Relative performance compared to scalar C:
Neoverse N1: 3.67x
Neoverse N2: 2.57x
Neoverse V1: 2.31x
Neoverse V2: 2.35x
---
source/common/aarch64/loopfilter-prim.cpp | 75 +++++++++++++++++++++++
1 file changed, 75 insertions(+)
diff --git a/source/common/aarch64/loopfilter-prim.cpp
b/source/common/aarch64/loopfilter-prim.cpp
index 9242c859e..7fef25cbb 100644
--- a/source/common/aarch64/loopfilter-prim.cpp
+++ b/source/common/aarch64/loopfilter-prim.cpp
@@ -1,5 +1,6 @@
#include "common.h"
#include "loopfilter-prim.h"
+#include "mem-neon.h"
#define PIXEL_MIN 0
@@ -311,6 +312,79 @@ void pelFilterLumaStrong_V_neon(pixel *src, intptr_t
srcStep, intptr_t offset,
}
}
+void pelFilterLumaStrong_H_neon(pixel *src, intptr_t srcStep, intptr_t offset,
+ int32_t tcP, int32_t tcQ)
+{
+ assert(UNIT_SIZE == 4 && srcStep == 1);
+ (void)srcStep;
+
+ 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);
+
+ uint8x8_t m0 = vld1_u8(src - 4 * offset);
+ uint8x8_t m1 = vld1_u8(src - 3 * offset);
+ uint8x8_t m2 = vld1_u8(src - 2 * offset);
+ uint8x8_t m3 = vld1_u8(src - 1 * offset);
+ uint8x8_t m4 = vld1_u8(src - 0 * offset);
+ uint8x8_t m5 = vld1_u8(src + 1 * offset);
+ uint8x8_t m6 = vld1_u8(src + 2 * offset);
+ uint8x8_t m7 = vld1_u8(src + 3 * offset);
+
+ uint8x8_t m12 =
+ vreinterpret_u8_u32(vzip1_u32(vreinterpret_u32_u8(m1),
vreinterpret_u32_u8(m2)));
+ uint8x8_t m23 =
+ vreinterpret_u8_u32(vzip1_u32(vreinterpret_u32_u8(m2),
vreinterpret_u32_u8(m3)));
+ uint8x8_t m34 =
+ vreinterpret_u8_u32(vzip1_u32(vreinterpret_u32_u8(m3),
vreinterpret_u32_u8(m4)));
+ uint8x8_t m45 =
+ vreinterpret_u8_u32(vzip1_u32(vreinterpret_u32_u8(m4),
vreinterpret_u32_u8(m5)));
+ uint8x8_t m56 =
+ vreinterpret_u8_u32(vzip1_u32(vreinterpret_u32_u8(m5),
vreinterpret_u32_u8(m6)));
+
+ // src[-1 * offset], src[0 * offset]
+ uint16x8_t p0 = vaddl_u8(m23, m34);
+ p0 = vaddw_u8(p0, m45);
+ uint16x8_t t0 = vshlq_n_u16(p0, 1);
+ uint16x8_t t1 = vaddl_u8(m12, m56);
+ uint16x8_t t01 = vaddq_u16(t0, t1);
+ t01 = vrshrq_n_u16(t01, 3);
+ t01 = vsubw_u8(t01, m34);
+ t01 = vreinterpretq_u16_s16(
+ vminq_s16(tc_vec, vmaxq_s16(neg_tc_vec, vreinterpretq_s16_u16(t01))));
+ uint8x8_t d01 = vmovn_u16(t01);
+ d01 = vadd_u8(d01, m34);
+ store_u8x4_strided_xN<2>(&src[-1 * offset], 1 * offset, &d01);
+
+ uint8x8_t m16 =
+ vreinterpret_u8_u32(vzip1_u32(vreinterpret_u32_u8(m1),
vreinterpret_u32_u8(m6)));
+ uint8x8_t m25 =
+ vreinterpret_u8_u32(vzip1_u32(vreinterpret_u32_u8(m2),
vreinterpret_u32_u8(m5)));
+
+ // src[-2 * offset], src[1 * offset]
+ uint16x8_t p1 = vaddw_u8(p0, m16);
+ uint16x8_t t23 = vrshrq_n_u16(p1, 2);
+ t23 = vsubw_u8(t23, m25);
+ t23 = vreinterpretq_u16_s16(
+ vminq_s16(tc_vec, vmaxq_s16(neg_tc_vec, vreinterpretq_s16_u16(t23))));
+ uint8x8_t d23 = vmovn_u16(t23);
+ d23 = vadd_u8(d23, m25);
+ store_u8x4_strided_xN<2>(&src[-2 * offset], 3 * offset, &d23);
+
+ uint8x8_t m07 =
+ vreinterpret_u8_u32(vzip1_u32(vreinterpret_u32_u8(m0),
vreinterpret_u32_u8(m7)));
+
+ // src[-3 * offset], src[2 * offset]
+ uint16x8_t p2 = vaddl_u8(m07, m16);
+ uint16x8_t t45 = vmlaq_n_u16(p1, p2, 2);
+ t45 = vrshrq_n_u16(t45, 3);
+ t45 = vsubw_u8(t45, m16);
+ t45 = vreinterpretq_u16_s16(
+ vminq_s16(tc_vec, vmaxq_s16(neg_tc_vec, vreinterpretq_s16_u16(t45))));
+ uint8x8_t d45 = vmovn_u16(t45);
+ d45 = vadd_u8(d45, m16);
+ store_u8x4_strided_xN<2>(&src[-3 * offset], 5 * offset, &d45);
+}
+
} // namespace
namespace X265_NS
@@ -328,6 +402,7 @@ void setupLoopFilterPrimitives_neon(EncoderPrimitives &p)
p.sign = calSign_neon;
p.pelFilterLumaStrong[0] = pelFilterLumaStrong_V_neon;
+ p.pelFilterLumaStrong[1] = pelFilterLumaStrong_H_neon;
}
--
2.34.1
>From b8d26d05e75e64129c92a67a0c63ffe432856341 Mon Sep 17 00:00:00 2001
Message-Id: <b8d26d05e75e64129c92a67a0c63ffe432856341.1739282617.git.microdaryl.rob...@arm.com>
In-Reply-To: <cover.1739282617.git.microdaryl.rob...@arm.com>
References: <cover.1739282617.git.microdaryl.rob...@arm.com>
From: Micro Daryl Robles <microdaryl.rob...@arm.com>
Date: Wed, 14 Aug 2024 01:22:44 +0100
Subject: [PATCH 2/5] AArch64: Add Neon implementation of pelFilterLumaStrong_H
Relative performance compared to scalar C:
Neoverse N1: 3.67x
Neoverse N2: 2.57x
Neoverse V1: 2.31x
Neoverse V2: 2.35x
---
source/common/aarch64/loopfilter-prim.cpp | 75 +++++++++++++++++++++++
1 file changed, 75 insertions(+)
diff --git a/source/common/aarch64/loopfilter-prim.cpp b/source/common/aarch64/loopfilter-prim.cpp
index 9242c859e..7fef25cbb 100644
--- a/source/common/aarch64/loopfilter-prim.cpp
+++ b/source/common/aarch64/loopfilter-prim.cpp
@@ -1,5 +1,6 @@
#include "common.h"
#include "loopfilter-prim.h"
+#include "mem-neon.h"
#define PIXEL_MIN 0
@@ -311,6 +312,79 @@ void pelFilterLumaStrong_V_neon(pixel *src, intptr_t srcStep, intptr_t offset,
}
}
+void pelFilterLumaStrong_H_neon(pixel *src, intptr_t srcStep, intptr_t offset,
+ int32_t tcP, int32_t tcQ)
+{
+ assert(UNIT_SIZE == 4 && srcStep == 1);
+ (void)srcStep;
+
+ 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);
+
+ uint8x8_t m0 = vld1_u8(src - 4 * offset);
+ uint8x8_t m1 = vld1_u8(src - 3 * offset);
+ uint8x8_t m2 = vld1_u8(src - 2 * offset);
+ uint8x8_t m3 = vld1_u8(src - 1 * offset);
+ uint8x8_t m4 = vld1_u8(src - 0 * offset);
+ uint8x8_t m5 = vld1_u8(src + 1 * offset);
+ uint8x8_t m6 = vld1_u8(src + 2 * offset);
+ uint8x8_t m7 = vld1_u8(src + 3 * offset);
+
+ uint8x8_t m12 =
+ vreinterpret_u8_u32(vzip1_u32(vreinterpret_u32_u8(m1), vreinterpret_u32_u8(m2)));
+ uint8x8_t m23 =
+ vreinterpret_u8_u32(vzip1_u32(vreinterpret_u32_u8(m2), vreinterpret_u32_u8(m3)));
+ uint8x8_t m34 =
+ vreinterpret_u8_u32(vzip1_u32(vreinterpret_u32_u8(m3), vreinterpret_u32_u8(m4)));
+ uint8x8_t m45 =
+ vreinterpret_u8_u32(vzip1_u32(vreinterpret_u32_u8(m4), vreinterpret_u32_u8(m5)));
+ uint8x8_t m56 =
+ vreinterpret_u8_u32(vzip1_u32(vreinterpret_u32_u8(m5), vreinterpret_u32_u8(m6)));
+
+ // src[-1 * offset], src[0 * offset]
+ uint16x8_t p0 = vaddl_u8(m23, m34);
+ p0 = vaddw_u8(p0, m45);
+ uint16x8_t t0 = vshlq_n_u16(p0, 1);
+ uint16x8_t t1 = vaddl_u8(m12, m56);
+ uint16x8_t t01 = vaddq_u16(t0, t1);
+ t01 = vrshrq_n_u16(t01, 3);
+ t01 = vsubw_u8(t01, m34);
+ t01 = vreinterpretq_u16_s16(
+ vminq_s16(tc_vec, vmaxq_s16(neg_tc_vec, vreinterpretq_s16_u16(t01))));
+ uint8x8_t d01 = vmovn_u16(t01);
+ d01 = vadd_u8(d01, m34);
+ store_u8x4_strided_xN<2>(&src[-1 * offset], 1 * offset, &d01);
+
+ uint8x8_t m16 =
+ vreinterpret_u8_u32(vzip1_u32(vreinterpret_u32_u8(m1), vreinterpret_u32_u8(m6)));
+ uint8x8_t m25 =
+ vreinterpret_u8_u32(vzip1_u32(vreinterpret_u32_u8(m2), vreinterpret_u32_u8(m5)));
+
+ // src[-2 * offset], src[1 * offset]
+ uint16x8_t p1 = vaddw_u8(p0, m16);
+ uint16x8_t t23 = vrshrq_n_u16(p1, 2);
+ t23 = vsubw_u8(t23, m25);
+ t23 = vreinterpretq_u16_s16(
+ vminq_s16(tc_vec, vmaxq_s16(neg_tc_vec, vreinterpretq_s16_u16(t23))));
+ uint8x8_t d23 = vmovn_u16(t23);
+ d23 = vadd_u8(d23, m25);
+ store_u8x4_strided_xN<2>(&src[-2 * offset], 3 * offset, &d23);
+
+ uint8x8_t m07 =
+ vreinterpret_u8_u32(vzip1_u32(vreinterpret_u32_u8(m0), vreinterpret_u32_u8(m7)));
+
+ // src[-3 * offset], src[2 * offset]
+ uint16x8_t p2 = vaddl_u8(m07, m16);
+ uint16x8_t t45 = vmlaq_n_u16(p1, p2, 2);
+ t45 = vrshrq_n_u16(t45, 3);
+ t45 = vsubw_u8(t45, m16);
+ t45 = vreinterpretq_u16_s16(
+ vminq_s16(tc_vec, vmaxq_s16(neg_tc_vec, vreinterpretq_s16_u16(t45))));
+ uint8x8_t d45 = vmovn_u16(t45);
+ d45 = vadd_u8(d45, m16);
+ store_u8x4_strided_xN<2>(&src[-3 * offset], 5 * offset, &d45);
+}
+
} // namespace
namespace X265_NS
@@ -328,6 +402,7 @@ void setupLoopFilterPrimitives_neon(EncoderPrimitives &p)
p.sign = calSign_neon;
p.pelFilterLumaStrong[0] = pelFilterLumaStrong_V_neon;
+ p.pelFilterLumaStrong[1] = pelFilterLumaStrong_H_neon;
}
--
2.34.1
_______________________________________________
x265-devel mailing list
x265-devel@videolan.org
https://mailman.videolan.org/listinfo/x265-devel
