Greyscale library: Assembler optimised update for coldfire, ~73% speedup when using IRAM, ~35% speedup without IRAM. Slight optimisation for other targets as well.

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@16482 a1c6a512-1295-4272-9138-f99709370657

4 files changed, 149 insertions, 4 deletions

diff --git a/apps/plugins/lib/Makefile b/apps/plugins/lib/Makefile
index 39f1743..c75cca1 100644
--- a/apps/plugins/lib/Makefile
+++ b/apps/plugins/lib/Makefile
@@ -29,7 +29,8 @@ endif
 include $(TOOLSDIR)/makesrc.inc
 
 SOURCES = $(SRC)
-OBJS := $(SRC:%.c=$(OBJDIR)/%.o)
+OBJS2 := $(SRC:%.c=$(OBJDIR)/%.o)
+OBJS = $(patsubst %.S, $(OBJDIR)/%.o, $(OBJS2))
 DEPFILE = $(OBJDIR)/dep-pluginlib
 DIRS = .
 
diff --git a/apps/plugins/lib/SOURCES b/apps/plugins/lib/SOURCES
index c1f5a48..b638740 100644
--- a/apps/plugins/lib/SOURCES
+++ b/apps/plugins/lib/SOURCES
@@ -8,6 +8,9 @@ grey_core.c
 grey_draw.c
 grey_parm.c
 grey_scroll.c
+#ifdef CPU_COLDFIRE
+grey_coldfire.S
+#endif
 #endif
 highscore.c
 #ifndef SIMULATOR
diff --git a/apps/plugins/lib/grey_coldfire.S b/apps/plugins/lib/grey_coldfire.S
new file mode 100755
index 0000000..cf66ca6
--- /dev/null
+++ b/apps/plugins/lib/grey_coldfire.S
@@ -0,0 +1,128 @@
+/***************************************************************************
+*             __________               __   ___.
+*   Open      \______   \ ____   ____ |  | _\_ |__   _______  ___
+*   Source     |       _//  _ \_/ ___\|  |/ /| __ \ /  _ \  \/  /
+*   Jukebox    |    |   (  <_> )  \___|    < | \_\ (  <_> > <  <
+*   Firmware   |____|_  /\____/ \___  >__|_ \|___  /\____/__/\_ \
+*                     \/            \/     \/    \/            \/
+* $Id: grey_draw.c 16080 2008-01-13 18:39:09Z amiconn $
+*
+* New greyscale framework
+* Coldfire assembler routines
+*
+* This is a generic framework to display 129 shades of grey on low-depth
+* bitmap LCDs (Archos b&w, Iriver & Ipod 4-grey) within plugins.
+*
+* Copyright (C) 2008 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"
+/* Plugins should not normally do this, but we need to check a macro, and
+ * plugin.h would confuse the assembler. */
+ 
+    .text
+    .global     _grey_line1
+    .type       _grey_line1, @function
+    
+#if (LCD_PIXELFORMAT == VERTICAL_PACKING) && (LCD_DEPTH == 2)
+    
+/****************************************************************************
+ * void _grey_line1(int width,
+ *                  unsigned char *dst,
+ *                  const unsigned char *src,
+ *                  const unsigned char *lut);
+ */
+
+_grey_line1:
+    lea.l   (-2*4, %sp), %sp
+    movem.l %d2/%a2, (%sp)
+    movem.l (2*4+4, %sp), %d2/%a0-%a2
+    clr.l   %d0
+
+    move.l  %a1, %d1
+    and.l   #1, %d1
+    beq.s   .p1_h_end
+
+    move.b  (%a1)+, %d0
+    move.b  (%d0.l, %a2), (%a0)
+    addq.l  #4, %a0
+    subq.l  #1, %d2
+.p1_h_end:
+
+    cmp.l   #2, %d2
+    blo.s   .p2_t_end
+    move.l  %a1, %d1
+    and.l   #2, %d1
+    beq.s   .p2_h_end
+    
+    move.w  (%a1)+, %d1
+    move.w  %d1, %d0
+    lsr.l   #8, %d0
+    move.b  (%d0.l, %a2), (%a0)
+    addq.l  #4, %a0
+    move.b  %d1, %d0
+    move.b  (%d0.l, %a2), (%a0)
+    addq.l  #4, %a0
+    subq.l  #2, %d2
+.p2_h_end:
+
+    subq.l  #4, %d2
+    blo.s   .p4_end
+    
+.p4_loop:
+    move.l  (%a1)+, %d1
+    swap    %d1
+    move.w  %d1, %d0
+    lsr.l   #8, %d0
+    move.b  (%d0.l, %a2), (%a0)
+    addq.l  #4, %a0
+    move.b  %d1, %d0
+    move.b  (%d0.l, %a2), (%a0)
+    addq.l  #4, %a0
+    swap    %d1
+    move.w  %d1, %d0
+    lsr.l   #8, %d0
+    move.b  (%d0.l, %a2), (%a0)
+    addq.l  #4, %a0
+    move.b  %d1, %d0
+    move.b  (%d0.l, %a2), (%a0)
+    addq.l  #4, %a0
+    subq.l  #4, %d2
+    bhs.s   .p4_loop
+    
+.p4_end:
+    addq.l  #4, %d2
+    cmp.l   #2, %d2
+    blo.s   .p2_t_end
+    
+    move.w  (%a1)+, %d1
+    move.w  %d1, %d0
+    lsr.l   #8, %d0
+    move.b  (%d0.l, %a2), (%a0)
+    addq.l  #4, %a0
+    move.b  %d1, %d0
+    move.b  (%d0.l, %a2), (%a0)
+    addq.l  #4, %a0
+    subq.l  #2, %d2
+.p2_t_end:
+
+    tst.l   %d2
+    beq.s   .p1_t_end
+
+    move.b  (%a1)+, %d0
+    move.b  (%d0.l, %a2), (%a0)
+.p1_t_end:
+
+    movem.l (%sp), %d2/%a2
+    lea.l   (2*4, %sp), %sp
+    rts
+    .size   _grey_line1, . - _grey_line1
+
+#endif
diff --git a/apps/plugins/lib/grey_draw.c b/apps/plugins/lib/grey_draw.c
index 335d6d1..6df5556 100644
--- a/apps/plugins/lib/grey_draw.c
+++ b/apps/plugins/lib/grey_draw.c
@@ -586,11 +586,17 @@ void grey_ub_clear_display(void)
 #endif
 }
 
+/* Assembler optimised helper function for copying a single line to the 
+ * greyvalue buffer. */
+void _grey_line1(int width, unsigned char *dst, const unsigned char *src,
+                 const unsigned char *lut);
+
 /* Draw a partial greyscale bitmap, canonical format */
 void grey_ub_gray_bitmap_part(const unsigned char *src, int src_x, int src_y,
                               int stride, int x, int y, int width, int height)
 {
     int yc, ye;
+    unsigned char *dst;
 
     /* nothing to draw? */
     if ((width <= 0) || (height <= 0) || (x >= _grey_info.width)
@@ -618,16 +624,22 @@ void grey_ub_gray_bitmap_part(const unsigned char *src, int src_x, int src_y,
     src += _GREY_MULUQ(stride, src_y) + src_x; /* move starting point */ 
     yc = y;
     ye = y + height;
+    dst = _grey_info.values + (x << _GREY_BSHIFT);
 
     do
     {
 #if LCD_PIXELFORMAT == HORIZONTAL_PACKING
-        int idx = _GREY_MULUQ(_grey_info.width, yc) + x;
+        int idx = _GREY_MULUQ(_grey_info.width, yc);
 #else
         int idx = _GREY_MULUQ(_grey_info.width, yc & ~_GREY_BMASK)
-                + (x << _GREY_BSHIFT) + (~yc & _GREY_BMASK);
+                + (~yc & _GREY_BMASK);
 #endif /* LCD_PIXELFORMAT */
-        unsigned char *dst_row = _grey_info.values + idx;
+
+#if (LCD_PIXELFORMAT == VERTICAL_PACKING) && (LCD_DEPTH == 2) \
+ && defined(CPU_COLDFIRE)
+        _grey_line1(width, dst + idx, src, _grey_info.gvalue);
+#else
+        unsigned char *dst_row = dst + idx;
         const unsigned char *src_row = src;
         const unsigned char *src_end = src + width;
 
@@ -637,6 +649,7 @@ void grey_ub_gray_bitmap_part(const unsigned char *src, int src_x, int src_y,
             dst_row += _GREY_BSIZE;
         }
         while (src_row < src_end);
+#endif
         
         src += stride;
     }