HDD6330: implement lcd_yuv_blit() function. The inner loop is written in assembler and the entire function is about 20% faster than the original from the ipod color.

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@28737 a1c6a512-1295-4272-9138-f99709370657

3 files changed, 241 insertions, 10 deletions

diff --git a/firmware/SOURCES b/firmware/SOURCES
index 349155f8ce..7f21a89771 100644
--- a/firmware/SOURCES
+++ b/firmware/SOURCES
@@ -735,6 +735,7 @@ target/arm/philips/fmradio_i2c-hdd.c
 target/arm/philips/hdd6330/backlight-hdd6330.c
 target/arm/philips/hdd6330/button-hdd6330.c
 target/arm/philips/hdd6330/lcd-hdd6330.c
+target/arm/philips/hdd6330/lcd-as-hdd6330.S
 target/arm/philips/hdd6330/powermgmt-hdd6330.c
 target/arm/usb-fw-pp502x.c
 #endif /* SIMULATOR */
diff --git a/firmware/target/arm/philips/hdd6330/lcd-as-hdd6330.S b/firmware/target/arm/philips/hdd6330/lcd-as-hdd6330.S
new file mode 100644
index 0000000000..364eb5b08f
--- /dev/null
+++ b/firmware/target/arm/philips/hdd6330/lcd-as-hdd6330.S
@@ -0,0 +1,149 @@
+/***************************************************************************
+ *             __________               __   ___.
+ *   Open      \______   \ ____   ____ |  | _\_ |__   _______  ___
+ *   Source     |       _//  _ \_/ ___\|  |/ /| __ \ /  _ \  \/  /
+ *   Jukebox    |    |   (  <_> )  \___|    < | \_\ (  <_> > <  <
+ *   Firmware   |____|_  /\____/ \___  >__|_ \|___  /\____/__/\_ \
+ *                     \/            \/     \/    \/            \/
+ * $Id:$
+ *
+ * Copyright (C) 2010 by Szymon Dziok
+ *
+ * Philips Gogear HDD6330 LCD assembly routine
+ *
+ * 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"
+#include "cpu.h"
+
+/****************************************************************************
+   void lcd_yuv_write_inner_loop(unsigned char const * const ysrc,
+                                 unsigned char const * const usrc,
+                                 unsigned char const * const vsrc,
+                                 int width);
+*/
+    .section    .icode, "ax", %progbits
+    .align      2
+    .global     lcd_yuv_write_inner_loop
+    .type       lcd_yuv_write_inner_loop, %function
+lcd_yuv_write_inner_loop:
+                                        @ r0 = ysrc
+                                        @ r1 = usrc
+                                        @ r2 = vsrc
+                                        @ r3 = width
+    stmfd sp!, { r4-r11, lr }           @ save regs
+    mov r4, #0x70000000                 @ r4 = LCD2_BLOCK_CTRL - 0x20
+    add r4, r4, #0x8a00                 @
+    add r5, r4, #0x100                  @ r5 = LCD2_BLOCK_DATA
+10:                                     @ loop
+
+    ldrb r7, [r1], #1                   @ *usrc++
+    ldrb r8, [r2], #1                   @ *vsrc++
+
+    add r10, r8, r8, asl #2             @ 101* vsrc
+    add r10, r10, r8, asl #5
+    add r10, r10, r8, asl #6
+    sub r10, r10, #0x3600
+    sub r10, r10, #0x0020               @ -13856 (ROUNDOFFSR)
+
+    add r11, r8, r8, asl #1             @ 51*vsrc + 24*usrc
+    add r11, r11, r11, asl #4
+    add r11, r11, r7, asl #3
+    add r11, r11, r7, asl #4
+    mov r12, #0x2100
+    add r12, r12, #0x60                 @ +8544 (ROUNDOFFSG) - r11
+    rsb r11, r11, r12
+
+    mov r12, r7, asl #7                 @ 128 * usrc
+    sub r12, r12, #0x4300
+    sub r12, r12, #0x00a0               @ -17312 (ROUNDOFFSB)
+
+@ pixel_1
+    ldrb r8, [r0], #1                   @ *ysrc++
+    mov r7, r8, asl #1
+    add r7, r7, r8, asl #3
+    add r7, r7, r8, asl #6              @ ysrc * 74
+
+    add r9, r10, r7
+    mov r9, r9, asr #9                  @ R
+
+    add r8, r11, r7
+    mov r8, r8, asr #8                  @ G
+
+    add r7, r12, r7
+    mov r7, r7, asr #9                  @ B
+
+    cmp r9, #31                         @ clamp R
+    mvnhi r9, r9, asr #31
+    andhi r9, r9, #31
+
+    cmp r8, #63                         @ clamp G
+    mvnhi r8, r8, asr #31
+    andhi r8, r8, #63
+
+    cmp r7, #31                         @ clamp B
+    mvnhi r7, r7, asr #31
+    andhi r7, r7, #31
+
+    orr r6, r7, r8, lsl #5              @ pack pixel
+    orr r6, r6, r9, lsl #11
+
+    mov r7, r6, lsl #8                  @ swap bytes
+    and r7, r7, #0xff00
+    add r6, r7, r6, lsr #8
+
+@ pixel_2
+    ldrb r8, [r0], #1                   @ *ysrc++
+    mov r7, r8, asl #1
+    add r7, r7, r8, asl #3
+    add r7, r7, r8, asl #6              @ ysrc * 74
+
+    add r9, r10, r7
+    mov r9, r9, asr #9                  @ R
+
+    add r8, r11, r7
+    mov r8, r8, asr #8                  @ G
+
+    add r7, r12, r7
+    mov r7, r7, asr #9                  @ B
+
+    cmp r9, #31                         @ clamp R
+    mvnhi r9, r9, asr #31
+    andhi r9, r9, #31
+
+    cmp r8, #63                         @ clamp G
+    mvnhi r8, r8, asr #31
+    andhi r8, r8, #63
+
+    cmp r7, #31                         @ clamp B
+    mvnhi r7, r7, asr #31
+    andhi r7, r7, #31
+
+    orr r7, r7, r8, lsl #5              @ pack pixel
+    orr r7, r7, r9, lsl #11
+
+    orr r6, r6, r7, lsl #24             @ swap bytes and add pixels simultaneously
+    mov r7, r7, lsr #8
+    orr r6, r6, r7, lsl #16
+
+11:                                     @ while (!(LCD2_BLOCK_CTRL & LCD2_BLOCK_TXOK));
+    ldr r11, [r4, #0x20]                @
+    tst r11, #0x1000000                 @
+    beq 11b                             @
+
+    str r6, [r5]                        @ send two pixels
+
+    subs r3, r3, #2                     @ decrease width
+    bgt 10b                             @ loop
+
+    ldmpc regs=r4-r11                   @ restore regs
+    .ltorg                              @ dump constant pool
+    .size   lcd_yuv_write_inner_loop, .-lcd_yuv_write_inner_loop
diff --git a/firmware/target/arm/philips/hdd6330/lcd-hdd6330.c b/firmware/target/arm/philips/hdd6330/lcd-hdd6330.c
index c8127c4450..53c0be1d6e 100644
--- a/firmware/target/arm/philips/hdd6330/lcd-hdd6330.c
+++ b/firmware/target/arm/philips/hdd6330/lcd-hdd6330.c
@@ -80,21 +80,102 @@ void lcd_yuv_set_options(unsigned options)
     lcd_yuv_options = options;
 }
 
+#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
+*/
+
+extern void lcd_yuv_write_inner_loop(unsigned char const * const ysrc,
+                                     unsigned char const * const usrc,
+                                     unsigned char const * const vsrc,
+                                     int width);
+
 /* Performance function to 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)
 {
-    (void)src;
-    (void)src_x;
-    (void)src_y;
-    (void)stride;
-    (void)x;
-    (void)y;
-    (void)width;
-    (void)height;
+    int h;
+
+    width = (width + 1) & ~1;
+
+    lcd_send_reg(0x01);
+    lcd_send_data(0x48);
+
+    lcd_send_reg(0x05);
+    lcd_send_data(0x0f);
+
+    lcd_send_reg(0x08);
+    lcd_send_data(y);
+
+    lcd_send_reg(0x09);
+    lcd_send_data(y + height - 1);
+
+    lcd_send_reg(0x0a);
+    lcd_send_data(x + 16);
+
+    lcd_send_reg(0x0b);
+    lcd_send_data(x + width - 1 + 16);
+
+    lcd_send_reg(0x06);
+
+    const int stride_div_csub_x = stride/CSUB_X;
+
+    h=0;
+    while (1)
+    {
+        /* 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);
+
+        const unsigned char *usrc = src[1] + uvoffset;
+        const unsigned char *vsrc = src[2] + uvoffset;
+
+        int pixels_to_write;
+
+        if (h==0)
+        {
+            while (!(LCD2_BLOCK_CTRL & LCD2_BLOCK_READY));
+            LCD2_BLOCK_CONFIG = 0;
+
+            if (height == 0) break;
+
+            pixels_to_write = (width * height) * 2;
+            h = height;
+
+            /* calculate how much we can do in one go */
+            if (pixels_to_write > 0x10000)
+            {
+                h = (0x10000/2) / width;
+                pixels_to_write = (width * h) * 2;
+            }
+
+            height -= h;
+            LCD2_BLOCK_CTRL = 0x10000080;
+            LCD2_BLOCK_CONFIG = 0xc0010000 | (pixels_to_write - 1);
+            LCD2_BLOCK_CTRL = 0x34000000;
+        }
+
+        lcd_yuv_write_inner_loop(ysrc,usrc,vsrc,width);
+
+        src_y++;
+        h--;
+    }
+
+    while (!(LCD2_BLOCK_CTRL & LCD2_BLOCK_READY));
+    LCD2_BLOCK_CONFIG = 0;
 }
- 
+
 /* Update the display.
    This must be called after all other LCD functions that change the display. */
 void lcd_update(void)
@@ -146,7 +227,7 @@ void lcd_update_rect(int x, int y, int width, int height)
 
     addr = (unsigned long*)&lcd_framebuffer[y][x];
 
-    while (height > 0) 
+    while (height > 0)
     {
         int c, r;
         int h, pixels_to_write;