Improve dma copies on the gpu

This commit is contained in:
gdkchan 2018-08-14 15:47:43 -03:00
parent 17f54b5d78
commit 6e4f43446b
3 changed files with 33 additions and 5 deletions

View file

@ -691,6 +691,19 @@ namespace ChocolArm64.Memory
Marshal.Copy(Data, 0, (IntPtr)(RamPtr + (uint)Position), Data.Length);
}
public void CopyBytes(long Src, long Dst, long Size)
{
EnsureRangeIsValid(Src, Size, AMemoryPerm.Read);
EnsureRangeIsValid(Dst, Size, AMemoryPerm.Write);
Buffer.MemoryCopy(
RamPtr + (uint)Src,
RamPtr + (uint)Dst,
Size,
Size);
}
private void EnsureRangeIsValid(long Position, long Size, AMemoryPerm Perm)
{
long EndPos = Position + Size;

View file

@ -58,6 +58,9 @@ namespace Ryujinx.HLE.Gpu.Engines
int SrcPitch = ReadRegister(NvGpuEngineDmaReg.SrcPitch);
int DstPitch = ReadRegister(NvGpuEngineDmaReg.DstPitch);
int XCount = ReadRegister(NvGpuEngineDmaReg.XCount);
int YCount = ReadRegister(NvGpuEngineDmaReg.YCount);
int DstBlkDim = ReadRegister(NvGpuEngineDmaReg.DstBlkDim);
int DstSizeX = ReadRegister(NvGpuEngineDmaReg.DstSizeX);
int DstSizeY = ReadRegister(NvGpuEngineDmaReg.DstSizeY);
@ -103,8 +106,17 @@ namespace Ryujinx.HLE.Gpu.Engines
DstSwizzle = new BlockLinearSwizzle(DstSizeX, 1, DstBlockHeight);
}
for (int Y = 0; Y < DstSizeY; Y++)
for (int X = 0; X < DstSizeX; X++)
if (SrcLinear == DstLinear)
{
long Src = Vmm.GetPhysicalAddress(SrcAddress);
long Dst = Vmm.GetPhysicalAddress(DstAddress);
Vmm.Memory.CopyBytes(Src, Dst, XCount * YCount);
}
else
{
for (int Y = 0; Y < YCount; Y++)
for (int X = 0; X < XCount; X++)
{
long SrcOffset = SrcAddress + (uint)SrcSwizzle.GetSwizzleOffset(X, Y);
long DstOffset = DstAddress + (uint)DstSwizzle.GetSwizzleOffset(X, Y);
@ -112,6 +124,7 @@ namespace Ryujinx.HLE.Gpu.Engines
Vmm.WriteByte(DstOffset, Vmm.ReadByte(SrcOffset));
}
}
}
private long MakeInt64From2xInt32(NvGpuEngineDmaReg Reg)
{

View file

@ -6,6 +6,8 @@ namespace Ryujinx.HLE.Gpu.Engines
DstAddress = 0x102,
SrcPitch = 0x104,
DstPitch = 0x105,
XCount = 0x106,
YCount = 0x107,
DstBlkDim = 0x1c3,
DstSizeX = 0x1c4,
DstSizeY = 0x1c5,