1  using Ryujinx.Common;
  2  using Ryujinx.Graphics.Device;
  3  using Ryujinx.Graphics.Texture;
  4  using Ryujinx.Graphics.Video;
  5  using System;
  6  using System.Runtime.Intrinsics;
  7  using System.Runtime.Intrinsics.X86;
  8  using static Ryujinx.Graphics.Nvdec.Image.SurfaceCommon;
  9  using static Ryujinx.Graphics.Nvdec.MemoryExtensions;
 10  
 11  namespace Ryujinx.Graphics.Nvdec.Image
 12  {
 13      static class SurfaceWriter
 14      {
 15          public static void Write(DeviceMemoryManager mm, ISurface surface, uint lumaOffset, uint chromaOffset)
 16          {
 17              int lumaSize = GetBlockLinearSize(surface.Width, surface.Height, 1);
 18  
 19              using var luma = mm.GetWritableRegion(ExtendOffset(lumaOffset), lumaSize);
 20  
 21              WriteLuma(
 22                  luma.Memory.Span,
 23                  surface.YPlane.AsSpan(),
 24                  surface.Stride,
 25                  surface.Width,
 26                  surface.Height);
 27  
 28              int chromaSize = GetBlockLinearSize(surface.UvWidth, surface.UvHeight, 2);
 29  
 30              using var chroma = mm.GetWritableRegion(ExtendOffset(chromaOffset), chromaSize);
 31  
 32              WriteChroma(
 33                  chroma.Memory.Span,
 34                  surface.UPlane.AsSpan(),
 35                  surface.VPlane.AsSpan(),
 36                  surface.UvStride,
 37                  surface.UvWidth,
 38                  surface.UvHeight);
 39          }
 40  
 41          public static void WriteInterlaced(
 42              DeviceMemoryManager mm,
 43              ISurface surface,
 44              uint lumaTopOffset,
 45              uint chromaTopOffset,
 46              uint lumaBottomOffset,
 47              uint chromaBottomOffset)
 48          {
 49              int lumaSize = GetBlockLinearSize(surface.Width, surface.Height / 2, 1);
 50  
 51              using var lumaTop = mm.GetWritableRegion(ExtendOffset(lumaTopOffset), lumaSize);
 52              using var lumaBottom = mm.GetWritableRegion(ExtendOffset(lumaBottomOffset), lumaSize);
 53  
 54              WriteLuma(
 55                  lumaTop.Memory.Span,
 56                  surface.YPlane.AsSpan(),
 57                  surface.Stride * 2,
 58                  surface.Width,
 59                  surface.Height / 2);
 60  
 61              WriteLuma(
 62                  lumaBottom.Memory.Span,
 63                  surface.YPlane.AsSpan()[surface.Stride..],
 64                  surface.Stride * 2,
 65                  surface.Width,
 66                  surface.Height / 2);
 67  
 68              int chromaSize = GetBlockLinearSize(surface.UvWidth, surface.UvHeight / 2, 2);
 69  
 70              using var chromaTop = mm.GetWritableRegion(ExtendOffset(chromaTopOffset), chromaSize);
 71              using var chromaBottom = mm.GetWritableRegion(ExtendOffset(chromaBottomOffset), chromaSize);
 72  
 73              WriteChroma(
 74                  chromaTop.Memory.Span,
 75                  surface.UPlane.AsSpan(),
 76                  surface.VPlane.AsSpan(),
 77                  surface.UvStride * 2,
 78                  surface.UvWidth,
 79                  surface.UvHeight / 2);
 80  
 81              WriteChroma(
 82                  chromaBottom.Memory.Span,
 83                  surface.UPlane.AsSpan()[surface.UvStride..],
 84                  surface.VPlane.AsSpan()[surface.UvStride..],
 85                  surface.UvStride * 2,
 86                  surface.UvWidth,
 87                  surface.UvHeight / 2);
 88          }
 89  
 90          private static void WriteLuma(Span<byte> dst, ReadOnlySpan<byte> src, int srcStride, int width, int height)
 91          {
 92              LayoutConverter.ConvertLinearToBlockLinear(dst, width, height, srcStride, 1, 2, src);
 93          }
 94  
 95          private unsafe static void WriteChroma(
 96              Span<byte> dst,
 97              ReadOnlySpan<byte> srcU,
 98              ReadOnlySpan<byte> srcV,
 99              int srcStride,
100              int width,
101              int height)
102          {
103              OffsetCalculator calc = new(width, height, 0, false, 2, 2);
104  
105              if (Sse2.IsSupported)
106              {
107                  int strideTrunc64 = BitUtils.AlignDown(width * 2, 64);
108  
109                  int inStrideGap = srcStride - width;
110  
111                  fixed (byte* outputPtr = dst, srcUPtr = srcU, srcVPtr = srcV)
112                  {
113                      byte* inUPtr = srcUPtr;
114                      byte* inVPtr = srcVPtr;
115  
116                      for (int y = 0; y < height; y++)
117                      {
118                          calc.SetY(y);
119  
120                          for (int x = 0; x < strideTrunc64; x += 64, inUPtr += 32, inVPtr += 32)
121                          {
122                              byte* offset = outputPtr + calc.GetOffsetWithLineOffset64(x);
123                              byte* offset2 = offset + 0x20;
124                              byte* offset3 = offset + 0x100;
125                              byte* offset4 = offset + 0x120;
126  
127                              Vector128<byte> value = *(Vector128<byte>*)inUPtr;
128                              Vector128<byte> value2 = *(Vector128<byte>*)inVPtr;
129                              Vector128<byte> value3 = *(Vector128<byte>*)(inUPtr + 16);
130                              Vector128<byte> value4 = *(Vector128<byte>*)(inVPtr + 16);
131  
132                              Vector128<byte> uv0 = Sse2.UnpackLow(value, value2);
133                              Vector128<byte> uv1 = Sse2.UnpackHigh(value, value2);
134                              Vector128<byte> uv2 = Sse2.UnpackLow(value3, value4);
135                              Vector128<byte> uv3 = Sse2.UnpackHigh(value3, value4);
136  
137                              *(Vector128<byte>*)offset = uv0;
138                              *(Vector128<byte>*)offset2 = uv1;
139                              *(Vector128<byte>*)offset3 = uv2;
140                              *(Vector128<byte>*)offset4 = uv3;
141                          }
142  
143                          for (int x = strideTrunc64 / 2; x < width; x++, inUPtr++, inVPtr++)
144                          {
145                              byte* offset = outputPtr + calc.GetOffset(x);
146  
147                              *offset = *inUPtr;
148                              *(offset + 1) = *inVPtr;
149                          }
150  
151                          inUPtr += inStrideGap;
152                          inVPtr += inStrideGap;
153                      }
154                  }
155              }
156              else
157              {
158                  for (int y = 0; y < height; y++)
159                  {
160                      int srcBaseOffset = y * srcStride;
161  
162                      calc.SetY(y);
163  
164                      for (int x = 0; x < width; x++)
165                      {
166                          int dstOffset = calc.GetOffset(x);
167  
168                          dst[dstOffset + 0] = srcU[srcBaseOffset + x];
169                          dst[dstOffset + 1] = srcV[srcBaseOffset + x];
170                      }
171                  }
172              }
173          }
174      }
175  }