Mirrors/Ryujinx
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Support compute uniform buffers emulated with global memory (#924)

Browse source
This commit is contained in:
gdkchan 2020-02-10 21:10:05 -03:00 • committed by GitHub
parent 2e6080ccbb
commit 7e4d986a73
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
5 changed files with 133 additions and 16 deletions
Download patch file Download diff file Expand all files Collapse all files

79
Ryujinx.Graphics.Shader/Translation/Optimizations/GlobalToStorage.cs
View file

@ -31,8 +31,27 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
if (storageIndex >= 0)
{
// Storage buffers are implemented using global memory access.
// If we know from where the base address of the access is loaded,
// we can guess which storage buffer it is accessing.
// We can then replace the global memory access with a storage
// buffer access.
node = ReplaceGlobalWithStorage(node, config, storageIndex);
}
else if (config.Stage == ShaderStage.Compute && operation.Inst == Instruction.LoadGlobal)
{
// Here we effectively try to replace a LDG instruction with LDC.
// The hardware only supports a limited amount of constant buffers
// so NVN "emulates" more constant buffers using global memory access.
// Here we try to replace the global access back to a constant buffer
// load.
storageIndex = SearchForStorageBase(asgOperation, UbeBaseOffset, UbeBaseOffset + UbeDescsSize);
if (storageIndex >= 0)
{
node = ReplaceLdgWithLdc(node, config, storageIndex);
}
}
}
}
}
@ -42,8 +61,6 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
{
Operation operation = (Operation)node.Value;
Operation storageOp;
Operand GetStorageOffset()
{
Operand addrLow = operation.GetSource(0);
@ -80,6 +97,8 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
sources[index] = operation.GetSource(index);
}
Operation storageOp;
if (operation.Inst.IsAtomic())
{
Instruction inst = (operation.Inst & ~Instruction.MrMask) | Instruction.MrStorage;
@ -109,6 +128,62 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
return node;
}
private static LinkedListNode<INode> ReplaceLdgWithLdc(LinkedListNode<INode> node, ShaderConfig config, int storageIndex)
{
Operation operation = (Operation)node.Value;
Operand GetCbufOffset()
{
Operand addrLow = operation.GetSource(0);
Operand baseAddrLow = Cbuf(0, UbeBaseOffset + storageIndex * StorageDescSize);
Operand baseAddrTrunc = Local();
Operand alignMask = Const(-config.QueryInfo(QueryInfoName.StorageBufferOffsetAlignment));
Operation andOp = new Operation(Instruction.BitwiseAnd, baseAddrTrunc, baseAddrLow, alignMask);
node.List.AddBefore(node, andOp);
Operand byteOffset = Local();
Operand wordOffset = Local();
Operation subOp = new Operation(Instruction.Subtract, byteOffset, addrLow, baseAddrTrunc);
Operation shrOp = new Operation(Instruction.ShiftRightU32, wordOffset, byteOffset, Const(2));
node.List.AddBefore(node, subOp);
node.List.AddBefore(node, shrOp);
return wordOffset;
}
Operand[] sources = new Operand[operation.SourcesCount];
sources[0] = Const(UbeFirstCbuf + storageIndex);
sources[1] = GetCbufOffset();
for (int index = 2; index < operation.SourcesCount; index++)
{
sources[index] = operation.GetSource(index);
}
Operation ldcOp = new Operation(Instruction.LoadConstant, operation.Dest, sources);
for (int index = 0; index < operation.SourcesCount; index++)
{
operation.SetSource(index, null);
}
LinkedListNode<INode> oldNode = node;
node = node.List.AddBefore(node, ldcOp);
node.List.Remove(oldNode);
return node;
}
private static int SearchForStorageBase(Operation operation, int sbStart, int sbEnd)
{
Queue<Operation> assignments = new Queue<Operation>();