Add SPECK-64 SSE intrinsics

Performance went from about 11.9 cpb (C++) to about 4.5 cpb (SSE)

1 files changed, 35 insertions, 1 deletions

diff --git a/speck.cpp b/speck.cpp
index 00e73a0c..f6762094 100644
--- a/speck.cpp
+++ b/speck.cpp
@@ -7,9 +7,10 @@
 #include "misc.h"

 #include "cpu.h"

 

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

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

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

@@ -176,6 +177,9 @@ ANONYMOUS_NAMESPACE_END
 NAMESPACE_BEGIN(CryptoPP)

 

 #if defined(CRYPTOPP_ARM_NEON_AVAILABLE)

+extern size_t SPECK64_Enc_AdvancedProcessBlocks_NEON(const word32* subKeys, size_t rounds,

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

+

 extern size_t SPECK128_Enc_AdvancedProcessBlocks_NEON(const word64* subKeys, size_t rounds,

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

 

@@ -183,6 +187,14 @@ extern size_t SPECK128_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 SPECK64_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 SPECK64_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 SPECK128_Enc_AdvancedProcessBlocks_SSSE3(const word64* subKeys, size_t rounds,

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

@@ -353,6 +365,28 @@ void SPECK128::Dec::ProcessAndXorBlock(const byte *inBlock, const byte *xorBlock
 }

 

 #if defined(CRYPTOPP_SPECK_ADVANCED_PROCESS_BLOCKS)

+size_t SPECK64::Enc::AdvancedProcessBlocks(const byte *inBlocks, const byte *xorBlocks,

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

+{

+#if defined(CRYPTOPP_SSE41_AVAILABLE)

+    if (HasSSE41())

+        return SPECK64_Enc_AdvancedProcessBlocks_SSE41(m_rkeys, (size_t)m_rounds,

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

+#endif

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

+}

+

+size_t SPECK64::Dec::AdvancedProcessBlocks(const byte *inBlocks, const byte *xorBlocks,

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

+{

+#if defined(CRYPTOPP_SSE41_AVAILABLE)

+    if (HasSSE41())

+        return SPECK64_Dec_AdvancedProcessBlocks_SSE41(m_rkeys, (size_t)m_rounds,

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

+#endif

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

+}

+

 size_t SPECK128::Enc::AdvancedProcessBlocks(const byte *inBlocks, const byte *xorBlocks,

         byte *outBlocks, size_t length, word32 flags) const

 {