1
0
Fork 0
mirror of https://github.com/Ryujinx/Ryujinx.git synced 2024-10-01 12:30:00 +02:00
Ryujinx/Ryujinx.Graphics/Shader/Translation/Translator.cs
gdkchan 6b23a2c125 New shader translator implementation (#654)
* Start implementing a new shader translator

* Fix shift instructions and a typo

* Small refactoring on StructuredProgram, move RemovePhis method to a separate class

* Initial geometry shader support

* Implement TLD4

* Fix -- There's no negation on FMUL32I

* Add constant folding and algebraic simplification optimizations, nits

* Some leftovers from constant folding

* Avoid cast for constant assignments

* Add a branch elimination pass, and misc small fixes

* Remove redundant branches, add expression propagation and other improvements on the code

* Small leftovers -- add missing break and continue, remove unused properties, other improvements

* Add null check to handle empty block cases on block visitor

* Add HADD2 and HMUL2 half float shader instructions

* Optimize pack/unpack sequences, some fixes related to half float instructions

* Add TXQ, TLD, TLDS and TLD4S shader texture instructions, and some support for bindless textures, some refactoring on codegen

* Fix copy paste mistake that caused RZ to be ignored on the AST instruction

* Add workaround for conditional exit, and fix half float instruction with constant buffer

* Add missing 0.0 source for TLDS.LZ variants

* Simplify the switch for TLDS.LZ

* Texture instructions related fixes

* Implement the HFMA instruction, and some misc. fixes

* Enable constant folding on UnpackHalf2x16 instructions

* Refactor HFMA to use OpCode* for opcode decoding rather than on the helper methods

* Remove the old shader translator

* Remove ShaderDeclInfo and other unused things

* Add dual vertex shader support

* Add ShaderConfig, used to pass shader type and maximum cbuffer size

* Move and rename some instruction enums

* Move texture instructions into a separate file

* Move operand GetExpression and locals management to OperandManager

* Optimize opcode decoding using a simple list and binary search

* Add missing condition for do-while on goto elimination

* Misc. fixes on texture instructions

* Simplify TLDS switch

* Address PR feedback, and a nit
2019-04-18 09:57:08 +10:00

219 lines
No EOL
7.3 KiB
C#

using Ryujinx.Graphics.Gal;
using Ryujinx.Graphics.Shader.CodeGen.Glsl;
using Ryujinx.Graphics.Shader.Decoders;
using Ryujinx.Graphics.Shader.Instructions;
using Ryujinx.Graphics.Shader.IntermediateRepresentation;
using Ryujinx.Graphics.Shader.StructuredIr;
using Ryujinx.Graphics.Shader.Translation.Optimizations;
using System.Collections.Generic;
using static Ryujinx.Graphics.Shader.IntermediateRepresentation.OperandHelper;
namespace Ryujinx.Graphics.Shader.Translation
{
public static class Translator
{
public static ShaderProgram Translate(IGalMemory memory, ulong address, ShaderConfig config)
{
return Translate(memory, address, 0, config);
}
public static ShaderProgram Translate(
IGalMemory memory,
ulong address,
ulong addressB,
ShaderConfig config)
{
Operation[] shaderOps = DecodeShader(memory, address, config.Type);
if (addressB != 0)
{
//Dual vertex shader.
Operation[] shaderOpsB = DecodeShader(memory, addressB, config.Type);
shaderOps = Combine(shaderOps, shaderOpsB);
}
BasicBlock[] irBlocks = ControlFlowGraph.MakeCfg(shaderOps);
Dominance.FindDominators(irBlocks[0], irBlocks.Length);
Dominance.FindDominanceFrontiers(irBlocks);
Ssa.Rename(irBlocks);
Optimizer.Optimize(irBlocks);
StructuredProgramInfo sInfo = StructuredProgram.MakeStructuredProgram(irBlocks);
GlslProgram program = GlslGenerator.Generate(sInfo, config);
ShaderProgramInfo spInfo = new ShaderProgramInfo(
program.CBufferDescriptors,
program.TextureDescriptors);
return new ShaderProgram(spInfo, program.Code);
}
private static Operation[] DecodeShader(IGalMemory memory, ulong address, GalShaderType shaderType)
{
ShaderHeader header = new ShaderHeader(memory, address);
Block[] cfg = Decoder.Decode(memory, address);
EmitterContext context = new EmitterContext(shaderType, header);
for (int blkIndex = 0; blkIndex < cfg.Length; blkIndex++)
{
Block block = cfg[blkIndex];
context.CurrBlock = block;
context.MarkLabel(context.GetLabel(block.Address));
for (int opIndex = 0; opIndex < block.OpCodes.Count; opIndex++)
{
OpCode op = block.OpCodes[opIndex];
if (op.NeverExecute)
{
continue;
}
Operand predSkipLbl = null;
bool skipPredicateCheck = op.Emitter == InstEmit.Bra;
if (op is OpCodeSync opSync)
{
//If the instruction is a SYNC instruction with only one
//possible target address, then the instruction is basically
//just a simple branch, we can generate code similar to branch
//instructions, with the condition check on the branch itself.
skipPredicateCheck |= opSync.Targets.Count < 2;
}
if (!(op.Predicate.IsPT || skipPredicateCheck))
{
Operand label;
if (opIndex == block.OpCodes.Count - 1 && block.Next != null)
{
label = context.GetLabel(block.Next.Address);
}
else
{
label = Label();
predSkipLbl = label;
}
Operand pred = Register(op.Predicate);
if (op.InvertPredicate)
{
context.BranchIfTrue(label, pred);
}
else
{
context.BranchIfFalse(label, pred);
}
}
context.CurrOp = op;
op.Emitter(context);
if (predSkipLbl != null)
{
context.MarkLabel(predSkipLbl);
}
}
}
return context.GetOperations();
}
private static Operation[] Combine(Operation[] a, Operation[] b)
{
//Here we combine two shaders.
//For shader A:
//- All user attribute stores on shader A are turned into copies to a
//temporary variable. It's assumed that shader B will consume them.
//- All return instructions are turned into branch instructions, the
//branch target being the start of the shader B code.
//For shader B:
//- All user attribute loads on shader B are turned into copies from a
//temporary variable, as long that attribute is written by shader A.
List<Operation> output = new List<Operation>(a.Length + b.Length);
Operand[] temps = new Operand[AttributeConsts.UserAttributesCount * 4];
Operand lblB = Label();
for (int index = 0; index < a.Length; index++)
{
Operation operation = a[index];
if (IsUserAttribute(operation.Dest))
{
int tIndex = (operation.Dest.Value - AttributeConsts.UserAttributeBase) / 4;
Operand temp = temps[tIndex];
if (temp == null)
{
temp = Local();
temps[tIndex] = temp;
}
operation.Dest = temp;
}
if (operation.Inst == Instruction.Return)
{
output.Add(new Operation(Instruction.Branch, lblB));
}
else
{
output.Add(operation);
}
}
output.Add(new Operation(Instruction.MarkLabel, lblB));
for (int index = 0; index < b.Length; index++)
{
Operation operation = b[index];
for (int srcIndex = 0; srcIndex < operation.SourcesCount; srcIndex++)
{
Operand src = operation.GetSource(srcIndex);
if (IsUserAttribute(src))
{
Operand temp = temps[(src.Value - AttributeConsts.UserAttributeBase) / 4];
if (temp != null)
{
operation.SetSource(srcIndex, temp);
}
}
}
output.Add(operation);
}
return output.ToArray();
}
private static bool IsUserAttribute(Operand operand)
{
return operand != null &&
operand.Type == OperandType.Attribute &&
operand.Value >= AttributeConsts.UserAttributeBase &&
operand.Value < AttributeConsts.UserAttributeEnd;
}
}
}