Sansa AMS: Time has shown that switching between 16 and 32bit mode costs much time (due to the micro delay needed), so do 32bit transfers unconditionally for lcd updates at the cost of updating slightly larger rectangles (gives upto 15% speed up, nearly at maximum now).

Unify this optimized dbop transfer function and re-use it more often (it still handles 16bit transfers).

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@24198 a1c6a512-1295-4272-9138-f99709370657

4 files changed, 76 insertions, 93 deletions

diff --git a/firmware/target/arm/as3525/dbop-as3525.c b/firmware/target/arm/as3525/dbop-as3525.c
index 938eee2fb1..b54ad17a6c 100644
--- a/firmware/target/arm/as3525/dbop-as3525.c
+++ b/firmware/target/arm/as3525/dbop-as3525.c
@@ -20,6 +20,7 @@
  ****************************************************************************/
 
 #include "config.h"
+#include <inttypes.h>
 #include "as3525.h"
 #include "dbop-as3525.h"
 
@@ -75,3 +76,50 @@ unsigned short dbop_debug(void)
 {
     return dbop_input_value;
 }
+
+static inline void dbop_set_mode(int mode)
+{
+    int delay = 10;
+    if (mode == 32 && (!(DBOP_CTRL & (1<<13|1<<14))))
+        DBOP_CTRL |= (1<<13|1<<14);
+    else if (mode == 16 && (DBOP_CTRL & (1<<13|1<<14)))
+        DBOP_CTRL &= ~(1<<14|1<<13);
+    else
+        return;
+    while(delay--) asm volatile("nop");
+}
+
+void dbop_write_data(const int16_t* p_bytes, int count)
+{
+    
+    const int32_t *data;
+    if ((intptr_t)p_bytes & 0x3 || count == 1)
+    {   /* need to do a single 16bit write beforehand if the address is
+         * not word aligned or count is 1, switch to 16bit mode if needed */
+        dbop_set_mode(16);
+        DBOP_DOUT16 = *p_bytes++;
+        if (!(--count))
+            return;
+    }
+    /* from here, 32bit transfers are save
+     * set it to transfer 4*(outputwidth) units at a time,
+     * if bit 12 is set it only does 2 halfwords though (we never set it)
+     * switch to 32bit output if needed */
+    dbop_set_mode(32);
+    data = (int32_t*)p_bytes;
+    while (count > 1)
+    {
+        DBOP_DOUT32 = *data++;
+        count -= 2;
+
+        /* Wait if push fifo is full */
+        while ((DBOP_STAT & (1<<6)) != 0);
+    }
+    /* While push fifo is not empty */
+    while ((DBOP_STAT & (1<<10)) == 0);
+
+    /* due to the 32bit alignment requirement or uneven count,
+     * we possibly need to do a 16bit transfer at the end also */
+    if (count > 0)
+        dbop_write_data((int16_t*)data, 1);
+}
diff --git a/firmware/target/arm/as3525/dbop-as3525.h b/firmware/target/arm/as3525/dbop-as3525.h
index 63156031c4..3db62bd3e1 100644
--- a/firmware/target/arm/as3525/dbop-as3525.h
+++ b/firmware/target/arm/as3525/dbop-as3525.h
@@ -19,5 +19,17 @@
  *
  ****************************************************************************/
 
+
+#ifndef __DBOP_AS3525_H__
+#define __DBOP_AS3525_H__
+#include <inttypes.h>
 unsigned short int  dbop_read_input(void);
 unsigned short      dbop_debug(void);
+
+/*
+ * Write any data to dbop
+ * switches to 32bit transfers if possible,
+ * for best performance pre-align count and data
+ **/
+void dbop_write_data(const int16_t *data, int count);
+#endif
diff --git a/firmware/target/arm/as3525/sansa-e200v2/lcd-e200v2.c b/firmware/target/arm/as3525/sansa-e200v2/lcd-e200v2.c
index f4d1a7cf56..54aa22e398 100644
--- a/firmware/target/arm/as3525/sansa-e200v2/lcd-e200v2.c
+++ b/firmware/target/arm/as3525/sansa-e200v2/lcd-e200v2.c
@@ -120,19 +120,11 @@ static void as3525_dbop_init(void)
     /* TODO: The OF calls some other functions here, but maybe not important */
 }
 
-static void lcd_write_value16(unsigned short value)
-{
-    DBOP_CTRL &= ~(1<<14|1<<13);
-    lcd_delay(10);
-    DBOP_DOUT16 = value;
-    while ((DBOP_STAT & (1<<10)) == 0);
-}
-
-static void lcd_write_cmd(int cmd)
+static void lcd_write_cmd(short cmd)
 {
     /* Write register */
     DBOP_TIMPOL_23 = 0xa167006e;
-    lcd_write_value16(cmd);
+    dbop_write_data(&cmd, 1);
 
     /* Wait for fifo to empty */
     while ((DBOP_STAT & (1<<10)) == 0);
@@ -144,43 +136,13 @@ static void lcd_write_cmd(int cmd)
     DBOP_TIMPOL_23 = 0xa167e06f;
 }
 
-void lcd_write_data(const fb_data* p_bytes, int count)
-{
-    const long *data;
-    if ((int)p_bytes & 0x3)
-    {   /* need to do a single 16bit write beforehand if the address is
-         * not word aligned */
-        lcd_write_value16(*p_bytes);
-        count--;p_bytes++;
-    }
-    /* from here, 32bit transfers are save
-     * set it to transfer 4*(outputwidth) units at a time,
-     * if bit 12 is set it only does 2 halfwords though */
-    DBOP_CTRL |= (1<<13|1<<14);
-    data = (long*)p_bytes;
-    while (count > 1)
-    {
-        DBOP_DOUT32 = *data++;
-        count -= 2;
-
-        /* Wait if push fifo is full */
-        while ((DBOP_STAT & (1<<6)) != 0);
-    }
-    /* While push fifo is not empty */
-    while ((DBOP_STAT & (1<<10)) == 0);
-
-    /* due to the 32bit alignment requirement or uneven count,
-     * we possibly need to do a 16bit transfer at the end also */
-    if (count > 0)
-        lcd_write_value16(*(fb_data*)data);
-}
 
 static void lcd_write_reg(int reg, int value)
 {
-    fb_data data = value;
+    unsigned short data = value;
 
     lcd_write_cmd(reg);
-    lcd_write_value16(data);
+    dbop_write_data(&data, 1);
 }
 
 /*** hardware configuration ***/
@@ -486,7 +448,7 @@ void lcd_update(void)
 
     lcd_write_cmd(R_WRITE_DATA_2_GRAM);
 
-    lcd_write_data((fb_data*)lcd_framebuffer, LCD_WIDTH*LCD_HEIGHT);
+    dbop_write_data((fb_data*)lcd_framebuffer, LCD_WIDTH*LCD_HEIGHT);
 }
 
 /* Update a fraction of the display. */
@@ -526,7 +488,7 @@ void lcd_update_rect(int x, int y, int width, int height)
 
     do
     {
-        lcd_write_data(ptr, width);
+        dbop_write_data(ptr, width);
         ptr += LCD_WIDTH;
     }
     while (--height > 0);
diff --git a/firmware/target/arm/as3525/sansa-fuze/lcd-fuze.c b/firmware/target/arm/as3525/sansa-fuze/lcd-fuze.c
index 3332e0c78c..bdf1c704e0 100644
--- a/firmware/target/arm/as3525/sansa-fuze/lcd-fuze.c
+++ b/firmware/target/arm/as3525/sansa-fuze/lcd-fuze.c
@@ -124,66 +124,23 @@ static void as3525_dbop_init(void)
     /* TODO: The OF calls some other functions here, but maybe not important */
 }
 
-static void lcd_write_value16(unsigned short value)
-{
-    DBOP_CTRL &= ~(1<<14|1<<13);
-    lcd_delay(10);
-    DBOP_DOUT16 = value;
-    while ((DBOP_STAT & (1<<10)) == 0);
-}
-
-static void lcd_write_cmd(int cmd)
+static void lcd_write_cmd(short cmd)
 {
     /* Write register */
     DBOP_TIMPOL_23 = 0xa167006e;
-    lcd_write_value16(cmd);
-
-    /* Wait for fifo to empty */
-    while ((DBOP_STAT & (1<<10)) == 0);
+    dbop_write_data(&cmd, 1);
 
     lcd_delay(4);
 
     DBOP_TIMPOL_23 = 0xa167e06f;
 }
 
-void lcd_write_data(const fb_data* p_bytes, int count)
-{
-    const long *data;
-    if ((int)p_bytes & 0x3)
-    {   /* need to do a single 16bit write beforehand if the address is
-         * not word aligned */
-        lcd_write_value16(*p_bytes);
-        count--;p_bytes++;
-    }
-    /* from here, 32bit transfers are save
-     * set it to transfer 4*(outputwidth) units at a time,
-     * if bit 12 is set it only does 2 halfwords though */
-    DBOP_CTRL |= (1<<13|1<<14);
-    lcd_delay(10);
-    data = (long*)p_bytes;
-    while (count > 1)
-    {
-        DBOP_DOUT32 = *data++;
-        count -= 2;
-
-        /* Wait if push fifo is full */
-        while ((DBOP_STAT & (1<<6)) != 0);
-    }
-    /* While push fifo is not empty */
-    while ((DBOP_STAT & (1<<10)) == 0);
-
-    /* due to the 32bit alignment requirement or uneven count,
-     * we possibly need to do a 16bit transfer at the end also */
-    if (count > 0)
-        lcd_write_value16(*(fb_data*)data);
-}
-
 static void lcd_write_reg(int reg, int value)
 {
-    fb_data data = value;
+    int16_t data = value;
 
     lcd_write_cmd(reg);
-    lcd_write_value16(data);
+    dbop_write_data(&data, 1);
 }
 
 /*** hardware configuration ***/
@@ -437,7 +394,7 @@ void lcd_update(void)
 
     lcd_write_cmd(R_WRITE_DATA_2_GRAM);
 
-    lcd_write_data((fb_data*)lcd_framebuffer, LCD_WIDTH*LCD_HEIGHT);
+    dbop_write_data((fb_data*)lcd_framebuffer, LCD_WIDTH*LCD_HEIGHT);
 }
 
 /* Update a fraction of the display. */
@@ -470,6 +427,10 @@ void lcd_update_rect(int x, int y, int width, int height)
 
     lcd_write_reg(R_ENTRY_MODE, r_entry_mode);
 
+    /* we need to make x and width even to enable 32bit transfers */
+    width = (width + (x & 1) + 1) & ~1;
+    x &= ~1;
+
     lcd_window_x(x, x + width - 1);
     lcd_window_y(y, y + height -1);
 
@@ -479,7 +440,7 @@ void lcd_update_rect(int x, int y, int width, int height)
 
     do
     {
-        lcd_write_data(ptr, width);
+        dbop_write_data(ptr, width);
         ptr += LCD_WIDTH;
     }
     while (--height > 0);