Gigabeat: properly confined framebuffer copies and a few pendantic changes to lcd_yuv_blit. No difference full screen but quite a speedup to copy only the required bit: 534->1062 fps for 1/4 screen update using test_fps.

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@13821 a1c6a512-1295-4272-9138-f99709370657

2 files changed, 146 insertions, 41 deletions

diff --git a/firmware/target/arm/s3c2440/gigabeat-fx/lcd-as-meg-fx.S b/firmware/target/arm/s3c2440/gigabeat-fx/lcd-as-meg-fx.S
index d431c95f29..f4805fd1e1 100644
--- a/firmware/target/arm/s3c2440/gigabeat-fx/lcd-as-meg-fx.S
+++ b/firmware/target/arm/s3c2440/gigabeat-fx/lcd-as-meg-fx.S
@@ -21,6 +21,87 @@
 #include "cpu.h"
 
 /****************************************************************************
+ * void lcd_copy_buffer_rect(fb_data *dst, fb_data *src, int width,
+ *                           int height);
+ */
+     .section    .icode, "ax", %progbits
+     .align      2
+     .global     lcd_copy_buffer_rect
+     .type       lcd_copy_buffer_rect, %function
+                                        @ r0 = dst
+                                        @ r1 = src
+                                        @ r2 = width
+                                        @ r3 = height
+lcd_copy_buffer_rect:                   @
+    stmfd   sp!, { r4-r12, lr }         @ save non-scratch regs
+    mov     r5, r2                      @ r5 = cached width
+    rsb     r4, r2, #LCD_WIDTH          @ r4 = LCD_WIDTH - width
+10: @ copy line                         @
+    subs    r2, r5, #1                  @ r2 = width - 1
+    beq     40f @ finish line           @ one halfword? skip to trailing copy
+    tst     r0, #2                      @ word aligned?
+    beq     20f @ rem copy              @ yes? skip to word copy
+    ldrh    r6, [r1], #2                @ copy leading halfword
+    subs    r2, r2, #1                  @
+    strh    r6, [r0], #2                @
+    ble     40f @ finish line           @ next line if lt or finish
+                                        @ trailing halfword if eq
+20: @ rem copy                          @
+    add     r14, r2, #1                 @ get remaining width mod 16 after word
+                                        @ align (rw)
+    and     r14, r14, #0xe              @ r14 = 0 (16), 2, 4, 6, 8, 10, 12, 14
+    add     pc, pc, r14, lsl #3         @ branch to 32-byte align
+    nop                                 @ 
+    b       30f                         @ rw % 16 = 0 or 1? use octword loop
+    nop                                 @
+    nop                                 @
+    nop                                 @
+    ldr     r6, [r1], #4                @ rw % 16 = 2 or 3
+    subs    r2, r2, #2                  @
+    str     r6, [r0], #4                @
+    b       25f @ copy up done          @
+    ldmia   r1!, { r6-r7 }              @ rw % 16 = 4 or 5
+    subs    r2, r2, #4                  @
+    stmia   r0!, { r6-r7 }              @
+    b       25f @ copy up done          @
+    ldmia   r1!, { r6-r8 }              @ rw % 16 = 6 or 7
+    subs    r2, r2, #6                  @
+    stmia   r0!, { r6-r8 }              @
+    b       25f @ copy up done          @
+    ldmia   r1!, { r6-r9 }              @ rw % 16 = 8 or 9
+    subs    r2, r2, #8                  @
+    stmia   r0!, { r6-r9 }              @
+    b       25f @ copy up done          @
+    ldmia   r1!, { r6-r10 }             @ rw % 16 = 10 or 11
+    subs    r2, r2, #10                 @
+    stmia   r0!, { r6-r10 }             @
+    b       25f @ copy up done          @
+    ldmia   r1!, { r6-r11 }             @ rw % 16 = 12 or 13
+    subs    r2, r2, #12                 @
+    stmia   r0!, { r6-r11 }             @
+    b       25f @ copy up done          @
+    ldmia   r1!, { r6-r12 }             @ rw % 16 = 14 or 15
+    subs    r2, r2, #14                 @
+    stmia   r0!, { r6-r12 }             @
+25: @ copy up done                      @
+    ble     40f @ finish line           @ no 32-byte segments remaining?
+30: @ octword loop                      @ copy 16 pixels per loop
+    ldmia   r1!, { r6-r12, r14 }        @
+    subs    r2, r2, #16                 @
+    stmia   r0!, { r6-r12, r14 }        @
+    bgt     30b @ octword loop          @ 
+40: @ finish line                       @
+    ldreqh  r6, [r1], #2                @ finish last halfword if eq ...
+    add     r1, r1, r4, lsl #1          @
+    streqh  r6, [r0], #2                @ ...
+    add     r0, r0, r4, lsl #1          @
+    subs    r3, r3, #1                  @ next line
+    bgt     10b @ copy line             @
+    ldmfd   sp!, { r4-r12, pc }         @ restore regs and return
+    .size   lcd_copy_buffer_rect, .-lcd_copy_buffer_rect
+
+
+/****************************************************************************
  * void lcd_write_yuv_420_lines(fb_data *dst,
  *                              unsigned char chroma_buf[LCD_HEIGHT/2*3],
                                 unsigned char const * const src[3],
@@ -45,8 +126,8 @@ lcd_write_yuv420_lines:
                                         @ r2   = yuv_src
                                         @ r3   = width
                                         @ [sp] = stride
-    stmdb       sp!, { r4-r12, lr }     @ save non-scratch
-    stmdb       sp!, { r0, r3 }         @ save dst and width
+    stmfd       sp!, { r4-r12, lr }     @ save non-scratch
+    stmfd       sp!, { r0, r3 }         @ save dst and width
     mov         r14, #74                @ r14 = Y factor
     ldmia       r2, { r4, r5, r6 }      @ r4 = yuv_src[0] = Y'_p
                                         @ r5 = yuv_src[1] = Cb_p
@@ -105,7 +186,7 @@ lcd_write_yuv420_lines:
     orr         r12, r2, r7, lsl #5     @ r4 |= (g << 5)
     ldrb        r2, [r4], #1            @ r2 = Y' = *Y'_p++
     orr         r12, r12, r11, lsl #11  @ r4 = b | (r << 11)
-    strh        r12, [r0], #240         @ store pixel
+    strh        r12, [r0], #LCD_WIDTH   @ store pixel
                                         @
     sub         r2, r2, #16             @ r7 = Y = (Y' - 16)*74
     mul         r7, r2, r14             @ next Y
@@ -133,14 +214,14 @@ lcd_write_yuv420_lines:
                                         @
     orr         r12, r2, r11, lsl #11   @ r4 = b | (r << 11)
     orr         r12, r12, r7, lsl #5    @ r4 |= (g << 5)
-    strh        r12, [r0, #240]!        @ store pixel
-    add         r0, r0, #2*240          @
+    strh        r12, [r0, #LCD_WIDTH]!  @ store pixel
+    add         r0, r0, #2*LCD_WIDTH    @
                                         @
     subs        r3, r3, #2              @
     bgt         10b @ loop line 1       @
                                         @ do second line
                                         @
-    ldmia       sp!, { r0, r3 }         @ pop dst and width
+    ldmfd       sp!, { r0, r3 }         @ pop dst and width
     sub         r0, r0, #2              @ set dst to start of next line
     sub         r1, r1, r3, asl #1      @ rewind chroma pointer...
     ldr         r2, [sp, #40]           @ r2 = stride
@@ -182,7 +263,7 @@ lcd_write_yuv420_lines:
     orr         r12, r2, r11, lsl #11   @ r4 = b | (r << 11)
     ldrb        r2, [r4], #1            @ r2 = Y' = *Y'_p++
     orr         r12, r12, r7, lsl #5    @ r4 |= (g << 5)
-    strh        r12, [r0], #240         @ store pixel
+    strh        r12, [r0], #LCD_WIDTH   @ store pixel
                                         @
                                         @ do second pixel
                                         @
@@ -212,11 +293,11 @@ lcd_write_yuv420_lines:
                                         @
     orr         r12, r2, r11, lsl #11   @ r4 = b | (r << 11)
     orr         r12, r12, r7, lsl #5    @ r4 |= (g << 5)
-    strh        r12, [r0, #240]!        @ store pixel
-    add         r0, r0, #2*240          @
+    strh        r12, [r0, #LCD_WIDTH]!  @ store pixel
+    add         r0, r0, #2*LCD_WIDTH    @
                                         @
     subs        r3, r3, #2              @
     bgt         20b @ loop line 2       @
                                         @
-    ldmia       sp!, { r4-r12, pc }     @ restore registers and return
+    ldmfd       sp!, { r4-r12, pc }     @ restore registers and return
     .size   lcd_write_yuv420_lines, .-lcd_write_yuv420_lines
diff --git a/firmware/target/arm/s3c2440/gigabeat-fx/lcd-meg-fx.c b/firmware/target/arm/s3c2440/gigabeat-fx/lcd-meg-fx.c
index c96e0d6d25..5e1110eb3d 100644
--- a/firmware/target/arm/s3c2440/gigabeat-fx/lcd-meg-fx.c
+++ b/firmware/target/arm/s3c2440/gigabeat-fx/lcd-meg-fx.c
@@ -15,6 +15,12 @@ volatile bool lcd_poweroff = false;
 extern unsigned fg_pattern;
 extern unsigned bg_pattern;
 
+/* Copies a rectangle from one framebuffer to another. Can be used in
+   single transfer mode with width = num pixels, and height = 1 which
+   allows a full-width rectangle to be copied more efficiently. */
+extern void lcd_copy_buffer_rect(fb_data *dst, const fb_data *src,
+                                 int width, int height);
+
 bool lcd_enabled()
 {
     return lcd_on;
@@ -36,7 +42,6 @@ unsigned int LCDBASEL(unsigned int address)
     return (address & ((1 << 22)-1)) >> 1;
 }
 
-
 /* LCD init */
 void lcd_init_device(void)
 {
@@ -122,17 +127,41 @@ void lcd_init_device(void)
 /* Update a fraction of the display. */
 void lcd_update_rect(int x, int y, int width, int height)
 {
-    (void)x;
-    (void)width;
-    (void)y;
-    (void)height;
+    fb_data *dst, *src;
 
-    if(!lcd_on)
-    {
-        sleep(200);
+    if (!lcd_on)
         return;
+
+    if (x + width > LCD_WIDTH)
+        width = LCD_WIDTH - x; /* Clip right */
+    if (x < 0)
+        width += x, x = 0; /* Clip left */
+    if (width <= 0)
+        return; /* nothing left to do */
+
+    if (y + height > LCD_HEIGHT)
+        height = LCD_HEIGHT - y; /* Clip bottom */
+    if (y < 0)
+        height += y, y = 0; /* Clip top */
+    if (height <= 0)
+        return; /* nothing left to do */
+
+    /* TODO: It may be faster to swap the addresses of lcd_driver_framebuffer
+     * and lcd_framebuffer */
+    dst = (fb_data *)FRAME + LCD_WIDTH*y + x;
+    src = &lcd_framebuffer[y][x];
+
+    /* Copy part of the Rockbox framebuffer to the second framebuffer */
+    if (width < LCD_WIDTH)
+    {
+        /* Not full width - do line-by-line */
+        lcd_copy_buffer_rect(dst, src, width, height);
+    }
+    else
+    {
+        /* Full width - copy as one line */
+        lcd_copy_buffer_rect(dst, src, LCD_WIDTH*height, 1);
     }
-    memcpy(((char*)FRAME) + (y * sizeof(fb_data) * LCD_WIDTH), ((char *)&lcd_framebuffer) + (y * sizeof(fb_data) * LCD_WIDTH), ((height * sizeof(fb_data) * LCD_WIDTH)));
 }
 
 void lcd_enable(bool state)
@@ -144,7 +173,7 @@ void lcd_enable(bool state)
         if(!lcd_on)
         {
             lcd_on = true;
-            memcpy(FRAME, lcd_framebuffer, LCD_WIDTH*LCD_HEIGHT*2);
+            lcd_update();
             LCDCON1 |= 1;
         }
     }
@@ -161,7 +190,11 @@ void lcd_enable(bool state)
    This must be called after all other LCD functions that change the display. */
 void lcd_update(void)
 {
-    lcd_update_rect(0, 0, LCD_WIDTH, LCD_HEIGHT);
+    if (!lcd_on)
+        return;
+
+    lcd_copy_buffer_rect((fb_data *)FRAME, &lcd_framebuffer[0][0],
+                         LCD_WIDTH*LCD_HEIGHT, 1);
 }
 
 void lcd_bitmap_transparent_part(const fb_data *src, int src_x, int src_y,
@@ -171,28 +204,19 @@ void lcd_bitmap_transparent_part(const fb_data *src, int src_x, int src_y,
     fb_data *dst, *dst_end;
     unsigned int transcolor;
 
-    /* nothing to draw? */
-    if ((width <= 0) || (height <= 0) || (x >= LCD_WIDTH) || (y >= LCD_HEIGHT)
-        || (x + width <= 0) || (y + height <= 0))
-        return;
-
-    /* clipping */
-    if (x < 0)
-    {
-        width += x;
-        src_x -= x;
-        x = 0;
-    }
-    if (y < 0)
-    {
-        height += y;
-        src_y -= y;
-        y = 0;
-    }
     if (x + width > LCD_WIDTH)
-        width = LCD_WIDTH - x;
+        width = LCD_WIDTH - x; /* Clip right */
+    if (x < 0)
+        width += x, x = 0; /* Clip left */
+    if (width <= 0)
+        return; /* nothing left to do */
+
     if (y + height > LCD_HEIGHT)
-        height = LCD_HEIGHT - y;
+        height = LCD_HEIGHT - y; /* Clip bottom */
+    if (y < 0)
+        height += y, y = 0; /* Clip top */
+    if (height <= 0)
+        return; /* nothing left to do */
 
     src += stride * src_y + src_x; /* move starting point */
     dst = &lcd_framebuffer[(y)][(x)];