Changeset 2850

Timestamp:

10/20/99 00:19:39 (14 years ago)

Author:

ucacoxh

Message:

Replaced divisions in interpolation converter that were not powers
of two with a multiplication scaling factor and a power of 2
denominator. Improves performance 4-6 fold for upsampling by 3,5,6.

i.e. have sequence that looked like:

x[i] = 2 * last + 1 * new; x[i] /= 3;
x[i] = 1 * last + 2 * new; x[i] /= 3;
x[i] = new;

since 1/3 approximates to 5 / 16, now looks like:

x[i] = 10 * last + 5 * new; x[i] /= 16;
x[i] = 5 * last + 10 * new; x[i] /= 16;
x[i] = 15 * new; x[i] /= 16;

Have to scale last entry so not put jumps in amplitude. I vaguely
recall reading that power of 2 div was quicker the rest. Took a good
while to trace it though since my initial guesses were arithmetic
pipelining and memory alignment ;-)

Not sure we really care about this since the time is trivial anyway,
but gave odd data points in performance measurement for thesis. For
the record, an UltraSPARC-II 296 MHz can upsample 330000 20ms frames
from 8k to 48k in 1 second. Okay, the sound quality is not great, but
that's still a staggering amount of maths.

Location:

rat/trunk

Files:

• MODS (modified) (1 diff)
• convert_linear.c (modified) (4 diffs)

rat/trunk/MODS

@@ -899 +899 @@
           update.
         - Unrolled loops in extrapolation converter.  
+        - Replaced divisions in interpolation converter that were not powers
+          of two with a multiplication scaling factor and a power of 2 
+          denominator.  Improves performance 4-6 fold for upsampling by 3,5,6.
 
 TODO -- They're features not bugs dammit!

rat/trunk/convert_linear.c

@@ -51 +51 @@
         loop = min(src_len/channels, dst_len/(channels*l->scale));
 
+        /* On some platforms divisions by powers of 2 is way quicker */
+        /* than other divisions.  To improve interpolation perf.     */
+        /* approximate fractions which are not powers of two         */
+        /* i.e. 1 / 3 ->  5 / 16                                     */
+        /*      1 / 5 ->  6 / 32                                     */
+        /*      1 / 6 -> 11 / 64                                     */
+
         switch (l->scale) {
         case 6:
@@ -56 +63 @@
                         register int il, ic;
                         il = *last; ic = *sp;
-                        r = 5 * il + 1 * ic; r /= 6; *dp = (sample)r; dp += channels;
-                        r = 4 * il + 2 * ic; r /= 6; *dp = (sample)r; dp += channels;
-                        r = 3 * il + 3 * ic; r /= 6; *dp = (sample)r; dp += channels;
-                        r = 2 * il + 4 * ic; r /= 6; *dp = (sample)r; dp += channels;
-                        r = 1 * il + 5 * ic; r /= 6; *dp = (sample)r; dp += channels;
-                        *dp = (sample)ic; dp += channels;
+
+                        r = 55 * il + 11 * ic; r /= 64; *dp = (sample)r; dp += channels;
+                        r = 44 * il + 22 * ic; r /= 64; *dp = (sample)r; dp += channels;
+                        r = 33 * il + 33 * ic; r /= 64; *dp = (sample)r; dp += channels;
+                        r = 22 * il + 44 * ic; r /= 64; *dp = (sample)r; dp += channels;
+                        r = 11 * il + 55 * ic; r /= 64; *dp = (sample)r; dp += channels;
+                        r =           66 * ic; r /= 64; *dp = (sample)r; dp += channels;
                         last = sp;
                         sp += channels;
@@ -70 +78 @@
                         register int il, ic;
                         il = *last; ic = *sp;
-                        r = 4 * il + 1 * ic; r /= 5; *dp = (sample)r; dp += channels;
-                        r = 3 * il + 2 * ic; r /= 5; *dp = (sample)r; dp += channels;
-                        r = 2 * il + 3 * ic; r /= 5; *dp = (sample)r; dp += channels;
-                        r = 1 * il + 4 * ic; r /= 5; *dp = (sample)r; dp += channels;
-                        *dp = (sample)ic; dp += channels;
+                        r = 24 * il +  6 * ic; r /= 32; *dp = (sample)r; dp += channels;
+                        r = 18 * il + 12 * ic; r /= 32; *dp = (sample)r; dp += channels;
+                        r = 12 * il + 18 * ic; r /= 32; *dp = (sample)r; dp += channels;
+                        r =  6 * il + 24 * ic; r /= 32; *dp = (sample)r; dp += channels;
+                        r =           30 * ic; r /= 32; *dp = (sample)r; dp += channels;
                         last = sp;
                         sp  += channels;
@@ -95 +103 @@
                         register int il, ic;
                         il = *last; ic = *sp;
-                        r = 2 * il + 1 * ic; r /= 3; *dp = (sample)r; dp += channels;
-                        r = 1 * il + 2 * ic; r /= 3; *dp = (sample)r; dp += channels;
-                        *dp = (sample)ic; dp += channels;
+                        r = 10 * il +  5 * ic; r /= 16; *dp = (sample)r; dp += channels;
+                        r =  5 * il + 10 * ic; r /= 16; *dp = (sample)r; dp += channels;
+                        r =           15 * ic; r /= 16; *dp = (sample)r; dp += channels;
                         last = sp;
                         sp  += channels;