %function
+lcd_write_yuv420_lines:
+                                      /* r0 = src = yuv_src */
+                                      /* r1 = dst = LCD_BASE */
+                                      /* r2 = width */
+                                      /* r3 = stride */                
+    stmfd       sp!, { r4-r10, lr }   /* save non-scratch */
+    ldmia       r0, { r9, r10, r12 }  /* r9 = yuv_src[0] = Y'_p */
+                                      /* r10 = yuv_src[1] = Cb_p */
+                                      /* r12 = yuv_src[2] = Cr_p */
+    add         r3, r9, r3            /* r3 = &ysrc[stride] */
+    add         r4, r2, r2, asr #1    /* chroma buffer lenght = width/2 *3 */
+    mov         r4, r4, asl #2        /*   use words for str/ldm possibility */
+    add         r4, r4, #19           /*   plus room for 4 additional words, */
+    bic         r4, r4, #3            /*   rounded up to multiples of 4 byte */
+    sub         sp, sp, r4            /*   and allocate on stack */
+    stmia       sp, {r1-r4}           /* LCD_BASE, width, &ysrc[stride], stack_alloc */
+
+    mov         r7, r2                /* r7 = loop count */
+    add         r8, sp, #16           /* chroma buffer */
+    mov         lr, r1                /* LCD data port = LCD_BASE */
+
+    /* 1st loop start */
+10:                                   /* loop start */
+
+    ldrb        r0, [r10], #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         r2, r1, r1, asl #1    /* r2 = Cr*51 + Cb*24 */
+    add         r2, r2, r2, asl #4   
+    add         r2, r2, r0, asl #3   
+    add         r2, r2, 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
+    rsb         r2, r2, #128          /* r2 = guv = (-r2 + 128) >> 8 */
+    mov         r2, r2, asr #8       
+    add         r0, r0, #2            /* r0 = bu = (Cb*128 + 256) >> 9 */
+    mov         r0, r0, asr #2       
+    stmia       r8!, {r0-r2}          /* store r0, r1 and r2 to chroma buffer */
+
+    /* 1st loop, first pixel */
+    ldrb        r5, [r9], #1          /* r5 = *ysrc++ = *Y'_p++ */
+    sub         r5, r5, #16           /* r5 = (Y'-16) * 74 */
+    add         r3, r5, r5, asl #2
+    add         r5, r3, r5, asl #5
+
+    add         r6, r1, r5, asr #8    /* r6 = r = (Y >> 9) + rv */
+    add         r3, r2, r5, asr #7    /* r3 = 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, r3, 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         r3, #63               /* clamp g */
+    mvnhi       r3, r3, asr #31
+    andhi       r3, r3, #63
+    cmp         r4, #31               /* clamp b */
+    mvnhi       r4, r4, asr #31         
+    andhi       r4, r4, #31          
+15:                                   /* no clamp */
+
+    /* calculate pixel_1 and save to r5 for later pixel packing */
+    orr         r4, r4, r3, lsl #5    /* pixel_1 = r<<11 | g<<5 | b */
+    orr         r5, r4, r6, lsl #11   /* r5 = pixel_1 */
+
+    /* 1st loop, second pixel */
+    ldrb        r4, [r9], #1          /* r4 = *ysrc++ = *Y'_p++ */
+    sub         r4, r4, #16           /* r4 = (Y'-16) * 74 */
+    add         r3, r4, r4, asl #2
+    add         r4, r3, r4, asl #5
+
+    add         r6, r1, r4, asr #8    /* r6 = r = (Y >> 9) + rv */
+    add         r3, r2, r4, asr #7    /* r3 = g = (Y >> 8) + guv */
+    add         r4, r0, r4, asr #8    /* r4 = b = (Y >> 9) + bu */   
+
+    orr         r0, r6, r4            /* check if clamping is needed... */
+    orr         r0, r0, r3, asr #1    /* ...at all */
+    cmp         r0, #31                 
+    bls         15f                   /* -> no clamp */
+    cmp         r6, #31               /* clamp r */
+    mvnhi       r6, r6, asr #31         
+    andhi       r6, r6, #31             
+    cmp         r3, #63               /* clamp g */
+    mvnhi       r3, r3, asr #31
+    andhi       r3, r3, #63
+    cmp         r4, #31               /* clamp b */
+    mvnhi       r4, r4, asr #31         
+    andhi       r4, r4, #31          
+15:                                   /* no clamp */
+
+    /* calculate pixel_2 and pack with pixel_1 before writing */
+    orr         r4, r4, r3, lsl #5    /* pixel_2 = r<<11 | g<<5 | b */
+    orr         r4, r4, r6, lsl #11   /* r4 = pixel_2 */
+
+    /* wait for FIFO half full */
+.fifo_wait1:
+    ldr         r3, [lr, #0x1C]       /* while (LCD_STATUS & 0x08); */
+    tst         r3, #0x8
+    bgt         .fifo_wait1
+
+    str         r5, [lr, #0x40]       /* write pixel_1 */
+    str         r4, [lr, #0x40]       /* write pixel_2 */
+
+    subs        r7, r7, #2            /* check for loop end */
+    bgt         10b                   /* back to beginning  */
+    /* 1st loop end */
+
+    /* Reload several registers for pointer rewinding for next loop */
+    add         r8, sp, #16           /* chroma buffer */
+    ldmia       sp, { r1, r7, r9}     /* r1  = LCD_BASE */
+                                      /* r7  = loop count */
+                                      /* r9 = &ysrc[stride] */   
+
+    /* 2nd loop start */
+20:                                   /* loop start */
+    /* restore r0 (bu), r1 (rv) and r2 (guv) from chroma buffer */
+    ldmia       r8!, {r0-r2}
+
+    /* 2nd loop, first pixel */
+    ldrb        r5, [r9], #1          /* r5 = *ysrc++ = *Y'_p++ */
+    sub         r5, r5, #16           /* r5 = (Y'-16) * 74 */
+    add         r3, r5, r5, asl #2
+    add         r5, r3, r5, asl #5
+
+    add         r6, r1, r5, asr #8    /* r6 = r = (Y >> 9) + rv */
+    add         r3, r2, r5, asr #7    /* r3 = 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, r3, 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         r3, #63               /* clamp g */
+    mvnhi       r3, r3, asr #31
+    andhi       r3, r3, #63
+    cmp         r4, #31               /* clamp b */
+    mvnhi       r4, r4, asr #31         
+    andhi       r4, r4, #31          
+15:                                   /* no clamp */
+    /* calculate pixel_1 and save to r5 for later pixel packing */
+    orr         r4, r4, r3, lsl #5    /* pixel_1 = r<<11 | g<<5 | b */
+    orr         r5, r4, r6, lsl #11   /* r5 = pixel_1 */
+
+    /* 2nd loop, second pixel */
+    ldrb        r4, [r9], #1          /* r4 = *ysrc++ = *Y'_p++ */
+    sub         r4, r4, #16           /* r4 = (Y'-16) * 74 */
+    add         r3, r4, r4, asl #2
+    add         r4, r3, r4, asl #5
+
+    add         r6, r1, r4, asr #8    /* r6 = r = (Y >> 9) + rv */
+    add         r3, r2, r4, asr #7    /* r3 = g = (Y >> 8) + guv */
+    add         r4, r0, r4, asr #8    /* r4 = b = (Y >> 9) + bu */
+
+    orr         r0, r6, r4            /* check if clamping is needed... */
+    orr         r0, r0, r3, asr #1    /* ...at all */
+    cmp         r0, #31                 
+    bls         15f                   /* -> no clamp */
+    cmp         r6, #31               /* clamp r */
+    mvnhi       r6, r6, asr #31         
+    andhi       r6, r6, #31             
+    cmp         r3, #63               /* clamp g */
+    mvnhi       r3, r3, asr #31
+    andhi       r3, r3, #63
+    cmp         r4, #31               /* clamp b */
+    mvnhi       r4, r4, asr #31         
+    andhi       r4, r4, #31          
+15:                                   /* no clamp */
+
+    /* calculate pixel_2 and pack with pixel_1 before writing */
+    orr         r4, r4, r3, lsl #5    /* pixel_2 = r<<11 | g<<5 | b */
+    orr         r4, r4, r6, lsl #11   /* r4 = pixel_2 */
+
+    /* wait for FIFO half full */
+.fifo_wait2:
+    ldr         r3, [lr, #0x1C]       /* while (LCD_STATUS & 0x08); */
+    tst         r3, #0x8
+    bgt         .fifo_wait2
+
+    str         r5, [lr, #0x40]       /* write pixel_1 */
+    str         r4, [lr, #0x40]       /* write pixel_2 */
+
+    subs        r7, r7, #2            /* check for loop end */
+    bgt         20b                   /* back to beginning  */
+    /* 2nd loop end */
+
+    ldr         r3, [sp, #12]
+    add         sp, sp, r3            /* deallocate buffer */
+    ldmpc       regs=r4-r10           /* restore registers */
+
+    .ltorg
+    .size   lcd_write_yuv420_lines, .-lcd_write_yuv420_lines
diff --git a/firmware/target/arm/s5l8700/ipodnano2g/lcd-nano2g.c b/firmware/target/arm/s5l8700/ipodnano2g/lcd-nano2g.c
index 071b68f..d1c41d6 100644
--- a/firmware/target/arm/s5l8700/ipodnano2g/lcd-nano2g.c
+++ b/firmware/target/arm/s5l8700/ipodnano2g/lcd-nano2g.c
@@ -400,48 +400,21 @@ void lcd_update_rect(int x, int y, int width, int height)
     }
 }
 
-/*** update functions ***/
-
-#define CSUB_X 2
-#define CSUB_Y 2
-
-/*   YUV- > RGB565 conversion
- *   |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
-*/
+/* Line write helper function for lcd_yuv_blit. Writes two lines of yuv420. */
+extern void lcd_write_yuv420_lines(unsigned char const * const src[3],
+                                   unsigned lcd_baseadress,
+                                   int width,
+                                   int stride);
 
-#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 */
+/* Blit a YUV bitmap directly to the LCD */
 void lcd_blit_yuv(unsigned char * const src[3],
                   int src_x, int src_y, int stride,
                   int x, int y, int width, int height)
 {
-    int h;
-    int y0, x0, y1, x1;
-
-    width = (width + 1) & ~1;
+    unsigned int z, y0, x0, y1, x1;;
+    unsigned char const * yuv_src[3];
+    
+    width = (width + 1) & ~1;       /* ensure width is even */
 
     x0 = x;                         /* start horiz */
     y0 = y;                         /* start vert */
@@ -471,79 +444,17 @@ void lcd_blit_yuv(unsigned char * const src[3],
         s5l_lcd_write_cmd(R_MEMORY_WRITE);
     }
 
-    const int stride_div_csub_x = stride/CSUB_X;
-
-    h = height;
-    while (h > 0) {
-        /* upsampling, YUV->RGB conversion and reduction to RGB565 in one go */
-        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);
+    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]);
 
-        const unsigned char *usrc = src[1] + uvoffset;
-        const unsigned char *vsrc = src[2] + uvoffset;
-        const unsigned char *row_end = ysrc + width;
+    height >>= 1;
 
-        int yp, up, vp;
-        int red1, green1, blue1;
-        int red2, green2, blue2;
-
-        int rc, gc, bc;
-
-        do
-        {
-            up = *usrc++;
-            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;
-            }
-
-            /* output 2 pixels */
-            while (LCD_STATUS & 0x08); /* wait while FIFO is half full */
-            lcd_write_pixel((red1 << 11) | (green1 << 5) | blue1);
-            lcd_write_pixel((red2 << 11) | (green2 << 5) | blue2);
-        }
-        while (ysrc < row_end);
-
-        src_y++;
-        h--;
-    }
+    do {
+        lcd_write_yuv420_lines(yuv_src, LCD_BASE, width, stride);
+        yuv_src[0] += stride << 1;
+        yuv_src[1] += stride >> 1; /* Skip down one chroma line */
+        yuv_src[2] += stride >> 1;
+    } while (--height > 0);
 }