mirrors/Ryubing
1
0
Fork 0
mirror of https://git.ryujinx.app/ryubing/ryujinx.git synced 2025-09-14 05:55:43 +00:00
Code Issues Packages Projects Releases Wiki Activity

Implement VP9 loop filtering (#550)

Browse Source 
Unmerged PR from OG Ryujinx (#4367). From @gdkchan:

> The main goal of this change is porting the loop filtering from
libvpx, which should fix the block artifacts on some VP9 videos on games
using NVDEC to decode them. In addition to that, there are two other
changes:
> 
> - The remaining decoder code required to decode a VP9 video (with
headers included) has been added. That was done because it's much better
to test the decoder standalone with a video file. I decided to keep that
code on the emulator, even if some of it is unused, since it makes
standalone testing easier in the future too, and we can include unit
tests with video files.
> - Large refactoring of both new and existing code to conform with our
conding [sic] styles, done by @TSRBerry (thanks!) Some of it has been
automated.
> 
> Since we had no loop filtering before, this change will make video
decoding slower. That may cause frame drop etc if the decoder is not
fast enough in some games. I plan to optimize the decoder more in the
future to make up for that, but if possible I'd prefer to not do it as
part of this PR, but if the perf loss is too severe I might consider.
> 
> This will need to be tested on games that had the block artifacts, it
would be nice to confirm if they match hardware now, and get some
before/after screenshots etc.

Comment from @Bjorn29512:

> Significantly improves the block artifacts in FE: Engage.
> 
> Before:
>
![](https://user-images.githubusercontent.com/110204265/216882414-ec88dbda-7544-4490-8a47-37f074056ae3.png)
> 
> After:
>
![](https://user-images.githubusercontent.com/110204265/216882478-4e81fead-1033-4877-b282-f9cac6d6aa3b.png)

---------

Co-authored-by: gdkchan <gab.dark.100@gmail.com>
Co-authored-by: TSR Berry <20988865+TSRBerry@users.noreply.github.com>
This commit is contained in:
Keaton
2025-02-18 20:59:36 -06:00
committed by GitHub
parent 920933bc9f
commit f91cd05260
79 changed files with 11343 additions and 3036 deletions
Download Patch File Download Diff File Expand all files Collapse all files

164
src/Ryujinx.Graphics.Nvdec.Vp9/Dsp/Convolve.cs
View File

@@ -1,4 +1,4 @@
using Ryujinx.Common.Memory;
using Ryujinx.Common.Memory;
using Ryujinx.Graphics.Nvdec.Vp9.Common;
using System.Diagnostics;
using System.Runtime.CompilerServices;
@@ -75,17 +75,16 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
Vector128<int> zero = Vector128<int>.Zero;
Vector128<int> const64 = Vector128.Create(64);
ulong x, y;
src -= SubpelTaps / 2 - 1;
src -= (SubpelTaps / 2) - 1;
fixed (Array8<short>* xFilter = xFilters)
{
Vector128<short> vfilter = Sse2.LoadVector128((short*)xFilter + (uint)(x0Q4 & SubpelMask) * 8);
Vector128<short> vfilter = Sse2.LoadVector128((short*)xFilter + ((uint)(x0Q4 & SubpelMask) * 8));
for (y = 0; y < (uint)h; ++y)
for (ulong y = 0; y < (uint)h; ++y)
{
ulong srcOffset = (uint)x0Q4 >> SubpelBits;
for (x = 0; x < (uint)w; x += 4)
for (ulong x = 0; x < (uint)w; x += 4)
{
Vector128<short> vsrc0 = Sse41.ConvertToVector128Int16(&src[srcOffset + x]);
Vector128<short> vsrc1 = Sse41.ConvertToVector128Int16(&src[srcOffset + x + 1]);
@@ -94,8 +93,10 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
Vector128<int> sum0123 = MultiplyAddAdjacent(vsrc0, vsrc1, vsrc2, vsrc3, vfilter, zero);
Sse.StoreScalar((float*)&dst[x], PackUnsignedSaturate(RoundShift(sum0123, const64), zero).AsSingle());
Sse.StoreScalar((float*)&dst[x],
PackUnsignedSaturate(RoundShift(sum0123, const64), zero).AsSingle());
}
src += srcStride;
dst += dstStride;
}
@@ -117,22 +118,20 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
if (Sse41.IsSupported && UseIntrinsics && xStepQ4 == 1 << SubpelBits)
{
ConvolveHorizSse41(src, srcStride, dst, dstStride, xFilters, x0Q4, w, h);
return;
}
int x, y;
src -= SubpelTaps / 2 - 1;
src -= (SubpelTaps / 2) - 1;
for (y = 0; y < h; ++y)
for (int y = 0; y < h; ++y)
{
int xQ4 = x0Q4;
for (x = 0; x < w; ++x)
for (int x = 0; x < w; ++x)
{
byte* srcX = &src[xQ4 >> SubpelBits];
ref Array8<short> xFilter = ref xFilters[xQ4 & SubpelMask];
int k, sum = 0;
for (k = 0; k < SubpelTaps; ++k)
int sum = 0;
for (int k = 0; k < SubpelTaps; ++k)
{
sum += srcX[k] * xFilter[k];
}
@@ -140,6 +139,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
dst[x] = BitUtils.ClipPixel(BitUtils.RoundPowerOfTwo(sum, FilterBits));
xQ4 += xStepQ4;
}
src += srcStride;
dst += dstStride;
}
@@ -156,25 +156,26 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
int w,
int h)
{
int x, y;
src -= SubpelTaps / 2 - 1;
src -= (SubpelTaps / 2) - 1;
for (y = 0; y < h; ++y)
for (int y = 0; y < h; ++y)
{
int xQ4 = x0Q4;
for (x = 0; x < w; ++x)
for (int x = 0; x < w; ++x)
{
byte* srcX = &src[xQ4 >> SubpelBits];
ref Array8<short> xFilter = ref xFilters[xQ4 & SubpelMask];
int k, sum = 0;
for (k = 0; k < SubpelTaps; ++k)
int sum = 0;
for (int k = 0; k < SubpelTaps; ++k)
{
sum += srcX[k] * xFilter[k];
}
dst[x] = (byte)BitUtils.RoundPowerOfTwo(dst[x] + BitUtils.ClipPixel(BitUtils.RoundPowerOfTwo(sum, FilterBits)), 1);
dst[x] = (byte)BitUtils.RoundPowerOfTwo(
dst[x] + BitUtils.ClipPixel(BitUtils.RoundPowerOfTwo(sum, FilterBits)), 1);
xQ4 += xStepQ4;
}
src += srcStride;
dst += dstStride;
}
@@ -203,18 +204,17 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
srcStride * 6,
srcStride * 7);
ulong x, y;
src -= srcStride * (SubpelTaps / 2 - 1);
src -= srcStride * ((SubpelTaps / 2) - 1);
fixed (Array8<short>* yFilter = yFilters)
{
Vector128<short> vfilter = Sse2.LoadVector128((short*)yFilter + (uint)(y0Q4 & SubpelMask) * 8);
Vector128<short> vfilter = Sse2.LoadVector128((short*)yFilter + ((uint)(y0Q4 & SubpelMask) * 8));
ulong srcBaseY = (uint)y0Q4 >> SubpelBits;
for (y = 0; y < (uint)h; ++y)
for (ulong y = 0; y < (uint)h; ++y)
{
ulong srcOffset = (srcBaseY + y) * (uint)srcStride;
for (x = 0; x < (uint)w; x += 4)
for (ulong x = 0; x < (uint)w; x += 4)
{
Vector256<int> vsrc = Avx2.GatherVector256((uint*)&src[srcOffset + x], indices, 1).AsInt32();
@@ -240,8 +240,10 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
Vector128<int> sum0123 = MultiplyAddAdjacent(vsrc0, vsrc1, vsrc2, vsrc3, vfilter, zero);
Sse.StoreScalar((float*)&dst[x], PackUnsignedSaturate(RoundShift(sum0123, const64), zero).AsSingle());
Sse.StoreScalar((float*)&dst[x],
PackUnsignedSaturate(RoundShift(sum0123, const64), zero).AsSingle());
}
dst += dstStride;
}
}
@@ -262,22 +264,20 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
if (Avx2.IsSupported && UseIntrinsics && yStepQ4 == 1 << SubpelBits)
{
ConvolveVertAvx2(src, srcStride, dst, dstStride, yFilters, y0Q4, w, h);
return;
}
int x, y;
src -= srcStride * (SubpelTaps / 2 - 1);
src -= srcStride * ((SubpelTaps / 2) - 1);
for (x = 0; x < w; ++x)
for (int x = 0; x < w; ++x)
{
int yQ4 = y0Q4;
for (y = 0; y < h; ++y)
for (int y = 0; y < h; ++y)
{
byte* srcY = &src[(yQ4 >> SubpelBits) * srcStride];
ref Array8<short> yFilter = ref yFilters[yQ4 & SubpelMask];
int k, sum = 0;
for (k = 0; k < SubpelTaps; ++k)
int sum = 0;
for (int k = 0; k < SubpelTaps; ++k)
{
sum += srcY[k * srcStride] * yFilter[k];
}
@@ -285,6 +285,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
dst[y * dstStride] = BitUtils.ClipPixel(BitUtils.RoundPowerOfTwo(sum, FilterBits));
yQ4 += yStepQ4;
}
++src;
++dst;
}
@@ -301,18 +302,17 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
int w,
int h)
{
int x, y;
src -= srcStride * (SubpelTaps / 2 - 1);
src -= srcStride * ((SubpelTaps / 2) - 1);
for (x = 0; x < w; ++x)
for (int x = 0; x < w; ++x)
{
int yQ4 = y0Q4;
for (y = 0; y < h; ++y)
for (int y = 0; y < h; ++y)
{
byte* srcY = &src[(yQ4 >> SubpelBits) * srcStride];
ref Array8<short> yFilter = ref yFilters[yQ4 & SubpelMask];
int k, sum = 0;
for (k = 0; k < SubpelTaps; ++k)
int sum = 0;
for (int k = 0; k < SubpelTaps; ++k)
{
sum += srcY[k * srcStride] * yFilter[k];
}
@@ -321,6 +321,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
dst[y * dstStride] + BitUtils.ClipPixel(BitUtils.RoundPowerOfTwo(sum, FilterBits)), 1);
yQ4 += yStepQ4;
}
++src;
++dst;
}
@@ -420,15 +421,16 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
// ==> yStepQ4 = 64. Since w and h are at most 16, the temp buffer is still
// big enough.
byte* temp = stackalloc byte[64 * 135];
int intermediateHeight = (((h - 1) * yStepQ4 + y0Q4) >> SubpelBits) + SubpelTaps;
int intermediateHeight = ((((h - 1) * yStepQ4) + y0Q4) >> SubpelBits) + SubpelTaps;
Debug.Assert(w <= 64);
Debug.Assert(h <= 64);
Debug.Assert(yStepQ4 <= 32 || (yStepQ4 <= 64 && h <= 32));
Debug.Assert(xStepQ4 <= 64);
ConvolveHoriz(src - srcStride * (SubpelTaps / 2 - 1), srcStride, temp, 64, filter, x0Q4, xStepQ4, w, intermediateHeight);
ConvolveVert(temp + 64 * (SubpelTaps / 2 - 1), 64, dst, dstStride, filter, y0Q4, yStepQ4, w, h);
ConvolveHoriz(src - (srcStride * ((SubpelTaps / 2) - 1)), srcStride, temp, 64, filter, x0Q4, xStepQ4, w,
intermediateHeight);
ConvolveVert(temp + (64 * ((SubpelTaps / 2) - 1)), 64, dst, dstStride, filter, y0Q4, yStepQ4, w, h);
}
public static unsafe void Convolve8Avg(
@@ -489,11 +491,9 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
int w,
int h)
{
int x, y;
for (y = 0; y < h; ++y)
for (int y = 0; y < h; ++y)
{
for (x = 0; x < w; ++x)
for (int x = 0; x < w; ++x)
{
dst[x] = (byte)BitUtils.RoundPowerOfTwo(dst[x] + src[x], 1);
}
@@ -611,18 +611,17 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
int h,
int bd)
{
int x, y;
src -= SubpelTaps / 2 - 1;
src -= (SubpelTaps / 2) - 1;
for (y = 0; y < h; ++y)
for (int y = 0; y < h; ++y)
{
int xQ4 = x0Q4;
for (x = 0; x < w; ++x)
for (int x = 0; x < w; ++x)
{
ushort* srcX = &src[xQ4 >> SubpelBits];
ref Array8<short> xFilter = ref xFilters[xQ4 & SubpelMask];
int k, sum = 0;
for (k = 0; k < SubpelTaps; ++k)
int sum = 0;
for (int k = 0; k < SubpelTaps; ++k)
{
sum += srcX[k] * xFilter[k];
}
@@ -630,6 +629,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
dst[x] = BitUtils.ClipPixelHighbd(BitUtils.RoundPowerOfTwo(sum, FilterBits), bd);
xQ4 += xStepQ4;
}
src += srcStride;
dst += dstStride;
}
@@ -647,25 +647,26 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
int h,
int bd)
{
int x, y;
src -= SubpelTaps / 2 - 1;
src -= (SubpelTaps / 2) - 1;
for (y = 0; y < h; ++y)
for (int y = 0; y < h; ++y)
{
int xQ4 = x0Q4;
for (x = 0; x < w; ++x)
for (int x = 0; x < w; ++x)
{
ushort* srcX = &src[xQ4 >> SubpelBits];
ref Array8<short> xFilter = ref xFilters[xQ4 & SubpelMask];
int k, sum = 0;
for (k = 0; k < SubpelTaps; ++k)
int sum = 0;
for (int k = 0; k < SubpelTaps; ++k)
{
sum += srcX[k] * xFilter[k];
}
dst[x] = (ushort)BitUtils.RoundPowerOfTwo(dst[x] + BitUtils.ClipPixelHighbd(BitUtils.RoundPowerOfTwo(sum, FilterBits), bd), 1);
dst[x] = (ushort)BitUtils.RoundPowerOfTwo(
dst[x] + BitUtils.ClipPixelHighbd(BitUtils.RoundPowerOfTwo(sum, FilterBits), bd), 1);
xQ4 += xStepQ4;
}
src += srcStride;
dst += dstStride;
}
@@ -683,18 +684,17 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
int h,
int bd)
{
int x, y;
src -= srcStride * (SubpelTaps / 2 - 1);
src -= srcStride * ((SubpelTaps / 2) - 1);
for (x = 0; x < w; ++x)
for (int x = 0; x < w; ++x)
{
int yQ4 = y0Q4;
for (y = 0; y < h; ++y)
for (int y = 0; y < h; ++y)
{
ushort* srcY = &src[(yQ4 >> SubpelBits) * srcStride];
ref Array8<short> yFilter = ref yFilters[yQ4 & SubpelMask];
int k, sum = 0;
for (k = 0; k < SubpelTaps; ++k)
int sum = 0;
for (int k = 0; k < SubpelTaps; ++k)
{
sum += srcY[k * srcStride] * yFilter[k];
}
@@ -702,6 +702,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
dst[y * dstStride] = BitUtils.ClipPixelHighbd(BitUtils.RoundPowerOfTwo(sum, FilterBits), bd);
yQ4 += yStepQ4;
}
++src;
++dst;
}
@@ -719,26 +720,27 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
int h,
int bd)
{
int x, y;
src -= srcStride * (SubpelTaps / 2 - 1);
src -= srcStride * ((SubpelTaps / 2) - 1);
for (x = 0; x < w; ++x)
for (int x = 0; x < w; ++x)
{
int yQ4 = y0Q4;
for (y = 0; y < h; ++y)
for (int y = 0; y < h; ++y)
{
ushort* srcY = &src[(yQ4 >> SubpelBits) * srcStride];
ref Array8<short> yFilter = ref yFilters[yQ4 & SubpelMask];
int k, sum = 0;
for (k = 0; k < SubpelTaps; ++k)
int sum = 0;
for (int k = 0; k < SubpelTaps; ++k)
{
sum += srcY[k * srcStride] * yFilter[k];
}
dst[y * dstStride] = (ushort)BitUtils.RoundPowerOfTwo(
dst[y * dstStride] + BitUtils.ClipPixelHighbd(BitUtils.RoundPowerOfTwo(sum, FilterBits), bd), 1);
dst[y * dstStride] + BitUtils.ClipPixelHighbd(BitUtils.RoundPowerOfTwo(sum, FilterBits), bd),
1);
yQ4 += yStepQ4;
}
++src;
++dst;
}
@@ -771,15 +773,17 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
// --Require an additional SubpelTaps rows for the 8-tap filter tails.
// --((64 - 1) * 32 + 15) >> 4 + 8 = 135.
ushort* temp = stackalloc ushort[64 * 135];
int intermediateHeight = (((h - 1) * yStepQ4 + y0Q4) >> SubpelBits) + SubpelTaps;
int intermediateHeight = ((((h - 1) * yStepQ4) + y0Q4) >> SubpelBits) + SubpelTaps;
Debug.Assert(w <= 64);
Debug.Assert(h <= 64);
Debug.Assert(yStepQ4 <= 32);
Debug.Assert(xStepQ4 <= 32);
HighbdConvolveHoriz(src - srcStride * (SubpelTaps / 2 - 1), srcStride, temp, 64, filter, x0Q4, xStepQ4, w, intermediateHeight, bd);
HighbdConvolveVert(temp + 64 * (SubpelTaps / 2 - 1), 64, dst, dstStride, filter, y0Q4, yStepQ4, w, h, bd);
HighbdConvolveHoriz(src - (srcStride * ((SubpelTaps / 2) - 1)), srcStride, temp, 64, filter, x0Q4, xStepQ4,
w, intermediateHeight, bd);
HighbdConvolveVert(temp + (64 * ((SubpelTaps / 2) - 1)), 64, dst, dstStride, filter, y0Q4, yStepQ4, w, h,
bd);
}
public static unsafe void HighbdConvolve8Horiz(
@@ -928,11 +932,9 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
int h,
int bd)
{
int x, y;
for (y = 0; y < h; ++y)
for (int y = 0; y < h; ++y)
{
for (x = 0; x < w; ++x)
for (int x = 0; x < w; ++x)
{
dst[x] = (ushort)BitUtils.RoundPowerOfTwo(dst[x] + src[x], 1);
}
@@ -942,4 +944,4 @@ namespace Ryujinx.Graphics.Nvdec.Vp9.Dsp
}
}
}
}
}