Micro optimizations for the amp filter.

Use memcpy in the special case amp==0 (no amplification) and optimize the code in
the performance-critical loop. Intrestingly, using the likely()/unlikely() macros made the
code slower.

Results (three runs on identical input data on a 32bit x86 machine under Linux, gcc-4.4.0):

old with --amp 3:
        0m0.776s 0m0.790s 0m0.812s, avg: 792
new with --amp 3:
        0m0.456s 0m0.492s 0m0.477s, avg: 475
speedup: 1.67

old with --amp 0:
        0m0.791s 0m0.808s 0m0.810s, avg: 803
new with --amp 0:
        0m0.100s 0m0.103s 0m0.094s, avg: 99
speedup: 8.1

diff --git a/amp_filter.c b/amp_filter.c
index 63b1b9a618bcb561e9d73f2baf215dcea2e75221..e180db437c6a26fe8c978fefccdd51841cd97fd0 100644 (file)
--- a/amp_filter.c
+++ b/amp_filter.c
@@ -32,19 +32,29 @@ struct private_amp_data {
 
 static ssize_t amp_convert(char *inbuf, size_t inbuf_len, struct filter_node *fn)
 {
-       size_t i, length = PARA_MIN((inbuf_len / 2) * 2,
-               (fn->bufsize - fn->loaded) / 2 * 2);
+       size_t i, length = PARA_MIN((inbuf_len / 2),
+               (fn->bufsize - fn->loaded) / 2);
        struct private_amp_data *pad = fn->private_data;
        int16_t *ip = (int16_t *)inbuf, *op = (int16_t *)(fn->buf + fn->loaded);
+       int factor = 64 + pad->amp;
 
        if (!length)
                return 0;
-       for (i = 0; i < length / 2; i++) {
-               int x = (PARA_ABS(*ip) * (64 + pad->amp)) >> 6;
-               *op++ = *ip++ > 0? PARA_MIN(x, 32767) : PARA_MAX(-x, -32768);
+
+       if (pad->amp == 0) {
+               memcpy(op, ip, length * 2);
+               goto out;
+       }
+       for (i = 0; i < length; i++) {
+               int x = (ip[i] * factor) >> 6;
+
+               op[i] = x;
+               if (op[i] != x)
+                       op[i] = (x >= 32768)? 32767 : -32768;
        }
-       fn->loaded += length;
-       return length;
+out:
+       fn->loaded += length * 2;
+       return length * 2;
 }
 
 static void amp_close(struct filter_node *fn)