Just incorporate the X64 ASM version of Salsa20/12 for X64 platforms. This gives us (for example) 1.5gb/sec encryption on a Core i5 2.8ghz.

ext/x64-salsa2012-asm/README.md

@@ -0,0 +1,6 @@
+Blazingly fast X64 ASM implementation of Salsa20/12
+======
+
+This is ripped from the [cnacl](https://github.com/cjdelisle/cnacl) source. The actual code is by Danial J. Bernstein and is in the public domain.
+
+This is included on Linux and Mac 64-bit builds and is significantly faster than the SSE intrinsics or C versions. It's used for packet encode/decode only since its use differs a bit from the regular Salsa20 C++ class. Specifically it lacks the ability to be called on multiple blocks, preferring instead to take a key and a single stream to encrypt and that's it.

ext/x64-salsa2012-asm/salsa2012.h

@@ -0,0 +1,13 @@
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// output, outlen, nonce, key (256-bit / 32-byte)
+extern int zt_salsa2012_amd64_xmm6(unsigned char *,unsigned long long,const unsigned char *,const unsigned char *);
+
+// ciphertext, message, mlen, nonce, key
+extern int zt_salsa2012_amd64_xmm6_xor(unsigned char *,const unsigned char *,unsigned long long,const unsigned char *,const unsigned char *);
+
+#ifdef __cplusplus
+}
+#endif