private static void FillBlock(ShaderConfig config, Block block, ulong limitAddress, ulong startAddress)
{
IGpuAccessor gpuAccessor = config.GpuAccessor;
ulong address = block.Address;
int bufferOffset = 0;
ReadOnlySpan <ulong> buffer = ReadOnlySpan <ulong> .Empty;
InstOp op = default;
do
{
if (address + 7 >= limitAddress)
{
break;
}
// Ignore scheduling instructions, which are written every 32 bytes.
if ((address & 0x1f) == 0)
{
address += 8;
bufferOffset++;
continue;
}
if (bufferOffset >= buffer.Length)
{
buffer = gpuAccessor.GetCode(startAddress + address, 8);
bufferOffset = 0;
}
ulong opCode = buffer[bufferOffset++];
op = InstTable.GetOp(address, opCode);
if (op.Props.HasFlag(InstProps.TexB))
{
config.SetUsedFeature(FeatureFlags.Bindless);
}
if (op.Name == InstName.Ald || op.Name == InstName.Ast || op.Name == InstName.Ipa)
{
SetUserAttributeUses(config, op.Name, opCode);
}
else if (op.Name == InstName.Ssy || op.Name == InstName.Pbk)
{
block.AddPushOp(op);
}
block.OpCodes.Add(op);
address += 8;
}while (!op.Props.HasFlag(InstProps.Bra));
block.EndAddress = address;
}
public ShaderHeader(IGpuAccessor gpuAccessor, ulong address)
{
ReadOnlySpan <int> header = MemoryMarshal.Cast <ulong, int>(gpuAccessor.GetCode(address, 0x50));
int commonWord0 = header[0];
int commonWord1 = header[1];
int commonWord2 = header[2];
int commonWord3 = header[3];
int commonWord4 = header[4];
SphType = commonWord0.Extract(0, 5);
Version = commonWord0.Extract(5, 5);
Stage = (ShaderStage)commonWord0.Extract(10, 4);
// Invalid.
if (Stage == ShaderStage.Compute)
{
Stage = ShaderStage.Vertex;
}
MrtEnable = commonWord0.Extract(14);
KillsPixels = commonWord0.Extract(15);
DoesGlobalStore = commonWord0.Extract(16);
SassVersion = commonWord0.Extract(17, 4);
GpPassthrough = commonWord0.Extract(24);
DoesLoadOrStore = commonWord0.Extract(26);
DoesFp64 = commonWord0.Extract(27);
StreamOutMask = commonWord0.Extract(28, 4);
ShaderLocalMemoryLowSize = commonWord1.Extract(0, 24);
PerPatchAttributeCount = commonWord1.Extract(24, 8);
ShaderLocalMemoryHighSize = commonWord2.Extract(0, 24);
ThreadsPerInputPrimitive = commonWord2.Extract(24, 8);
ShaderLocalMemoryCrsSize = commonWord3.Extract(0, 24);
OutputTopology = (OutputTopology)commonWord3.Extract(24, 4);
MaxOutputVertexCount = commonWord4.Extract(0, 12);
StoreReqStart = commonWord4.Extract(12, 8);
StoreReqEnd = commonWord4.Extract(24, 8);
ImapTypes = new ImapPixelType[32];
for (int i = 0; i < 32; i++)
{
byte imap = (byte)(header[6 + (i >> 2)] >> ((i & 3) * 8));
ImapTypes[i] = new ImapPixelType(
(PixelImap)((imap >> 0) & 3),
(PixelImap)((imap >> 2) & 3),
(PixelImap)((imap >> 4) & 3),
(PixelImap)((imap >> 6) & 3));
}
int type2OmapTarget = header[18];
int type2Omap = header[19];
OmapTargets = type2OmapTarget;
OmapSampleMask = type2Omap.Extract(0);
OmapDepth = type2Omap.Extract(1);
}