Add SIMON-64 SSE intrinsics

Performance went from about 29 cpb (C++) to about 11.1 cpb (SSE)

1 files changed, 43 insertions, 2 deletions

diff --git a/simon.cpp b/simon.cpp
index 565a5205..a966a4ca 100644
--- a/simon.cpp
+++ b/simon.cpp
@@ -7,9 +7,10 @@
 #include "misc.h"

 #include "cpu.h"

 

-// Uncomment for benchmarking C++ against SSE2 or NEON.

-// Do so in both speck.cpp and speck-simd.cpp.

+// Uncomment for benchmarking C++ against SSE or NEON.

+// Do so in both simon.cpp and simon-simd.cpp.

 // #undef CRYPTOPP_SSSE3_AVAILABLE

+// #undef CRYPTOPP_SSE41_AVAILABLE

 // #undef CRYPTOPP_ARM_NEON_AVAILABLE

 

 // Disable NEON/ASIMD for Cortex-A53 and A57. The shifts are too slow and C/C++ is about

@@ -206,6 +207,14 @@ extern size_t SIMON128_Dec_AdvancedProcessBlocks_NEON(const word64* subKeys, siz
     const byte *inBlocks, const byte *xorBlocks, byte *outBlocks, size_t length, word32 flags);

 #endif

 

+#if defined(CRYPTOPP_SSE41_AVAILABLE)

+extern size_t SIMON64_Enc_AdvancedProcessBlocks_SSE41(const word32* subKeys, size_t rounds,

+    const byte *inBlocks, const byte *xorBlocks, byte *outBlocks, size_t length, word32 flags);

+

+extern size_t SIMON64_Dec_AdvancedProcessBlocks_SSE41(const word32* subKeys, size_t rounds,

+    const byte *inBlocks, const byte *xorBlocks, byte *outBlocks, size_t length, word32 flags);

+#endif

+

 #if defined(CRYPTOPP_SSSE3_AVAILABLE)

 extern size_t SIMON128_Enc_AdvancedProcessBlocks_SSSE3(const word64* subKeys, size_t rounds,

     const byte *inBlocks, const byte *xorBlocks, byte *outBlocks, size_t length, word32 flags);

@@ -376,6 +385,38 @@ void SIMON128::Dec::ProcessAndXorBlock(const byte *inBlock, const byte *xorBlock
 }

 

 #if defined(CRYPTOPP_SIMON_ADVANCED_PROCESS_BLOCKS)

+size_t SIMON64::Enc::AdvancedProcessBlocks(const byte *inBlocks, const byte *xorBlocks,

+        byte *outBlocks, size_t length, word32 flags) const

+{

+#if defined(CRYPTOPP_SSE41_AVAILABLE)

+    if (HasSSE41())

+        return SIMON64_Enc_AdvancedProcessBlocks_SSE41(m_rkeys, (size_t)m_rounds,

+            inBlocks, xorBlocks, outBlocks, length, flags);

+#endif

+#if defined(CRYPTOPP_ARM_NEON_AVAILABLE)

+    if (HasNEON())

+        return SIMON64_Enc_AdvancedProcessBlocks_NEON(m_rkeys, (size_t)m_rounds,

+            inBlocks, xorBlocks, outBlocks, length, flags);

+#endif

+    return BlockTransformation::AdvancedProcessBlocks(inBlocks, xorBlocks, outBlocks, length, flags);

+}

+

+size_t SIMON64::Dec::AdvancedProcessBlocks(const byte *inBlocks, const byte *xorBlocks,

+        byte *outBlocks, size_t length, word32 flags) const

+{

+#if defined(CRYPTOPP_SSE41_AVAILABLE)

+    if (HasSSE41())

+        return SIMON64_Dec_AdvancedProcessBlocks_SSE41(m_rkeys, (size_t)m_rounds,

+            inBlocks, xorBlocks, outBlocks, length, flags);

+#endif

+#if defined(CRYPTOPP_ARM_NEON_AVAILABLE)

+    if (HasNEON())

+        return SIMON64_Dec_AdvancedProcessBlocks_NEON(m_rkeys, (size_t)m_rounds,

+            inBlocks, xorBlocks, outBlocks, length, flags);

+#endif

+    return BlockTransformation::AdvancedProcessBlocks(inBlocks, xorBlocks, outBlocks, length, flags);

+}

+

 size_t SIMON128::Enc::AdvancedProcessBlocks(const byte *inBlocks, const byte *xorBlocks,

         byte *outBlocks, size_t length, word32 flags) const

 {