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

Refactor shader GPU state and memory access (#1203)

Browse source 
* Refactor shader GPU state and memory access

* Fix NVDEC project build

* Address PR feedback and add missing XML comments
This commit is contained in:
gdkchan 2020-05-05 22:02:28 -03:00 • committed by GitHub
parent 7f500e7cae
commit b8eb6abecc
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
35 changed files with 633 additions and 684 deletions
Download patch file Download diff file Expand all files Collapse all files

6
Ryujinx.Graphics.Shader/Translation/Lowering.cs
View file

@ -79,7 +79,7 @@ namespace Ryujinx.Graphics.Shader.Translation
sbSlot = PrependOperation(Instruction.ConditionalSelect, inRange, Const(slot), sbSlot);
}
Operand alignMask = Const(-config.QueryInfo(QueryInfoName.StorageBufferOffsetAlignment));
Operand alignMask = Const(-config.GpuAccessor.QueryStorageBufferOffsetAlignment());
Operand baseAddrTrunc = PrependOperation(Instruction.BitwiseAnd, sbBaseAddrLow, alignMask);
Operand byteOffset = PrependOperation(Instruction.Subtract, addrLow, baseAddrTrunc);
@ -131,9 +131,9 @@ namespace Ryujinx.Graphics.Shader.Translation
bool hasOffset = (texOp.Flags & TextureFlags.Offset) != 0;
bool hasOffsets = (texOp.Flags & TextureFlags.Offsets) != 0;
bool hasInvalidOffset = (hasOffset || hasOffsets) && !config.QueryInfoBool(QueryInfoName.SupportsNonConstantTextureOffset);
bool hasInvalidOffset = (hasOffset || hasOffsets) && !config.GpuAccessor.QuerySupportsNonConstantTextureOffset();
bool isRect = config.QueryInfoBool(QueryInfoName.IsTextureRectangle, texOp.Handle);
bool isRect = config.GpuAccessor.QueryIsTextureRectangle(texOp.Handle);
if (!(hasInvalidOffset || isRect))
{

4
Ryujinx.Graphics.Shader/Translation/Optimizations/GlobalToStorage.cs
View file

@ -69,7 +69,7 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
Operand baseAddrTrunc = Local();
Operand alignMask = Const(-config.QueryInfo(QueryInfoName.StorageBufferOffsetAlignment));
Operand alignMask = Const(-config.GpuAccessor.QueryStorageBufferOffsetAlignment());
Operation andOp = new Operation(Instruction.BitwiseAnd, baseAddrTrunc, baseAddrLow, alignMask);
@ -140,7 +140,7 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
Operand baseAddrTrunc = Local();
Operand alignMask = Const(-config.QueryInfo(QueryInfoName.StorageBufferOffsetAlignment));
Operand alignMask = Const(-config.GpuAccessor.QueryStorageBufferOffsetAlignment());
Operation andOp = new Operation(Instruction.BitwiseAnd, baseAddrTrunc, baseAddrLow, alignMask);

58
Ryujinx.Graphics.Shader/Translation/ShaderConfig.cs
View file

@ -18,11 +18,11 @@ namespace Ryujinx.Graphics.Shader.Translation
public bool OmapSampleMask { get; }
public bool OmapDepth { get; }
public IGpuAccessor GpuAccessor { get; }
public TranslationFlags Flags { get; }
private TranslatorCallbacks _callbacks;
public ShaderConfig(TranslationFlags flags, TranslatorCallbacks callbacks)
public ShaderConfig(IGpuAccessor gpuAccessor, TranslationFlags flags)
{
Stage = ShaderStage.Compute;
OutputTopology = OutputTopology.PointList;
@ -32,11 +32,11 @@ namespace Ryujinx.Graphics.Shader.Translation
OmapTargets = null;
OmapSampleMask = false;
OmapDepth = false;
GpuAccessor = gpuAccessor;
Flags = flags;
_callbacks = callbacks;
}
public ShaderConfig(ShaderHeader header, TranslationFlags flags, TranslatorCallbacks callbacks)
public ShaderConfig(ShaderHeader header, IGpuAccessor gpuAccessor, TranslationFlags flags)
{
Stage = header.Stage;
OutputTopology = header.OutputTopology;
@ -46,8 +46,8 @@ namespace Ryujinx.Graphics.Shader.Translation
OmapTargets = header.OmapTargets;
OmapSampleMask = header.OmapSampleMask;
OmapDepth = header.OmapDepth;
GpuAccessor = gpuAccessor;
Flags = flags;
_callbacks = callbacks;
}
public int GetDepthRegister()
@ -68,51 +68,5 @@ namespace Ryujinx.Graphics.Shader.Translation
// The depth register is always two registers after the last color output.
return count + 1;
}
public bool QueryInfoBool(QueryInfoName info, int index = 0)
{
return Convert.ToBoolean(QueryInfo(info, index));
}
public int QueryInfo(QueryInfoName info, int index = 0)
{
if (_callbacks.QueryInfo != null)
{
return _callbacks.QueryInfo(info, index);
}
else
{
switch (info)
{
case QueryInfoName.ComputeLocalSizeX:
case QueryInfoName.ComputeLocalSizeY:
case QueryInfoName.ComputeLocalSizeZ:
return 1;
case QueryInfoName.ComputeLocalMemorySize:
return 0x1000;
case QueryInfoName.ComputeSharedMemorySize:
return 0xc000;
case QueryInfoName.IsTextureBuffer:
return Convert.ToInt32(false);
case QueryInfoName.IsTextureRectangle:
return Convert.ToInt32(false);
case QueryInfoName.PrimitiveTopology:
return (int)InputTopology.Points;
case QueryInfoName.StorageBufferOffsetAlignment:
return 16;
case QueryInfoName.SupportsNonConstantTextureOffset:
return Convert.ToInt32(true);
case QueryInfoName.TextureFormat:
return (int)TextureFormat.R8G8B8A8Unorm;
}
}
return 0;
}
public void PrintLog(string message)
{
_callbacks.PrintLog?.Invoke(message);
}
}
}

36
Ryujinx.Graphics.Shader/Translation/ShaderHeader.cs
View file

@ -1,6 +1,5 @@
using Ryujinx.Graphics.Shader.Decoders;
using System;
using System.Runtime.InteropServices;
namespace Ryujinx.Graphics.Shader.Translation
{
@ -110,15 +109,13 @@ namespace Ryujinx.Graphics.Shader.Translation
public bool OmapSampleMask { get; }
public bool OmapDepth { get; }
public ShaderHeader(ReadOnlySpan<byte> code)
public ShaderHeader(IGpuAccessor gpuAccessor, ulong address)
{
ReadOnlySpan<int> header = MemoryMarshal.Cast<byte, int>(code);
int commonWord0 = header[0];
int commonWord1 = header[1];
int commonWord2 = header[2];
int commonWord3 = header[3];
int commonWord4 = header[4];
int commonWord0 = gpuAccessor.MemoryRead<int>(address + 0);
int commonWord1 = gpuAccessor.MemoryRead<int>(address + 4);
int commonWord2 = gpuAccessor.MemoryRead<int>(address + 8);
int commonWord3 = gpuAccessor.MemoryRead<int>(address + 12);
int commonWord4 = gpuAccessor.MemoryRead<int>(address + 16);
SphType = commonWord0.Extract(0, 5);
Version = commonWord0.Extract(5, 5);
@ -163,22 +160,19 @@ namespace Ryujinx.Graphics.Shader.Translation
ImapTypes = new ImapPixelType[32];
for (int i = 0; i < 8; i++)
for (ulong i = 0; i < 32; i++)
{
for (int j = 0; j < 4; j++)
{
byte imap = (byte)(header[6 + i] >> (j * 8));
byte imap = gpuAccessor.MemoryRead<byte>(address + 0x18 + i);
ImapTypes[i * 4 + j] = new ImapPixelType(
(PixelImap)((imap >> 0) & 3),
(PixelImap)((imap >> 2) & 3),
(PixelImap)((imap >> 4) & 3),
(PixelImap)((imap >> 6) & 3));
}
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];
int type2OmapTarget = gpuAccessor.MemoryRead<int>(address + 0x48);
int type2Omap = gpuAccessor.MemoryRead<int>(address + 0x4c);
OmapTargets = new OmapTarget[8];

62
Ryujinx.Graphics.Shader/Translation/Translator.cs
View file

@ -14,46 +14,22 @@ namespace Ryujinx.Graphics.Shader.Translation
{
private const int HeaderSize = 0x50;
public static ReadOnlySpan<byte> ExtractCode(ReadOnlySpan<byte> code, bool compute, out int headerSize)
public static ShaderProgram Translate(ulong address, IGpuAccessor gpuAccessor, TranslationFlags flags)
{
headerSize = compute ? 0 : HeaderSize;
Block[] cfg = Decoder.Decode(code, (ulong)headerSize);
if (cfg == null)
{
return code;
}
ulong endAddress = 0;
foreach (Block block in cfg)
{
if (endAddress < block.EndAddress)
{
endAddress = block.EndAddress;
}
}
return code.Slice(0, headerSize + (int)endAddress);
}
public static ShaderProgram Translate(ReadOnlySpan<byte> code, TranslatorCallbacks callbacks, TranslationFlags flags)
{
Operation[] ops = DecodeShader(code, callbacks, flags, out ShaderConfig config, out int size);
Operation[] ops = DecodeShader(address, gpuAccessor, flags, out ShaderConfig config, out int size);
return Translate(ops, config, size);
}
public static ShaderProgram Translate(ReadOnlySpan<byte> vpACode, ReadOnlySpan<byte> vpBCode, TranslatorCallbacks callbacks, TranslationFlags flags)
public static ShaderProgram Translate(ulong addressA, ulong addressB, IGpuAccessor gpuAccessor, TranslationFlags flags)
{
Operation[] vpAOps = DecodeShader(vpACode, callbacks, flags, out _, out _);
Operation[] vpBOps = DecodeShader(vpBCode, callbacks, flags, out ShaderConfig config, out int sizeB);
Operation[] opsA = DecodeShader(addressA, gpuAccessor, flags, out _, out int sizeA);
Operation[] opsB = DecodeShader(addressB, gpuAccessor, flags, out ShaderConfig config, out int sizeB);
return Translate(Combine(vpAOps, vpBOps), config, sizeB);
return Translate(Combine(opsA, opsB), config, sizeB, sizeA);
}
private static ShaderProgram Translate(Operation[] ops, ShaderConfig config, int size)
private static ShaderProgram Translate(Operation[] ops, ShaderConfig config, int size, int sizeA = 0)
{
BasicBlock[] blocks = ControlFlowGraph.MakeCfg(ops);
@ -83,34 +59,34 @@ namespace Ryujinx.Graphics.Shader.Translation
string glslCode = program.Code;
return new ShaderProgram(spInfo, config.Stage, glslCode, size);
return new ShaderProgram(spInfo, config.Stage, glslCode, size, sizeA);
}
private static Operation[] DecodeShader(
ReadOnlySpan<byte> code,
TranslatorCallbacks callbacks,
TranslationFlags flags,
out ShaderConfig config,
out int size)
ulong address,
IGpuAccessor gpuAccessor,
TranslationFlags flags,
out ShaderConfig config,
out int size)
{
Block[] cfg;
if ((flags & TranslationFlags.Compute) != 0)
{
config = new ShaderConfig(flags, callbacks);
config = new ShaderConfig(gpuAccessor, flags);
cfg = Decoder.Decode(code, 0);
cfg = Decoder.Decode(gpuAccessor, address);
}
else
{
config = new ShaderConfig(new ShaderHeader(code), flags, callbacks);
config = new ShaderConfig(new ShaderHeader(gpuAccessor, address), gpuAccessor, flags);
cfg = Decoder.Decode(code, HeaderSize);
cfg = Decoder.Decode(gpuAccessor, address + HeaderSize);
}
if (cfg == null)
{
config.PrintLog("Invalid branch detected, failed to build CFG.");
gpuAccessor.Log("Invalid branch detected, failed to build CFG.");
size = 0;
@ -150,7 +126,7 @@ namespace Ryujinx.Graphics.Shader.Translation
{
instName = "???";
config.PrintLog($"Invalid instruction at 0x{op.Address:X6} (0x{op.RawOpCode:X16}).");
gpuAccessor.Log($"Invalid instruction at 0x{op.Address:X6} (0x{op.RawOpCode:X16}).");
}
string dbgComment = $"0x{op.Address:X6}: 0x{op.RawOpCode:X16} {instName}";

17
Ryujinx.Graphics.Shader/Translation/TranslatorCallbacks.cs
View file

@ -1,17 +0,0 @@
using System;
namespace Ryujinx.Graphics.Shader.Translation
{
public struct TranslatorCallbacks
{
internal Func<QueryInfoName, int, int> QueryInfo { get; }
internal Action<string> PrintLog { get; }
public TranslatorCallbacks(Func<QueryInfoName, int, int> queryInfoCallback, Action<string> printLogCallback)
{
QueryInfo = queryInfoCallback;
PrintLog = printLogCallback;
}
}
}