PR #21189 opened by Rémi Denis-Courmont (Courmisch)
URL: https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/21189
Patch URL: https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/21189.patch

See https://godbolt.org/z/6d9venez7


From aa6c8d167f43c17fd7a324125f7bf668aec6d08a Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?R=C3=A9mi=20Denis-Courmont?= <[email protected]>
Date: Sun, 14 Dec 2025 11:06:13 +0200
Subject: [PATCH 1/2] lavc/mathops: simplify mid_pred()

This reduces mid_pred() (i.e. median of 3) down to the minimum:
- 3 comparisons and 4 conditional moves, or
- 4 min/max,
whilst eliminating all branches.

The same algorithm is already implemented via inline assembler for some
architectures, but notably not Arm64 and RVA22. Besides, using C code
allows the compiler to schedule instruction properly.
---
 libavcodec/mathops.h | 29 +++++++++++++++--------------
 1 file changed, 15 insertions(+), 14 deletions(-)

diff --git a/libavcodec/mathops.h b/libavcodec/mathops.h
index aa0bdfe956..4411d138b4 100644
--- a/libavcodec/mathops.h
+++ b/libavcodec/mathops.h
@@ -93,23 +93,24 @@ static av_always_inline unsigned UMULH(unsigned a, unsigned 
b){
 #endif
 
 /* median of 3 */
-#ifndef mid_pred
-#define mid_pred mid_pred
-static inline av_const int mid_pred(int a, int b, int c)
+static inline av_const int median3_c(int a, int b, int c)
 {
-    if(a>b){
-        if(c>b){
-            if(c>a) b=a;
-            else    b=c;
-        }
-    }else{
-        if(b>c){
-            if(c>a) b=c;
-            else    b=a;
-        }
+    int max2, min2, m;
+
+    if (a >= b) {
+        max2 = a;
+        min2 = b;
+    } else {
+        max2 = b;
+        min2 = a;
     }
-    return b;
+    m = (c >= max2) ? max2 : c;
+
+    return (m >= min2) ? m : min2;
 }
+
+#ifndef mid_pred
+#define mid_pred median3_c
 #endif
 
 #ifndef median4
-- 
2.49.1


From d7fa8218ddee4aa0c6c92dd39c1859ad6dcef325 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?R=C3=A9mi=20Denis-Courmont?= <[email protected]>
Date: Sun, 14 Dec 2025 11:19:21 +0200
Subject: [PATCH 2/2] lavc/mathops: remove bespoke Arm mid_pred()

The C codegen is as good if not slightly better than the assembler at
this point.
---
 libavcodec/arm/mathops.h | 22 ----------------------
 1 file changed, 22 deletions(-)

diff --git a/libavcodec/arm/mathops.h b/libavcodec/arm/mathops.h
index dc57c5571c..9d0385db6c 100644
--- a/libavcodec/arm/mathops.h
+++ b/libavcodec/arm/mathops.h
@@ -81,28 +81,6 @@ static inline av_const int MUL16(int ra, int rb)
 
 #endif
 
-#define mid_pred mid_pred
-static inline av_const int mid_pred(int a, int b, int c)
-{
-    int m;
-    __asm__ (
-        "mov   %0, %2  \n\t"
-        "cmp   %1, %2  \n\t"
-        "itt   gt      \n\t"
-        "movgt %0, %1  \n\t"
-        "movgt %1, %2  \n\t"
-        "cmp   %1, %3  \n\t"
-        "it    le      \n\t"
-        "movle %1, %3  \n\t"
-        "cmp   %0, %1  \n\t"
-        "it    gt      \n\t"
-        "movgt %0, %1  \n\t"
-        : "=&r"(m), "+r"(a)
-        : "r"(b), "r"(c)
-        : "cc");
-    return m;
-}
-
 #endif /* HAVE_INLINE_ASM */
 
 #endif /* AVCODEC_ARM_MATHOPS_H */
-- 
2.49.1

_______________________________________________
ffmpeg-devel mailing list -- [email protected]
To unsubscribe send an email to [email protected]

Reply via email to