EC-JERASURE: rewrite region-xor function using vector operations to get ~ x1.5 speedu...

diff --git a/src/osd/ErasureCodePluginJerasure/ErasureCodeJerasure.cc b/src/osd/ErasureCodePluginJerasure/ErasureCodeJerasure.cc
index fe656e58ee00f3a109b8a4bcff5fe08574711928..84c82c97151fcb9870c52fcb568aea770ba4b499 100644 (file)
--- a/src/osd/ErasureCodePluginJerasure/ErasureCodeJerasure.cc
+++ b/src/osd/ErasureCodePluginJerasure/ErasureCodeJerasure.cc
@@ -18,6 +18,7 @@
 #include <algorithm>
 #include "common/debug.h"
 #include "ErasureCodeJerasure.h"
+#include "vectorop.h"
 extern "C" {
 #include "jerasure.h"
 #include "reed_sol.h"
@@ -192,7 +193,11 @@ int ErasureCodeJerasureReedSolomonVandermonde::jerasure_decode(int *erasures,
 
 unsigned ErasureCodeJerasureReedSolomonVandermonde::get_alignment()
 {
-  return k*w*sizeof(int);
+  unsigned alignment = k*w*sizeof(int);
+  if ( ((w*sizeof(int))%LARGEST_VECTOR_WORDSIZE) )
+    alignment = k*w*LARGEST_VECTOR_WORDSIZE;
+  return alignment;
+
 }
 
 void ErasureCodeJerasureReedSolomonVandermonde::parse(const map<std::string,std::string> &parameters)
@@ -232,7 +237,10 @@ int ErasureCodeJerasureReedSolomonRAID6::jerasure_decode(int *erasures,
 
 unsigned ErasureCodeJerasureReedSolomonRAID6::get_alignment()
 {
-  return k*w*sizeof(int);
+  unsigned alignment = k*w*sizeof(int);
+  if ( ((w*sizeof(int))%LARGEST_VECTOR_WORDSIZE) )
+    alignment = k*w*LARGEST_VECTOR_WORDSIZE;
+  return alignment;
 }
 
 void ErasureCodeJerasureReedSolomonRAID6::parse(const map<std::string,std::string> &parameters)
@@ -274,7 +282,10 @@ int ErasureCodeJerasureCauchy::jerasure_decode(int *erasures,
 
 unsigned ErasureCodeJerasureCauchy::get_alignment()
 {
-  return k*w*packetsize*sizeof(int);
+  unsigned alignment = k*w*packetsize*sizeof(int);
+  if ( ((w*packetsize*sizeof(int))%LARGEST_VECTOR_WORDSIZE) )
+    alignment = k*w*packetsize*LARGEST_VECTOR_WORDSIZE;
+  return alignment;
 }
 
 void ErasureCodeJerasureCauchy::parse(const map<std::string,std::string> &parameters)
@@ -341,7 +352,10 @@ int ErasureCodeJerasureLiberation::jerasure_decode(int *erasures,
 
 unsigned ErasureCodeJerasureLiberation::get_alignment()
 {
-  return k*w*packetsize*sizeof(int);
+  unsigned alignment = k*w*packetsize*sizeof(int);
+  if ( ((w*packetsize*sizeof(int))%LARGEST_VECTOR_WORDSIZE) )
+    alignment = k*w*packetsize*LARGEST_VECTOR_WORDSIZE;
+  return alignment;
 }
 
 void ErasureCodeJerasureLiberation::parse(const map<std::string,std::string> &parameters)

diff --git a/src/osd/ErasureCodePluginJerasure/galois.c b/src/osd/ErasureCodePluginJerasure/galois.c
index 0de6fbd334cfe012df15ae028d2dfc6caeacde8b..f1ec347fadc0bd5343f84a5133142631cda626be 100755 (executable)
--- a/src/osd/ErasureCodePluginJerasure/galois.c
+++ b/src/osd/ErasureCodePluginJerasure/galois.c
@@ -49,8 +49,10 @@ POSSIBILITY OF SUCH DAMAGE.
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
+#include <assert.h>
 
 #include "galois.h"
+#include "vectorop.h"
 
 #define NONE (10)
 #define TABLE (11)
@@ -737,31 +739,38 @@ void galois_w32_region_multiply(char *region,      /* Region to multiply */
     }
   }
   return;
-
 }
 
+
 void galois_region_xor(           char *r1,         /* Region 1 */
                                   char *r2,         /* Region 2 */
                                   char *r3,         /* Sum region (r3 = r1 ^ r2) -- can be r1 or r2 */
                                   int nbytes)       /* Number of bytes in region */
 {
-  long *l1;
-  long *l2;
-  long *l3;
-  long *ltop;
-  char *ctop;
-  
-  ctop = r1 + nbytes;
-  ltop = (long *) ctop;
-  l1 = (long *) r1;
-  l2 = (long *) r2;
-  l3 = (long *) r3;
- 
-  while (l1 < ltop) {
-    *l3 = ((*l1)  ^ (*l2));
-    l1++;
-    l2++;
-    l3++;
+  if (nbytes%VECTOR_WORDSIZE) {
+    assert(!((long long)r1%sizeof(long)));
+    assert(!((long long)r2%sizeof(long)));
+    assert(!((long long)r3%sizeof(long)));
+    long* l1 = (long*)r1;
+    long* l2 = (long*)r2;
+    long* l3 = (long*)r3;
+    char *ctop = r1 + nbytes;
+    long* ltop = (long*)ctop;
+    while (l1 < ltop) {
+      *l3++ = ((*l1++)  ^ (*l2++));
+    }
+  } else {
+    assert(!((long long)r1%VECTOR_WORDSIZE));
+    assert(!((long long)r2%VECTOR_WORDSIZE));
+    assert(!((long long)r3%VECTOR_WORDSIZE));
+    vector_op_t* l1 = (vector_op_t*)r1;
+    vector_op_t* l2 = (vector_op_t*)r2;
+    vector_op_t* l3 = (vector_op_t*)r3;
+    char *ctop = r1 + nbytes;
+    vector_op_t* ltop = (vector_op_t*)ctop;
+    while (l1 < ltop) {
+      *l3++ = ((*l1++)  ^ (*l2++));
+    }
   }
 }
 

diff --git a/src/osd/ErasureCodePluginJerasure/vectorop.h b/src/osd/ErasureCodePluginJerasure/vectorop.h
new file mode 100644 (file)
index 0000000..5135f82
--- /dev/null
+++ b/src/osd/ErasureCodePluginJerasure/vectorop.h
@@ -0,0 +1,51 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*- 
+// vim: ts=8 sw=2 smarttab
+/*
+ * Ceph - scalable distributed file system
+ *
+ * Copyright (C) 2014 CERN/Switzerland
+ *               
+ *
+ * Authors: Andreas-Joachim Peters <andreas.joachim.peters@cern.ch> 
+ *          
+ *
+ *  This library is free software; you can redistribute it and/or
+ *  modify it under the terms of the GNU Lesser General Public
+ *  License as published by the Free Software Foundation; either
+ *  version 2.1 of the License, or (at your option) any later version.
+ * 
+ */
+
+#ifndef CEPH_VECTOROP_H
+#define        CEPH_VECTOROP_H
+
+// -------------------------------------------------------------------------
+// constant used in the block alignment function to allow for vector ops
+// -------------------------------------------------------------------------
+#define LARGEST_VECTOR_WORDSIZE 16
+
+// -------------------------------------------------------------------------
+// switch to 128-bit XOR operations if possible 
+// -------------------------------------------------------------------------
+#if __GNUC__ > 4 || \
+  (__GNUC__ == 4 && (__GNUC_MINOR__ >= 4) ) || \
+  (__clang__ == 1 )
+
+#ifdef VECTOROP_DEBUG
+#pragma message "* using 128-bit vector operations in " __FILE__ 
+#endif
+// -------------------------------------------------------------------------
+// use 128-bit pointer
+// -------------------------------------------------------------------------
+typedef long vector_op_t __attribute__ ((vector_size (16)));
+#define VECTOR_WORDSIZE 16
+#else
+// -------------------------------------------------------------------------
+// use 64-bit pointer
+// -------------------------------------------------------------------------
+typedef unsigned long long vector_op_t;
+#define VECTOR_WORDSIZE 8
+#endif
+
+#endif /* CEPH_VECTOROP_H */
+