Reactor: Rework optimization flags.
Replace the old 'runOptimizations' flag on Nucleus::acquireRoutine() and Nucleus::acquireCoroutine() with a new OptimizationLevel enumerator.
runOptimizations was always true, and there is already the rr::optimization extern to control the optimization passes performed.
All Function and Coroutines are now passed the new vk::ReactorOptimizationLevel. This will be changed to a non-Default value in another CL.
Bug: b/135609394
Change-Id: I1154da05d413b18a471a3818fbb03f356a3d0e6c
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33482
Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Ben Clayton <bclayton@google.com>
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>
diff --git a/src/Device/Blitter.cpp b/src/Device/Blitter.cpp
index 30bcf94..bdcfe0a 100644
--- a/src/Device/Blitter.cpp
+++ b/src/Device/Blitter.cpp
@@ -1534,7 +1534,7 @@
}
}
- return function("BlitRoutine");
+ return function(vk::ReactorOptimizationLevel, "BlitRoutine");
}
Routine *Blitter::getBlitRoutine(const State &state)
@@ -1889,7 +1889,7 @@
}
}
- return function("BlitRoutine");
+ return function(vk::ReactorOptimizationLevel, "BlitRoutine");
}
void Blitter::updateBorders(vk::Image* image, const VkImageSubresourceLayers& subresourceLayers)
diff --git a/src/Device/PixelProcessor.cpp b/src/Device/PixelProcessor.cpp
index 898da83..51f4517 100644
--- a/src/Device/PixelProcessor.cpp
+++ b/src/Device/PixelProcessor.cpp
@@ -238,7 +238,7 @@
{
QuadRasterizer *generator = new PixelProgram(state, pipelineLayout, pixelShader, descriptorSets);
generator->generate();
- routine = (*generator)("PixelRoutine_%0.8X", state.shaderID);
+ routine = (*generator)(vk::ReactorOptimizationLevel, "PixelRoutine_%0.8X", state.shaderID);
delete generator;
routineCache->add(state, routine);
diff --git a/src/Device/VertexProcessor.cpp b/src/Device/VertexProcessor.cpp
index 2f489bb..76daf76 100644
--- a/src/Device/VertexProcessor.cpp
+++ b/src/Device/VertexProcessor.cpp
@@ -105,7 +105,7 @@
{
VertexRoutine *generator = new VertexProgram(state, pipelineLayout, vertexShader, descriptorSets);
generator->generate();
- routine = (*generator)("VertexRoutine_%0.8X", state.shaderID);
+ routine = (*generator)(vk::ReactorOptimizationLevel, "VertexRoutine_%0.8X", state.shaderID);
delete generator;
routineCache->add(state, routine);
diff --git a/src/Pipeline/SetupRoutine.cpp b/src/Pipeline/SetupRoutine.cpp
index a87aac6..1973a71 100644
--- a/src/Pipeline/SetupRoutine.cpp
+++ b/src/Pipeline/SetupRoutine.cpp
@@ -453,7 +453,7 @@
Return(1);
}
- routine = function("SetupRoutine");
+ routine = function(vk::ReactorOptimizationLevel, "SetupRoutine");
}
void SetupRoutine::setupGradient(Pointer<Byte> &primitive, Pointer<Byte> &triangle, Float4 &w012, Float4 (&m)[3], Pointer<Byte> &v0, Pointer<Byte> &v1, Pointer<Byte> &v2, int attribute, int planeEquation, bool flat, bool perspective, int component)
diff --git a/src/Pipeline/SpirvShaderSampling.cpp b/src/Pipeline/SpirvShaderSampling.cpp
index a54eb8d..6255726 100644
--- a/src/Pipeline/SpirvShaderSampling.cpp
+++ b/src/Pipeline/SpirvShaderSampling.cpp
@@ -231,7 +231,7 @@
}
}
- return (ImageSampler*)function("sampler")->getEntry();
+ return (ImageSampler*)function(vk::ReactorOptimizationLevel, "sampler")->getEntry();
}
sw::TextureType SpirvShader::convertTextureType(VkImageViewType imageViewType)
diff --git a/src/Reactor/Coroutine.hpp b/src/Reactor/Coroutine.hpp
index 3d88048..993e7e0 100644
--- a/src/Reactor/Coroutine.hpp
+++ b/src/Reactor/Coroutine.hpp
@@ -133,7 +133,7 @@
// called without building a new rr::Function or rr::Coroutine.
// While automatically called by operator(), finalize() should be called
// as early as possible to release the global Reactor mutex lock.
- inline void finalize();
+ inline void finalize(OptimizationLevel optLevel = OptimizationLevel::Default);
// Starts execution of the coroutine and returns a unique_ptr to a
// Stream<> that exposes the await() function for obtaining yielded
@@ -147,7 +147,7 @@
};
template<typename Return, typename... Arguments>
- Coroutine<Return(Arguments...)>::Coroutine() : routine{}
+ Coroutine<Return(Arguments...)>::Coroutine()
{
core.reset(new Nucleus());
@@ -164,11 +164,11 @@
}
template<typename Return, typename... Arguments>
- void Coroutine<Return(Arguments...)>::finalize()
+ void Coroutine<Return(Arguments...)>::finalize(OptimizationLevel optLevel /* = OptimizationLevel::Default */)
{
if(core != nullptr)
{
- routine.reset(core->acquireCoroutine("coroutine", true));
+ routine.reset(core->acquireCoroutine("coroutine", optLevel));
core.reset(nullptr);
}
}
diff --git a/src/Reactor/LLVMReactor.cpp b/src/Reactor/LLVMReactor.cpp
index d6f530d..af049c6 100644
--- a/src/Reactor/LLVMReactor.cpp
+++ b/src/Reactor/LLVMReactor.cpp
@@ -219,7 +219,11 @@
using ObjLayer = llvm::orc::RTDyldObjectLinkingLayer;
using CompileLayer = llvm::orc::IRCompileLayer<ObjLayer, llvm::orc::SimpleCompiler>;
public:
- JITRoutine(std::unique_ptr<llvm::Module> module, llvm::Function **funcs, size_t count) :
+ JITRoutine(
+ std::unique_ptr<llvm::Module> module,
+ llvm::Function **funcs,
+ size_t count,
+ rr::OptimizationLevel optLevel) :
resolver(createLegacyLookupResolver(
session,
[&](const std::string &name) {
@@ -241,6 +245,8 @@
targetMachine(llvm::EngineBuilder()
#ifdef ENABLE_RR_DEBUG_INFO
.setOptLevel(llvm::CodeGenOpt::None)
+#else
+ .setOptLevel(toLLVM(optLevel))
#endif // ENABLE_RR_DEBUG_INFO
.setMArch(JITGlobals::get()->arch)
.setMAttrs(JITGlobals::get()->mattrs)
@@ -306,6 +312,19 @@
}
private:
+ static ::llvm::CodeGenOpt::Level toLLVM(rr::OptimizationLevel level)
+ {
+ switch (level)
+ {
+ case rr::OptimizationLevel::None: return ::llvm::CodeGenOpt::None;
+ case rr::OptimizationLevel::Less: return ::llvm::CodeGenOpt::Less;
+ case rr::OptimizationLevel::Default: return ::llvm::CodeGenOpt::Default;
+ case rr::OptimizationLevel::Aggressive: return ::llvm::CodeGenOpt::Aggressive;
+ default: UNREACHABLE("Unknown OptimizationLevel %d", int(level));
+ }
+ return ::llvm::CodeGenOpt::Default;
+ }
+
std::shared_ptr<llvm::orc::SymbolResolver> resolver;
std::unique_ptr<llvm::TargetMachine> targetMachine;
llvm::orc::ExecutionSession session;
@@ -361,10 +380,10 @@
passManager->run(*module);
}
- rr::Routine *acquireRoutine(llvm::Function **funcs, size_t count)
+ rr::Routine *acquireRoutine(llvm::Function **funcs, size_t count, rr::OptimizationLevel optLevel)
{
ASSERT(module);
- return new JITRoutine(std::move(module), funcs, count);
+ return new JITRoutine(std::move(module), funcs, count, optLevel);
}
llvm::LLVMContext context;
@@ -1118,7 +1137,7 @@
::codegenMutex.unlock();
}
- Routine *Nucleus::acquireRoutine(const char *name, bool runOptimizations)
+ Routine *Nucleus::acquireRoutine(const char *name, OptimizationLevel optimizationLevel)
{
if(jit->builder->GetInsertBlock()->empty() || !jit->builder->GetInsertBlock()->back().isTerminator())
{
@@ -1156,10 +1175,7 @@
}
#endif // defined(ENABLE_RR_LLVM_IR_VERIFICATION) || !defined(NDEBUG)
- if(runOptimizations)
- {
- optimize();
- }
+ optimize();
if(false)
{
@@ -1168,7 +1184,7 @@
jit->module->print(file, 0);
}
- auto routine = jit->acquireRoutine(&jit->function, 1);
+ auto routine = jit->acquireRoutine(&jit->function, 1, optimizationLevel);
jit.reset();
return routine;
@@ -4656,7 +4672,7 @@
jit->builder->SetInsertPoint(resumeBlock);
}
-Routine* Nucleus::acquireCoroutine(const char *name, bool runOptimizations)
+Routine* Nucleus::acquireCoroutine(const char *name, OptimizationLevel optimizationLevel)
{
ASSERT_MSG(jit->coroutine.id != nullptr, "acquireCoroutine() called without a call to createCoroutine()");
@@ -4685,10 +4701,7 @@
pm.add(llvm::createCoroCleanupPass());
pm.run(*jit->module);
- if(runOptimizations)
- {
- optimize();
- }
+ optimize();
if(false)
{
@@ -4701,7 +4714,7 @@
funcs[Nucleus::CoroutineEntryBegin] = jit->function;
funcs[Nucleus::CoroutineEntryAwait] = jit->coroutine.await;
funcs[Nucleus::CoroutineEntryDestroy] = jit->coroutine.destroy;
- Routine *routine = jit->acquireRoutine(funcs, Nucleus::CoroutineEntryCount);
+ auto routine = jit->acquireRoutine(funcs, Nucleus::CoroutineEntryCount, optimizationLevel);
jit.reset();
return routine;
diff --git a/src/Reactor/Nucleus.hpp b/src/Reactor/Nucleus.hpp
index 3362fb4..1b4e2db 100644
--- a/src/Reactor/Nucleus.hpp
+++ b/src/Reactor/Nucleus.hpp
@@ -21,6 +21,10 @@
#include <vector>
#include <atomic>
+#ifdef None
+#undef None // b/127920555
+#endif
+
namespace rr
{
class Type;
@@ -47,6 +51,14 @@
extern Optimization optimization[10];
+ enum class OptimizationLevel
+ {
+ None,
+ Less,
+ Default,
+ Aggressive,
+ };
+
class Nucleus
{
public:
@@ -54,7 +66,7 @@
virtual ~Nucleus();
- Routine *acquireRoutine(const char *name, bool runOptimizations = true);
+ Routine *acquireRoutine(const char *name, OptimizationLevel optimizationLevel);
static Value *allocateStackVariable(Type *type, int arraySize = 0);
static BasicBlock *createBasicBlock();
@@ -81,7 +93,7 @@
};
static void createCoroutine(Type *ReturnType, std::vector<Type*> &Params);
- Routine *acquireCoroutine(const char *name, bool runOptimizations = true);
+ Routine *acquireCoroutine(const char *name, OptimizationLevel optimizationLevel);
static void yield(Value*);
// Terminators
diff --git a/src/Reactor/Reactor.hpp b/src/Reactor/Reactor.hpp
index d9419a8..53853c0 100644
--- a/src/Reactor/Reactor.hpp
+++ b/src/Reactor/Reactor.hpp
@@ -2465,6 +2465,7 @@
}
Routine *operator()(const char *name, ...);
+ Routine *operator()(OptimizationLevel optLevel, const char *name, ...);
protected:
Nucleus *core;
@@ -3039,7 +3040,20 @@
vsnprintf(fullName, 1024, name, vararg);
va_end(vararg);
- return core->acquireRoutine(fullName, true);
+ return core->acquireRoutine(fullName, OptimizationLevel::Default);
+ }
+
+ template<typename Return, typename... Arguments>
+ Routine *Function<Return(Arguments...)>::operator()(OptimizationLevel optLevel, const char *name, ...)
+ {
+ char fullName[1024 + 1];
+
+ va_list vararg;
+ va_start(vararg, name);
+ vsnprintf(fullName, 1024, name, vararg);
+ va_end(vararg);
+
+ return core->acquireRoutine(fullName, optLevel);
}
template<class T, class S>
diff --git a/src/Reactor/SubzeroReactor.cpp b/src/Reactor/SubzeroReactor.cpp
index d655c93..e8cc581 100644
--- a/src/Reactor/SubzeroReactor.cpp
+++ b/src/Reactor/SubzeroReactor.cpp
@@ -585,7 +585,7 @@
::codegenMutex.unlock();
}
- Routine *Nucleus::acquireRoutine(const char *name, bool runOptimizations)
+ Routine *Nucleus::acquireRoutine(const char *name, OptimizationLevel optimizationLevel)
{
if(basicBlock->getInsts().empty() || basicBlock->getInsts().back().getKind() != Ice::Inst::Ret)
{
@@ -3506,7 +3506,7 @@
void FlushDebug() {}
void Nucleus::createCoroutine(Type *YieldType, std::vector<Type*> &Params) { UNIMPLEMENTED("createCoroutine"); }
- Routine* Nucleus::acquireCoroutine(const char *name, bool runOptimizations) { UNIMPLEMENTED("acquireCoroutine"); return nullptr; }
+ Routine* Nucleus::acquireCoroutine(const char *name, OptimizationLevel optimizationLevel) { UNIMPLEMENTED("acquireCoroutine"); return nullptr; }
void Nucleus::yield(Value* val) { UNIMPLEMENTED("Yield"); }
}
diff --git a/src/Vulkan/VkConfig.h b/src/Vulkan/VkConfig.h
index 157f34e..adaa353 100644
--- a/src/Vulkan/VkConfig.h
+++ b/src/Vulkan/VkConfig.h
@@ -17,6 +17,8 @@
#include "Version.h"
+#include "Reactor/Nucleus.hpp" // ReactorOptimizationLevel
+
#include <Vulkan/VulkanPlatform.h>
namespace vk
@@ -77,6 +79,9 @@
MAX_POINT_SIZE = 1, // Large points are not supported. If/when we turn this on, must be >= 64.
};
+// Optimization level to use for JIT functions.
+static constexpr auto ReactorOptimizationLevel = rr::OptimizationLevel::Default;
+
}
#endif // VK_CONFIG_HPP_
diff --git a/src/Vulkan/VkPipeline.cpp b/src/Vulkan/VkPipeline.cpp
index cf4764b..d54fddc 100644
--- a/src/Vulkan/VkPipeline.cpp
+++ b/src/Vulkan/VkPipeline.cpp
@@ -242,7 +242,7 @@
// TODO(b/119409619): use allocator.
auto program = std::make_shared<sw::ComputeProgram>(key.getShader(), key.getLayout(), descriptorSets);
program->generate();
- program->finalize();
+ program->finalize(vk::ReactorOptimizationLevel);
return program;
}
