Rename SSE2 stuff to be generally SIMD since it supports at least SSE2

and Altivec (though Altivec is seemingly slower ATM...)

3 files changed, 168 insertions, 0 deletions

diff --git a/src/block/xtea_simd/info.txt b/src/block/xtea_simd/info.txt
new file mode 100644
index 000000000..0cdce14ef
--- /dev/null
+++ b/src/block/xtea_simd/info.txt
@@ -0,0 +1,16 @@
+realname "XTEA (SIMD)"
+
+define XTEA_SIMD
+
+load_on auto
+
+<add>
+xtea_simd.cpp
+xtea_simd.h
+</add>
+
+<requires>
+xtea
+simd_32
+simd_engine
+</requires>
diff --git a/src/block/xtea_simd/xtea_simd.cpp b/src/block/xtea_simd/xtea_simd.cpp
new file mode 100644
index 000000000..6151c355c
--- /dev/null
+++ b/src/block/xtea_simd/xtea_simd.cpp
@@ -0,0 +1,124 @@
+/*
+* XTEA in SIMD
+* (C) 2009 Jack Lloyd
+*
+* Distributed under the terms of the Botan license
+*/
+
+#include <botan/xtea_simd.h>
+#include <botan/loadstor.h>
+#include <botan/simd_32.h>
+
+namespace Botan {
+
+namespace {
+
+void xtea_encrypt_8(const byte in[64], byte out[64], const u32bit EK[64])
+   {
+   SIMD_32 L0 = SIMD_32::load_be(in     );
+   SIMD_32 R0 = SIMD_32::load_be(in + 16);
+   SIMD_32 L1 = SIMD_32::load_be(in + 32);
+   SIMD_32 R1 = SIMD_32::load_be(in + 48);
+
+   SIMD_32::transpose(L0, R0, L1, R1);
+
+   for(u32bit i = 0; i != 32; i += 2)
+      {
+      SIMD_32 K0(EK[2*i  ]);
+      SIMD_32 K1(EK[2*i+1]);
+      SIMD_32 K2(EK[2*i+2]);
+      SIMD_32 K3(EK[2*i+3]);
+
+      L0 += (((R0 << 4) ^ (R0 >> 5)) + R0) ^ K0;
+      L1 += (((R1 << 4) ^ (R1 >> 5)) + R1) ^ K0;
+
+      R0 += (((L0 << 4) ^ (L0 >> 5)) + L0) ^ K1;
+      R1 += (((L1 << 4) ^ (L1 >> 5)) + L1) ^ K1;
+
+      L0 += (((R0 << 4) ^ (R0 >> 5)) + R0) ^ K2;
+      L1 += (((R1 << 4) ^ (R1 >> 5)) + R1) ^ K2;
+
+      R0 += (((L0 << 4) ^ (L0 >> 5)) + L0) ^ K3;
+      R1 += (((L1 << 4) ^ (L1 >> 5)) + L1) ^ K3;
+      }
+
+   SIMD_32::transpose(L0, R0, L1, R1);
+
+   L0.store_be(out);
+   R0.store_be(out + 16);
+   L1.store_be(out + 32);
+   R1.store_be(out + 48);
+   }
+
+void xtea_decrypt_8(const byte in[64], byte out[64], const u32bit EK[64])
+   {
+   SIMD_32 L0 = SIMD_32::load_be(in     );
+   SIMD_32 R0 = SIMD_32::load_be(in + 16);
+   SIMD_32 L1 = SIMD_32::load_be(in + 32);
+   SIMD_32 R1 = SIMD_32::load_be(in + 48);
+
+   SIMD_32::transpose(L0, R0, L1, R1);
+
+   for(u32bit i = 0; i != 32; i += 2)
+      {
+      SIMD_32 K0(EK[63 - 2*i]);
+      SIMD_32 K1(EK[62 - 2*i]);
+      SIMD_32 K2(EK[61 - 2*i]);
+      SIMD_32 K3(EK[60 - 2*i]);
+
+      R0 -= (((L0 << 4) ^ (L0 >> 5)) + L0) ^ K0;
+      R1 -= (((L1 << 4) ^ (L1 >> 5)) + L1) ^ K0;
+
+      L0 -= (((R0 << 4) ^ (R0 >> 5)) + R0) ^ K1;
+      L1 -= (((R1 << 4) ^ (R1 >> 5)) + R1) ^ K1;
+
+      R0 -= (((L0 << 4) ^ (L0 >> 5)) + L0) ^ K2;
+      R1 -= (((L1 << 4) ^ (L1 >> 5)) + L1) ^ K2;
+
+      L0 -= (((R0 << 4) ^ (R0 >> 5)) + R0) ^ K3;
+      L1 -= (((R1 << 4) ^ (R1 >> 5)) + R1) ^ K3;
+      }
+
+   SIMD_32::transpose(L0, R0, L1, R1);
+
+   L0.store_be(out);
+   R0.store_be(out + 16);
+   L1.store_be(out + 32);
+   R1.store_be(out + 48);
+   }
+
+}
+
+/*
+* XTEA Encryption
+*/
+void XTEA_SIMD::encrypt_n(const byte in[], byte out[], u32bit blocks) const
+   {
+   while(blocks >= 8)
+      {
+      xtea_encrypt_8(in, out, this->EK);
+      in += 8 * BLOCK_SIZE;
+      out += 8 * BLOCK_SIZE;
+      blocks -= 8;
+      }
+
+   XTEA::encrypt_n(in, out, blocks);
+   }
+
+/*
+* XTEA Decryption
+*/
+void XTEA_SIMD::decrypt_n(const byte in[], byte out[], u32bit blocks) const
+   {
+   while(blocks >= 8)
+      {
+      xtea_decrypt_8(in, out, this->EK);
+      in += 8 * BLOCK_SIZE;
+      out += 8 * BLOCK_SIZE;
+      blocks -= 8;
+      }
+
+   XTEA::decrypt_n(in, out, blocks);
+   }
+
+}
diff --git a/src/block/xtea_simd/xtea_simd.h b/src/block/xtea_simd/xtea_simd.h
new file mode 100644
index 000000000..e4ce734ed
--- /dev/null
+++ b/src/block/xtea_simd/xtea_simd.h
@@ -0,0 +1,28 @@
+/*
+* XTEA in SIMD
+* (C) 2009 Jack Lloyd
+*
+* Distributed under the terms of the Botan license
+*/
+
+#ifndef BOTAN_XTEA_SIMD_H__
+#define BOTAN_XTEA_SIMD_H__
+
+#include <botan/xtea.h>
+
+namespace Botan {
+
+/*
+* XTEA (SIMD variant)
+*/
+class BOTAN_DLL XTEA_SIMD : public XTEA
+   {
+   public:
+      void encrypt_n(const byte in[], byte out[], u32bit blocks) const;
+      void decrypt_n(const byte in[], byte out[], u32bit blocks) const;
+      BlockCipher* clone() const { return new XTEA_SIMD; }
+   };
+
+}
+
+#endif