AVX2 optimizations, loop unrolling

dotnet/WhiteRabbit.UnmanagedBridge/WhiteRabbit.UnmanagedBridge.cpp

@@ -10,7 +10,7 @@ void WhiteRabbitUnmanagedBridge::MD5Unmanaged::ComputeMD5(unsigned __int32 * inp
 {
 #if AVX2
     md5(input + 0 * 8 * 8, output + 0 * 8);
-    md5(input + 1 * 8 * 8, output + 1 * 8);
+    //md5(input + 1 * 8 * 8, output + 1 * 8);
 #elif SIMD
     md5(input + 0 * 8 * 4, output + 0 * 4);
     md5(input + 1 * 8 * 4, output + 1 * 4);
@@ -29,5 +29,5 @@ void WhiteRabbitUnmanagedBridge::MD5Unmanaged::ComputeMD5(unsigned __int32 * inp
 
 void WhiteRabbitUnmanagedBridge::MD5Unmanaged::FillPhraseSet(__int64* bufferPointer, __int64* allWordsPointer, __int32* wordIndexes, unsigned __int64* permutationsPointer, int permutationOffset, int numberOfCharacters, int numberOfWords)
 {
-    fillPhraseSet(bufferPointer, allWordsPointer, wordIndexes, permutationsPointer, permutationOffset, numberOfCharacters, numberOfWords);
+    fillPhraseSet(bufferPointer, (unsigned __int64*)allWordsPointer, wordIndexes, permutationsPointer, permutationOffset, numberOfCharacters, numberOfWords);
 }

dotnet/WhiteRabbit.UnmanagedBridge/constants.h

@@ -1,3 +1,3 @@
 #pragma once
 
-#define PHRASES_PER_SET 16
+#define PHRASES_PER_SET 8

dotnet/WhiteRabbit.UnmanagedBridge/phraseset.cpp

@@ -6,105 +6,50 @@
 
 #pragma unmanaged
 
-void fillPhraseSet(__int64* bufferPointer, __int64* allWordsPointer, __int32* wordIndexes, unsigned __int64* permutationsPointer, int permutationOffset, int numberOfCharacters, int numberOfWords)
+#define REPEAT(macro) \
+    macro(0); \
+    macro(1); \
+    macro(2); \
+    macro(3); \
+    macro(4); \
+    macro(5); \
+    macro(6); \
+    macro(7);
+
+
+void fillPhraseSet(__int64* bufferPointer, unsigned __int64* allWordsPointer, __int32* wordIndexes, unsigned __int64* permutationsPointer, int permutationOffset, int numberOfCharacters, int numberOfWords)
 {
     unsigned __int64 permutations[PHRASES_PER_SET];
-    __int32 cumulativeWordOffsets[PHRASES_PER_SET];
+    unsigned __int64 cumulativeWordOffsets = 0;
-    __int32 permutationOffsetInBytes = permutationOffset * sizeof(*permutations);
+
+    auto avx2buffer = (__m256i*)bufferPointer;
 
 #define INIT_DATA(phraseNumber) \
-    permutations[phraseNumber] = *(unsigned __int64*)(((char*)permutationsPointer) + permutationOffsetInBytes + phraseNumber * sizeof(*permutations)); \
+    permutations[phraseNumber] = permutationsPointer[permutationOffset + phraseNumber]; \
-    cumulativeWordOffsets[phraseNumber] = 0;
 
-    INIT_DATA(0);
+    REPEAT(INIT_DATA);
-    INIT_DATA(1);
-    INIT_DATA(2);
-    INIT_DATA(3);
-    INIT_DATA(4);
-    INIT_DATA(5);
-    INIT_DATA(6);
-    INIT_DATA(7);
-    INIT_DATA(8);
-    INIT_DATA(9);
-    INIT_DATA(10);
-    INIT_DATA(11);
-    INIT_DATA(12);
-    INIT_DATA(13);
-    INIT_DATA(14);
-    INIT_DATA(15);
 
 #define PROCESS_WORD(phraseNumber) \
     { \
-        auto currentWord = allWordsPointer + wordIndexes[permutations[phraseNumber] & 15] * 128; \
+        auto currentWord = allWordsPointer + wordIndexes[permutations[phraseNumber] % 16] * 128; \
-        permutations[phraseNumber] = permutations[phraseNumber] >> 4; \
+        permutations[phraseNumber] >>= 4; \
-        bufferPointer[phraseNumber * 4 + 0] |= currentWord[cumulativeWordOffsets[phraseNumber] + 0]; \
+        avx2buffer[phraseNumber] = _mm256_or_si256(avx2buffer[phraseNumber], *(__m256i*)(currentWord + ((cumulativeWordOffsets >> (8 * (phraseNumber % 8))) % 256))); \
-        bufferPointer[phraseNumber * 4 + 1] |= currentWord[cumulativeWordOffsets[phraseNumber] + 1]; \
+        cumulativeWordOffsets += (((unsigned __int64*)currentWord)[127]) << (8 * (phraseNumber % 8)); \
-        bufferPointer[phraseNumber * 4 + 2] |= currentWord[cumulativeWordOffsets[phraseNumber] + 2]; \
-        bufferPointer[phraseNumber * 4 + 3] |= currentWord[cumulativeWordOffsets[phraseNumber] + 3]; \
-        cumulativeWordOffsets[phraseNumber] += (__int32)currentWord[127]; \
     }
 
     for (auto j = 0; j < numberOfWords; j++)
     {
-        PROCESS_WORD(0);
+        REPEAT(PROCESS_WORD);
-        PROCESS_WORD(1);
-        PROCESS_WORD(2);
-        PROCESS_WORD(3);
-        PROCESS_WORD(4);
-        PROCESS_WORD(5);
-        PROCESS_WORD(6);
-        PROCESS_WORD(7);
-        PROCESS_WORD(8);
-        PROCESS_WORD(9);
-        PROCESS_WORD(10);
-        PROCESS_WORD(11);
-        PROCESS_WORD(12);
-        PROCESS_WORD(13);
-        PROCESS_WORD(14);
-        PROCESS_WORD(15);
     }
 
     auto length = numberOfCharacters + numberOfWords - 1;
     auto lengthInBits = (unsigned __int32)(length << 3);
 
-#define FILL_PHRASE_LAST_BYTE(phraseNumber, byteBuffer) ((unsigned char*)bufferPointer)[length + phraseNumber * 32] = 128;
+#define FILL_PHRASE_LAST_BYTE(phraseNumber) ((unsigned char*)bufferPointer)[length + phraseNumber * 32] = 128;
-#define FILL_PHRASE_SET_LENGTH(phraseNumber, uintBuffer, lengthInBits) ((unsigned __int32*)bufferPointer)[7 + phraseNumber * 8] = lengthInBits;
+#define FILL_PHRASE_SET_LENGTH(phraseNumber) ((unsigned __int32*)bufferPointer)[7 + phraseNumber * 8] = lengthInBits;
-
-    FILL_PHRASE_LAST_BYTE(0, byteBuffer);
-    FILL_PHRASE_LAST_BYTE(1, byteBuffer);
-    FILL_PHRASE_LAST_BYTE(2, byteBuffer);
-    FILL_PHRASE_LAST_BYTE(3, byteBuffer);
-    FILL_PHRASE_SET_LENGTH(0, uintBuffer, lengthInBits);
-    FILL_PHRASE_SET_LENGTH(1, uintBuffer, lengthInBits);
-    FILL_PHRASE_SET_LENGTH(2, uintBuffer, lengthInBits);
-    FILL_PHRASE_SET_LENGTH(3, uintBuffer, lengthInBits);
-
-    FILL_PHRASE_LAST_BYTE(4, byteBuffer);
-    FILL_PHRASE_LAST_BYTE(5, byteBuffer);
-    FILL_PHRASE_LAST_BYTE(6, byteBuffer);
-    FILL_PHRASE_LAST_BYTE(7, byteBuffer);
-    FILL_PHRASE_SET_LENGTH(4, uintBuffer, lengthInBits);
-    FILL_PHRASE_SET_LENGTH(5, uintBuffer, lengthInBits);
-    FILL_PHRASE_SET_LENGTH(6, uintBuffer, lengthInBits);
-    FILL_PHRASE_SET_LENGTH(7, uintBuffer, lengthInBits);
-
-    FILL_PHRASE_LAST_BYTE(8, byteBuffer);
-    FILL_PHRASE_LAST_BYTE(9, byteBuffer);
-    FILL_PHRASE_LAST_BYTE(10, byteBuffer);
-    FILL_PHRASE_LAST_BYTE(11, byteBuffer);
-    FILL_PHRASE_SET_LENGTH(8, uintBuffer, lengthInBits);
-    FILL_PHRASE_SET_LENGTH(9, uintBuffer, lengthInBits);
-    FILL_PHRASE_SET_LENGTH(10, uintBuffer, lengthInBits);
-    FILL_PHRASE_SET_LENGTH(11, uintBuffer, lengthInBits);
 
-    FILL_PHRASE_LAST_BYTE(12, byteBuffer);
+    REPEAT(FILL_PHRASE_LAST_BYTE);
-    FILL_PHRASE_LAST_BYTE(13, byteBuffer);
+    REPEAT(FILL_PHRASE_SET_LENGTH);
-    FILL_PHRASE_LAST_BYTE(14, byteBuffer);
-    FILL_PHRASE_LAST_BYTE(15, byteBuffer);
-    FILL_PHRASE_SET_LENGTH(12, uintBuffer, lengthInBits);
-    FILL_PHRASE_SET_LENGTH(13, uintBuffer, lengthInBits);
-    FILL_PHRASE_SET_LENGTH(14, uintBuffer, lengthInBits);
-    FILL_PHRASE_SET_LENGTH(15, uintBuffer, lengthInBits);
 }
 
 #pragma managed

dotnet/WhiteRabbit.UnmanagedBridge/phraseset.h

@@ -1,3 +1,3 @@
 #pragma once
 
-void fillPhraseSet(__int64* bufferPointer, __int64* allWordsPointer, __int32* wordIndexes, unsigned __int64* permutationsPointer, int permutationOffset, int numberOfCharacters, int numberOfWords);
+void fillPhraseSet(__int64* bufferPointer, unsigned __int64* allWordsPointer, __int32* wordIndexes, unsigned __int64* permutationsPointer, int permutationOffset, int numberOfCharacters, int numberOfWords);