Ryujinx/Ryujinx.Graphics.Vic/Scaler.cs
gdkchan 1402d8391d
Support NVDEC H264 interlaced video decoding and VIC deinterlacing (#3225)
* Support NVDEC H264 interlaced video decoding and VIC deinterlacing

* Remove unused code
2022-03-23 17:09:32 -03:00

124 lines
5 KiB
C#

using System;
using System.Runtime.Intrinsics;
using System.Runtime.Intrinsics.X86;
namespace Ryujinx.Graphics.Vic
{
static class Scaler
{
public static void DeinterlaceWeave(Span<byte> data, ReadOnlySpan<byte> prevData, int width, int fieldSize, bool isTopField)
{
// Prev I Curr I Curr P
// TTTTTTTT BBBBBBBB TTTTTTTT
// -------- -------- BBBBBBBB
if (isTopField)
{
for (int offset = 0; offset < data.Length; offset += fieldSize * 2)
{
prevData.Slice(offset >> 1, width).CopyTo(data.Slice(offset + fieldSize, width));
}
}
else
{
for (int offset = 0; offset < data.Length; offset += fieldSize * 2)
{
prevData.Slice(offset >> 1, width).CopyTo(data.Slice(offset, width));
}
}
}
public static void DeinterlaceBob(Span<byte> data, int width, int fieldSize, bool isTopField)
{
// Curr I Curr P
// TTTTTTTT TTTTTTTT
// -------- TTTTTTTT
if (isTopField)
{
for (int offset = 0; offset < data.Length; offset += fieldSize * 2)
{
data.Slice(offset, width).CopyTo(data.Slice(offset + fieldSize, width));
}
}
else
{
for (int offset = 0; offset < data.Length; offset += fieldSize * 2)
{
data.Slice(offset + fieldSize, width).CopyTo(data.Slice(offset, width));
}
}
}
public unsafe static void DeinterlaceMotionAdaptive(
Span<byte> data,
ReadOnlySpan<byte> prevData,
ReadOnlySpan<byte> nextData,
int width,
int fieldSize,
bool isTopField)
{
// Very simple motion adaptive algorithm.
// If the pixel changed between previous and next frame, use Bob, otherwise use Weave.
//
// Example pseudo code:
// C_even = (P_even == N_even) ? P_even : C_odd
// Where: C is current frame, P is previous frame and N is next frame, and even/odd are the fields.
//
// Note: This does not fully match the hardware algorithm.
// The motion adaptive deinterlacing implemented on hardware is considerably more complex,
// and hard to implement accurately without proper documentation as for example, the
// method used for motion estimation is unknown.
int start = isTopField ? fieldSize : 0;
int otherFieldOffset = isTopField ? -fieldSize : fieldSize;
fixed (byte* pData = data, pPrevData = prevData, pNextData = nextData)
{
for (int offset = start; offset < data.Length; offset += fieldSize * 2)
{
int refOffset = (offset - start) >> 1;
int x = 0;
if (Avx2.IsSupported)
{
for (; x < (width & ~0x1f); x += 32)
{
Vector256<byte> prevPixels = Avx.LoadVector256(pPrevData + refOffset + x);
Vector256<byte> nextPixels = Avx.LoadVector256(pNextData + refOffset + x);
Vector256<byte> bob = Avx.LoadVector256(pData + offset + otherFieldOffset + x);
Vector256<byte> diff = Avx2.CompareEqual(prevPixels, nextPixels);
Avx.Store(pData + offset + x, Avx2.BlendVariable(bob, prevPixels, diff));
}
}
else if (Sse41.IsSupported)
{
for (; x < (width & ~0xf); x += 16)
{
Vector128<byte> prevPixels = Sse2.LoadVector128(pPrevData + refOffset + x);
Vector128<byte> nextPixels = Sse2.LoadVector128(pNextData + refOffset + x);
Vector128<byte> bob = Sse2.LoadVector128(pData + offset + otherFieldOffset + x);
Vector128<byte> diff = Sse2.CompareEqual(prevPixels, nextPixels);
Sse2.Store(pData + offset + x, Sse41.BlendVariable(bob, prevPixels, diff));
}
}
for (; x < width; x++)
{
byte prevPixel = prevData[refOffset + x];
byte nextPixel = nextData[refOffset + x];
if (nextPixel != prevPixel)
{
data[offset + x] = data[offset + otherFieldOffset + x];
}
else
{
data[offset + x] = prevPixel;
}
}
}
}
}
}
}