[engine, dma] Adjustment of the execution mask handling

src/video_core/engines/engine_interface.h

@@ -6,7 +6,7 @@
 
 #pragma once
 
-#include <bitset>
+#include <array>
 #include <limits>
 #include <vector>
 
@@ -26,6 +26,11 @@ class EngineInterface {
 public:
     virtual ~EngineInterface() = default;
 
+    static constexpr size_t ExecutionMaskBits = (std::numeric_limits<u16>::max)();
+    static constexpr size_t ExecutionMaskWordBits = 64;
+    static constexpr size_t ExecutionMaskWords =
+        (ExecutionMaskBits + ExecutionMaskWordBits - 1) / ExecutionMaskWordBits;
+
     /// Write the value to the register identified by method.
     virtual void CallMethod(u32 method, u32 method_argument, bool is_last_call) = 0;
 
@@ -40,7 +45,33 @@ public:
         ConsumeSinkImpl();
     }
 
-    std::bitset<(std::numeric_limits<u16>::max)()> execution_mask{};
+    void ClearExecutionMask() {
+        execution_mask_words.fill(0);
+    }
+
+    void SetExecutionMaskBit(u32 method, bool value = true) {
+        if (method >= ExecutionMaskBits) {
+            return;
+        }
+        const size_t word_index = method >> 6;
+        const u64 bit_mask = 1ULL << (method & 63);
+        if (value) {
+            execution_mask_words[word_index] |= bit_mask;
+        } else {
+            execution_mask_words[word_index] &= ~bit_mask;
+        }
+    }
+
+    bool IsExecutionMaskSet(u32 method) const {
+        if (method >= ExecutionMaskBits) {
+            return false;
+        }
+        const size_t word_index = method >> 6;
+        const u64 bit_mask = 1ULL << (method & 63);
+        return (execution_mask_words[word_index] & bit_mask) != 0;
+    }
+
+    std::array<u64, ExecutionMaskWords> execution_mask_words{};
     std::vector<std::pair<u32, u32>> method_sink{};
     bool current_dirty{};
     GPUVAddr current_dma_segment;