Add experimental AVX base32 algo, modularise base32 out

1 files changed, 102 insertions, 0 deletions

diff --git a/onion_base32.c b/onion_base32.c
new file mode 100644
index 0000000..d70facf
--- /dev/null
+++ b/onion_base32.c
@@ -0,0 +1,102 @@
+#include <stdlib.h>
+#include <string.h>
+
+static const char base32_lookup[] = "abcdefghijklmnopqrstuvwxyz234567";
+
+int
+check_base32(char *subject) {
+	size_t offset = 0;
+
+	if ((offset = strspn(subject, base32_lookup)) != strlen(subject)) {
+		return offset;
+	}
+	return 0;
+}
+
+/* Simple and reliable base32 algorithm - "old trusty"
+ * Note: This is not a general base32 algorithm; it outputs only the
+ * first 16 base32 symbols of the input buffer, using only the first
+ * 20 bytes of that buffer.
+ */
+void
+onion_base32(char output[16], unsigned char sum[20]) {
+	size_t c = 0;
+	int i = 0;
+
+	for (i = 0; i < 10; i+=5) {
+		output[c++] = base32_lookup[sum[i] >> 3];
+		output[c++] = base32_lookup[((sum[i] & 0x07) << 2) | (sum[i+1] >> 6)];
+		output[c++] = base32_lookup[(sum[i+1] >> 1) & 0x1F];
+		output[c++] = base32_lookup[((sum[i+1] & 1) << 4) | (sum[i+2] >> 4)];
+		output[c++] = base32_lookup[((sum[i+2] & 0x0F) << 1) | ((sum[i+3] & 0x80) >> 7)];
+		output[c++] = base32_lookup[(sum[i+3] >> 2) & 0x1F];
+		output[c++] = base32_lookup[((sum[i+3] & 0x03) << 3) | (sum[i+4] >> 5)];
+		output[c++] = base32_lookup[sum[i+4] & 0x1F];
+	}
+}
+
+#ifdef AVX_ONION_BASE32
+#include <immintrin.h>
+
+/* A slightly-parallel base32 algorithm using AVX
+ * Note: This is not a general base32 algorithm; it outputs only the
+ * first 16 base32 symbols of the input buffer, using only the first
+ * 20 bytes of that buffer.
+ *
+ * Somewhat inspired by http://www.alfredklomp.com/programming/sse-base64/
+ */
+void
+onion_base32_avx(char output[16], unsigned char sum[20]) {
+	__m128i res;
+	__m128i ssum;
+	__m128i masklow5;
+	__m128i lmask, l;
+	__m128i nmask, n;
+
+	ssum = _mm_loadu_si128((__m128i*)sum);
+
+	/* FIXME seems a little hacky */
+	masklow5 = _mm_set1_epi32(0x1F000000);
+	masklow5 = _mm_slli_epi64(masklow5, 32);
+
+	ssum = _mm_shuffle_epi8(ssum,
+		_mm_setr_epi8(9,9,9,9,8,7,6,5,4,4,4,4,3,2,1,0 ));
+
+	/* remember how I said "slightly parallel" ? */
+	res = _mm_srli_epi64(ssum, 3) & masklow5;
+	masklow5 = _mm_srli_epi64(masklow5, 8);
+
+	res |= _mm_srli_epi64(ssum, 6) & masklow5;
+	masklow5 = _mm_srli_epi64(masklow5, 8);
+
+	res |= _mm_srli_epi64(ssum, 9) & masklow5;
+	masklow5 = _mm_srli_epi64(masklow5, 8);
+
+	res |= _mm_srli_epi64(ssum, 12) & masklow5;
+	masklow5 = _mm_srli_epi64(masklow5, 8);
+
+	res |= _mm_srli_epi64(ssum, 15) & masklow5;
+	masklow5 = _mm_srli_epi64(masklow5, 8);
+
+	res |= _mm_srli_epi64(ssum, 18) & masklow5;
+	masklow5 = _mm_srli_epi64(masklow5, 8);
+
+	res |= _mm_srli_epi64(ssum, 21) & masklow5;
+	masklow5 = _mm_srli_epi64(masklow5, 8);
+
+	res |= _mm_srli_epi64(ssum, 24) & masklow5;
+	masklow5 = _mm_srli_epi64(masklow5, 8);
+
+
+	res = _mm_shuffle_epi8(res,
+		_mm_setr_epi8(15,14,13,12,11,10,9,8,7,6,5,4,3,2,1,0));
+
+	lmask = _mm_cmplt_epi8(res, _mm_set1_epi8(26));
+	nmask = _mm_andnot_si128(lmask, _mm_cmplt_epi8(res, _mm_set1_epi8(32)));
+
+	l = lmask & _mm_add_epi8(res, _mm_set1_epi8('a'));
+	n = nmask & _mm_add_epi8(res, _mm_set1_epi8('2' - 26));
+
+	_mm_storeu_si128((__m128i*)output, l|n);
+}
+#endif /* ifdef AVX_ONION_BASE32 */