[dxvk] Use memcmp replacement for pipeline state lookup

Measured to be over twice as fast as memcmp on Ryzen for the
512-byte graphics pipeline state struct, achieving two cycles
per iteration.

src/dxvk/dxvk_graphics_state.h

@@ -641,11 +641,11 @@ namespace dxvk {
     }
     
     bool operator == (const DxvkGraphicsPipelineStateInfo& other) const {
-      return !std::memcmp(this, &other, sizeof(*this));
+      return bit::bcmpeq(this, &other);
     }
 
     bool operator != (const DxvkGraphicsPipelineStateInfo& other) const {
-      return std::memcmp(this, &other, sizeof(*this));
+      return !bit::bcmpeq(this, &other);
     }
 
     bool useDynamicStencilRef() const {
@@ -709,11 +709,11 @@ namespace dxvk {
     }
     
     bool operator == (const DxvkComputePipelineStateInfo& other) const {
-      return !std::memcmp(this, &other, sizeof(*this));
+      return bit::bcmpeq(this, &other);
     }
 
     bool operator != (const DxvkComputePipelineStateInfo& other) const {
-      return std::memcmp(this, &other, sizeof(*this));
+      return !bit::bcmpeq(this, &other);
     }
     
     DxvkBindingMask         bsBindingMask;

src/util/util_bit.h

@@ -80,4 +80,56 @@ namespace dxvk::bit {
     return shift > Bits ? shift - Bits : 0;
   }
 
+  /**
+   * \brief Compares two aligned structs bit by bit
+   *
+   * \param [in] a First struct
+   * \param [in] b Second struct
+   * \returns \c true if the structs are equal
+   */
+  template<typename T>
+  bool bcmpeq(const T* a, const T* b) {
+    static_assert(alignof(T) >= 16);
+    #if defined(__GNUC__) || defined(__clang__) || defined(_MSC_VER)
+    auto ai = reinterpret_cast<const __m128i*>(a);
+    auto bi = reinterpret_cast<const __m128i*>(b);
+
+    size_t i = 0;
+
+    #if defined(__clang__)
+    #pragma nounroll
+    #elif defined(__GNUC__)
+    #pragma GCC unroll 0
+    #endif
+
+    for ( ; i < 2 * (sizeof(T) / 32); i += 2) {
+      __m128i eq0 = _mm_cmpeq_epi8(
+        _mm_load_si128(ai + i),
+        _mm_load_si128(bi + i));
+      __m128i eq1 = _mm_cmpeq_epi8(
+        _mm_load_si128(ai + i + 1),
+        _mm_load_si128(bi + i + 1));
+      __m128i eq = _mm_and_si128(eq0, eq1);
+
+      int mask = _mm_movemask_epi8(eq);
+      if (mask != 0xFFFF)
+        return false;
+    }
+
+    for ( ; i < sizeof(T) / 16; i++) {
+      __m128i eq = _mm_cmpeq_epi8(
+        _mm_load_si128(ai + i),
+        _mm_load_si128(bi + i));
+
+      int mask = _mm_movemask_epi8(eq);
+      if (mask != 0xFFFF)
+        return false;
+    }
+
+    return true;
+    #else
+    return !std::memcmp(a, b, sizeof(T));
+    #endif
+  }
+  
 }