shader: Implement MEMBAR.GL

Implement using memoryBarrier in GLSL and OpMemoryBarrier on SPIR-V.

externals/sirit

@@ -1 +1 @@
-Subproject commit e1a6729df7f11e33f6dc0939b18995a57c8bf3d8
+Subproject commit 12f40a80324d7c154f19f25c448a5ce27d38cd18

src/video_core/engines/shader_bytecode.h

@@ -384,6 +384,15 @@ enum class IsberdMode : u64 {
 
 enum class IsberdShift : u64 { None = 0, U16 = 1, B32 = 2 };
 
+enum class MembarType : u64 {
+    CTA = 0,
+    GL = 1,
+    SYS = 2,
+    VC = 3,
+};
+
+enum class MembarUnknown : u64 { Default = 0, IVALLD = 1, IVALLT = 2, IVALLTD = 3 };
+
 enum class HalfType : u64 {
     H0_H1 = 0,
     F32 = 1,
@@ -1545,6 +1554,11 @@ union Instruction {
         BitField<47, 2, IsberdShift> shift;
     } isberd;
 
+    union {
+        BitField<8, 2, MembarType> type;
+        BitField<0, 2, MembarUnknown> unknown;
+    } membar;
+
     union {
         BitField<48, 1, u64> signed_a;
         BitField<38, 1, u64> is_byte_chunk_a;
@@ -1669,6 +1683,7 @@ public:
         IPA,
         OUT_R, // Emit vertex/primitive
         ISBERD,
+        MEMBAR,
         VMAD,
         VSETP,
         FFMA_IMM, // Fused Multiply and Add
@@ -1930,7 +1945,7 @@ private:
             INST("111000100100----", Id::BRA, Type::Flow, "BRA"),
             INST("111000100101----", Id::BRX, Type::Flow, "BRX"),
             INST("1111000011111---", Id::SYNC, Type::Flow, "SYNC"),
-            INST("111000110100---", Id::BRK, Type::Flow, "BRK"),
+            INST("111000110100----", Id::BRK, Type::Flow, "BRK"),
             INST("111000110000----", Id::EXIT, Type::Flow, "EXIT"),
             INST("1111000011110---", Id::DEPBAR, Type::Synch, "DEPBAR"),
             INST("0101000011011---", Id::VOTE, Type::Warp, "VOTE"),
@@ -1969,6 +1984,7 @@ private:
             INST("11100000--------", Id::IPA, Type::Trivial, "IPA"),
             INST("1111101111100---", Id::OUT_R, Type::Trivial, "OUT_R"),
             INST("1110111111010---", Id::ISBERD, Type::Trivial, "ISBERD"),
+            INST("1110111110011---", Id::MEMBAR, Type::Trivial, "MEMBAR"),
             INST("01011111--------", Id::VMAD, Type::Video, "VMAD"),
             INST("0101000011110---", Id::VSETP, Type::Video, "VSETP"),
             INST("0011001-1-------", Id::FFMA_IMM, Type::Ffma, "FFMA_IMM"),

src/video_core/renderer_opengl/gl_shader_decompiler.cpp

@@ -1992,6 +1992,11 @@ private:
         return {fmt::format("readInvocationARB({}, {})", value, index), Type::Float};
     }
 
+    Expression MemoryBarrierGL(Operation) {
+        code.AddLine("memoryBarrier();");
+        return {};
+    }
+
     struct Func final {
         Func() = delete;
         ~Func() = delete;
@@ -2173,6 +2178,8 @@ private:
 
         &GLSLDecompiler::ThreadId,
         &GLSLDecompiler::ShuffleIndexed,
+
+        &GLSLDecompiler::MemoryBarrierGL,
     };
     static_assert(operation_decompilers.size() == static_cast<std::size_t>(OperationCode::Amount));
 

src/video_core/renderer_vulkan/vk_shader_decompiler.cpp

@@ -1971,6 +1971,18 @@ private:
         return {OpSubgroupReadInvocationKHR(t_float, value, index), Type::Float};
     }
 
+    Expression MemoryBarrierGL(Operation) {
+        const auto scope = spv::Scope::Device;
+        const auto semantics =
+            spv::MemorySemanticsMask::AcquireRelease | spv::MemorySemanticsMask::UniformMemory |
+            spv::MemorySemanticsMask::WorkgroupMemory |
+            spv::MemorySemanticsMask::AtomicCounterMemory | spv::MemorySemanticsMask::ImageMemory;
+
+        OpMemoryBarrier(Constant(t_uint, static_cast<u32>(scope)),
+                        Constant(t_uint, static_cast<u32>(semantics)));
+        return {};
+    }
+
     Id DeclareBuiltIn(spv::BuiltIn builtin, spv::StorageClass storage, Id type, std::string name) {
         const Id id = OpVariable(type, storage);
         Decorate(id, spv::Decoration::BuiltIn, static_cast<u32>(builtin));
@@ -2374,6 +2386,8 @@ private:
 
         &SPIRVDecompiler::ThreadId,
         &SPIRVDecompiler::ShuffleIndexed,
+
+        &SPIRVDecompiler::MemoryBarrierGL,
     };
     static_assert(operation_decompilers.size() == static_cast<std::size_t>(OperationCode::Amount));
 

src/video_core/shader/decode/other.cpp

@@ -257,6 +257,12 @@ u32 ShaderIR::DecodeOther(NodeBlock& bb, u32 pc) {
         SetRegister(bb, instr.gpr0, GetRegister(instr.gpr8));
         break;
     }
+    case OpCode::Id::MEMBAR: {
+        UNIMPLEMENTED_IF(instr.membar.type != Tegra::Shader::MembarType::GL);
+        UNIMPLEMENTED_IF(instr.membar.unknown != Tegra::Shader::MembarUnknown::Default);
+        bb.push_back(Operation(OperationCode::MemoryBarrierGL));
+        break;
+    }
     case OpCode::Id::DEPBAR: {
         LOG_DEBUG(HW_GPU, "DEPBAR instruction is stubbed");
         break;

src/video_core/shader/node.h

@@ -189,6 +189,8 @@ enum class OperationCode {
     ThreadId,       /// () -> uint
     ShuffleIndexed, /// (uint value, uint index) -> uint
 
+    MemoryBarrierGL, /// () -> void
+
     Amount,
 };