Changed the hash algorithm to CRC32 utilising the SSE4.2 instruction. The algorithm will automatically be used for the Accurate Texture Cache, EFB to RAM and texture id's when a SSE4.2 capable CPU is detected. It will fallback to the old algorithm if SSE4.2 is not detected. Using CRC32 speeds up the hash algorithm by around 2X.

git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@7060 8ced0084-cf51-0410-be5f-012b33b47a6e

Source/Core/Common/Src/Common.h

@@ -155,7 +155,7 @@ private:
 #elif __GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 3)
 #define _M_SSE 0x301
 #elif _MSC_VER >= 1500 // Visual Studio 2008
-#define _M_SSE 0x401
+#define _M_SSE 0x402
 #endif
 
 // Host communication.

Source/Core/Common/Src/Hash.cpp

@@ -16,6 +16,10 @@
 // http://code.google.com/p/dolphin-emu/
 
 #include "Hash.h"
+#if _M_SSE >= 0x402
+#include "CPUDetect.h"
+#include <nmmintrin.h>
+#endif
 
 // uint32_t
 // WARNING - may read one more byte!
@@ -105,12 +109,44 @@ u32 HashEctor(const u8* ptr, int length)
 }
 
 #ifdef _M_X64
+// CRC32 hash using the SSE4.2 instruction
+u64 GetCRC32(const u8 *src, int len, u32 samples)
+{
+#if _M_SSE >= 0x402
+	u64 h = len;
+	u32 Step = (len / 8);
+	const u64 *data = (const u64 *)src;
+	const u64 *end = data + Step;
+	if(samples == 0) samples = Step;
+	Step = Step / samples;
+	if(Step < 1) Step = 1;
+	while(data < end)
+	{
+		h = _mm_crc32_u64(h, data[0]);
+		data += Step;
+	}
+
+	const u8 *data2 = (const u8*)end;
+	return _mm_crc32_u64(h, u64(data2[0]));
+#else
+	return 0;
+#endif
+}
+
 u64 GetHash64(const u8 *src, int len, u32 samples)
 {
 	const u64 m = 0xc6a4a7935bd1e995;
-	const int r = 47;
-
 	u64 h = len * m;
+
+#if _M_SSE >= 0x402
+	if (cpu_info.bSSE4_2)
+	{
+		h = GetCRC32(src, len, samples);
+	}
+	else
+#endif
+	{
+		const int r = 47;
 		u32 Step = (len / 8);
 		const u64 *data = (const u64 *)src;
 		const u64 *end = data + Step;
@@ -145,14 +181,47 @@ u64 GetHash64(const u8 *src, int len, u32 samples)
 		h ^= h >> r;
 		h *= m;
 		h ^= h >> r;
+	}
 
 	return h;
 } 
-
 #else
+// CRC32 hash using the SSE4.2 instruction
+u64 GetCRC32(const u8 *src, int len, u32 samples)
+{
+#if _M_SSE >= 0x402
+	u32 h = len;
+	u32 Step = (len/4);
+	const u32 *data = (const u32 *)src;
+	const u32 *end = data + Step;
+	if(samples == 0) samples = Step;
+	Step  = Step / samples;
+	if(Step < 1) Step = 1;
+	while(data < end)
+	{
+		h = _mm_crc32_u32(h, data[0]);
+		data += Step;
+	}
+
+	const u8 *data2 = (const u8*)end;
+	return (u64)_mm_crc32_u32(h, u32(data2[0]));
+#else
+	return 0;
+#endif
+}
+
 u64 GetHash64(const u8 *src, int len, u32 samples)
 {
 	const u32 m = 0x5bd1e995;
+	u64 h = 0;
+#if _M_SSE >= 0x402
+	if (cpu_info.bSSE4_2)
+	{
+		h = GetCRC32(src, len, samples);
+	}
+	else
+#endif
+	{
 		const int r = 24;
 		
 		u32 h1 = len;
@@ -210,12 +279,10 @@ u64 GetHash64(const u8 *src, int len, u32 samples)
 		h1 ^= h2 >> 17; h1 *= m;
 		h2 ^= h1 >> 19; h2 *= m;
 
-	u64 h = h1;
+		h = h1;
 
 		h = (h << 32) | h2;
-
+	}
 	return h;
 }
-
-
 #endif

Source/Core/Common/Src/Hash.h

@@ -24,5 +24,6 @@ u32 HashFletcher(const u8* data_u8, size_t length);  // FAST. Length & 1 == 0.
 u32 HashAdler32(const u8* data, size_t len);         // Fairly accurate, slightly slower
 u32 HashFNV(const u8* ptr, int length);              // Another fast and decent hash
 u32 HashEctor(const u8* ptr, int length);            // JUNK. DO NOT USE FOR NEW THINGS
+u64 GetCRC32(const u8 *src, int len, u32 samples); // SSE4.2 version of CRC32
 u64 GetHash64(const u8 *src, int len, u32 samples);
 #endif // _HASH_H_