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
2019-04-17 23:57:08 +00:00
|
|
|
using Ryujinx.Graphics.Shader.CodeGen.Glsl;
|
|
|
|
using Ryujinx.Graphics.Shader.Decoders;
|
|
|
|
using Ryujinx.Graphics.Shader.IntermediateRepresentation;
|
|
|
|
using Ryujinx.Graphics.Shader.StructuredIr;
|
|
|
|
using Ryujinx.Graphics.Shader.Translation.Optimizations;
|
2019-10-13 06:02:07 +00:00
|
|
|
using System;
|
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
2019-04-17 23:57:08 +00:00
|
|
|
using System.Collections.Generic;
|
|
|
|
|
|
|
|
using static Ryujinx.Graphics.Shader.IntermediateRepresentation.OperandHelper;
|
|
|
|
|
|
|
|
namespace Ryujinx.Graphics.Shader.Translation
|
|
|
|
{
|
|
|
|
public static class Translator
|
|
|
|
{
|
2019-10-13 06:02:07 +00:00
|
|
|
private const int HeaderSize = 0x50;
|
|
|
|
|
2020-05-06 01:02:28 +00:00
|
|
|
public static ShaderProgram Translate(ulong address, IGpuAccessor gpuAccessor, TranslationFlags flags)
|
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
2019-04-17 23:57:08 +00:00
|
|
|
{
|
2020-07-07 02:41:07 +00:00
|
|
|
Operation[] ops = DecodeShader(address, gpuAccessor, flags, out ShaderConfig config, out int size, out FeatureFlags featureFlags);
|
|
|
|
|
|
|
|
config.UsedFeatures = featureFlags;
|
2019-10-13 06:02:07 +00:00
|
|
|
|
2019-11-14 17:20:30 +00:00
|
|
|
return Translate(ops, config, size);
|
2019-11-08 20:29:41 +00:00
|
|
|
}
|
|
|
|
|
2020-05-06 01:02:28 +00:00
|
|
|
public static ShaderProgram Translate(ulong addressA, ulong addressB, IGpuAccessor gpuAccessor, TranslationFlags flags)
|
2019-11-08 20:29:41 +00:00
|
|
|
{
|
2020-07-07 02:41:07 +00:00
|
|
|
Operation[] opsA = DecodeShader(addressA, gpuAccessor, flags | TranslationFlags.VertexA, out _, out int sizeA, out FeatureFlags featureFlagsA);
|
|
|
|
Operation[] opsB = DecodeShader(addressB, gpuAccessor, flags, out ShaderConfig config, out int sizeB, out FeatureFlags featureFlagsB);
|
|
|
|
|
|
|
|
config.UsedFeatures = featureFlagsA | featureFlagsB;
|
2019-11-08 20:29:41 +00:00
|
|
|
|
2020-05-06 01:02:28 +00:00
|
|
|
return Translate(Combine(opsA, opsB), config, sizeB, sizeA);
|
2019-11-08 20:29:41 +00:00
|
|
|
}
|
|
|
|
|
2020-05-06 01:02:28 +00:00
|
|
|
private static ShaderProgram Translate(Operation[] ops, ShaderConfig config, int size, int sizeA = 0)
|
2019-11-08 20:29:41 +00:00
|
|
|
{
|
2019-12-01 02:53:09 +00:00
|
|
|
BasicBlock[] blocks = ControlFlowGraph.MakeCfg(ops);
|
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
2019-04-17 23:57:08 +00:00
|
|
|
|
2019-12-01 02:53:09 +00:00
|
|
|
if (blocks.Length > 0)
|
2019-11-15 03:01:54 +00:00
|
|
|
{
|
2019-12-01 02:53:09 +00:00
|
|
|
Dominance.FindDominators(blocks[0], blocks.Length);
|
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
2019-04-17 23:57:08 +00:00
|
|
|
|
2019-12-01 02:53:09 +00:00
|
|
|
Dominance.FindDominanceFrontiers(blocks);
|
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
2019-04-17 23:57:08 +00:00
|
|
|
|
2019-12-01 02:53:09 +00:00
|
|
|
Ssa.Rename(blocks);
|
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
2019-04-17 23:57:08 +00:00
|
|
|
|
2019-12-01 02:53:09 +00:00
|
|
|
Optimizer.RunPass(blocks, config);
|
|
|
|
|
|
|
|
Lowering.RunPass(blocks, config);
|
2019-11-15 03:01:54 +00:00
|
|
|
}
|
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
2019-04-17 23:57:08 +00:00
|
|
|
|
2019-12-01 02:53:09 +00:00
|
|
|
StructuredProgramInfo sInfo = StructuredProgram.MakeStructuredProgram(blocks, config);
|
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
2019-04-17 23:57:08 +00:00
|
|
|
|
|
|
|
GlslProgram program = GlslGenerator.Generate(sInfo, config);
|
|
|
|
|
|
|
|
ShaderProgramInfo spInfo = new ShaderProgramInfo(
|
|
|
|
program.CBufferDescriptors,
|
2019-10-13 06:02:07 +00:00
|
|
|
program.SBufferDescriptors,
|
|
|
|
program.TextureDescriptors,
|
2019-10-18 02:41:18 +00:00
|
|
|
program.ImageDescriptors,
|
2019-10-13 06:02:07 +00:00
|
|
|
sInfo.UsesInstanceId);
|
|
|
|
|
|
|
|
string glslCode = program.Code;
|
|
|
|
|
2020-05-06 01:02:28 +00:00
|
|
|
return new ShaderProgram(spInfo, config.Stage, glslCode, size, sizeA);
|
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
2019-04-17 23:57:08 +00:00
|
|
|
}
|
|
|
|
|
2019-11-14 17:20:30 +00:00
|
|
|
private static Operation[] DecodeShader(
|
2020-05-06 01:02:28 +00:00
|
|
|
ulong address,
|
|
|
|
IGpuAccessor gpuAccessor,
|
|
|
|
TranslationFlags flags,
|
|
|
|
out ShaderConfig config,
|
2020-07-07 02:41:07 +00:00
|
|
|
out int size,
|
|
|
|
out FeatureFlags featureFlags)
|
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
2019-04-17 23:57:08 +00:00
|
|
|
{
|
2019-10-13 06:02:07 +00:00
|
|
|
Block[] cfg;
|
|
|
|
|
2019-12-06 22:37:00 +00:00
|
|
|
if ((flags & TranslationFlags.Compute) != 0)
|
2019-10-13 06:02:07 +00:00
|
|
|
{
|
2020-05-06 01:02:28 +00:00
|
|
|
config = new ShaderConfig(gpuAccessor, flags);
|
2019-10-13 06:02:07 +00:00
|
|
|
|
2020-05-06 01:02:28 +00:00
|
|
|
cfg = Decoder.Decode(gpuAccessor, address);
|
2019-10-13 06:02:07 +00:00
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
2020-05-06 01:02:28 +00:00
|
|
|
config = new ShaderConfig(new ShaderHeader(gpuAccessor, address), gpuAccessor, flags);
|
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
2019-04-17 23:57:08 +00:00
|
|
|
|
2020-05-06 01:02:28 +00:00
|
|
|
cfg = Decoder.Decode(gpuAccessor, address + HeaderSize);
|
2019-10-13 06:02:07 +00:00
|
|
|
}
|
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
2019-04-17 23:57:08 +00:00
|
|
|
|
2019-11-15 03:01:54 +00:00
|
|
|
if (cfg == null)
|
|
|
|
{
|
2020-05-06 01:02:28 +00:00
|
|
|
gpuAccessor.Log("Invalid branch detected, failed to build CFG.");
|
2019-11-15 23:41:06 +00:00
|
|
|
|
2019-11-15 03:01:54 +00:00
|
|
|
size = 0;
|
|
|
|
|
2020-07-07 02:41:07 +00:00
|
|
|
featureFlags = FeatureFlags.None;
|
|
|
|
|
2020-01-01 15:39:09 +00:00
|
|
|
return Array.Empty<Operation>();
|
2019-11-15 03:01:54 +00:00
|
|
|
}
|
|
|
|
|
2019-12-16 04:59:46 +00:00
|
|
|
EmitterContext context = new EmitterContext(config);
|
|
|
|
|
2019-11-14 17:20:30 +00:00
|
|
|
ulong maxEndAddress = 0;
|
|
|
|
|
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
2019-04-17 23:57:08 +00:00
|
|
|
for (int blkIndex = 0; blkIndex < cfg.Length; blkIndex++)
|
|
|
|
{
|
|
|
|
Block block = cfg[blkIndex];
|
|
|
|
|
2019-11-14 17:20:30 +00:00
|
|
|
if (maxEndAddress < block.EndAddress)
|
|
|
|
{
|
|
|
|
maxEndAddress = block.EndAddress;
|
|
|
|
}
|
|
|
|
|
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
2019-04-17 23:57:08 +00:00
|
|
|
context.CurrBlock = block;
|
|
|
|
|
|
|
|
context.MarkLabel(context.GetLabel(block.Address));
|
|
|
|
|
|
|
|
for (int opIndex = 0; opIndex < block.OpCodes.Count; opIndex++)
|
|
|
|
{
|
|
|
|
OpCode op = block.OpCodes[opIndex];
|
|
|
|
|
2019-12-06 22:37:00 +00:00
|
|
|
if ((flags & TranslationFlags.DebugMode) != 0)
|
2019-10-13 06:02:07 +00:00
|
|
|
{
|
|
|
|
string instName;
|
|
|
|
|
|
|
|
if (op.Emitter != null)
|
|
|
|
{
|
|
|
|
instName = op.Emitter.Method.Name;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
instName = "???";
|
2020-01-01 15:39:09 +00:00
|
|
|
|
2020-05-06 01:02:28 +00:00
|
|
|
gpuAccessor.Log($"Invalid instruction at 0x{op.Address:X6} (0x{op.RawOpCode:X16}).");
|
2019-10-13 06:02:07 +00:00
|
|
|
}
|
|
|
|
|
2019-11-14 17:20:30 +00:00
|
|
|
string dbgComment = $"0x{op.Address:X6}: 0x{op.RawOpCode:X16} {instName}";
|
2019-10-13 06:02:07 +00:00
|
|
|
|
|
|
|
context.Add(new CommentNode(dbgComment));
|
|
|
|
}
|
|
|
|
|
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
2019-04-17 23:57:08 +00:00
|
|
|
if (op.NeverExecute)
|
|
|
|
{
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
Operand predSkipLbl = null;
|
|
|
|
|
2019-11-27 03:38:56 +00:00
|
|
|
bool skipPredicateCheck = op is OpCodeBranch opBranch && !opBranch.PushTarget;
|
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
2019-04-17 23:57:08 +00:00
|
|
|
|
2019-11-14 17:20:30 +00:00
|
|
|
if (op is OpCodeBranchPop opBranchPop)
|
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
2019-04-17 23:57:08 +00:00
|
|
|
{
|
2019-11-27 03:38:56 +00:00
|
|
|
// If the instruction is a SYNC or BRK instruction with only one
|
2019-07-02 02:39:22 +00:00
|
|
|
// 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.
|
2019-11-27 03:38:56 +00:00
|
|
|
skipPredicateCheck = opBranchPop.Targets.Count < 2;
|
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
2019-04-17 23:57:08 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
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;
|
|
|
|
|
2020-01-01 15:39:09 +00:00
|
|
|
op.Emitter?.Invoke(context);
|
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
2019-04-17 23:57:08 +00:00
|
|
|
|
|
|
|
if (predSkipLbl != null)
|
|
|
|
{
|
|
|
|
context.MarkLabel(predSkipLbl);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2019-12-06 22:37:00 +00:00
|
|
|
size = (int)maxEndAddress + (((flags & TranslationFlags.Compute) != 0) ? 0 : HeaderSize);
|
2019-11-14 17:20:30 +00:00
|
|
|
|
2020-07-07 02:41:07 +00:00
|
|
|
featureFlags = context.UsedFeatures;
|
|
|
|
|
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
2019-04-17 23:57:08 +00:00
|
|
|
return context.GetOperations();
|
|
|
|
}
|
|
|
|
|
|
|
|
private static Operation[] Combine(Operation[] a, Operation[] b)
|
|
|
|
{
|
2019-07-02 02:39:22 +00:00
|
|
|
// 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.
|
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
2019-04-17 23:57:08 +00:00
|
|
|
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;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|