Switch ImGui's ImHash function to use Special K's SSE4.2 implementation of crc32c

Author: Kaldaien

include/SpecialK/crc32.h

@@ -1,4 +1,4 @@
-/**
+/**
  * This file is part of Special K.
  *
  * Special K is free software : you can redistribute it
@@ -37,13 +37,13 @@ SK_File_GetCRC32C ( const wchar_t* wszFile,
 
 uint32_t
 __cdecl
-crc32       (uint32_t crc, _Notnull_ const void *buf, size_t size) ;
+crc32       (uint32_t crc, _Notnull_ const void *buf, size_t size) noexcept;
 
 // Returns the value of crc if attempting to checksum this memory throws an
 //   access violation exception
 uint32_t
 __cdecl
-safe_crc32c (uint32_t crc, _Notnull_ const void *buf, size_t size);
+safe_crc32c (uint32_t crc, _Notnull_ const void *buf, size_t size) noexcept;
 
 /*
     Computes CRC-32C (Castagnoli) checksum. Uses Intel's CRC32 instruction if it is available.
@@ -52,37 +52,37 @@ safe_crc32c (uint32_t crc, _Notnull_ const void *buf, size_t size);
 uint32_t
 __cdecl
 crc32c (
-    uint32_t    crc,            // Initial CRC value. Typically it's 0.
-                                // You can supply non-trivial initial value here.
-                                // Initial value can be used to chain CRC from multiple buffers.
+    uint32_t    crc,              // Initial CRC value. Typically it's 0.
+                                  // You can supply non-trivial initial value here.
+                                  // Initial value can be used to chain CRC from multiple buffers.
 
     _Notnull_
-    const void *input,          // Data to be put through the CRC algorithm.
-    size_t      length);        // Length of the data in the input buffer.
+    const void *input,            // Data to be put through the CRC algorithm.
+    size_t      length) noexcept; // Length of the data in the input buffer.
 
 
-void __cdecl __crc32_init (void);
+void __cdecl __crc32_init (void) noexcept;
 
 /*
 	Software fallback version of CRC-32C (Castagnoli) checksum.
 */
 uint32_t
 __cdecl
-crc32c_append_sw (uint32_t crc, const void *input, size_t length) ;
+crc32c_append_sw (uint32_t crc, const void *input, size_t length) noexcept;
 
 /*
 	Hardware version of CRC-32C (Castagnoli) checksum. Will fail, if CPU does
   not support related instructions. Use a crc32c_append version instead of.
 */
 uint32_t
 __cdecl
-crc32c_append_hw (uint32_t crc, const void *input, size_t length) ;
+crc32c_append_hw (uint32_t crc, const void *input, size_t length) noexcept;
 
 /*
 	Checks is hardware version of CRC-32C is available.
 */
 int
 __cdecl
-crc32c_hw_available (void) ;
+crc32c_hw_available (void) noexcept;
 
 SK_INCLUDE_END (CRC32)

include/imgui/imgui_internal.h

@@ -1,4 +1,4 @@
-// dear imgui, v1.90 WIP
+// dear imgui, v1.90 WIP
 // (internal structures/api)
 
 // You may use this file to debug, understand or extend Dear ImGui features but we don't provide any guarantee of forward compatibility.
@@ -350,8 +350,8 @@ namespace ImStb
 //-----------------------------------------------------------------------------
 
 // Helpers: Hashing
-IMGUI_API ImGuiID       ImHashData(const void* data, size_t data_size, ImGuiID seed = 0);
-IMGUI_API ImGuiID       ImHashStr(const char* data, size_t data_size = 0, ImGuiID seed = 0);
+IMGUI_API ImGuiID       ImHashData(const void* data, size_t data_size, ImGuiID seed = 0) noexcept;
+IMGUI_API ImGuiID       ImHashStr(const char* data, size_t data_size = 0, ImGuiID seed = 0) noexcept;
 
 // Helpers: Sorting
 #ifndef ImQsort

src/SpecialK.cpp

@@ -1695,6 +1695,8 @@ SK_Attach (DLL_ROLE role)
 
       if (SK_DLL_IsAttached ())
       {
+        __crc32_init ();
+
         SK_MinHook_Init ();
 
         return TRUE;

src/crc32.cpp

@@ -97,7 +97,7 @@ std::unique_ptr <InstructionSet::InstructionSet_Internal> InstructionSet::CPU_Re
 extern "C"
 uint32_t
 __cdecl
-crc32 (uint32_t crc, _Notnull_ const void *buf, size_t size)
+crc32 (uint32_t crc, _Notnull_ const void *buf, size_t size) noexcept
 {
   const auto *p =
        static_cast <const uint8_t *> (buf);
@@ -155,15 +155,15 @@ static uint32_t short_shifts [ 4][256];
 
 static bool _tableInitialized = false;
 
-extern "C" void __cdecl calculate_table (void) ;
+extern "C" void __cdecl calculate_table (void) noexcept;
 
 /* Table-driven software version as a fall-back.  This is about 15 times slower
    than using the hardware instructions.  This assumes little-endian integers,
    as is the case on Intel processors that the assembler code here is for. */
 extern "C"
 uint32_t
 __cdecl
-crc32c_append_sw (uint32_t crci, const void *input, size_t length)
+crc32c_append_sw (uint32_t crci, const void *input, size_t length) noexcept
 {
   auto next =
     static_cast <buffer> (input);
@@ -247,7 +247,7 @@ crc32c_append_sw (uint32_t crci, const void *input, size_t length)
 static
 inline
 uint32_t
-shift_crc ( const uint32_t shift_table[][256], uint32_t crc )
+shift_crc ( const uint32_t shift_table[][256], uint32_t crc ) noexcept
 {
   return shift_table [0][ crc        & 0xff]
        ^ shift_table [1][(crc >> 8)  & 0xff]
@@ -259,7 +259,7 @@ shift_crc ( const uint32_t shift_table[][256], uint32_t crc )
 extern "C"
 uint32_t
 __cdecl
-crc32c_append_hw (uint32_t crc, const void *buf, size_t len)
+crc32c_append_hw (uint32_t crc, const void *buf, size_t len) noexcept
 {
   if (buf == nullptr || len < 1)
     return crc;
@@ -432,7 +432,7 @@ crc32c_append_hw (uint32_t crc, const void *buf, size_t len)
 extern "C"
 int
 __cdecl
-crc32c_hw_available (void)
+crc32c_hw_available (void) noexcept
 {
   int              info [4] = { 0 };
 #ifndef SK_BUILT_BY_CLANG
@@ -447,7 +447,7 @@ crc32c_hw_available (void)
 extern "C"
 void
 __cdecl
-calculate_table (void)
+calculate_table (void) noexcept
 {
   for (int i = 0; i < 256; i++)
   {
@@ -469,7 +469,7 @@ calculate_table (void)
 extern "C"
 void
 __cdecl
-calculate_table_hw (void)
+calculate_table_hw (void) noexcept
 {
   for (unsigned int i = 0; i < 256UL; i++)
   {
@@ -520,7 +520,7 @@ volatile LONG
 
 extern "C"
 void __cdecl
-__crc32_init (void)
+__crc32_init (void) noexcept
 {
   typedef
     uint32_t (__cdecl *appendfunc_pfn)(       uint32_t,
@@ -565,25 +565,8 @@ extern "C"
 uint32_t __cdecl
 crc32c (        uint32_t crc,
 _Notnull_ const void    *input,
-                size_t   length )
+                size_t   length ) noexcept
 {
-  if (append_func == nullptr)
-  {
-    static volatile LONG __init = 0;
-
-    if (InterlockedCompareExchange (&__init, 1, 0) == 0)
-    {
-      __crc32_init ();
-
-      InterlockedIncrement (&__init);
-    }
-
-    else
-    {
-      SK_Thread_SpinUntilAtomicMin (&__init, 2);
-    }
-  }
-
   if ( input       != nullptr &&
        length      >  0       &&
        append_func != nullptr )

src/imgui/imgui.cpp

@@ -2022,17 +2022,28 @@ static const ImU32 GCrc32LookupTable[256] =
     0xBDBDF21C,0xCABAC28A,0x53B39330,0x24B4A3A6,0xBAD03605,0xCDD70693,0x54DE5729,0x23D967BF,0xB3667A2E,0xC4614AB8,0x5D681B02,0x2A6F2B94,0xB40BBE37,0xC30C8EA1,0x5A05DF1B,0x2D02EF8D,
 };
 
+extern "C"
+uint32_t __cdecl
+crc32c (        uint32_t crc,
+_Notnull_ const void    *input,
+                size_t   length ) noexcept;
+
 // Known size hash
 // It is ok to call ImHashData on a string with known length but the ### operator won't be supported.
 // FIXME-OPT: Replace with e.g. FNV1a hash? CRC32 pretty much randomly access 1KB. Need to do proper measurements.
-ImGuiID ImHashData(const void* data_p, size_t data_size, ImGuiID seed)
+ImGuiID ImHashData(const void* data_p, size_t data_size, ImGuiID seed) noexcept
 {
     ImU32 crc = ~seed;
     const unsigned char* data = (const unsigned char*)data_p;
+#ifdef USE_ORIGINAL_IMHASH
     const ImU32* crc32_lut = GCrc32LookupTable;
     while (data_size-- != 0)
         crc = (crc >> 8) ^ crc32_lut[(crc & 0xFF) ^ *data++];
     return ~crc;
+#else
+    crc32c (crc, data, static_cast <unsigned int> (data_size));
+    return ~crc;
+#endif
 }
 
 // Zero-terminated string hash, with support for ### to reset back to seed value
@@ -2041,11 +2052,12 @@ ImGuiID ImHashData(const void* data_p, size_t data_size, ImGuiID seed)
 // - If we reach ### in the string we discard the hash so far and reset to the seed.
 // - We don't do 'current += 2; continue;' after handling ### to keep the code smaller/faster (measured ~10% diff in Debug build)
 // FIXME-OPT: Replace with e.g. FNV1a hash? CRC32 pretty much randomly access 1KB. Need to do proper measurements.
-ImGuiID ImHashStr(const char* data_p, size_t data_size, ImGuiID seed)
+ImGuiID ImHashStr(const char* data_p, size_t data_size, ImGuiID seed) noexcept
 {
     seed = ~seed;
     ImU32 crc = seed;
     const unsigned char* data = (const unsigned char*)data_p;
+#ifdef USE_ORIGINAL_IMHASH
     const ImU32* crc32_lut = GCrc32LookupTable;
     if (data_size != 0)
     {
@@ -2066,6 +2078,39 @@ ImGuiID ImHashStr(const char* data_p, size_t data_size, ImGuiID seed)
             crc = (crc >> 8) ^ crc32_lut[(crc & 0xFF) ^ c];
         }
     }
+#else
+    size_t               unparsed_size = data_size;
+    const unsigned char* start_of_data = data;
+
+    if (data_size != 0)
+    {
+        while (data_size-- != 0)
+        {
+            if (*data++ == '#' && data_size >= 2 && data[0] == '#' && data[1] == '#')
+                start_of_data = &data[2]; // Skip "###" in the hash
+        }
+    }
+    else
+    {
+        size_t str_size = 0;
+        while (unsigned char c = *data++)
+        {
+            str_size++;
+
+            if (c == '#' && data[0] == '#' && data[1] == '#') {
+                start_of_data = &data[2]; // Skip "###" in the hash
+                str_size      = 0;
+            }
+        }
+        unparsed_size = str_size;
+    }
+
+    if (start_of_data <= data + unparsed_size)
+    {
+      crc =
+        crc32c (crc, start_of_data, static_cast <unsigned int> (data + unparsed_size - start_of_data));
+    }
+#endif
     return ~crc;
 }
 

src/render/d3d9/d3d9_texmgr.cpp

@@ -1603,7 +1603,7 @@ SK::D3D9::TextureManager::loadQueuedTextures (void)
 
 uint32_t
 __cdecl
-safe_crc32c (uint32_t seed, _Notnull_ const void *buf, size_t size)
+safe_crc32c (uint32_t seed, _Notnull_ const void *buf, size_t size) noexcept
 {
   // Current limit == 2 GiB
   if (size > (1024ULL * 1024ULL * 1024ULL * 2) || buf == nullptr)