[x265] [PATCH 1/2] AArch64: Add Neon implementation of 4x4 intra_pred_planar

[Micro Daryl Robles]

Hi Chen,

Thank you for the feedback.

For the vcreate_u8 vs vld1(const array):
  I checked the asm and the compiler just puts it in a constant pool and load it, which is effectively the same as your suggestion.

Thanks,
Micro

From: chen <chenm003 at 163.com>
Date: Friday, 24 January 2025 at 05:59
To: Development for x265 <x265-devel at videolan.org>
Cc: nd <nd at arm.com>, Micro Daryl Robles <MicroDaryl.Robles at arm.com>
Subject: Re:[x265] [PATCH 1/2] AArch64: Add Neon implementation of 4x4 intra_pred_planar

Thank for the patches.



The algorithm kernel in my understand below, there have two vcreate_u8, it may merge into constant array c[2] to avoid compiler performance issue, other part looks good to me.



+    uint16x8_t t = vmull_u8(topRight, vget_high_u8(c0));                                       ;  topRight     *[4 3 2 1 4 3 2 1]

+    t = vmlal_u8(t, above, vget_low_u8(c1));                                                   ; +above        *[1 1 1 1 3 3 3 3]

+    t = vmlal_u8(t, bottomLeft, vget_high_u8(c1));                                             ; +bottomLeft   *[3 3 3 3 1 1 1 1]

+

+    uint8x8_t index02 = vcreate_u8(0x0A0A0A0A08080808);

+    uint8x8_t left02 = vqtbl1_u8(src, index02);                                                ; left02 = [L2 L2 L2 L2 L0 L0 L0 L0]

+    uint16x8_t t02 = vmlal_u8(t, left02, vget_low_u8(c0));                                     ; t02 = [ L2*[0 1 2 3] L0*[0 1 2 3] ]

+    uint8x8_t d02 = vrshrn_n_u16(t02, log2Size + 1);




At 2025-01-23 22:15:30, "Micro Daryl Robles" <microdaryl.robles at arm.com> wrote:

>Relative performance compared to scalar C:

>

> Neoverse N1: 2.65x

> Neoverse N2: 2.27x

> Neoverse V1: 2.67x

> Neoverse V2: 2.82x

>---

> source/common/aarch64/intrapred-prim.cpp | 47 ++++++++++++++++++++++++

> 1 file changed, 47 insertions(+)

>

>diff --git a/source/common/aarch64/intrapred-prim.cpp b/source/common/aarch64/intrapred-prim.cpp

>index 68a9f26ac..c5d47fe0d 100644

>--- a/source/common/aarch64/intrapred-prim.cpp

>+++ b/source/common/aarch64/intrapred-prim.cpp

>@@ -4,6 +4,7 @@

>

> #if HAVE_NEON

> #include "arm64-utils.h"

>+#include "mem-neon.h"

> #include <arm_neon.h>

>

> using namespace X265_NS;

>@@ -399,6 +400,51 @@ void planar_pred_neon(pixel * dst, intptr_t dstStride, const pixel * srcPix, int

>         }

> }

>

>+#if !HIGH_BIT_DEPTH

>+void intra_pred_planar4_neon(pixel *dst, intptr_t dstStride, const pixel *srcPix,

>+                             int /*dirMode*/, int /*bFilter*/)

>+{

>+    const int log2Size = 2;

>+    const int blkSize = 1 << log2Size;

>+

>+    uint8x16_t src = vld1q_u8(srcPix + 1);

>+

>+    uint8x8_t above =

>+        vreinterpret_u8_u32(vdup_laneq_u32(vreinterpretq_u32_u8(src), 0));

>+

>+    uint8x8_t topRight = vdup_laneq_u8(src, blkSize);

>+    uint8x8_t bottomLeft = vdup_laneq_u8(src, 3 * blkSize);

>+

>+    const uint8_t c[2][16] =

>+    {

>+        {3, 2, 1, 0, 3, 2, 1, 0, 1, 2, 3, 4, 1, 2, 3, 4},

>+        {3, 3, 3, 3, 1, 1, 1, 1, 1, 1, 1, 1, 3, 3, 3, 3}

>+    };

>+

>+    const uint8x16_t c0 = vld1q_u8(c[0]);

>+    const uint8x16_t c1 = vld1q_u8(c[1]);

>+

>+    uint16x8_t t = vmull_u8(topRight, vget_high_u8(c0));

>+    t = vmlal_u8(t, above, vget_low_u8(c1));

>+    t = vmlal_u8(t, bottomLeft, vget_high_u8(c1));

>+

>+    uint8x8_t index02 = vcreate_u8(0x0A0A0A0A08080808);

>+    uint8x8_t left02 = vqtbl1_u8(src, index02);

>+    uint16x8_t t02 = vmlal_u8(t, left02, vget_low_u8(c0));

>+    uint8x8_t d02 = vrshrn_n_u16(t02, log2Size + 1);

>+

>+    uint8x8_t index13 = vcreate_u8(0x0B0B0B0B09090909);

>+    uint8x8_t left13 = vqtbl1_u8(src, index13);

>+    uint16x8_t t13 = vmlal_u8(t, left13, vget_low_u8(c0));

>+    uint16x8_t sub_bottomLeft_above = vsubl_u8(bottomLeft, above);

>+    t13 = vaddq_u16(t13, sub_bottomLeft_above);

>+    uint8x8_t d13 = vrshrn_n_u16(t13, log2Size + 1);

>+

>+    store_u8x4_strided_xN<2>(dst + 0 * dstStride, 2 * dstStride, &d02);

>+    store_u8x4_strided_xN<2>(dst + 1 * dstStride, 2 * dstStride, &d13);

>+}

>+#endif

>+

> static void dcPredFilter(const pixel* above, const pixel* left, pixel* dst, intptr_t dststride, int size)

> {

>     // boundary pixels processing

>@@ -576,6 +622,7 @@ void setupIntraPrimitives_neon(EncoderPrimitives &p)

>     p.cu[BLOCK_32x32].intra_pred_allangs = all_angs_pred_neon<5>;

>

> #if !HIGH_BIT_DEPTH

>+    p.cu[BLOCK_4x4].intra_pred[PLANAR_IDX] = intra_pred_planar4_neon;

>     p.cu[BLOCK_8x8].intra_pred[PLANAR_IDX] = PFX(intra_pred_planar8_neon);

>     p.cu[BLOCK_16x16].intra_pred[PLANAR_IDX] = PFX(intra_pred_planar16_neon);

> #endif

>--

>2.34.1

>

IMPORTANT NOTICE: The contents of this email and any attachments are confidential and may also be privileged. If you are not the intended recipient, please notify the sender immediately and do not disclose the contents to any other person, use it for any purpose, or store or copy the information in any medium. Thank you.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.videolan.org/pipermail/x265-devel/attachments/20250124/7a57c0fd/attachment-0001.htm>