Merge branch 'master' into Texture2DArray

3 files changed, 45 insertions, 13 deletions

diff --git a/src/video_core/engines/maxwell_3d.cpp b/src/video_core/engines/maxwell_3d.cpp
index 7357d20d1..d79c50919 100644
--- a/src/video_core/engines/maxwell_3d.cpp
+++ b/src/video_core/engines/maxwell_3d.cpp
@@ -43,15 +43,17 @@ void Maxwell3D::CallMacroMethod(u32 method, std::vector<u32> parameters) {
     // Reset the current macro.
     executing_macro = 0;
 
-    // The requested macro must have been uploaded already.
-    auto macro_code = uploaded_macros.find(method);
-    if (macro_code == uploaded_macros.end()) {
-        LOG_ERROR(HW_GPU, "Macro {:04X} was not uploaded", method);
+    // Lookup the macro offset
+    const u32 entry{(method - MacroRegistersStart) >> 1};
+    const auto& search{macro_offsets.find(entry)};
+    if (search == macro_offsets.end()) {
+        LOG_CRITICAL(HW_GPU, "macro not found for method 0x{:X}!", method);
+        UNREACHABLE();
         return;
     }
 
     // Execute the current macro.
-    macro_interpreter.Execute(macro_code->second, std::move(parameters));
+    macro_interpreter.Execute(search->second, std::move(parameters));
 }
 
 void Maxwell3D::WriteReg(u32 method, u32 value, u32 remaining_params) {
@@ -97,6 +99,10 @@ void Maxwell3D::WriteReg(u32 method, u32 value, u32 remaining_params) {
         ProcessMacroUpload(value);
         break;
     }
+    case MAXWELL3D_REG_INDEX(macros.bind): {
+        ProcessMacroBind(value);
+        break;
+    }
     case MAXWELL3D_REG_INDEX(const_buffer.cb_data[0]):
     case MAXWELL3D_REG_INDEX(const_buffer.cb_data[1]):
     case MAXWELL3D_REG_INDEX(const_buffer.cb_data[2]):
@@ -158,9 +164,13 @@ void Maxwell3D::WriteReg(u32 method, u32 value, u32 remaining_params) {
 }
 
 void Maxwell3D::ProcessMacroUpload(u32 data) {
-    // Store the uploaded macro code to interpret them when they're called.
-    auto& macro = uploaded_macros[regs.macros.entry * 2 + MacroRegistersStart];
-    macro.push_back(data);
+    ASSERT_MSG(regs.macros.upload_address < macro_memory.size(),
+               "upload_address exceeded macro_memory size!");
+    macro_memory[regs.macros.upload_address++] = data;
+}
+
+void Maxwell3D::ProcessMacroBind(u32 data) {
+    macro_offsets[regs.macros.entry] = data;
 }
 
 void Maxwell3D::ProcessQueryGet() {
diff --git a/src/video_core/engines/maxwell_3d.h b/src/video_core/engines/maxwell_3d.h
index 443affc36..50873813e 100644
--- a/src/video_core/engines/maxwell_3d.h
+++ b/src/video_core/engines/maxwell_3d.h
@@ -475,12 +475,13 @@ public:
                 INSERT_PADDING_WORDS(0x45);
 
                 struct {
-                    INSERT_PADDING_WORDS(1);
+                    u32 upload_address;
                     u32 data;
                     u32 entry;
+                    u32 bind;
                 } macros;
 
-                INSERT_PADDING_WORDS(0x189);
+                INSERT_PADDING_WORDS(0x188);
 
                 u32 tfb_enabled;
 
@@ -994,12 +995,25 @@ public:
     /// Returns the texture information for a specific texture in a specific shader stage.
     Texture::FullTextureInfo GetStageTexture(Regs::ShaderStage stage, std::size_t offset) const;
 
+    /// Memory for macro code - it's undetermined how big this is, however 1MB is much larger than
+    /// we've seen used.
+    using MacroMemory = std::array<u32, 0x40000>;
+
+    /// Gets a reference to macro memory.
+    const MacroMemory& GetMacroMemory() const {
+        return macro_memory;
+    }
+
 private:
     void InitializeRegisterDefaults();
 
     VideoCore::RasterizerInterface& rasterizer;
 
-    std::unordered_map<u32, std::vector<u32>> uploaded_macros;
+    /// Start offsets of each macro in macro_memory
+    std::unordered_map<u32, u32> macro_offsets;
+
+    /// Memory for macro code
+    MacroMemory macro_memory;
 
     /// Macro method that is currently being executed / being fed parameters.
     u32 executing_macro = 0;
@@ -1022,9 +1036,12 @@ private:
      */
     void CallMacroMethod(u32 method, std::vector<u32> parameters);
 
-    /// Handles writes to the macro uploading registers.
+    /// Handles writes to the macro uploading register.
     void ProcessMacroUpload(u32 data);
 
+    /// Handles writes to the macro bind register.
+    void ProcessMacroBind(u32 data);
+
     /// Handles a write to the CLEAR_BUFFERS register.
     void ProcessClearBuffers();
 
diff --git a/src/video_core/engines/shader_bytecode.h b/src/video_core/engines/shader_bytecode.h
index b84da512f..83a6fd875 100644
--- a/src/video_core/engines/shader_bytecode.h
+++ b/src/video_core/engines/shader_bytecode.h
@@ -578,6 +578,10 @@ union Instruction {
     } fmul32;
 
     union {
+        BitField<52, 1, u64> generates_cc;
+    } op_32;
+
+    union {
         BitField<48, 1, u64> is_signed;
     } shift;
 
@@ -1231,6 +1235,7 @@ union Instruction {
     BitField<60, 1, u64> is_b_gpr;
     BitField<59, 1, u64> is_c_gpr;
     BitField<20, 24, s64> smem_imm;
+    BitField<0, 5, ControlCode> flow_control_code;
 
     Attribute attribute;
     Sampler sampler;
@@ -1658,4 +1663,4 @@ private:
     }
 };
 
-} // namespace Tegra::Shader
\ No newline at end of file
+} // namespace Tegra::Shader