Generate vulkanMemoryModel shader

Author: Omar Ahmed

docs/OpenCLCOnVulkan.md

@@ -199,10 +199,15 @@ shader is as follows:
   ID is reported in the descriptor map file, generated via the `-descriptormap`
   option.
 
-The shaders produced use the GLSL450 memory model. As such, there is an assumption of
+#### Memory Model
+
+The default shaders produced use the GLSL450 memory model. As such, there is an assumption of
 no aliasing by default. The compiler does not generate *Aliased* decorations
 currently. Users should be aware of this and ensure they are not relying on aliasing.
 
+`vulkan-memory-model` option could be used to produce Vulkan memory model shaders. Memory scopes for loads and stores in this mode will be decided
+according to the pointer type.
+
 #### Embedded Reflection Instructions
 
 Clspv embeds reflection information via use of the [NonSemantic.ClspvReflection](http://htmlpreview.github.io/?https://github.com/KhronosGroup/SPIRV-Registry/blob/master/nonsemantic/NonSemantic.ClspvReflection.html)

include/clspv/Option.h

@@ -260,6 +260,9 @@ Vec3ToVec4SupportClass Vec3ToVec4();
 // Returns true if opaque pointers are enabled
 bool OpaquePointers();
 
+// Returns true if vulkan memory model is enabled
+bool VulkanMemoryModel();
+
 // Returns true if the debug information should be generated
 bool DebugInfo();
 

lib/Option.cpp

@@ -381,6 +381,11 @@ static llvm::cl::opt<bool>
                     llvm::cl::desc("Use opaque pointers"),
                     llvm::cl::init(false));
 
+static llvm::cl::opt<bool>
+    vulkan_memory_model("vulkan-memory-model",
+                        llvm::cl::desc("Use vulkan memory model"),
+                        llvm::cl::init(false));
+
 static llvm::cl::opt<bool>
     debug_info("g", llvm::cl::init(false),
                llvm::cl::desc("Produce debug information."));
@@ -500,6 +505,8 @@ Vec3ToVec4SupportClass Vec3ToVec4() {
 
 bool OpaquePointers() { return opaque_pointers; }
 
+bool VulkanMemoryModel() { return vulkan_memory_model; }
+
 bool DebugInfo() { return debug_info; }
 
 std::set<FeatureMacro> EnabledFeatureMacros() { return enabled_feature_macros; }

lib/ReplaceOpenCLBuiltinPass.cpp

@@ -3510,9 +3510,21 @@ bool ReplaceOpenCLBuiltinPass::replaceAtomics(Function &F, spv::Op Op) {
 
     // We need to map the OpenCL constants to the SPIR-V equivalents.
     const auto ConstantScopeDevice = ConstantInt::get(IntTy, spv::ScopeDevice);
-    const auto ConstantMemorySemantics = ConstantInt::get(
-        IntTy, spv::MemorySemanticsUniformMemoryMask |
-                   spv::MemorySemanticsSequentiallyConsistentMask);
+    auto EqualMemorySemanticsMask =
+        clspv::Option::VulkanMemoryModel()
+            ? spv::MemorySemanticsAcquireReleaseMask
+            : spv::MemorySemanticsSequentiallyConsistentMask;
+    auto UnEqualMemorySemanticsMask =
+        clspv::Option::VulkanMemoryModel()
+            ? spv::MemorySemanticsAcquireMask
+            : spv::MemorySemanticsSequentiallyConsistentMask;
+
+    const auto EqualConstantMemorySemantics =
+        ConstantInt::get(IntTy, spv::MemorySemanticsUniformMemoryMask |
+                                    EqualMemorySemanticsMask);
+    const auto UnEqualConstantMemorySemantics =
+        ConstantInt::get(IntTy, spv::MemorySemanticsUniformMemoryMask |
+                                    UnEqualMemorySemanticsMask);
 
     SmallVector<Value *, 5> Params;
 
@@ -3523,11 +3535,11 @@ bool ReplaceOpenCLBuiltinPass::replaceAtomics(Function &F, spv::Op Op) {
     Params.push_back(ConstantScopeDevice);
 
     // The memory semantics.
-    Params.push_back(ConstantMemorySemantics);
+    Params.push_back(EqualConstantMemorySemantics);
 
     if (2 < CI->arg_size()) {
       // The unequal memory semantics.
-      Params.push_back(ConstantMemorySemantics);
+      Params.push_back(UnEqualConstantMemorySemantics);
 
       // The value.
       Params.push_back(CI->getArgOperand(2));
@@ -3548,9 +3560,12 @@ bool ReplaceOpenCLBuiltinPass::replaceAtomics(Function &F,
   return replaceCallsWithValue(F, [&](CallInst *CI) {
     auto align = F.getParent()->getDataLayout().getABITypeAlign(
         CI->getArgOperand(1)->getType());
+    auto MemorySemantics = clspv::Option::VulkanMemoryModel()
+                               ? AtomicOrdering::AcquireRelease
+                               : AtomicOrdering::SequentiallyConsistent;
+
     return new AtomicRMWInst(Op, CI->getArgOperand(0), CI->getArgOperand(1),
-                             align, AtomicOrdering::SequentiallyConsistent,
-                             SyncScope::System, CI);
+                             align, MemorySemantics, SyncScope::System, CI);
   });
 }