Greyscale library: Changed the internal data format once more (separated pixel values and phases), allowing for further optimisation of drawing, scrolling etc. * Optimised grey phase blitting in the core reduces CPU load on all architectures, most significantly on coldfire. Previous version was too slow to keep up at 45MHz, leading to unwanted graininess (update frequency was halved). Also fixed screendump on 2bpp targets with vertical pixel packing.

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@16043 a1c6a512-1295-4272-9138-f99709370657

2 files changed, 66 insertions, 71 deletions

diff --git a/firmware/target/sh/archos/lcd-archos-bitmap.c b/firmware/target/sh/archos/lcd-archos-bitmap.c
index 17c4d76092..18d48f4ead 100644
--- a/firmware/target/sh/archos/lcd-archos-bitmap.c
+++ b/firmware/target/sh/archos/lcd-archos-bitmap.c
@@ -155,10 +155,13 @@ void lcd_blit(const unsigned char* data, int x, int by, int width,
     }
 }
 
+/* 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 by and bheight are in 8-pixel units! */
-void lcd_grey_phase_blit(const struct grey_data *data, int x, int by,
-                         int width, int bheight, int stride)
+void lcd_grey_phase_blit(unsigned char *values, unsigned char *phases,
+                         int x, int by, int width, int bheight, int stride)
 {
     stride <<= 3; /* 8 pixels per block */
     while (bheight--)
@@ -167,8 +170,9 @@ void lcd_grey_phase_blit(const struct grey_data *data, int x, int by,
         lcd_write_command (LCD_CNTL_HIGHCOL | (((x+xoffset)>>4) & 0xf));
         lcd_write_command (LCD_CNTL_LOWCOL | ((x+xoffset) & 0xf));
 
-        lcd_grey_data(data, width);
-        data += stride;
+        lcd_grey_data(values, phases, width);
+        values += stride;
+        phases += stride;
     }
 }
 
diff --git a/firmware/target/sh/archos/lcd-as-archos-bitmap.S b/firmware/target/sh/archos/lcd-as-archos-bitmap.S
index 492413fdb7..a60ea9aad1 100644
--- a/firmware/target/sh/archos/lcd-as-archos-bitmap.S
+++ b/firmware/target/sh/archos/lcd-as-archos-bitmap.S
@@ -210,28 +210,35 @@ _lcd_write_data:
  * one or multiple pixels.
  *
  * Arguments:
- *   r4 - data address, (phase,value)-pairs
- *   r5 - pixel block count
+ *   r4 - pixel value data address
+ *   r5 - pixel phase data address
+ *   r6 - pixel block count
  *
  * Register usage:
- *   r0 - current pixel value
+ *   r0 - scratch / phase signs mask
  *   r1 - scratch
  *   r2 - precalculated port value (CS and SC low, DS and SD high),
  *        negated (neg)!
  *   r3 - lcd port address
- *   r5 - end address
- *   r6/r7 - current/next pixel phase
- *   r8 - current block address (for writing back phase)
- *   r9 - 0x80 (for phase modification)
+ *   r4 - current value address
+ *   r5 - current phase address
+ *   r6 - end address
+ *   r7/r8 - current/next pixel phase
+ *   r9 - current pixel value
+ *   r10  - 0x00000080 \
+ *   r11 - 0x00008000  > for phase sign check
+ *   r12 - 0x00800000 /
  */
 
 _lcd_grey_data:
     mov.l   r8, @-r15       /* save r8 */
-    shll2   r5              /*    v */
+    shll2   r6              /*    v */
     mov.l   r9, @-r15       /* save r9 */
-    shll2   r5              /* r5 *= 16; (8 pixel per block * 2 bytes/pixel) */
+    shll    r6              /* r6 *= 8; (8 pixels per block) */
+    mov.l   r10, @-r15      /* save r10 */
+    add     r4, r6          /* end address */
     mov.l   .lcdr, r3       /* put lcd data port address in r3 */
-    add     r4, r5          /* end address */
+    nop                     /* keep alignment */
 
     /* This code will fail if an interrupt changes the contents of PBDRL.
      * If so, we must disable the interrupt here. If disabling interrupts
@@ -240,110 +247,90 @@ _lcd_grey_data:
      * this would significantly decrease performance. */
 
     mov.b   @r3, r0         /* r0 = PBDRL */
-    mov     #0x80, r9       /* for phase modification - "or #imm,xx" only allows r0 */
-    mov.b   @r4+, r6        /* fetch first pixel phase */
     or      #(LCD_DS|LCD_SD), r0    /* r0 |= LCD_DS|LCD_SD */
+    mov.l   r11, @-r15      /* save r11 */
     and     #(~(LCD_CS|LCD_SC)), r0 /* r0 &= ~(LCD_CS|LCD_SC) */
+    mov.l   r12, @-r15      /* save r12 */
     neg     r0, r2          /* r2 = 0 - r0 */
-    mov     #-3, r0         /* offset for storing phase */
 
     /* loop exploits that SD is on bit 0 for recorders and Ondios */
 
-.greyloop:
-    cmp/pz  r6              /* phase non-negative? */
-    mov.b   @r4+, r8        /* fetch pixel value */
-    negc    r2, r1          /* T -> SD, SC low */
-    mov.b   r1, @r3         /* set port */
-    or      r9, r6          /* r6 -= (r6 >= 0) ? 128 : 0; */
-    mov.b   @r4+, r7        /* fetch next pixel phase */
-    add     #(LCD_SC), r1   /* rise SC */
-    mov.b   r1, @r3         /* set port */
-    add     r8, r6          /* calculate new phase */
-    mov.b   r6, @(r0,r4)    /* store phase */
+    mov.w   .ptest, r10
+    swap.b  r10, r11
+    mov.l   @r5, r7
+    swap.w  r10, r12
+    mov.l   .pmask, r0
 
+.greyloop:
     cmp/pz  r7
-    mov.b   @r4+, r8
+    mov.l   @r4+, r9
     negc    r2, r1
     mov.b   r1, @r3
-    or      r9, r7
-    mov.b   @r4+, r6
     add     #(LCD_SC), r1
     mov.b   r1, @r3
-    add     r8, r7
-    mov.b   r7, @(r0,r4)
 
-    cmp/pz  r6
-    mov.b   @r4+, r8
+    tst     r12, r7
+    mov.l   @(4,r5), r8
     negc    r2, r1
     mov.b   r1, @r3
-    or      r9, r6
-    mov.b   @r4+, r7
     add     #(LCD_SC), r1
     mov.b   r1, @r3
-    add     r8, r6
-    mov.b   r6, @(r0,r4)
 
-    cmp/pz  r7
-    mov.b   @r4+, r8
+    tst     r11, r7
     negc    r2, r1
+    tst     r10, r7
     mov.b   r1, @r3
-    or      r9, r7
-    mov.b   @r4+, r6
     add     #(LCD_SC), r1
     mov.b   r1, @r3
-    add     r8, r7
-    mov.b   r7, @(r0,r4)
 
-    cmp/pz  r6
-    mov.b   @r4+, r8
+    or      r0, r7
+    sub     r9, r7
     negc    r2, r1
     mov.b   r1, @r3
-    or      r9, r6
-    mov.b   @r4+, r7
     add     #(LCD_SC), r1
     mov.b   r1, @r3
-    add     r8, r6
-    mov.b   r6, @(r0,r4)
 
-    cmp/pz  r7
-    mov.b   @r4+, r8
+    cmp/pz  r8
+    mov.l   r7, @r5
     negc    r2, r1
     mov.b   r1, @r3
-    or      r9, r7
-    mov.b   @r4+, r6
     add     #(LCD_SC), r1
     mov.b   r1, @r3
-    add     r8, r7
-    mov.b   r7, @(r0,r4)
 
-    cmp/pz  r6
-    mov.b   @r4+, r8
+    tst     r12, r8
+    mov.l   @r4+, r9
     negc    r2, r1
     mov.b   r1, @r3
-    or      r9, r6
-    mov.b   @r4+, r7
     add     #(LCD_SC), r1
     mov.b   r1, @r3
-    add     r8, r6
-    mov.b   r6, @(r0,r4)
 
-    cmp/pz  r7
-    mov.b   @r4+, r8
+    tst     r11, r8
+    mov.l   @(8,r5), r7
     negc    r2, r1
     mov.b   r1, @r3
-    or      r9, r7
-    mov.b   @r4+, r6
     add     #(LCD_SC), r1
     mov.b   r1, @r3
-    add     r8, r7
-    mov.b   r7, @(r0,r4)
 
-    cmp/hi  r4, r5          /* some blocks left? */
+    tst     r10, r8
+    or      r0, r8
+    negc    r2, r1
+    mov.b   r1, @r3
+    add     #(LCD_SC), r1
+    mov.b   r1, @r3
+
+    sub     r9, r8
+    mov.l   r8, @(4,r5)
+    
+    add     #8, r5
+    cmp/hi  r4, r6
     bt      .greyloop
 
+    mov.l   @r15+, r12      /* restore r12 */
     mov     #(LCD_CS|LCD_DS|LCD_SD|LCD_SC), r0
-    mov.l   @r15+, r9       /* restore r9 */
+    mov.l   @r15+, r11      /* restore r11 */
     or      r0, r1          /* restore port */
+    mov.l   @r15+, r10      /* restore r10 */
+    mov.l   @r15+, r9       /* restore r9 */
     mov.l   @r15+, r8       /* restore r8 */
     rts
     mov.b   r1, @r3
@@ -351,7 +338,11 @@ _lcd_grey_data:
     /* This is the place to reenable the interrupts, if we have disabled
      * them. See above. */
 
+.ptest:
+    .short  0x0080
 
     .align  2
 .lcdr:
     .long   LCDR
+.pmask:
+    .long   0x80808080