Convert the scalar C implementation of partialButterflyInverse4() to use Neon.
Relative performance compared to scalar C:
Neoverse N1: 3.38x
Neoverse V1: 2.84x
Neoverse V2: 2.82x
---
source/common/aarch64/dct-prim.cpp | 60 ++++++++++++++++++------------
1 file changed, 37 insertions(+), 23 deletions(-)
diff --git a/source/common/aarch64/dct-prim.cpp
b/source/common/aarch64/dct-prim.cpp
index 093392e06..c58e20113 100644
--- a/source/common/aarch64/dct-prim.cpp
+++ b/source/common/aarch64/dct-prim.cpp
@@ -739,29 +739,43 @@ static inline void partialButterfly8_neon(const int16_t
*src, int16_t *dst)
}
}
-static void partialButterflyInverse4(const int16_t *src, int16_t *dst, int
shift, int line)
+template<int shift>
+static inline void partialButterflyInverse4_neon(const int16_t *src, int16_t
*dst)
{
- int j;
- int E[2], O[2];
- int add = 1 << (shift - 1);
+ int16x4_t s0 = vld1_s16(src + 0);
+ int16x4_t s1 = vld1_s16(src + 4);
+ int16x4_t s2 = vld1_s16(src + 8);
+ int16x4_t s3 = vld1_s16(src + 12);
- for (j = 0; j < line; j++)
- {
- /* Utilizing symmetry properties to the maximum to minimize the number
of multiplications */
- O[0] = g_t4[1][0] * src[line] + g_t4[3][0] * src[3 * line];
- O[1] = g_t4[1][1] * src[line] + g_t4[3][1] * src[3 * line];
- E[0] = g_t4[0][0] * src[0] + g_t4[2][0] * src[2 * line];
- E[1] = g_t4[0][1] * src[0] + g_t4[2][1] * src[2 * line];
+ // Multiply and accumulate with g_t4 constants.
+ int32x4_t O[2];
+ O[0] = vmull_n_s16(s1, 83);
+ O[0] = vmlal_n_s16(O[0], s3, 36);
+ O[1] = vmull_n_s16(s1, 36);
+ O[1] = vmlal_n_s16(O[1], s3, -83);
+
+ int32x4_t E[2];
+ E[0] = vaddl_s16(s0, s2);
+ E[0] = vmulq_n_s32(E[0], 64);
+ E[1] = vsubl_s16(s0, s2);
+ E[1] = vmulq_n_s32(E[1], 64);
+
+ int32x4_t t0 = vaddq_s32(E[0], O[0]);
+ int32x4_t t1 = vaddq_s32(E[1], O[1]);
+ int32x4_t t2 = vsubq_s32(E[1], O[1]);
+ int32x4_t t3 = vsubq_s32(E[0], O[0]);
- /* Combining even and odd terms at each hierarchy levels to calculate
the final spatial domain vector */
- dst[0] = (int16_t)(x265_clip3(-32768, 32767, (E[0] + O[0] + add) >>
shift));
- dst[1] = (int16_t)(x265_clip3(-32768, 32767, (E[1] + O[1] + add) >>
shift));
- dst[2] = (int16_t)(x265_clip3(-32768, 32767, (E[1] - O[1] + add) >>
shift));
- dst[3] = (int16_t)(x265_clip3(-32768, 32767, (E[0] - O[0] + add) >>
shift));
+ int16x4_t d0 = vqrshrn_n_s32(t0, shift);
+ int16x4_t d1 = vqrshrn_n_s32(t1, shift);
+ int16x4_t d2 = vqrshrn_n_s32(t2, shift);
+ int16x4_t d3 = vqrshrn_n_s32(t3, shift);
- src++;
- dst += 4;
- }
+ transpose_4x4_s16(d0, d1, d2, d3);
+
+ vst1_s16(dst + 0, d0);
+ vst1_s16(dst + 4, d1);
+ vst1_s16(dst + 8, d2);
+ vst1_s16(dst + 12, d3);
}
@@ -1222,14 +1236,14 @@ void idst4_neon(const int16_t *src, int16_t *dst,
intptr_t dstStride)
void idct4_neon(const int16_t *src, int16_t *dst, intptr_t dstStride)
{
- const int shift_1st = 7;
- const int shift_2nd = 12 - (X265_DEPTH - 8);
+ const int shift_pass1 = 7;
+ const int shift_pass2 = 12 - (X265_DEPTH - 8);
ALIGN_VAR_32(int16_t, coef[4 * 4]);
ALIGN_VAR_32(int16_t, block[4 * 4]);
- partialButterflyInverse4(src, coef, shift_1st, 4); // Forward DST BY FAST
ALGORITHM, block input, coef output
- partialButterflyInverse4(coef, block, shift_2nd, 4); // Forward DST BY
FAST ALGORITHM, coef input, coeff output
+ partialButterflyInverse4_neon<shift_pass1>(src, coef);
+ partialButterflyInverse4_neon<shift_pass2>(coef, block);
for (int i = 0; i < 4; i++)
{
--
2.34.1
>From 9a7a9a2288c66e6a532f73de1d70cf437c98f8bd Mon Sep 17 00:00:00 2001
Message-Id: <9a7a9a2288c66e6a532f73de1d70cf437c98f8bd.1732619564.git.microdaryl.rob...@arm.com>
In-Reply-To: <cover.1732619564.git.microdaryl.rob...@arm.com>
References: <cover.1732619564.git.microdaryl.rob...@arm.com>
From: Micro Daryl Robles <microdaryl.rob...@arm.com>
Date: Tue, 3 Sep 2024 09:58:28 +0100
Subject: [PATCH 4/7] AArch64: Add Neon implementation of 4x4 IDCT
Convert the scalar C implementation of partialButterflyInverse4() to use Neon.
Relative performance compared to scalar C:
Neoverse N1: 3.38x
Neoverse V1: 2.84x
Neoverse V2: 2.82x
---
source/common/aarch64/dct-prim.cpp | 60 ++++++++++++++++++------------
1 file changed, 37 insertions(+), 23 deletions(-)
diff --git a/source/common/aarch64/dct-prim.cpp b/source/common/aarch64/dct-prim.cpp
index 093392e06..c58e20113 100644
--- a/source/common/aarch64/dct-prim.cpp
+++ b/source/common/aarch64/dct-prim.cpp
@@ -739,29 +739,43 @@ static inline void partialButterfly8_neon(const int16_t *src, int16_t *dst)
}
}
-static void partialButterflyInverse4(const int16_t *src, int16_t *dst, int shift, int line)
+template<int shift>
+static inline void partialButterflyInverse4_neon(const int16_t *src, int16_t *dst)
{
- int j;
- int E[2], O[2];
- int add = 1 << (shift - 1);
+ int16x4_t s0 = vld1_s16(src + 0);
+ int16x4_t s1 = vld1_s16(src + 4);
+ int16x4_t s2 = vld1_s16(src + 8);
+ int16x4_t s3 = vld1_s16(src + 12);
- for (j = 0; j < line; j++)
- {
- /* Utilizing symmetry properties to the maximum to minimize the number of multiplications */
- O[0] = g_t4[1][0] * src[line] + g_t4[3][0] * src[3 * line];
- O[1] = g_t4[1][1] * src[line] + g_t4[3][1] * src[3 * line];
- E[0] = g_t4[0][0] * src[0] + g_t4[2][0] * src[2 * line];
- E[1] = g_t4[0][1] * src[0] + g_t4[2][1] * src[2 * line];
+ // Multiply and accumulate with g_t4 constants.
+ int32x4_t O[2];
+ O[0] = vmull_n_s16(s1, 83);
+ O[0] = vmlal_n_s16(O[0], s3, 36);
+ O[1] = vmull_n_s16(s1, 36);
+ O[1] = vmlal_n_s16(O[1], s3, -83);
+
+ int32x4_t E[2];
+ E[0] = vaddl_s16(s0, s2);
+ E[0] = vmulq_n_s32(E[0], 64);
+ E[1] = vsubl_s16(s0, s2);
+ E[1] = vmulq_n_s32(E[1], 64);
+
+ int32x4_t t0 = vaddq_s32(E[0], O[0]);
+ int32x4_t t1 = vaddq_s32(E[1], O[1]);
+ int32x4_t t2 = vsubq_s32(E[1], O[1]);
+ int32x4_t t3 = vsubq_s32(E[0], O[0]);
- /* Combining even and odd terms at each hierarchy levels to calculate the final spatial domain vector */
- dst[0] = (int16_t)(x265_clip3(-32768, 32767, (E[0] + O[0] + add) >> shift));
- dst[1] = (int16_t)(x265_clip3(-32768, 32767, (E[1] + O[1] + add) >> shift));
- dst[2] = (int16_t)(x265_clip3(-32768, 32767, (E[1] - O[1] + add) >> shift));
- dst[3] = (int16_t)(x265_clip3(-32768, 32767, (E[0] - O[0] + add) >> shift));
+ int16x4_t d0 = vqrshrn_n_s32(t0, shift);
+ int16x4_t d1 = vqrshrn_n_s32(t1, shift);
+ int16x4_t d2 = vqrshrn_n_s32(t2, shift);
+ int16x4_t d3 = vqrshrn_n_s32(t3, shift);
- src++;
- dst += 4;
- }
+ transpose_4x4_s16(d0, d1, d2, d3);
+
+ vst1_s16(dst + 0, d0);
+ vst1_s16(dst + 4, d1);
+ vst1_s16(dst + 8, d2);
+ vst1_s16(dst + 12, d3);
}
@@ -1222,14 +1236,14 @@ void idst4_neon(const int16_t *src, int16_t *dst, intptr_t dstStride)
void idct4_neon(const int16_t *src, int16_t *dst, intptr_t dstStride)
{
- const int shift_1st = 7;
- const int shift_2nd = 12 - (X265_DEPTH - 8);
+ const int shift_pass1 = 7;
+ const int shift_pass2 = 12 - (X265_DEPTH - 8);
ALIGN_VAR_32(int16_t, coef[4 * 4]);
ALIGN_VAR_32(int16_t, block[4 * 4]);
- partialButterflyInverse4(src, coef, shift_1st, 4); // Forward DST BY FAST ALGORITHM, block input, coef output
- partialButterflyInverse4(coef, block, shift_2nd, 4); // Forward DST BY FAST ALGORITHM, coef input, coeff output
+ partialButterflyInverse4_neon<shift_pass1>(src, coef);
+ partialButterflyInverse4_neon<shift_pass2>(coef, block);
for (int i = 0; i < 4; i++)
{
--
2.34.1
_______________________________________________
x265-devel mailing list
x265-devel@videolan.org
https://mailman.videolan.org/listinfo/x265-devel
