From 6d39f2624ca780432d491cb9ec9dc9576e26beca Mon Sep 17 00:00:00 2001
From: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
Date: Sat, 22 Mar 2025 07:53:34 +0100
Subject: [PATCH 80/81] avcodec/ppc/mpegvideo_altivec: Don't process coeffs as
scalars
block_last_index and nCoeffs is an optimization designed
to avoid processing unnecessarily many coefficients; yet
it would be legal to always process all coefficients
(all coefficients beyond nCoeffs are zero anyway and
zeros are always unquantized to zeros). Therefore
one does not need a scalar tail.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
---
libavcodec/ppc/mpegvideo_altivec.c | 21 ++-------------------
1 file changed, 2 insertions(+), 19 deletions(-)
diff --git a/libavcodec/ppc/mpegvideo_altivec.c b/libavcodec/ppc/mpegvideo_altivec.c
index 7e3db6f76a..a455cadc4b 100644
--- a/libavcodec/ppc/mpegvideo_altivec.c
+++ b/libavcodec/ppc/mpegvideo_altivec.c
@@ -42,7 +42,7 @@
static void dct_unquantize_h263_altivec(MpegEncContext *s,
int16_t *block, int n, int qscale)
{
- int i, level, qmul, qadd;
+ int i, qmul, qadd;
int nCoeffs;
qadd = (qscale - 1) | 1;
@@ -74,7 +74,6 @@ static void dct_unquantize_h263_altivec(MpegEncContext *s,
register vector signed short blockv, qmulv, qaddv, nqaddv, temp1;
register vector bool short blockv_null, blockv_neg;
register short backup_0 = block[0];
- register int j = 0;
qmulv = vec_splat((vec_s16)vec_lde(0, &qmul8), 0);
qaddv = vec_splat((vec_s16)vec_lde(0, &qadd8), 0);
@@ -82,7 +81,7 @@ static void dct_unquantize_h263_altivec(MpegEncContext *s,
// vectorize all the 16 bytes-aligned blocks
// of 8 elements
- for(; (j + 7) <= nCoeffs ; j+=8) {
+ for (register int j = 0; j <= nCoeffs ; j += 8) {
blockv = vec_ld(j << 1, block);
blockv_neg = vec_cmplt(blockv, vczero);
blockv_null = vec_cmpeq(blockv, vczero);
@@ -95,22 +94,6 @@ static void dct_unquantize_h263_altivec(MpegEncContext *s,
vec_st(blockv, j << 1, block);
}
- // if nCoeffs isn't a multiple of 8, finish the job
- // using good old scalar units.
- // (we could do it using a truncated vector,
- // but I'm not sure it's worth the hassle)
- for(; j <= nCoeffs ; j++) {
- level = block[j];
- if (level) {
- if (level < 0) {
- level = level * qmul - qadd;
- } else {
- level = level * qmul + qadd;
- }
- block[j] = level;
- }
- }
-
if (i == 1) {
// cheat. this avoid special-casing the first iteration
block[0] = backup_0;
--
2.45.2