Changeset 85638 in vbox for trunk

Timestamp:

Aug 6, 2020 4:17:54 PM (5 years ago)

Author:

vboxsync

Message:

IPRT/sha3: Some VS2019 performance tweaks. bugref:9734

Location:

trunk/src/VBox/Runtime

Files:

• common/checksum/alt-sha3.cpp (modified) (3 diffs)
• testcase/tstRTDigest-2.cpp (modified) (1 diff)

trunk/src/VBox/Runtime/common/checksum/alt-sha3.cpp

@@ -33 +33 @@
 
 /** @def RTSHA3_FULL_UNROLL
- * Do full loop unrolling unless we're using VS2019 as it seems to degrate
- * performances there for some reason.  With gcc 10.2.1 on a recent Intel system
- * (10890XE), this results SHA3-512 throughput (tstRTDigest-2) increasing from
- * 83532 KiB/s to 194942 KiB/s against a text size jump from 5913 to 6929 bytes.
+ * Do full loop unrolling.
+ *
+ * With gcc 10.2.1 on a recent Intel system (10890XE), this results SHA3-512
+ * throughput (tstRTDigest-2) increasing from 83532 KiB/s to 194942 KiB/s
+ * against a text size jump from 5913 to 6929 bytes, i.e. +1016 bytes.
+ *
+ * With VS2019 on a half decent AMD system (3990X), this results in SHA3-512
+ * speedup from 147676 KiB/s to about 192770 KiB/s.  The text cost is +612 bytes
+ * (4496 to 5108).  When disabling the unrolling of Rho+Pi we get a little
+ * increase 196591 KiB/s (+3821) for some reason, saving 22 bytes of code.
  *
  * For comparison, openssl 1.1.1g assembly code (AMD64) achives 264915 KiB/s,
@@ -42 +48 @@
  * KECCAK_2X without ROL optimizations (they improve it to 203493 KiB/s).
  */
-#if !defined(_MSC_VER) || defined(DOXYGEN_RUNNING)
+#if !defined(IN_SUP_HARDENED_R3) || defined(DOXYGEN_RUNNING)
 # define RTSHA3_FULL_UNROLL
 #endif
@@ -147 +153 @@
          */
         {
-#ifndef RTSHA3_FULL_UNROLL
+#if !defined(RTSHA3_FULL_UNROLL) || defined(_MSC_VER) /* VS2019 is slightly slow with this section unrolled. go figure */
             static uint8_t const s_aidxState[] = {10,7,11,17,18,  3, 5,16, 8,21, 24, 4,15,23,19, 13,12, 2,20,14, 22, 9, 6, 1};
             static uint8_t const s_acRotate[]  = { 1,3, 6,10,15, 21,28,36,45,55,  2,14,27,41,56,  8,25,43,62,18, 39,61,20,44};

trunk/src/VBox/Runtime/testcase/tstRTDigest-2.cpp

@@ -2048 +2048 @@
 }
 
-int main()
+
+static unsigned checkArgs(int cArgs, char **papszArgs, const char *pszName, const char *pszFamily)
+{
+    if (cArgs <= 1)
+        return 1;
+    size_t const cchName   = strlen(pszName);
+    size_t const cchFamily = strlen(pszFamily);
+    for (int i = 1; i < cArgs; i++)
+    {
+        const char  *pszArg = papszArgs[i];
+        const char  *pszSep = strpbrk(pszArg, ":=");
+        size_t const cchCur = pszSep ? (size_t)(pszSep - pszArg) : strlen(pszArg);
+        if (   (cchCur == cchName   && RTStrNICmp(pszArg, pszName, cchCur) == 0)
+            || (cchCur == cchFamily && RTStrNICmp(pszArg, pszFamily, cchCur) == 0) )
+        {
+            if (!pszSep || pszSep[1] == '\0')
+                return 1;
+            return RTStrToUInt32(pszSep + 1);
+        }
+    }
+    return 0;
+}
+
+
+int main(int argc, char **argv)
 {
     RTTEST hTest;
-    int rc = RTTestInitAndCreate("tstRTDigest-2", &hTest);
-    if (rc)
-        return rc;
+    RTEXITCODE rcExit = RTTestInitExAndCreate(argc, &argv, 0, "tstRTDigest-2", &hTest);
+    if (rcExit != RTEXITCODE_SUCCESS)
+        return rcExit;
     RTTestBanner(hTest);
 
-    testMd2();
-    testMd4();
-    testMd5();
-    testSha1();
-    testSha256();
-    testSha224();
-    testSha512();
-    testSha384();
+#define DO(a_szName, a_szFamily, a_fnTestExpr) do { \
+            unsigned const cTimes = checkArgs(argc, argv, a_szName, a_szFamily); \
+            for (unsigned i = 0; i < cTimes; i++) { a_fnTestExpr; } \
+        } while (0)
+    DO("MD2",    "MD",   testMd2());
+    DO("MD4",    "MD",   testMd4());
+    DO("MD5",    "MD",   testMd5());
+    DO("SHA1",   "SHA",  testSha1());
+    DO("SHA256", "SHA2", testSha256());
+    DO("SHA224", "SHA2", testSha224());
+    DO("SHA512", "SHA2", testSha512());
+    DO("SHA384", "SHA2", testSha384());
 #ifndef IPRT_WITHOUT_SHA512T224
-    testSha512t224();
+    DO("SHA512T224", "SHA2", testSha512t224());
 #endif
 #ifndef IPRT_WITHOUT_SHA512T256
-    testSha512t256();
+    DO("SHA512T256", "SHA2", testSha512t256());
 #endif
-    testSha3_224();
-    testSha3_256();
-    testSha3_384();
-    testSha3_512();
+    DO("SHA3-224", "SHA3", testSha3_224());
+    DO("SHA3-256", "SHA3", testSha3_256());
+    DO("SHA3-384", "SHA3", testSha3_384());
+    DO("SHA3-512", "SHA3", testSha3_512());
 
     return RTTestSummaryAndDestroy(hTest);