Port SIMON64 to Altivec

SIMON64 runs about 4x faster than C++ for POWER4 and friends. If POWER7 is available it goes back to full speed due to efficient unaligned loads

1 files changed, 26 insertions, 12 deletions

diff --git a/simon.cpp b/simon.cpp
index 4c48efd2..0e777306 100644
--- a/simon.cpp
+++ b/simon.cpp
@@ -7,6 +7,14 @@
 #include "misc.h"

 #include "cpu.h"

 

+#ifndef CRYPTOPP_INLINE

+# if defined(CRYPTOPP_DEBUG)

+#  define CRYPTOPP_INLINE static

+# else

+#  define CRYPTOPP_INLINE inline

+# endif

+#endif

+

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

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

 // #undef CRYPTOPP_SSSE3_AVAILABLE

@@ -228,11 +236,11 @@ extern size_t SIMON128_Dec_AdvancedProcessBlocks_SSSE3(const word64* subKeys, si
     const byte *inBlocks, const byte *xorBlocks, byte *outBlocks, size_t length, word32 flags);

 #endif

 

-#if (CRYPTOPP_POWER7_AVAILABLE)

-extern size_t SIMON64_Enc_AdvancedProcessBlocks_POWER7(const word32* subKeys, size_t rounds,

+#if (CRYPTOPP_ALTIVEC_AVAILABLE)

+extern size_t SIMON64_Enc_AdvancedProcessBlocks_ALTIVEC(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_POWER7(const word32* subKeys, size_t rounds,

+extern size_t SIMON64_Dec_AdvancedProcessBlocks_ALTIVEC(const word32* subKeys, size_t rounds,

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

 #endif

 

@@ -255,9 +263,9 @@ std::string SIMON64::Base::AlgorithmProvider() const
     if (HasNEON())

         return "NEON";

 # endif

-# if (CRYPTOPP_POWER7_AVAILABLE)

-    if (HasPower7())

-        return "Power7";

+# if (CRYPTOPP_ALTIVEC_AVAILABLE)

+    if (HasAltivec())

+        return "Altivec";

 # endif

 #endif

     return "C++";

@@ -292,6 +300,12 @@ void SIMON64::Base::UncheckedSetKey(const byte *userKey, unsigned int keyLength,
     default:

         CRYPTOPP_ASSERT(0);;

     }

+

+    // Altivec loads the current subkey as a 16-byte vector

+    // The extra elements ensure memory backs the last subkey.

+#if CRYPTOPP_ALTIVEC_AVAILABLE

+    m_rkeys.Grow(m_rkeys.size()+4);

+#endif

 }

 

 void SIMON64::Enc::ProcessAndXorBlock(const byte *inBlock, const byte *xorBlock, byte *outBlock) const

@@ -463,9 +477,9 @@ size_t SIMON64::Enc::AdvancedProcessBlocks(const byte *inBlocks, const byte *xor
         return SIMON64_Enc_AdvancedProcessBlocks_NEON(m_rkeys, (size_t)m_rounds,

             inBlocks, xorBlocks, outBlocks, length, flags);

 #endif

-#if (CRYPTOPP_POWER7_AVAILABLE)

-    if (HasPower7())

-        return SIMON64_Enc_AdvancedProcessBlocks_POWER7(m_rkeys, (size_t)m_rounds,

+#if (CRYPTOPP_ALTIVEC_AVAILABLE)

+    if (HasAltivec())

+        return SIMON64_Enc_AdvancedProcessBlocks_ALTIVEC(m_rkeys, (size_t)m_rounds,

             inBlocks, xorBlocks, outBlocks, length, flags);

 #endif

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

@@ -484,9 +498,9 @@ size_t SIMON64::Dec::AdvancedProcessBlocks(const byte *inBlocks, const byte *xor
         return SIMON64_Dec_AdvancedProcessBlocks_NEON(m_rkeys, (size_t)m_rounds,

             inBlocks, xorBlocks, outBlocks, length, flags);

 #endif

-#if (CRYPTOPP_POWER7_AVAILABLE)

-    if (HasPower7())

-        return SIMON64_Dec_AdvancedProcessBlocks_POWER7(m_rkeys, (size_t)m_rounds,

+#if (CRYPTOPP_ALTIVEC_AVAILABLE)

+    if (HasAltivec())

+        return SIMON64_Dec_AdvancedProcessBlocks_ALTIVEC(m_rkeys, (size_t)m_rounds,

             inBlocks, xorBlocks, outBlocks, length, flags);

 #endif

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