Optimizations for doom: coldfire asm drawspan routine = not much, fixed point multiply changes = not much, H300 asm lcd update = some, IRAM sound updates and simplifications = more

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@9747 a1c6a512-1295-4272-9138-f99709370657

4 files changed, 107 insertions, 50 deletions

diff --git a/apps/plugins/doom/i_sound.c b/apps/plugins/doom/i_sound.c
index 7579ada5b5..271444527f 100644
--- a/apps/plugins/doom/i_sound.c
+++ b/apps/plugins/doom/i_sound.c
@@ -48,11 +48,11 @@
 //  mixing buffer, and the samplerate of the raw data.
 
 // Needed for calling the actual sound output.
-#define SAMPLECOUNT  512
+#define SAMPLECOUNT     512
 
-#define NUM_CHANNELS  16
+#define NUM_CHANNELS    16
 // It is 2 for 16bit, and 2 for two channels.
-#define BUFMUL                  4
+#define BUFMUL          2
 #define MIXBUFFERSIZE  (SAMPLECOUNT*BUFMUL)
 
 #if (CONFIG_KEYPAD == IPOD_3G_PAD) || (CONFIG_KEYPAD == IPOD_4G_PAD)
@@ -66,7 +66,7 @@
 //  Basically, samples from all active internal channels
 //  are modifed and added, and stored in the buffer
 //  that is submitted to the audio device.
-signed short *mixbuffer=NULL;
+signed short mixbuffer[MIXBUFFERSIZE] IBSS_ATTR;
 
 typedef struct {
    // SFX id of the playing sound effect.
@@ -91,7 +91,7 @@ typedef struct {
    int *rightvol_lookup;
 } channel_info_t;
 
-channel_info_t channelinfo[NUM_CHANNELS];
+channel_info_t channelinfo[NUM_CHANNELS] IBSS_ATTR;
 
 int *vol_lookup; // Volume lookups.
 
@@ -355,13 +355,6 @@ int I_SoundIsPlaying(int handle)
 // This function currently supports only 16bit.
 //
 
-bool swap=0;
-bool lastswap=1;
-   // Pointers in global mixbuffer, left, right, end.
-   signed short*  leftout;
-   signed short*  rightout;
-   signed short*  leftend;
-
 void I_UpdateSound( void )
 {
    // Mix current sound data.
@@ -370,25 +363,26 @@ void I_UpdateSound( void )
    register int  dl;
    register int  dr;
 
+   // Pointers in global mixbuffer, left, right, end.
+   signed short*  leftout;
+   signed short*  rightout;
+   signed short*  leftend;
+
    // Step in mixbuffer, left and right, thus two.
    int    step;
 
    // Mixing channel index.
    int    chan;
 
-   if(lastswap==swap)
-      return;
-   lastswap=swap;
-
    // Left and right channel
    //  are in global mixbuffer, alternating.
-   leftout = (swap ? mixbuffer : mixbuffer + SAMPLECOUNT*2);
-   rightout = (swap ? mixbuffer : mixbuffer + SAMPLECOUNT*2)+1;
+   leftout = mixbuffer;
+   rightout = mixbuffer +1;
    step = 2;
 
    // Determine end, for left channel only
    //  (right channel is implicit).
-   leftend = (swap ? mixbuffer : mixbuffer + SAMPLECOUNT*2) + SAMPLECOUNT*step;
+   leftend = mixbuffer + SAMPLECOUNT*step;
 
    // Mix sounds into the mixing buffer.
    // Loop over step*SAMPLECOUNT,
@@ -467,15 +461,10 @@ void I_UpdateSound( void )
 
 void get_more(unsigned char** start, size_t* size)
 {
-   // This code works fine, the only problem is that doom runs slower then the sound
-   // updates (sometimes).  This code forces the update if the sound hasn't been
-   // remixed.
-   if(lastswap!=swap)
-      I_UpdateSound(); // Force sound update (We don't want stutters)
+   I_UpdateSound(); // Force sound update
 
-   *start = (unsigned char*)((swap ? mixbuffer : mixbuffer + SAMPLECOUNT*2));
+   *start = (unsigned char*)(mixbuffer);
    *size = SAMPLECOUNT*2*sizeof(short);
-   swap=!swap;
 }
 
 
@@ -520,9 +509,6 @@ void I_InitSound()
 
    printf( " pre-cached all sound data\n");
 
-   if(mixbuffer==NULL)
-      mixbuffer=malloc(sizeof(short)*MIXBUFFERSIZE);
-
    // Now initialize mixbuffer with zero.
    for ( i = 0; i< MIXBUFFERSIZE; i++ )
       mixbuffer[i] = 0;
diff --git a/apps/plugins/doom/i_video.c b/apps/plugins/doom/i_video.c
index a4db5f2672..db6adfa75e 100644
--- a/apps/plugins/doom/i_video.c
+++ b/apps/plugins/doom/i_video.c
@@ -16,7 +16,10 @@
  * GNU General Public License for more details.
  *
  * $Log$
- * Revision 1.15  2006/04/16 23:14:04  kkurbjun
+ * Revision 1.16  2006/04/20 19:39:56  kkurbjun
+ * Optimizations for doom: coldfire asm drawspan routine = not much, fixed point multiply changes = not much, H300 asm lcd update = some, IRAM sound updates and simplifications = more
+ *
+ * Revision 1.15  2006-04-16 23:14:04  kkurbjun
  * Fix run so that it stays enabled across level loads.  Removed some unused code and added some back in for hopeful future use.
  *
  * Revision 1.14  2006-04-15 22:08:36  kkurbjun
@@ -359,10 +362,43 @@ static void I_UploadNewPalette(int pal)
 void I_FinishUpdate (void)
 {
 #if (CONFIG_LCD == LCD_H300) && !defined(SIMULATOR)
-   /*
-      Lookup tables are no longer needed (H300 specific, decreases timedemo
-      by about 500 tics)
-   */
+
+#if 1
+   /* ASM screen update (drops 600 tics (100 asm)) */
+   asm (
+      "move.w #33,(%[LCD])    \n" /* Setup the LCD controller */
+      "clr.w (%[LCD2])        \n"
+      "move.w #34,(%[LCD])    \n" /* End LCD controller setup */
+      "move.l #220,%%d2       \n"
+      "move.l #176,%%d3       \n"
+      "clr.l %%d1             \n"
+      "widthloop:             \n"
+      "move.b (%[screenptr])+, %%d1             \n" /* Unrolled by 5 */
+      "move.w (%[palette], %%d1.l:2), (%[LCD2]) \n"
+      "move.b (%[screenptr])+, %%d1             \n"
+      "move.w (%[palette], %%d1.l:2), (%[LCD2]) \n"
+      "move.b (%[screenptr])+, %%d1             \n"
+      "move.w (%[palette], %%d1.l:2), (%[LCD2]) \n"
+      "move.b (%[screenptr])+, %%d1             \n"
+      "move.w (%[palette], %%d1.l:2), (%[LCD2]) \n"
+      "move.b (%[screenptr])+, %%d1             \n"
+      "move.w (%[palette], %%d1.l:2), (%[LCD2]) \n"
+      "subq.l #5,%%d2         \n"
+      "bne widthloop          \n"
+      "move.w #220,%%d2       \n"
+      "subq.l #1,%%d3         \n"
+      "bne widthloop          \n"
+   : /* outputs */
+   : /* inputs */
+      [screenptr] "a" (d_screens[0]),
+      [palette]   "a" (palette),
+      [LCD]       "a" (0xf0000000),
+      [LCD2]      "a" (0xf0000002)
+   : /* clobbers */
+      "d1", "d2", "d3"
+   );
+#else
+   /* C version of above (drops 500 tics) */
 
    // Start the write
    *(volatile unsigned short *) 0xf0000000 = 0x21; // register
@@ -383,6 +419,8 @@ void I_FinishUpdate (void)
       wcnt=0;
       hcnt++;
    }
+#endif
+
 #else
    unsigned char paletteIndex;
    int x, y;
diff --git a/apps/plugins/doom/m_fixed.h b/apps/plugins/doom/m_fixed.h
index 3c922e8f50..e29933befd 100644
--- a/apps/plugins/doom/m_fixed.h
+++ b/apps/plugins/doom/m_fixed.h
@@ -47,15 +47,15 @@ inline static int FixedMul( int a, int b )
 #if defined(CPU_COLDFIRE) && !defined(SIMULATOR)
    // Code contributed by Thom Johansen
    register int result;
-   asm volatile (
+   asm (
       "mac.l   %[x],%[y],%%acc0  \n"   /* multiply */
       "move.l  %[y],%%d2         \n"
       "mulu.l  %[x],%%d2         \n"   /* get lower half, avoid emac stall */
       "movclr.l %%acc0,%[result] \n"   /* get higher half */
-      "moveq.l #15,%%d1          \n"
-      "asl.l   %%d1,%[result]    \n"   /* hi <<= 15, plus one free */
-      "moveq.l #16,%%d1          \n"
-      "lsr.l   %%d1,%%d2         \n"   /* (unsigned)lo >>= 16 */
+      "asl.l   #8,%[result]      \n"   /* hi <<= 15, plus one free */
+      "asl.l   #7,%[result]      \n"   /* hi <<= 15, plus one free */
+      "lsr.l   #8,%%d2           \n"   /* (unsigned)lo >>= 16 */
+      "lsr.l   #8,%%d2           \n"   /* (unsigned)lo >>= 16 */
       "or.l    %%d2 ,%[result]   \n"   /* combine result */
    : /* outputs */
       [result]"=&d"(result)
@@ -63,7 +63,7 @@ inline static int FixedMul( int a, int b )
       [x] "d" (a),
       [y] "d" (b)
    : /* clobbers */
-      "d1", "d2"
+      "d2"
    );
    return result;
 #else
diff --git a/apps/plugins/doom/r_draw.c b/apps/plugins/doom/r_draw.c
index a6bc21e420..5f45323a36 100644
--- a/apps/plugins/doom/r_draw.c
+++ b/apps/plugins/doom/r_draw.c
@@ -526,16 +526,48 @@ byte *ds_source IBSS_ATTR;
 
 void R_DrawSpan (void)
 {
-   register unsigned count,xfrac = ds_xfrac,yfrac = ds_yfrac;
-
-   byte *source;
-   byte *colormap;
-   byte *dest;
-
-   source = ds_source;
-   colormap = ds_colormap;
-   dest = topleft + ds_y*SCREENWIDTH + ds_x1;
-   count = ds_x2 - ds_x1 + 1;
+#if defined(CPU_COLDFIRE) && !defined(SIMULATOR)
+   // only slightly faster
+   asm volatile (
+      "tst %[count]                             \n"
+      "beq endspanloop                          \n"
+      "clr.l %%d4                               \n"
+      "spanloop:                                \n"
+      "move.l %[xfrac], %%d1                    \n"
+      "move.l %[yfrac], %%d2                    \n"
+      "lsr.l #8,%%d1                            \n"
+      "lsr.l #8,%%d2                            \n"
+      "lsr.l #8,%%d1                            \n"
+      "lsr.l #2,%%d2                            \n"
+      "and.l #63,%%d1                           \n"
+      "and.l #4032,%%d2                         \n"
+      "or.l %%d2, %%d1                          \n"
+      "move.b (%[source], %%d1), %%d4           \n"
+      "add.l %[ds_xstep], %[xfrac]              \n"
+      "add.l %[ds_ystep], %[yfrac]              \n"
+      "move.b (%[colormap],%%d4.l), (%[dest])+  \n"
+      "subq.l #1, %[count]                      \n"
+      "bne spanloop                             \n"
+      "endspanloop:                             \n"
+   : /* outputs */
+   : /* inputs */
+      [count] "d" (ds_x2-ds_x1+1),
+      [xfrac] "d" (ds_xfrac),
+      [yfrac] "d" (ds_yfrac),
+      [source] "a" (ds_source),
+      [colormap] "a" (ds_colormap),
+      [dest] "a" (topleft+ds_y*SCREENWIDTH +ds_x1),
+      [ds_xstep] "d" (ds_xstep),
+      [ds_ystep] "d" (ds_ystep)
+   : /* clobbers */
+      "d1", "d2", "d4"
+   );
+#else
+   register unsigned count = ds_x2 - ds_x1 + 1,xfrac = ds_xfrac,yfrac = ds_yfrac;
+
+   register byte *source = ds_source;
+   register byte *colormap = ds_colormap;
+   register byte *dest = topleft + ds_y*SCREENWIDTH + ds_x1;
 
    while (count)
    {
@@ -550,6 +582,7 @@ void R_DrawSpan (void)
       *dest++ = colormap[source[spot]];
       count--;
    }
+#endif
 }
 
 //