Relative performance compared to scalar C:
Neoverse N1: 2.70x
Neoverse N2: 2.06x
Neoverse V1: 2.32x
Neoverse V2: 2.16x
---
source/common/aarch64/loopfilter-prim.cpp | 59 ++++++++++++++++++++++-
1 file changed, 57 insertions(+), 2 deletions(-)
diff --git a/source/common/aarch64/loopfilter-prim.cpp
b/source/common/aarch64/loopfilter-prim.cpp
index 8769d53ec..44b5b1fcd 100644
--- a/source/common/aarch64/loopfilter-prim.cpp
+++ b/source/common/aarch64/loopfilter-prim.cpp
@@ -3,7 +3,7 @@
#define PIXEL_MIN 0
-
+using namespace X265_NS;
#if !(HIGH_BIT_DEPTH) && defined(HAVE_NEON)
#include<arm_neon.h>
@@ -255,9 +255,63 @@ static void processSaoCUB0_neon(pixel *rec, const int8_t
*offset, int ctuWidth,
}
}
-}
+void pelFilterLumaStrong_V_neon(pixel *src, intptr_t srcStep, intptr_t offset,
+ int32_t tcP, int32_t tcQ)
+{
+ X265_CHECK(offset == 1, "Offset value must be 1 for LumaStrong
Vertical\n");
+
+ src -= offset * 4;
+
+ const int16x8_t tc_vec = vcombine_s16(vdup_n_s16(tcP), vdup_n_s16(tcQ));
+ const int16x8_t neg_tc_vec = vnegq_s16(tc_vec);
+ static const uint8_t filter[3][8] =
+ {
+ { 0, 2, 1, 2, 2, 1, 1, 0 },
+ { 0, 3, 1, 2, 2, 1, 3, 0 },
+ { 0, 1, 1, 2, 2, 1, 2, 0 },
+ };
+
+ const uint8x8_t f0 = vld1_u8(filter[0]);
+ const uint8x8_t f1 = vld1_u8(filter[1]);
+ const uint8x8_t f2 = vld1_u8(filter[2]);
+
+ // -1 index means value is zero because TBL instructions
+ // zero elements that have out of range indices.
+ const uint8x8_t idx0 = { -1, 0, 1, 2, 3, 4, 5, -1 };
+ const uint8x8_t idx1 = { -1, 1, 2, 3, 4, 5, 6, -1 };
+ const uint8x8_t idx2 = { -1, 2, 3, 4, 5, 6, 7, -1 };
+ const uint8x8_t idx3 = { -1, 3, 4, 5, 6, -1, 3, -1 };
+ const uint8x8_t idx4 = { -1, 4, -1, 1, 2, 3, 4, -1 };
+
+ const int16x8_t neg_shift = { 0, -3, -2, -3, -3, -2, -3, 0 };
+
+ for (int i = 0; i < UNIT_SIZE; i++, src += srcStep)
+ {
+ uint8x8_t s = vld1_u8(src);
+ uint8x8_t s0 = vtbl1_u8(s, idx0);
+ uint8x8_t s1 = vtbl1_u8(s, idx1);
+ uint8x8_t s2 = vtbl1_u8(s, idx2);
+ uint8x8_t s3 = vtbl1_u8(s, idx3);
+ uint8x8_t s4 = vtbl1_u8(s, idx4);
+
+ uint16x8_t s34 = vaddl_u8(s3, s4);
+ uint16x8_t sum = vmlal_u8(s34, s0, f0);
+ sum = vmlal_u8(sum, s1, f1);
+ sum = vmlal_u8(sum, s2, f2);
+
+ sum = vrshlq_u16(sum, neg_shift);
+ sum = vsubw_u8(sum, s1);
+ sum = vreinterpretq_u16_s16(
+ vminq_s16(tc_vec, vmaxq_s16(neg_tc_vec,
vreinterpretq_s16_u16(sum))));
+
+ uint8x8_t d = vmovn_u16(sum);
+ d = vadd_u8(d, s);
+ vst1_u8(src, d);
+ }
+}
+} // namespace
namespace X265_NS
{
@@ -273,6 +327,7 @@ void setupLoopFilterPrimitives_neon(EncoderPrimitives &p)
p.saoCuOrgB0 = processSaoCUB0_neon;
p.sign = calSign_neon;
+ p.pelFilterLumaStrong[0] = pelFilterLumaStrong_V_neon;
}
--
2.34.1
>From 57845bbe888bff5ce2b10b13c177c9ddeb52b07c Mon Sep 17 00:00:00 2001
Message-Id: <57845bbe888bff5ce2b10b13c177c9ddeb52b07c.1739886653.git.microdaryl.rob...@arm.com>
In-Reply-To: <cover.1739886653.git.microdaryl.rob...@arm.com>
References: <cover.1739886653.git.microdaryl.rob...@arm.com>
From: Micro Daryl Robles <microdaryl.rob...@arm.com>
Date: Fri, 9 Aug 2024 03:30:05 +0100
Subject: [PATCH v3 1/5] AArch64: Add Neon implementation of
pelFilterLumaStrong_V
Relative performance compared to scalar C:
Neoverse N1: 2.70x
Neoverse N2: 2.06x
Neoverse V1: 2.32x
Neoverse V2: 2.16x
---
source/common/aarch64/loopfilter-prim.cpp | 59 ++++++++++++++++++++++-
1 file changed, 57 insertions(+), 2 deletions(-)
diff --git a/source/common/aarch64/loopfilter-prim.cpp b/source/common/aarch64/loopfilter-prim.cpp
index 8769d53ec..44b5b1fcd 100644
--- a/source/common/aarch64/loopfilter-prim.cpp
+++ b/source/common/aarch64/loopfilter-prim.cpp
@@ -3,7 +3,7 @@
#define PIXEL_MIN 0
-
+using namespace X265_NS;
#if !(HIGH_BIT_DEPTH) && defined(HAVE_NEON)
#include<arm_neon.h>
@@ -255,9 +255,63 @@ static void processSaoCUB0_neon(pixel *rec, const int8_t *offset, int ctuWidth,
}
}
-}
+void pelFilterLumaStrong_V_neon(pixel *src, intptr_t srcStep, intptr_t offset,
+ int32_t tcP, int32_t tcQ)
+{
+ X265_CHECK(offset == 1, "Offset value must be 1 for LumaStrong Vertical\n");
+
+ src -= offset * 4;
+
+ const int16x8_t tc_vec = vcombine_s16(vdup_n_s16(tcP), vdup_n_s16(tcQ));
+ const int16x8_t neg_tc_vec = vnegq_s16(tc_vec);
+ static const uint8_t filter[3][8] =
+ {
+ { 0, 2, 1, 2, 2, 1, 1, 0 },
+ { 0, 3, 1, 2, 2, 1, 3, 0 },
+ { 0, 1, 1, 2, 2, 1, 2, 0 },
+ };
+
+ const uint8x8_t f0 = vld1_u8(filter[0]);
+ const uint8x8_t f1 = vld1_u8(filter[1]);
+ const uint8x8_t f2 = vld1_u8(filter[2]);
+
+ // -1 index means value is zero because TBL instructions
+ // zero elements that have out of range indices.
+ const uint8x8_t idx0 = { -1, 0, 1, 2, 3, 4, 5, -1 };
+ const uint8x8_t idx1 = { -1, 1, 2, 3, 4, 5, 6, -1 };
+ const uint8x8_t idx2 = { -1, 2, 3, 4, 5, 6, 7, -1 };
+ const uint8x8_t idx3 = { -1, 3, 4, 5, 6, -1, 3, -1 };
+ const uint8x8_t idx4 = { -1, 4, -1, 1, 2, 3, 4, -1 };
+
+ const int16x8_t neg_shift = { 0, -3, -2, -3, -3, -2, -3, 0 };
+
+ for (int i = 0; i < UNIT_SIZE; i++, src += srcStep)
+ {
+ uint8x8_t s = vld1_u8(src);
+ uint8x8_t s0 = vtbl1_u8(s, idx0);
+ uint8x8_t s1 = vtbl1_u8(s, idx1);
+ uint8x8_t s2 = vtbl1_u8(s, idx2);
+ uint8x8_t s3 = vtbl1_u8(s, idx3);
+ uint8x8_t s4 = vtbl1_u8(s, idx4);
+
+ uint16x8_t s34 = vaddl_u8(s3, s4);
+ uint16x8_t sum = vmlal_u8(s34, s0, f0);
+ sum = vmlal_u8(sum, s1, f1);
+ sum = vmlal_u8(sum, s2, f2);
+
+ sum = vrshlq_u16(sum, neg_shift);
+ sum = vsubw_u8(sum, s1);
+ sum = vreinterpretq_u16_s16(
+ vminq_s16(tc_vec, vmaxq_s16(neg_tc_vec, vreinterpretq_s16_u16(sum))));
+
+ uint8x8_t d = vmovn_u16(sum);
+ d = vadd_u8(d, s);
+ vst1_u8(src, d);
+ }
+}
+} // namespace
namespace X265_NS
{
@@ -273,6 +327,7 @@ void setupLoopFilterPrimitives_neon(EncoderPrimitives &p)
p.saoCuOrgB0 = processSaoCUB0_neon;
p.sign = calSign_neon;
+ p.pelFilterLumaStrong[0] = pelFilterLumaStrong_V_neon;
}
--
2.34.1
_______________________________________________
x265-devel mailing list
x265-devel@videolan.org
https://mailman.videolan.org/listinfo/x265-devel
