add some notes into imdct_standalone.c

diff --git a/media/audio/mp3/imdct36_standalone.c b/media/audio/mp3/imdct36_standalone.c
index 04cc9838f73c631beea05332208ad983d3a9a653..fe275f211b1d6ddeb1ac8585887ab2f93d719784 100644 (file)
--- a/media/audio/mp3/imdct36_standalone.c
+++ b/media/audio/mp3/imdct36_standalone.c
@@ -86,11 +86,35 @@ void imdct36(float *out, float *buf, float *in, const float *win)
     float t0, t1, t2, t3, s0, s1, s2, s3;
     float tmp[18], *tmp1, *in1;
 
+    /* straight sv.fadds/mr for this one, should do a pascal's triangle
+       (prefix sum)
+        sv.fadds/mrr 10.v, 11.v, 10.v
+     */
     for (i = 17; i >= 1; i--)
         in[i] += in[i-1];
+
+    /* this can use mrr mode with predicate 0b010101010101,
+       it will do the job, something like:
+        li r30, 0b010101010101
+        setvl 16
+        sv.fadds/mrr/m=r30 10.v, 12.v, 10.v
+        which will issue adds *in reverse* order
+              fadds 24, 26, 24  - not predicated (active)
+              fadds 23, 25, 23  - predicated (masked out)
+              fadds 22, 24, 22  - not predicated (active)
+              fadds 21, 23, 21  - predicated (masked out)
+              ...
+        should result in pascal's triangle for this one but
+        skipping every other element
+    */
     for (i = 17; i >= 3; i -= 2)
         in[i] += in[i-2];
 
+    /* all of these are independent, should be possible to do with
+        setvl 2.  probably possible to break these down into some
+        sort of much smaller loop but the hand-coded nature of the
+        algorithm makes it difficult to spot the pattern.
+    */
     for (j = 0; j < 2; j++) {
         tmp1 = tmp + j;
         in1 = in + j;
@@ -122,6 +146,9 @@ void imdct36(float *out, float *buf, float *in, const float *win)
         tmp1[ 8] = t3 - t1 - t0;
     }
 
+    /* this can be done setvl 5 and with a predicate 0b01111 for the
+       items marked "if predicate"
+     */
     i = 0;
     for (j = 0; j < 5; j++) {
         predicate = (j != 4);