[x265] [PATCH 2/2] AArch64: Add Neon implementation of 32x32 intra_pred_planar

[chen]

The 32x32 algorithm looks good to me

At 2025-01-23 22:15:45, "Micro Daryl Robles" <microdaryl.robles at arm.com> wrote:
>Relative performance compared to scalar C:
>
> Neoverse N1: 4.53x
> Neoverse N2: 4.19x
> Neoverse V1: 4.26x
> Neoverse V2: 3.93x
>---
> source/common/aarch64/intrapred-prim.cpp | 86 ++++++++++++++++++++++++
> 1 file changed, 86 insertions(+)
>
>diff --git a/source/common/aarch64/intrapred-prim.cpp b/source/common/aarch64/intrapred-prim.cpp
>index c5d47fe0d..3d4b4769f 100644
>--- a/source/common/aarch64/intrapred-prim.cpp
>+++ b/source/common/aarch64/intrapred-prim.cpp
>@@ -445,6 +445,91 @@ void intra_pred_planar4_neon(pixel *dst, intptr_t dstStride, const pixel *srcPix
> }
> #endif
> 
>+#if !HIGH_BIT_DEPTH
>+void intra_pred_planar32_neon(pixel *dst, intptr_t dstStride, const pixel *srcPix,
>+                              int /*dirMode*/, int /*bFilter*/)
>+{
>+    const int log2Size = 5;
>+    const int blkSize = 1 << log2Size;
>+
>+    const pixel *src0 = srcPix + 1;
>+    const pixel *src1 = srcPix + 2 * blkSize + 1;
>+
>+    uint8x8_t above0 = vld1_u8(src0 + 0 * 8);
>+    uint8x8_t above1 = vld1_u8(src0 + 1 * 8);
>+    uint8x8_t above2 = vld1_u8(src0 + 2 * 8);
>+    uint8x8_t above3 = vld1_u8(src0 + 3 * 8);
>+
>+    uint8x8_t topRight = vdup_n_u8(src0[blkSize]);
>+    uint8x8_t bottomLeft = vdup_n_u8(src1[blkSize]);
>+
>+    const uint8_t c[2][32] =
>+    {
>+        {31, 30, 29, 28, 27, 26, 25, 24, 23, 22, 21, 20, 19, 18, 17, 16,
>+         15, 14, 13, 12, 11, 10,  9,  8,  7,  6,  5,  4,  3,  2,  1,  0},
>+        { 1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12, 13, 14, 15, 16,
>+         17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32}
>+    };
>+
>+    // left constant
>+    const uint8x8_t l0 = vld1_u8(c[0] + 0 * 8);
>+    const uint8x8_t l1 = vld1_u8(c[0] + 1 * 8);
>+    const uint8x8_t l2 = vld1_u8(c[0] + 2 * 8);
>+    const uint8x8_t l3 = vld1_u8(c[0] + 3 * 8);
>+
>+    // topRight constant
>+    const uint8x8_t tR0 = vld1_u8(c[1] + 0 * 8);
>+    const uint8x8_t tR1 = vld1_u8(c[1] + 1 * 8);
>+    const uint8x8_t tR2 = vld1_u8(c[1] + 2 * 8);
>+    const uint8x8_t tR3 = vld1_u8(c[1] + 3 * 8);
>+
>+    const uint16x8_t offset = vdupq_n_u16(blkSize);
>+    const uint16x8_t offset_bottomLeft = vaddw_u8(offset, bottomLeft);
>+
>+    const uint8x8_t c31 = vdup_n_u8(31);
>+
>+    uint16x8_t t0 = vmlal_u8(offset_bottomLeft, topRight, tR0);
>+    t0 = vmlal_u8(t0, above0, c31);
>+
>+    uint16x8_t t1 = vmlal_u8(offset_bottomLeft, topRight, tR1);
>+    t1 = vmlal_u8(t1, above1, c31);
>+
>+    uint16x8_t t2 = vmlal_u8(offset_bottomLeft, topRight, tR2);
>+    t2 = vmlal_u8(t2, above2, c31);
>+
>+    uint16x8_t t3 = vmlal_u8(offset_bottomLeft, topRight, tR3);
>+    t3 = vmlal_u8(t3, above3, c31);
>+
>+    uint16x8_t sub_bottomLeft_above0 = vsubl_u8(bottomLeft, above0);
>+    uint16x8_t sub_bottomLeft_above1 = vsubl_u8(bottomLeft, above1);
>+    uint16x8_t sub_bottomLeft_above2 = vsubl_u8(bottomLeft, above2);
>+    uint16x8_t sub_bottomLeft_above3 = vsubl_u8(bottomLeft, above3);
>+
>+    for (int y = 0; y < 32; y++)
>+    {
>+        uint8x8_t left = vdup_n_u8(src1[y]);
>+
>+        uint16x8_t r0 = vmlal_u8(t0, left, l0);
>+        uint16x8_t r1 = vmlal_u8(t1, left, l1);
>+        uint16x8_t r2 = vmlal_u8(t2, left, l2);
>+        uint16x8_t r3 = vmlal_u8(t3, left, l3);
>+
>+        uint8x8_t d[4];
>+        d[0] = vshrn_n_u16(r0, log2Size + 1);
>+        d[1] = vshrn_n_u16(r1, log2Size + 1);
>+        d[2] = vshrn_n_u16(r2, log2Size + 1);
>+        d[3] = vshrn_n_u16(r3, log2Size + 1);
>+
>+        store_u8x8xn<4>(dst + y * dstStride, 8, d);
>+
>+        t0 = vaddq_u16(t0, sub_bottomLeft_above0);
>+        t1 = vaddq_u16(t1, sub_bottomLeft_above1);
>+        t2 = vaddq_u16(t2, sub_bottomLeft_above2);
>+        t3 = vaddq_u16(t3, sub_bottomLeft_above3);
>+    }
>+}
>+#endif
>+
> static void dcPredFilter(const pixel* above, const pixel* left, pixel* dst, intptr_t dststride, int size)
> {
>     // boundary pixels processing
>@@ -625,6 +710,7 @@ void setupIntraPrimitives_neon(EncoderPrimitives &p)
>     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);
>+    p.cu[BLOCK_32x32].intra_pred[PLANAR_IDX] = intra_pred_planar32_neon;
> #endif
> 
>     p.cu[BLOCK_4x4].intra_pred[DC_IDX] = intra_pred_dc_neon<4>;
>-- 
>2.34.1
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.videolan.org/pipermail/x265-devel/attachments/20250124/94cbaedd/attachment.htm>