From 6b23a2c125b9c48b5ebea92716004ef68698bb0f Mon Sep 17 00:00:00 2001 From: gdkchan Date: Wed, 17 Apr 2019 20:57:08 -0300 Subject: 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 --- .../Shader/Translation/Optimizations/Optimizer.cs | 172 +++++++++++++++++++++ 1 file changed, 172 insertions(+) create mode 100644 Ryujinx.Graphics/Shader/Translation/Optimizations/Optimizer.cs (limited to 'Ryujinx.Graphics/Shader/Translation/Optimizations/Optimizer.cs') diff --git a/Ryujinx.Graphics/Shader/Translation/Optimizations/Optimizer.cs b/Ryujinx.Graphics/Shader/Translation/Optimizations/Optimizer.cs new file mode 100644 index 00000000..88118e3a --- /dev/null +++ b/Ryujinx.Graphics/Shader/Translation/Optimizations/Optimizer.cs @@ -0,0 +1,172 @@ +using Ryujinx.Graphics.Shader.IntermediateRepresentation; +using System.Collections.Generic; +using System.Linq; + +namespace Ryujinx.Graphics.Shader.Translation.Optimizations +{ + static class Optimizer + { + public static void Optimize(BasicBlock[] blocks) + { + bool modified; + + do + { + modified = false; + + for (int blkIndex = 0; blkIndex < blocks.Length; blkIndex++) + { + BasicBlock block = blocks[blkIndex]; + + LinkedListNode node = block.Operations.First; + + while (node != null) + { + LinkedListNode nextNode = node.Next; + + bool isUnused = IsUnused(node.Value); + + if (!(node.Value is Operation operation) || isUnused) + { + if (isUnused) + { + RemoveNode(block, node); + + modified = true; + } + + node = nextNode; + + continue; + } + + ConstantFolding.Fold(operation); + + Simplification.Simplify(operation); + + if (DestIsLocalVar(operation)) + { + if (operation.Inst == Instruction.Copy) + { + PropagateCopy(operation); + + RemoveNode(block, node); + + modified = true; + } + else if (operation.Inst == Instruction.PackHalf2x16 && PropagatePack(operation)) + { + if (operation.Dest.UseOps.Count == 0) + { + RemoveNode(block, node); + } + + modified = true; + } + } + + node = nextNode; + } + + if (BranchElimination.Eliminate(block)) + { + RemoveNode(block, block.Operations.Last); + + modified = true; + } + } + } + while (modified); + } + + private static void PropagateCopy(Operation copyOp) + { + //Propagate copy source operand to all uses of + //the destination operand. + Operand dest = copyOp.Dest; + Operand src = copyOp.GetSource(0); + + INode[] uses = dest.UseOps.ToArray(); + + foreach (INode useNode in uses) + { + for (int index = 0; index < useNode.SourcesCount; index++) + { + if (useNode.GetSource(index) == dest) + { + useNode.SetSource(index, src); + } + } + } + } + + private static bool PropagatePack(Operation packOp) + { + //Propagate pack source operands to uses by unpack + //instruction. The source depends on the unpack instruction. + bool modified = false; + + Operand dest = packOp.Dest; + Operand src0 = packOp.GetSource(0); + Operand src1 = packOp.GetSource(1); + + INode[] uses = dest.UseOps.ToArray(); + + foreach (INode useNode in uses) + { + if (!(useNode is Operation operation) || operation.Inst != Instruction.UnpackHalf2x16) + { + continue; + } + + if (operation.GetSource(0) == dest) + { + operation.TurnIntoCopy(operation.ComponentIndex == 1 ? src1 : src0); + + modified = true; + } + } + + return modified; + } + + private static void RemoveNode(BasicBlock block, LinkedListNode llNode) + { + //Remove a node from the nodes list, and also remove itself + //from all the use lists on the operands that this node uses. + block.Operations.Remove(llNode); + + Queue nodes = new Queue(); + + nodes.Enqueue(llNode.Value); + + while (nodes.TryDequeue(out INode node)) + { + for (int index = 0; index < node.SourcesCount; index++) + { + Operand src = node.GetSource(index); + + if (src.Type != OperandType.LocalVariable) + { + continue; + } + + if (src.UseOps.Remove(node) && src.UseOps.Count == 0) + { + nodes.Enqueue(src.AsgOp); + } + } + } + } + + private static bool IsUnused(INode node) + { + return DestIsLocalVar(node) && node.Dest.UseOps.Count == 0; + } + + private static bool DestIsLocalVar(INode node) + { + return node.Dest != null && node.Dest.Type == OperandType.LocalVariable; + } + } +} \ No newline at end of file -- cgit v1.2.3