Unrolling provides 9-14% uplift on Arm Neoverse server platforms.
---
source/common/loopfilter.cpp | 46 +++++++++++++++++++++++++++++++++++-
1 file changed, 45 insertions(+), 1 deletion(-)
diff --git a/source/common/loopfilter.cpp b/source/common/loopfilter.cpp
index f4cd65389..3aad542b8 100644
--- a/source/common/loopfilter.cpp
+++ b/source/common/loopfilter.cpp
@@ -172,6 +172,50 @@ static void pelFilterChroma_c(pixel* src, intptr_t
srcStep, intptr_t offset, int
src[0] = x265_clip(m4 - (delta & maskQ));
}
}
+
+void pelFilterChroma_V_c(pixel *src, intptr_t srcStep, intptr_t offset,
int32_t tc,
+ int32_t maskP, int32_t maskQ)
+{
+ assert(offset == 1);
+ (void)offset;
+
+ int16_t m2 = (int16_t)src[0 * srcStep - 2];
+ int16_t m3 = (int16_t)src[0 * srcStep - 1];
+ int16_t m4 = (int16_t)src[0 * srcStep + 0];
+ int16_t m5 = (int16_t)src[0 * srcStep + 1];
+
+ int32_t delta = x265_clip3(-tc, tc, ((m4 - m3) * 4 + m2 - m5 + 4) >> 3);
+ src[0 * srcStep - 1] = x265_clip(m3 + (delta & maskP));
+ src[0 * srcStep + 0] = x265_clip(m4 - (delta & maskQ));
+
+ m2 = (int16_t)src[1 * srcStep - 2];
+ m3 = (int16_t)src[1 * srcStep - 1];
+ m4 = (int16_t)src[1 * srcStep + 0];
+ m5 = (int16_t)src[1 * srcStep + 1];
+
+ delta = x265_clip3(-tc, tc, ((m4 - m3) * 4 + m2 - m5 + 4) >> 3);
+ src[1 * srcStep - 1] = x265_clip(m3 + (delta & maskP));
+ src[1 * srcStep + 0] = x265_clip(m4 - (delta & maskQ));
+
+ m2 = (int16_t)src[2 * srcStep - 2];
+ m3 = (int16_t)src[2 * srcStep - 1];
+ m4 = (int16_t)src[2 * srcStep + 0];
+ m5 = (int16_t)src[2 * srcStep + 1];
+
+ delta = x265_clip3(-tc, tc, ((m4 - m3) * 4 + m2 - m5 + 4) >> 3);
+ src[2 * srcStep - 1] = x265_clip(m3 + (delta & maskP));
+ src[2 * srcStep + 0] = x265_clip(m4 - (delta & maskQ));
+
+ m2 = (int16_t)src[3 * srcStep - 2];
+ m3 = (int16_t)src[3 * srcStep - 1];
+ m4 = (int16_t)src[3 * srcStep + 0];
+ m5 = (int16_t)src[3 * srcStep + 1];
+
+ delta = x265_clip3(-tc, tc, ((m4 - m3) * 4 + m2 - m5 + 4) >> 3);
+ src[3 * srcStep - 1] = x265_clip(m3 + (delta & maskP));
+ src[3 * srcStep + 0] = x265_clip(m4 - (delta & maskQ));
+}
+
}
namespace X265_NS {
@@ -190,7 +234,7 @@ void setupLoopFilterPrimitives_c(EncoderPrimitives &p)
// C code is same for EDGE_VER and EDGE_HOR only asm code is different
p.pelFilterLumaStrong[0] = pelFilterLumaStrong_c;
p.pelFilterLumaStrong[1] = pelFilterLumaStrong_c;
- p.pelFilterChroma[0] = pelFilterChroma_c;
+ p.pelFilterChroma[0] = pelFilterChroma_V_c;
p.pelFilterChroma[1] = pelFilterChroma_c;
}
}
--
2.34.1
>From a5042902b2d65a0d0328cd700d6112f00798d298 Mon Sep 17 00:00:00 2001
Message-Id: <a5042902b2d65a0d0328cd700d6112f00798d298.1739282617.git.microdaryl.rob...@arm.com>
In-Reply-To: <[email protected]>
References: <[email protected]>
From: Micro Daryl Robles <[email protected]>
Date: Mon, 19 Aug 2024 14:45:07 +0100
Subject: [PATCH 4/5] Unroll C implementation of pelFilterChroma_V
Unrolling provides 9-14% uplift on Arm Neoverse server platforms.
---
source/common/loopfilter.cpp | 46 +++++++++++++++++++++++++++++++++++-
1 file changed, 45 insertions(+), 1 deletion(-)
diff --git a/source/common/loopfilter.cpp b/source/common/loopfilter.cpp
index f4cd65389..3aad542b8 100644
--- a/source/common/loopfilter.cpp
+++ b/source/common/loopfilter.cpp
@@ -172,6 +172,50 @@ static void pelFilterChroma_c(pixel* src, intptr_t srcStep, intptr_t offset, int
src[0] = x265_clip(m4 - (delta & maskQ));
}
}
+
+void pelFilterChroma_V_c(pixel *src, intptr_t srcStep, intptr_t offset, int32_t tc,
+ int32_t maskP, int32_t maskQ)
+{
+ assert(offset == 1);
+ (void)offset;
+
+ int16_t m2 = (int16_t)src[0 * srcStep - 2];
+ int16_t m3 = (int16_t)src[0 * srcStep - 1];
+ int16_t m4 = (int16_t)src[0 * srcStep + 0];
+ int16_t m5 = (int16_t)src[0 * srcStep + 1];
+
+ int32_t delta = x265_clip3(-tc, tc, ((m4 - m3) * 4 + m2 - m5 + 4) >> 3);
+ src[0 * srcStep - 1] = x265_clip(m3 + (delta & maskP));
+ src[0 * srcStep + 0] = x265_clip(m4 - (delta & maskQ));
+
+ m2 = (int16_t)src[1 * srcStep - 2];
+ m3 = (int16_t)src[1 * srcStep - 1];
+ m4 = (int16_t)src[1 * srcStep + 0];
+ m5 = (int16_t)src[1 * srcStep + 1];
+
+ delta = x265_clip3(-tc, tc, ((m4 - m3) * 4 + m2 - m5 + 4) >> 3);
+ src[1 * srcStep - 1] = x265_clip(m3 + (delta & maskP));
+ src[1 * srcStep + 0] = x265_clip(m4 - (delta & maskQ));
+
+ m2 = (int16_t)src[2 * srcStep - 2];
+ m3 = (int16_t)src[2 * srcStep - 1];
+ m4 = (int16_t)src[2 * srcStep + 0];
+ m5 = (int16_t)src[2 * srcStep + 1];
+
+ delta = x265_clip3(-tc, tc, ((m4 - m3) * 4 + m2 - m5 + 4) >> 3);
+ src[2 * srcStep - 1] = x265_clip(m3 + (delta & maskP));
+ src[2 * srcStep + 0] = x265_clip(m4 - (delta & maskQ));
+
+ m2 = (int16_t)src[3 * srcStep - 2];
+ m3 = (int16_t)src[3 * srcStep - 1];
+ m4 = (int16_t)src[3 * srcStep + 0];
+ m5 = (int16_t)src[3 * srcStep + 1];
+
+ delta = x265_clip3(-tc, tc, ((m4 - m3) * 4 + m2 - m5 + 4) >> 3);
+ src[3 * srcStep - 1] = x265_clip(m3 + (delta & maskP));
+ src[3 * srcStep + 0] = x265_clip(m4 - (delta & maskQ));
+}
+
}
namespace X265_NS {
@@ -190,7 +234,7 @@ void setupLoopFilterPrimitives_c(EncoderPrimitives &p)
// C code is same for EDGE_VER and EDGE_HOR only asm code is different
p.pelFilterLumaStrong[0] = pelFilterLumaStrong_c;
p.pelFilterLumaStrong[1] = pelFilterLumaStrong_c;
- p.pelFilterChroma[0] = pelFilterChroma_c;
+ p.pelFilterChroma[0] = pelFilterChroma_V_c;
p.pelFilterChroma[1] = pelFilterChroma_c;
}
}
--
2.34.1
_______________________________________________
x265-devel mailing list
[email protected]
https://mailman.videolan.org/listinfo/x265-devel
