[svn] gcry - r1416 - in trunk: . cipher tests
svn author wk
cvs at cvs.gnupg.org
Fri Dec 11 17:32:05 CET 2009
Author: wk
Date: 2009-12-11 17:32:04 +0100 (Fri, 11 Dec 2009)
New Revision: 1416
Modified:
trunk/NEWS
trunk/cipher/ChangeLog
trunk/cipher/sha256.c
trunk/cipher/sha512.c
trunk/tests/benchmark.c
Log:
Boost performance of SHA-512 and SHA-256.
Modified: trunk/cipher/ChangeLog
===================================================================
--- trunk/cipher/ChangeLog 2009-12-10 17:27:55 UTC (rev 1415)
+++ trunk/cipher/ChangeLog 2009-12-11 16:32:04 UTC (rev 1416)
@@ -1,3 +1,14 @@
+2009-12-11 Werner Koch <wk at g10code.com>
+
+ * sha256.c (Cho, Maj, Sum0, Sum1): Turn macros into inline
+ functions.
+ (transform): Partly unroll to interweave the chain variables
+
+ * sha512.c (ROTR, Ch, Maj, Sum0, Sum1): Turn macros into inline
+ functions.
+ (transform): Partly Unroll to interweave the chain variables.
+ Suggested by Christian Grothoff.
+
2009-12-10 Werner Koch <wk at g10code.com>
* Makefile.am (o_flag_munging): New.
Modified: trunk/NEWS
===================================================================
--- trunk/NEWS 2009-12-10 17:27:55 UTC (rev 1415)
+++ trunk/NEWS 2009-12-11 16:32:04 UTC (rev 1416)
@@ -11,6 +11,9 @@
* Fix a hang on some W2000 machines.
+ * Boost SHA-512 performance by 30% on ia32 boxes and gcc 4.3; SHA-256
+ goes up by 25%.
+
* Interface changes relative to the 1.4.2 release:
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
GCRY_CIPHER_MODE_AESWRAP NEW.
Modified: trunk/cipher/sha256.c
===================================================================
--- trunk/cipher/sha256.c 2009-12-10 17:27:55 UTC (rev 1415)
+++ trunk/cipher/sha256.c 2009-12-11 16:32:04 UTC (rev 1416)
@@ -94,10 +94,6 @@
/*
Transform the message X which consists of 16 32-bit-words. See FIPS
180-2 for details. */
-#define Cho(x,y,z) (z ^ (x & (y ^ z))) /* (4.2) same as SHA-1's F1 */
-#define Maj(x,y,z) ((x & y) | (z & (x|y))) /* (4.3) same as SHA-1's F3 */
-#define Sum0(x) (ror ((x), 2) ^ ror ((x), 13) ^ ror ((x), 22)) /* (4.4) */
-#define Sum1(x) (ror ((x), 6) ^ ror ((x), 11) ^ ror ((x), 25)) /* (4.5) */
#define S0(x) (ror ((x), 7) ^ ror ((x), 18) ^ ((x) >> 3)) /* (4.6) */
#define S1(x) (ror ((x), 17) ^ ror ((x), 19) ^ ((x) >> 10)) /* (4.7) */
#define R(a,b,c,d,e,f,g,h,k,w) do \
@@ -113,6 +109,35 @@
b = a; \
a = t1 + t2; \
} while (0)
+
+/* (4.2) same as SHA-1's F1. */
+static inline u32
+Cho (u32 x, u32 y, u32 z)
+{
+ return (z ^ (x & (y ^ z)));
+}
+
+/* (4.3) same as SHA-1's F3 */
+static inline u32
+Maj (u32 x, u32 y, u32 z)
+{
+ return ((x & y) | (z & (x|y)));
+}
+
+/* (4.4) */
+static inline u32
+Sum0 (u32 x)
+{
+ return (ror (x, 2) ^ ror (x, 13) ^ ror (x, 22));
+}
+
+/* (4.5) */
+static inline u32
+Sum1 (u32 x)
+{
+ return (ror (x, 6) ^ ror (x, 11) ^ ror (x, 25));
+}
+
static void
transform (SHA256_CONTEXT *hd, const unsigned char *data)
@@ -171,9 +196,56 @@
for (; i < 64; i++)
w[i] = S1(w[i-2]) + w[i-7] + S0(w[i-15]) + w[i-16];
- for (i=0; i < 64; i++)
- R(a,b,c,d,e,f,g,h,K[i],w[i]);
+ for (i=0; i < 64;)
+ {
+#if 0
+ R(a,b,c,d,e,f,g,h,K[i],w[i]);
+ i++;
+#else
+ t1 = h + Sum1 (e) + Cho (e, f, g) + K[i] + w[i];
+ t2 = Sum0 (a) + Maj (a, b, c);
+ d += t1;
+ h = t1 + t2;
+ t1 = g + Sum1 (d) + Cho (d, e, f) + K[i+1] + w[i+1];
+ t2 = Sum0 (h) + Maj (h, a, b);
+ c += t1;
+ g = t1 + t2;
+
+ t1 = f + Sum1 (c) + Cho (c, d, e) + K[i+2] + w[i+2];
+ t2 = Sum0 (g) + Maj (g, h, a);
+ b += t1;
+ f = t1 + t2;
+
+ t1 = e + Sum1 (b) + Cho (b, c, d) + K[i+3] + w[i+3];
+ t2 = Sum0 (f) + Maj (f, g, h);
+ a += t1;
+ e = t1 + t2;
+
+ t1 = d + Sum1 (a) + Cho (a, b, c) + K[i+4] + w[i+4];
+ t2 = Sum0 (e) + Maj (e, f, g);
+ h += t1;
+ d = t1 + t2;
+
+ t1 = c + Sum1 (h) + Cho (h, a, b) + K[i+5] + w[i+5];
+ t2 = Sum0 (d) + Maj (d, e, f);
+ g += t1;
+ c = t1 + t2;
+
+ t1 = b + Sum1 (g) + Cho (g, h, a) + K[i+6] + w[i+6];
+ t2 = Sum0 (c) + Maj (c, d, e);
+ f += t1;
+ b = t1 + t2;
+
+ t1 = a + Sum1 (f) + Cho (f, g, h) + K[i+7] + w[i+7];
+ t2 = Sum0 (b) + Maj (b, c, d);
+ e += t1;
+ a = t1 + t2;
+
+ i += 8;
+#endif
+ }
+
hd->h0 += a;
hd->h1 += b;
hd->h2 += c;
@@ -183,10 +255,6 @@
hd->h6 += g;
hd->h7 += h;
}
-#undef Cho
-#undef Maj
-#undef Sum0
-#undef Sum1
#undef S0
#undef S1
#undef R
Modified: trunk/cipher/sha512.c
===================================================================
--- trunk/cipher/sha512.c 2009-12-10 17:27:55 UTC (rev 1415)
+++ trunk/cipher/sha512.c 2009-12-11 16:32:04 UTC (rev 1416)
@@ -98,6 +98,36 @@
}
+static inline u64
+ROTR (u64 x, u64 n)
+{
+ return ((x >> n) | (x << (64 - n)));
+}
+
+static inline u64
+Ch (u64 x, u64 y, u64 z)
+{
+ return ((x & y) ^ ( ~x & z));
+}
+
+static inline u64
+Maj (u64 x, u64 y, u64 z)
+{
+ return ((x & y) ^ (x & z) ^ (y & z));
+}
+
+static inline u64
+Sum0 (u64 x)
+{
+ return (ROTR (x, 28) ^ ROTR (x, 34) ^ ROTR (x, 39));
+}
+
+static inline u64
+Sum1 (u64 x)
+{
+ return (ROTR (x, 14) ^ ROTR (x, 18) ^ ROTR (x, 41));
+}
+
/****************
* Transform the message W which consists of 16 64-bit-words
*/
@@ -182,21 +212,26 @@
}
#endif
-#define ROTR(x,n) (((x)>>(n)) | ((x)<<(64-(n))))
-#define Ch(x,y,z) (((x) & (y)) ^ ((~(x)) & (z)))
-#define Maj(x,y,z) (((x) & (y)) ^ ((x) & (z)) ^ ((y) & (z)))
-#define Sum0(x) (ROTR((x),28) ^ ROTR((x),34) ^ ROTR((x),39))
-#define Sum1(x) (ROTR((x),14) ^ ROTR((x),18) ^ ROTR((x),41))
#define S0(x) (ROTR((x),1) ^ ROTR((x),8) ^ ((x)>>7))
#define S1(x) (ROTR((x),19) ^ ROTR((x),61) ^ ((x)>>6))
for (t = 16; t < 80; t++)
w[t] = S1 (w[t - 2]) + w[t - 7] + S0 (w[t - 15]) + w[t - 16];
- for (t = 0; t < 80; t++)
+
+ for (t = 0; t < 80; )
{
u64 t1, t2;
+ /* Performance on a AMD Athlon(tm) Dual Core Processor 4050e
+ with gcc 4.3.3 using gcry_md_hash_buffer of each 10000 bytes
+ initialized to 0,1,2,3...255,0,... and 1000 iterations:
+
+ Not unrolled with macros: 440ms
+ Unrolled with macros: 350ms
+ Unrolled with inline: 330ms
+ */
+#if 0 /* Not unrolled. */
t1 = h + Sum1 (e) + Ch (e, f, g) + k[t] + w[t];
t2 = Sum0 (a) + Maj (a, b, c);
h = g;
@@ -207,12 +242,53 @@
c = b;
b = a;
a = t1 + t2;
+ t++;
+#else /* Unrolled to interweave the chain variables. */
+ t1 = h + Sum1 (e) + Ch (e, f, g) + k[t] + w[t];
+ t2 = Sum0 (a) + Maj (a, b, c);
+ d += t1;
+ h = t1 + t2;
- /* printf("t=%d a=%016llX b=%016llX c=%016llX d=%016llX "
- "e=%016llX f=%016llX g=%016llX h=%016llX\n",t,a,b,c,d,e,f,g,h); */
+ t1 = g + Sum1 (d) + Ch (d, e, f) + k[t+1] + w[t+1];
+ t2 = Sum0 (h) + Maj (h, a, b);
+ c += t1;
+ g = t1 + t2;
+
+ t1 = f + Sum1 (c) + Ch (c, d, e) + k[t+2] + w[t+2];
+ t2 = Sum0 (g) + Maj (g, h, a);
+ b += t1;
+ f = t1 + t2;
+
+ t1 = e + Sum1 (b) + Ch (b, c, d) + k[t+3] + w[t+3];
+ t2 = Sum0 (f) + Maj (f, g, h);
+ a += t1;
+ e = t1 + t2;
+
+ t1 = d + Sum1 (a) + Ch (a, b, c) + k[t+4] + w[t+4];
+ t2 = Sum0 (e) + Maj (e, f, g);
+ h += t1;
+ d = t1 + t2;
+
+ t1 = c + Sum1 (h) + Ch (h, a, b) + k[t+5] + w[t+5];
+ t2 = Sum0 (d) + Maj (d, e, f);
+ g += t1;
+ c = t1 + t2;
+
+ t1 = b + Sum1 (g) + Ch (g, h, a) + k[t+6] + w[t+6];
+ t2 = Sum0 (c) + Maj (c, d, e);
+ f += t1;
+ b = t1 + t2;
+
+ t1 = a + Sum1 (f) + Ch (f, g, h) + k[t+7] + w[t+7];
+ t2 = Sum0 (b) + Maj (b, c, d);
+ e += t1;
+ a = t1 + t2;
+
+ t += 8;
+#endif
}
- /* update chaining vars */
+ /* Update chaining vars. */
hd->h0 += a;
hd->h1 += b;
hd->h2 += c;
Modified: trunk/tests/benchmark.c
===================================================================
--- trunk/tests/benchmark.c 2009-12-10 17:27:55 UTC (rev 1415)
+++ trunk/tests/benchmark.c 2009-12-11 16:32:04 UTC (rev 1416)
@@ -47,6 +47,9 @@
/* Number of cipher repetitions. */
static int cipher_repetitions;
+/* Number of hash repetitions. */
+static int hash_repetitions;
+
/* Whether fips mode was active at startup. */
static int in_fips_mode;
@@ -387,8 +390,10 @@
{
int algo;
gcry_md_hd_t hd;
- int i;
+ int i, repcount;
char buf[1000];
+ char *largebuf;
+ char digest[512/8];
gcry_error_t err = GPG_ERR_NO_ERROR;
if (!algoname)
@@ -421,30 +426,57 @@
printf ("%-12s", gcry_md_algo_name (algo));
start_timer ();
- for (i=0; i < 1000; i++)
- gcry_md_write (hd, buf, sizeof buf);
+ for (repcount=0; repcount < hash_repetitions; repcount++)
+ for (i=0; i < 1000; i++)
+ gcry_md_write (hd, buf, sizeof buf);
gcry_md_final (hd);
stop_timer ();
printf (" %s", elapsed_time ());
+ fflush (stdout);
gcry_md_reset (hd);
start_timer ();
- for (i=0; i < 10000; i++)
- gcry_md_write (hd, buf, sizeof buf/10);
+ for (repcount=0; repcount < hash_repetitions; repcount++)
+ for (i=0; i < 10000; i++)
+ gcry_md_write (hd, buf, sizeof buf/10);
gcry_md_final (hd);
stop_timer ();
printf (" %s", elapsed_time ());
+ fflush (stdout);
gcry_md_reset (hd);
start_timer ();
- for (i=0; i < 1000000; i++)
- gcry_md_write (hd, "", 1);
+ for (repcount=0; repcount < hash_repetitions; repcount++)
+ for (i=0; i < 1000000; i++)
+ gcry_md_write (hd, "", 1);
gcry_md_final (hd);
stop_timer ();
printf (" %s", elapsed_time ());
+ fflush (stdout);
gcry_md_close (hd);
+
+ /* Now 100 hash operations on 10000 bytes using the fast function.
+ We initialize the buffer so that all memory pages are committed
+ and we have repeatable values. */
+ if (gcry_md_get_algo_dlen (algo) > sizeof digest)
+ die ("digest buffer too short\n");
+
+ largebuf = malloc (10000);
+ if (!largebuf)
+ die ("out of core\n");
+ for (i=0; i < 10000; i++)
+ largebuf[i] = i;
+ start_timer ();
+ for (repcount=0; repcount < hash_repetitions; repcount++)
+ for (i=0; i < 100; i++)
+ gcry_md_hash_buffer (algo, digest, largebuf, 10000);
+ stop_timer ();
+ printf (" %s", elapsed_time ());
+ free (largebuf);
+
putchar ('\n');
+ fflush (stdout);
}
static void
@@ -1053,7 +1085,7 @@
large_buffers = 1;
argc--; argv++;
}
- else if (!strcmp (*argv, "--cipher-repetition"))
+ else if (!strcmp (*argv, "--cipher-repetitions"))
{
argc--; argv++;
if (argc)
@@ -1062,6 +1094,15 @@
argc--; argv++;
}
}
+ else if (!strcmp (*argv, "--hash-repetitions"))
+ {
+ argc--; argv++;
+ if (argc)
+ {
+ hash_repetitions = atoi(*argv);
+ argc--; argv++;
+ }
+ }
else if (!strcmp (*argv, "--fips"))
{
argc--; argv++;
@@ -1096,10 +1137,10 @@
gcry_control (GCRYCTL_INITIALIZATION_FINISHED, 0);
-
-
if (cipher_repetitions < 1)
cipher_repetitions = 1;
+ if (hash_repetitions < 1)
+ hash_repetitions = 1;
if ( !argc )
{
More information about the Gnupg-commits
mailing list