Greyscale ipod lcd driver: * Assembler optimised low level functions. PP5002 targets benefit most (lcd_update() speedup >50%, and the greyscale overlay no longer makes mp3 playback skip). * Consistent brace placement style.

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@16060 a1c6a512-1295-4272-9138-f99709370657

3 files changed, 338 insertions, 133 deletions

diff --git a/firmware/SOURCES b/firmware/SOURCES
index ca06d86..0f943a0 100644
--- a/firmware/SOURCES
+++ b/firmware/SOURCES
@@ -694,6 +694,7 @@ target/arm/ata-pp5020.c
 target/arm/ipod/adc-ipod-pcf.c
 target/arm/ipod/backlight-4g_color.c
 target/arm/ipod/button-clickwheel.c
+target/arm/ipod/lcd-as-gray.S
 target/arm/ipod/lcd-gray.c
 target/arm/ipod/power-ipod.c
 target/arm/ipod/powermgmt-ipod-pcf.c
@@ -762,6 +763,7 @@ target/arm/i2s-pp.c
 target/arm/ipod/adc-ipod-pcf.c
 target/arm/ipod/3g/backlight-3g.c
 target/arm/ipod/button-1g-3g.c
+target/arm/ipod/lcd-as-gray.S
 target/arm/ipod/lcd-gray.c
 target/arm/ipod/power-ipod.c
 target/arm/ipod/powermgmt-ipod-pcf.c
@@ -778,6 +780,7 @@ target/arm/ipod/1g2g/adc-ipod-1g2g.c
 target/arm/ipod/1g2g/backlight-1g2g.c
 target/arm/ipod/1g2g/powermgmt-1g2g.c
 target/arm/ipod/button-1g-3g.c
+target/arm/ipod/lcd-as-gray.S
 target/arm/ipod/lcd-gray.c
 target/arm/ipod/power-ipod.c
 target/arm/usb-fw-pp5002.c
@@ -794,6 +797,7 @@ target/arm/i2s-pp.c
 target/arm/ipod/adc-ipod-pcf.c
 target/arm/ipod/backlight-mini1g_mini2g.c
 target/arm/ipod/button-mini1g.c
+target/arm/ipod/lcd-as-gray.S
 target/arm/ipod/lcd-gray.c
 target/arm/ipod/power-ipod.c
 target/arm/ipod/powermgmt-ipod-pcf.c
@@ -811,6 +815,7 @@ target/arm/i2s-pp.c
 target/arm/ipod/adc-ipod-pcf.c
 target/arm/ipod/backlight-mini1g_mini2g.c
 target/arm/ipod/button-clickwheel.c
+target/arm/ipod/lcd-as-gray.S
 target/arm/ipod/lcd-gray.c
 target/arm/ipod/power-ipod.c
 target/arm/ipod/powermgmt-ipod-pcf.c
diff --git a/firmware/target/arm/ipod/lcd-as-gray.S b/firmware/target/arm/ipod/lcd-as-gray.S
new file mode 100644
index 0000000..d16d09b
--- /dev/null
+++ b/firmware/target/arm/ipod/lcd-as-gray.S
@@ -0,0 +1,290 @@
+/***************************************************************************
+ *             __________               __   ___.
+ *   Open      \______   \ ____   ____ |  | _\_ |__   _______  ___
+ *   Source     |       _//  _ \_/ ___\|  |/ /| __ \ /  _ \  \/  /
+ *   Jukebox    |    |   (  <_> )  \___|    < | \_\ (  <_> > <  <
+ *   Firmware   |____|_  /\____/ \___  >__|_ \|___  /\____/__/\_ \
+ *                     \/            \/     \/    \/            \/
+ * $Id$
+ *
+ * Copyright (C) 2008 by Jens Arnold
+ *
+ * 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.
+ *
+ ****************************************************************************/
+
+#include "config.h"
+#include "cpu.h"
+
+#if CONFIG_CPU == PP5002
+    .section    .icode,"ax",%progbits
+#else
+    .text
+#endif
+    .align      2
+
+
+    .global     lcd_write_data
+    .type       lcd_write_data,%function
+
+lcd_write_data:
+    stmfd   sp!, {r4, lr}
+    ldr     r2, =LCD1_BASE
+
+.loop:
+    ldrb    r3, [r0], #1
+
+#ifdef IPOD_MINI2G
+    ldrb    r4, [r0], #1
+    orr     r3, r4, r3, lsl #8
+    orr     r3, r3, #0x760000
+1:
+    ldr     r4, [r2]
+    tst     r4, #0x8000
+    bne     1b
+    str     r3, [r2, #0x08]
+#else
+1:
+    ldr     r4, [r2]
+    tst     r4, #0x8000
+    bne     1b
+    str     r3, [r2, #0x10]
+
+    ldrb    r3, [r0], #1
+1:
+    ldr     r4, [r2]
+    tst     r4, #0x8000
+    bne     1b
+    str     r3, [r2, #0x10]
+#endif
+
+    subs    r1, r1, #1
+    bne     .loop
+
+    ldmfd   sp!, {r4, pc}
+
+.wd_end:
+    .size   lcd_write_data,.wd_end-lcd_write_data
+    
+
+#ifdef IPOD_MINI2G
+
+    .global     lcd_write_data_shifted
+    .type       lcd_write_data_shifted,%function
+    
+lcd_write_data_shifted:
+    stmfd   sp!, {r4-r6, lr}
+    ldr     r2, =LCD1_BASE
+    mov     r6, #0x760000
+    ldrb    r3, [r0], #1
+
+.sloop:
+    ldrb    r4, [r0], #1
+    orr     r3, r4, r3, lsl #8
+    ldrb    r4, [r0], #1
+    orr     r3, r4, r3, lsl #8
+    mov     r5, r3, lsl #12
+    orr     r5, r6, r5, lsr #16
+1:
+    ldr     r4, [r2]
+    tst     r4, #0x8000
+    bne     1b
+    str     r5, [r2, #0x08]
+
+    subs    r1, r1, #1
+    bne     .sloop
+
+    ldmfd   sp!, {r4-r6, pc}
+
+.wds_end:
+    .size   lcd_write_data_shifted,.wds_end-lcd_write_data_shifted
+    
+#elif defined IPOD_MINI
+
+    .global     lcd_write_data_shifted
+    .type       lcd_write_data_shifted,%function
+    
+lcd_write_data_shifted:
+    stmfd   sp!, {r4, r5, lr}
+    ldr     r2, =LCD1_BASE
+    ldrb    r3, [r0], #1
+
+.sloop:
+    ldrb    r4, [r0], #1
+    orr     r3, r4, r3, lsl #8
+    mov     r5, r3, lsr #4
+1:
+    ldr     r4, [r2]
+    tst     r4, #0x8000
+    bne     1b
+    str     r5, [r2, #0x10]
+
+    ldrb    r4, [r0], #1
+    orr     r3, r4, r3, lsl #8
+    mov     r5, r3, lsr #4
+1:
+    ldr     r4, [r2]
+    tst     r4, #0x8000
+    bne     1b
+    str     r5, [r2, #0x10]
+
+    subs    r1, r1, #1
+    bne     .sloop
+
+    ldmfd   sp!, {r4, r5, pc}
+.wds_end:
+    .size   lcd_write_data_shifted,.wds_end-lcd_write_data_shifted
+    
+#endif
+
+    .global     lcd_mono_data
+    .type       lcd_mono_data,%function
+    
+lcd_mono_data:
+    stmfd   sp!, {r4-r6, lr}
+    ldr     r2, =LCD1_BASE
+    adr     r6, .dibits
+
+.mloop:
+    ldrb    r3, [r0], #1
+    mov     r4, r3, lsr #4
+    ldrb    r5, [r6, r4]
+
+#ifdef IPOD_MINI2G
+    and     r4, r3, #0x0f
+    ldrb    r4, [r6, r4]
+    orr     r5, r4, r5, lsl #8
+    orr     r5, r5, #0x760000
+1:
+    ldr     r4, [r2]
+    tst     r4, #0x8000
+    bne     1b
+    str     r5, [r2, #0x08]
+#else
+1:
+    ldr     r4, [r2]
+    tst     r4, #0x8000
+    bne     1b
+    str     r5, [r2, #0x10]
+
+    and     r4, r3, #0x0f
+    ldrb    r5, [r6, r4]
+1:
+    ldr     r4, [r2]
+    tst     r4, #0x8000
+    bne     1b
+    str     r5, [r2, #0x10]
+#endif
+
+    subs    r1, r1, #1
+    bne     .mloop
+
+    ldmfd   sp!, {r4-r6, pc}
+
+.dibits:
+    .byte   0x00, 0x03, 0x0C, 0x0F, 0x30, 0x33, 0x3C, 0x3F
+    .byte   0xC0, 0xC3, 0xCC, 0xCF, 0xF0, 0xF3, 0xFC, 0xFF
+
+.md_end:
+    .size   lcd_mono_data,.md_end-lcd_mono_data
+    
+
+    .global     lcd_grey_data
+    .type       lcd_grey_data,%function
+
+/* A high performance function to write grey phase data to the display,
+ * one or multiple pixels.
+ *
+ * Arguments:
+ *   r0 - pixel value data address
+ *   r1 - pixel phase data address
+ *   r2 - pixel block count
+ *
+ * Register usage:
+ *   r3/r4 - current block of phases
+ *   r5/r6 - current block of values
+ *   r7 - lcd data accumulator
+ *   r8 - phase signs mask
+ *   r9 - lcd bridge address
+ */
+ 
+lcd_grey_data:
+    stmfd   sp!, {r4-r9, lr}
+    mov     r8, #0x80
+    orr     r8, r8, r8, lsl #8
+    orr     r8, r8, r8, lsl #16
+    ldr     r9, =LCD1_BASE
+
+.greyloop:
+    ldmia   r1, {r3-r4}         /* Fetch 8 pixel phases */
+    ldmia   r0!, {r5-r6}        /* Fetch 8 pixel values */
+
+#ifdef IPOD_MINI2G              /* Serial bridge mode */
+    mov     r7, #0x760000
+    tst     r3, #0x80
+    orreq   r7, r7, #0xc000
+    tst     r3, #0x8000
+    orreq   r7, r7, #0x3000
+    tst     r3, #0x800000
+    orreq   r7, r7, #0x0c00
+    tst     r3, #0x80000000
+    orreq   r7, r7, #0x0300
+    bic     r3, r3, r8
+    add     r3, r3, r5
+#else                           /* Parallel bridge mode */
+    mov     r7, #0
+    tst     r3, #0x80
+    orreq   r7, r7, #0xc0
+    tst     r3, #0x8000
+    orreq   r7, r7, #0x30
+    tst     r3, #0x800000
+    orreq   r7, r7, #0x0c
+    tst     r3, #0x80000000
+    orreq   r7, r7, #0x03
+    bic     r3, r3, r8
+    add     r3, r3, r5
+
+1:
+    ldr     r5, [r9]
+    tst     r5, #0x8000
+    bne     1b
+
+    str     r7, [r9, #0x10]
+    mov     r7, #0
+#endif
+
+    tst     r4, #0x80
+    orreq   r7, r7, #0xc0
+    tst     r4, #0x8000
+    orreq   r7, r7, #0x30
+    tst     r4, #0x800000
+    orreq   r7, r7, #0x0c
+    tst     r4, #0x80000000
+    orreq   r7, r7, #0x03
+    bic     r4, r4, r8
+    add     r4, r4, r6
+
+    stmia   r1!, {r3-r4}
+
+1:
+    ldr     r5, [r9]
+    tst     r5, #0x8000
+    bne     1b
+#ifdef IPOD_MINI2G
+    str     r7, [r9, #0x08]
+#else
+    str     r7, [r9, #0x10]
+#endif
+
+    subs    r2, r2, #1
+    bne     .greyloop
+
+    ldmfd   sp!, {r4-r9, pc}
+
+.gd_end:
+    .size   lcd_grey_data,.gd_end-lcd_grey_data
+
diff --git a/firmware/target/arm/ipod/lcd-gray.c b/firmware/target/arm/ipod/lcd-gray.c
index 5734480..b77d3eb 100644
--- a/firmware/target/arm/ipod/lcd-gray.c
+++ b/firmware/target/arm/ipod/lcd-gray.c
@@ -74,56 +74,47 @@ static unsigned short contrast_reg_h;
 static int addr_offset;
 #if defined(IPOD_MINI) || defined(IPOD_MINI2G)
 static int pix_offset;
+void lcd_write_data_shifted(const fb_data* p_bytes, int count);
 #endif
 
-static const unsigned char dibits[16] ICONST_ATTR = {
-    0x00, 0x03, 0x0C, 0x0F, 0x30, 0x33, 0x3C, 0x3F,
-    0xC0, 0xC3, 0xCC, 0xCF, 0xF0, 0xF3, 0xFC, 0xFF
-};
-
 /* wait for LCD with timeout */
 static inline void lcd_wait_write(void)
 {
     while (LCD1_CONTROL & LCD1_BUSY_MASK);
 }
 
-/* send LCD data */
-#if CONFIG_CPU == PP5002
-STATICIRAM void ICODE_ATTR lcd_send_data(unsigned data)
-#else
-static void lcd_send_data(unsigned data)
-#endif
+/* send LCD command */
+static void lcd_prepare_cmd(unsigned cmd)
 {
     lcd_wait_write();
 #ifdef IPOD_MINI2G
-    LCD1_CMD = data | 0x760000;
+    LCD1_CMD = cmd | 0x740000;
 #else
-    LCD1_DATA = data >> 8;
+    LCD1_CMD = 0;
     lcd_wait_write();
-    LCD1_DATA = data & 0xff;
+    LCD1_CMD = cmd;
 #endif
 }
 
-/* send LCD command */
-static void lcd_prepare_cmd(unsigned cmd)
+/* send LCD command and data */
+static void lcd_cmd_and_data(unsigned cmd, unsigned data)
 {
     lcd_wait_write();
 #ifdef IPOD_MINI2G
     LCD1_CMD = cmd | 0x740000;
+    lcd_wait_write();
+    LCD1_CMD = data | 0x760000;
 #else
     LCD1_CMD = 0;
     lcd_wait_write();
     LCD1_CMD = cmd;
+    lcd_wait_write();
+    LCD1_DATA = data >> 8;
+    lcd_wait_write();
+    LCD1_DATA = data & 0xff;
 #endif
 }
 
-/* send LCD command and data */
-static void lcd_cmd_and_data(unsigned cmd, unsigned data)
-{
-    lcd_prepare_cmd(cmd);
-    lcd_send_data(data);
-}
-
 /* LCD init */
 void lcd_init_device(void)
 {
@@ -230,27 +221,29 @@ void lcd_set_invert_display(bool yesno)
 void lcd_set_flip(bool yesno)
 {
 #if defined(IPOD_MINI) || defined(IPOD_MINI2G)
-    if (yesno) {
-         /* 168x112, inverse COM order */
+    if (yesno) 
+    {    /* 168x112, inverse COM order */
         lcd_cmd_and_data(R_DRV_OUTPUT_CONTROL, 0x020d);
         lcd_cmd_and_data(R_1ST_SCR_DRV_POS, 0x8316);    /* 22..131 */
         addr_offset = (22 << 5) | (20 - 4);
         pix_offset = -2;
-    } else {
-        /* 168x112,  inverse SEG order */
+    }
+    else 
+    {   /* 168x112,  inverse SEG order */
         lcd_cmd_and_data(R_DRV_OUTPUT_CONTROL, 0x010d);
         lcd_cmd_and_data(R_1ST_SCR_DRV_POS, 0x6d00);    /* 0..109 */
         addr_offset = 20;
         pix_offset = 0;
     }
 #else
-    if (yesno) {
-        /* 168x128, inverse SEG & COM order */
+    if (yesno) 
+    {   /* 168x128, inverse SEG & COM order */
         lcd_cmd_and_data(R_DRV_OUTPUT_CONTROL, 0x030f);
         lcd_cmd_and_data(R_1ST_SCR_DRV_POS, 0x8304);    /* 4..131 */
         addr_offset = (4 << 5) | (20 - 1);
-    } else {
-        /* 168x128 */
+    } 
+    else 
+    {   /* 168x128 */
         lcd_cmd_and_data(R_DRV_OUTPUT_CONTROL, 0x000f);
         lcd_cmd_and_data(R_1ST_SCR_DRV_POS, 0x7f00);    /* 0..127 */
         addr_offset = 20;
@@ -279,108 +272,38 @@ void lcd_enable(bool on)
 
 /*** update functions ***/
 
+/* Helper function. */
+void lcd_mono_data(const unsigned char *data, int count);
+
 /* Performance function that works with an external buffer
    note that x, bwidtht and stride are in 8-pixel units! */
 void lcd_blit(const unsigned char* data, int bx, int y, int bwidth,
               int height, int stride)
 {
-    const unsigned char *src, *src_end;
-
-    while (height--) {
-        src = data;
-        src_end = data + bwidth;
+    while (height--)
+    {
         lcd_cmd_and_data(R_RAM_ADDR_SET, (y++ << 5) + addr_offset - bx);
         lcd_prepare_cmd(R_RAM_DATA);
-        do {
-            unsigned byte = *src++;
-            lcd_send_data((dibits[byte>>4] << 8) | dibits[byte&0x0f]);
-        } while (src < src_end);
+
+        lcd_mono_data(data, bwidth);
         data += stride;
     }
 }
 
+/* Helper function for lcd_grey_phase_blit(). */
+void lcd_grey_data(unsigned char *values, unsigned char *phases, int count);
+
 /* Performance function that works with an external buffer
    note that bx and bwidth are in 8-pixel units! */
 void lcd_grey_phase_blit(unsigned char *values, unsigned char *phases,
                          int bx, int y, int bwidth, int height, int stride)
 {
-    unsigned char *val, *ph;
-    int bw;
-
-    while (height--) {
+    while (height--) 
+    {
         lcd_cmd_and_data(R_RAM_ADDR_SET, (y++ << 5) + addr_offset - bx);
         lcd_prepare_cmd(R_RAM_DATA);
 
-        val = values;
-        ph = phases;
-        bw = bwidth;
-        asm volatile (
-        "10:                                 \n"
-            "ldmia   %[ph], {r0-r1}          \n" /* Fetch 8 pixel phases */
-            "ldmia   %[val]!, {r2-r3}        \n" /* Fetch 8 pixel values */
-#ifdef IPOD_MINI2G
-            "mov     r4, #0x7600             \n"
-#else
-            "mov     r4, #0                  \n"
-#endif
-            "tst     r0, #0x80               \n"
-            "orreq   r4, r4, #0xc0           \n"
-            "tst     r0, #0x8000             \n"
-            "orreq   r4, r4, #0x30           \n"
-            "tst     r0, #0x800000           \n"
-            "orreq   r4, r4, #0x0c           \n"
-            "tst     r0, #0x80000000         \n"
-            "orreq   r4, r4, #0x03           \n"
-            "bic     r0, r0, %[clbt]         \n"
-            "add     r0, r0, r2              \n"
-
-#ifdef IPOD_MINI2G
-            "mov     r4, r4, lsl #8          \n"
-#else
-        "1:                                  \n"
-            "ldr     r2, [%[lcdb]]           \n"
-            "tst     r2, #0x8000             \n"
-            "bne     1b                      \n"
-
-            "str     r4, [%[lcdb], #0x10]    \n"
-            "mov     r4, #0                  \n"
-#endif
-
-            "tst     r1, #0x80               \n"
-            "orreq   r4, r4, #0xc0           \n"
-            "tst     r1, #0x8000             \n"
-            "orreq   r4, r4, #0x30           \n"
-            "tst     r1, #0x800000           \n"
-            "orreq   r4, r4, #0x0c           \n"
-            "tst     r1, #0x80000000         \n"
-            "orreq   r4, r4, #0x03           \n"
-            "bic     r1, r1, %[clbt]         \n"
-            "add     r1, r1, r3              \n"
-
-            "stmia   %[ph]!, {r0-r1}         \n"
-
-        "1:                                  \n"
-            "ldr     r2, [%[lcdb]]           \n"
-            "tst     r2, #0x8000             \n"
-            "bne     1b                      \n"
-#ifdef IPOD_MINI2G
-            "str     r4, [%[lcdb], #0x08]    \n"
-#else
-            "str     r4, [%[lcdb], #0x10]    \n"
-#endif
-
-            "subs    %[bw], %[bw], #1        \n"
-            "bne     10b                     \n"
-            : /* outputs */
-            [val]"+r"(val),
-            [ph] "+r"(ph),
-            [bw] "+r"(bw)
-            : /* inputs */
-            [clbt]"r"(0x80808080),
-            [lcdb]"r"(LCD1_BASE)
-            : /* clobbers */
-            "r0", "r1", "r2", "r3", "r4"
-        );
+        lcd_grey_data(values, phases, bwidth);
         values += stride;
         phases += stride;
     }
@@ -407,30 +330,17 @@ void lcd_update_rect(int x, int y, int width, int height)
     x >>= 3;
     width = xmax - x + 1;
 
-    for (; y <= ymax; y++) {
-        unsigned char *data, *data_end;
-
+    for (; y <= ymax; y++) 
+    {
         lcd_cmd_and_data(R_RAM_ADDR_SET, (y << 5) + addr_offset - x);
         lcd_prepare_cmd(R_RAM_DATA);
 
-        data = &lcd_framebuffer[y][2*x];
-        data_end = data + 2 * width;
 #if defined(IPOD_MINI) || defined(IPOD_MINI2G)
-        if (pix_offset == -2) {
-            unsigned cur_word = *data++;
-            do {
-                cur_word = (cur_word << 8) | *data++;
-                cur_word = (cur_word << 8) | *data++;
-                lcd_send_data((cur_word >> 4) & 0xffff);
-            } while (data <= data_end);
-        } else
+        if (pix_offset == -2)
+            lcd_write_data_shifted(&lcd_framebuffer[y][2*x], width);
+        else
 #endif
-        {
-            do {
-                unsigned highbyte = *data++;
-                lcd_send_data((highbyte << 8) | *data++);
-            } while (data < data_end);
-        }
+            lcd_write_data(&lcd_framebuffer[y][2*x], width);
     }
 }