diff --git a/Appl/CEP/CS1/CS1_BGLProc/src/CacheAlignedAllocator.h b/Appl/CEP/CS1/CS1_BGLProc/src/CacheAlignedAllocator.h
deleted file mode 100644
index ebff5e89ef08fd7f51fa484d206e9ec6bcfeb620..0000000000000000000000000000000000000000
--- a/Appl/CEP/CS1/CS1_BGLProc/src/CacheAlignedAllocator.h
+++ /dev/null
@@ -1,41 +0,0 @@
-#ifndef LOFAR_APPL_CEP_CS1_CS1_BGL_PROC_CACHE_ALIGNED_ALLOCATOR_H
-#define LOFAR_APPL_CEP_CS1_CS1_BGL_PROC_CACHE_ALIGNED_ALLOCATOR_H
-
-#include <malloc.h>
-#include <memory>
-
-
-#if defined HAVE_BGL || defined HAVE_BGP
-#define CACHE_LINE_SIZE 32
-#define CACHE_ALIGNED __attribute__ ((aligned(CACHE_LINE_SIZE)))
-#else
-#define CACHE_LINE_SIZE 16
-#define CACHE_ALIGNED
-#endif
-
-
-namespace LOFAR {
-namespace CS1 {
-
-template <typename T> class CacheAlignedAllocator : public std::allocator<T>
-{
- public:
- typedef typename std::allocator<T>::size_type size_type;
- typedef typename std::allocator<T>::pointer pointer;
- typedef typename std::allocator<T>::const_pointer const_pointer;
-
- pointer allocate(size_type size, const_pointer /*hint*/ = 0)
- {
- return static_cast<pointer>(memalign(CACHE_LINE_SIZE, size * sizeof(T)));
- }
-
- void deallocate(pointer ptr, size_type /*size*/)
- {
- free(ptr);
- }
-};
-
-} // namespace CS1
-} // namespace LOFAR
-
-#endif
diff --git a/Appl/CEP/CS1/CS1_BGLProc/src/FCNP_ClientStream.cc b/Appl/CEP/CS1/CS1_BGLProc/src/FCNP_ClientStream.cc
index 75a4a952e07b950390dff83182413331950e4676..6297de0761ba93c19959c66a36d1294d001da180 100644
--- a/Appl/CEP/CS1/CS1_BGLProc/src/FCNP_ClientStream.cc
+++ b/Appl/CEP/CS1/CS1_BGLProc/src/FCNP_ClientStream.cc
@@ -26,11 +26,13 @@
#if defined HAVE_FCNP && defined HAVE_BGP
#include <Common/Timer.h>
+#include <CS1_Interface/AlignedStdAllocator.h>
#include <FCNP_ClientStream.h>
#include <fcnp_cn.h>
-//#include <algorithm>
+#include <cstring>
+#include <vector>
namespace LOFAR {
@@ -48,10 +50,10 @@ void FCNP_ClientStream::read(void *ptr, size_t size)
if (reinterpret_cast<size_t>(ptr) % 16 != 0 || size % 16 != 0) {
size_t alignedSize = (size + 15) & ~ (size_t) 15;
- char tmp[alignedSize] __attribute__ ((aligned(16)));
+ std::vector<char, AlignedStdAllocator<char, 16> > alignedBuffer(alignedSize);
- FCNP_CN::IONtoCN_ZeroCopy(tmp, alignedSize);
- memcpy(ptr, tmp, size);
+ FCNP_CN::IONtoCN_ZeroCopy(&alignedBuffer[0], alignedSize);
+ memcpy(ptr, &alignedBuffer[0], size);
} else {
FCNP_CN::IONtoCN_ZeroCopy(ptr, size);
}
@@ -64,10 +66,10 @@ void FCNP_ClientStream::write(const void *ptr, size_t size)
if (reinterpret_cast<size_t>(ptr) % 16 != 0 || size % 16 != 0) {
size_t alignedSize = (size + 15) & ~ (size_t) 15;
- char tmp[alignedSize] __attribute__ ((aligned(16)));
+ std::vector<char, AlignedStdAllocator<char, 16> > alignedBuffer(alignedSize);
- memcpy(tmp, ptr, size);
- FCNP_CN::CNtoION_ZeroCopy(tmp, alignedSize);
+ memcpy(&alignedBuffer[0], ptr, size);
+ FCNP_CN::CNtoION_ZeroCopy(&alignedBuffer[0], alignedSize);
} else {
FCNP_CN::CNtoION_ZeroCopy(ptr, size);
}
diff --git a/Appl/CEP/CS1/CS1_BGLProc/src/FIR.cc b/Appl/CEP/CS1/CS1_BGLProc/src/FIR.cc
index e84f1d4515f94a784f88c4010e7b48f836a852d8..ec265d09a161b8ee5b14c81df508333863b6d5cf 100644
--- a/Appl/CEP/CS1/CS1_BGLProc/src/FIR.cc
+++ b/Appl/CEP/CS1/CS1_BGLProc/src/FIR.cc
@@ -3,7 +3,6 @@
//# Includes
#include <FIR.h>
-#include <CacheAlignedAllocator.h>
namespace LOFAR {
namespace CS1 {
@@ -15,7 +14,7 @@ namespace CS1 {
// This is efficiently achieved by negating the FIR filter constants of all
// uneven FIR filters.
-const float FIR::weights[NR_SUBBAND_CHANNELS][NR_TAPS] CACHE_ALIGNED = {
+const float FIR::weights[NR_SUBBAND_CHANNELS][NR_TAPS] __attribute__ ((aligned(32))) = {
#if NR_SUBBAND_CHANNELS == 256 && NR_TAPS == 16
{ 0.011659500, -0.011535200, 0.005131880, 0.001219900,
-0.006891530, 0.011598600, -0.015420900, 1.000000000,
@@ -1047,7 +1046,7 @@ const float FIR::weights[NR_SUBBAND_CHANNELS][NR_TAPS] CACHE_ALIGNED = {
};
-const float FIR::bandPassCorrectionFactors[NR_SUBBAND_CHANNELS] CACHE_ALIGNED = {
+const float FIR::bandPassCorrectionFactors[NR_SUBBAND_CHANNELS] __attribute__ ((aligned(32))) = {
#if NR_SUBBAND_CHANNELS == 256 && NR_TAPS == 16
#if 1
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
diff --git a/Appl/CEP/CS1/CS1_BGLProc/src/PPF.h b/Appl/CEP/CS1/CS1_BGLProc/src/PPF.h
index ade78c9c437643f28a8d1a8c9f1ed067598a1dcd..fd72c65d5345169b96c0c35c7942793aa47829fe 100644
--- a/Appl/CEP/CS1/CS1_BGLProc/src/PPF.h
+++ b/Appl/CEP/CS1/CS1_BGLProc/src/PPF.h
@@ -9,7 +9,7 @@
#include <FIR.h>
#include <TransposedData.h>
#include <FilteredData.h>
-#include <CacheAlignedAllocator.h>
+#include <CS1_Interface/AlignedStdAllocator.h>
#include <boost/multi_array.hpp>
@@ -55,9 +55,9 @@ class PPF
boost::multi_array<FIR, 3> itsFIRs; //[itsNrStations][NR_POLARIZATIONS][NR_SUBBAND_CHANNELS]
boost::multi_array<fcomplex, 3> itsFFTinData; //[NR_TAPS - 1 + itsNrSamplesPerIntegration][NR_POLARIZATIONS][NR_SUBBAND_CHANNELS]
#else
- boost::multi_array<fcomplex, 2, CacheAlignedAllocator<fcomplex> > itsTmp; //[4][itsNrSamplesPerIntegration]
- boost::multi_array<fcomplex, 3, CacheAlignedAllocator<fcomplex> > itsFFTinData; //[itsNrSamplesPerIntegration][NR_POLARIZATIONS][NR_SUBBAND_CHANNELS + 4]
- boost::multi_array<fcomplex, 3, CacheAlignedAllocator<fcomplex> > itsFFToutData; //[2][NR_POLARIZATIONS][NR_SUBBAND_CHANNELS]
+ boost::multi_array<fcomplex, 2, AlignedStdAllocator<fcomplex, 32> > itsTmp; //[4][itsNrSamplesPerIntegration]
+ boost::multi_array<fcomplex, 3, AlignedStdAllocator<fcomplex, 32> > itsFFTinData; //[itsNrSamplesPerIntegration][NR_POLARIZATIONS][NR_SUBBAND_CHANNELS + 4]
+ boost::multi_array<fcomplex, 3, AlignedStdAllocator<fcomplex, 32> > itsFFToutData; //[2][NR_POLARIZATIONS][NR_SUBBAND_CHANNELS]
#endif
#if defined HAVE_FFTW3
diff --git a/Appl/CEP/CS1/CS1_IONProc/src/FCNP_ServerStream.cc b/Appl/CEP/CS1/CS1_IONProc/src/FCNP_ServerStream.cc
index 236431e226dcaac3de78edaf77687f31a919f2f4..03334ec7891007cf6b7abbfbd31560b5153ac720 100644
--- a/Appl/CEP/CS1/CS1_IONProc/src/FCNP_ServerStream.cc
+++ b/Appl/CEP/CS1/CS1_IONProc/src/FCNP_ServerStream.cc
@@ -26,11 +26,14 @@
#if defined HAVE_FCNP && defined __PPC__
#include <Common/Timer.h>
+#include <CS1_Interface/AlignedStdAllocator.h>
#include <FCNP_ServerStream.h>
#include <fcnp_ion.h>
#include <algorithm>
+#include <cstring>
+#include <vector>
namespace LOFAR {
@@ -83,10 +86,10 @@ void FCNP_ServerStream::write(const void *buf, size_t size)
if (reinterpret_cast<size_t>(buf) % 16 != 0 || size % 16 != 0) {
size_t alignedSize = (size + 15) & ~ (size_t) 15;
- char tmp[alignedSize] __attribute__ ((aligned(16)));
+ std::vector<char, AlignedStdAllocator<char, 16> > alignedBuffer(alignedSize);
- memcpy(tmp, buf, size);
- FCNP_ION::IONtoCN_ZeroCopy(itsCore, tmp, alignedSize);
+ memcpy(&alignedBuffer[0], buf, size);
+ FCNP_ION::IONtoCN_ZeroCopy(itsCore, &alignedBuffer[0], alignedSize);
} else {
FCNP_ION::IONtoCN_ZeroCopy(itsCore, const_cast<const void *>(buf), size);
}
@@ -99,10 +102,10 @@ void FCNP_ServerStream::read(void *buf, size_t size)
if (reinterpret_cast<size_t>(buf) % 16 != 0 || size % 16 != 0) {
size_t alignedSize = (size + 15) & ~ (size_t) 15;
- char tmp[alignedSize] __attribute__ ((aligned(16)));
+ std::vector<char, AlignedStdAllocator<char, 16> > alignedBuffer(alignedSize);
- FCNP_ION::CNtoION_ZeroCopy(itsCore, tmp, alignedSize);
- memcpy(buf, tmp, size);
+ FCNP_ION::CNtoION_ZeroCopy(itsCore, &alignedBuffer[0], alignedSize);
+ memcpy(buf, &alignedBuffer[0], size);
} else {
FCNP_ION::CNtoION_ZeroCopy(itsCore, buf, size);
}
diff --git a/Appl/CEP/CS1/CS1_IONProc/src/InputSection.cc b/Appl/CEP/CS1/CS1_IONProc/src/InputSection.cc
index 48e26068187988aef15d2f270eda4139f7ef1da5..b5379634e64cb4f16309625d39bc0dae4bda7b72 100644
--- a/Appl/CEP/CS1/CS1_IONProc/src/InputSection.cc
+++ b/Appl/CEP/CS1/CS1_IONProc/src/InputSection.cc
@@ -33,6 +33,7 @@
#include <InputThread.h>
#include <ION_Allocator.h>
//#include <TH_ZoidServer.h>
+#include <CS1_Interface/AlignedStdAllocator.h>
#include <CS1_Interface/BGL_Command.h>
#include <CS1_Interface/BGL_Mapping.h>
#include <CS1_Interface/SubbandMetaData.h>
@@ -381,7 +382,7 @@ void InputSection::process()
command.write(str);
- std::vector<SubbandMetaData> metaDataPerComputeNode(itsNrPsets);
+ std::vector<SubbandMetaData, AlignedStdAllocator<SubbandMetaData, 16> > metaDataPerComputeNode(itsNrPsets);
for (unsigned pset = 0; pset < itsNrPsets; pset ++) {
unsigned subband = itsNSubbandsPerPset * pset + subbandBase;
diff --git a/Appl/CEP/CS1/CS1_Interface/include/CS1_Interface/AlignedStdAllocator.h b/Appl/CEP/CS1/CS1_Interface/include/CS1_Interface/AlignedStdAllocator.h
new file mode 100644
index 0000000000000000000000000000000000000000..ae761abd98cee6e469e7947ab29d09c96d2afde6
--- /dev/null
+++ b/Appl/CEP/CS1/CS1_Interface/include/CS1_Interface/AlignedStdAllocator.h
@@ -0,0 +1,53 @@
+#ifndef LOFAR_APPL_CEP_CS1_CS1_INTERFACE_ALIGNED_ALLOCATOR_H
+#define LOFAR_APPL_CEP_CS1_CS1_INTERFACE_ALIGNED_ALLOCATOR_H
+
+#include <memory>
+#include <stdexcept>
+
+#if _POSIX_C_SOURCE >= 200112L || _XOPEN_SOURCE >= 600
+#include <cstdlib>
+#else
+#include <malloc.h>
+#endif
+
+
+namespace LOFAR {
+namespace CS1 {
+
+template <typename T, size_t ALIGNMENT> class AlignedStdAllocator : public std::allocator<T>
+{
+ public:
+ typedef typename std::allocator<T>::size_type size_type;
+ typedef typename std::allocator<T>::pointer pointer;
+ typedef typename std::allocator<T>::const_pointer const_pointer;
+
+ template <class U> struct rebind
+ {
+ typedef AlignedStdAllocator<U, ALIGNMENT> other;
+ };
+
+ pointer allocate(size_type size, const_pointer /*hint*/ = 0)
+ {
+ void *ptr;
+
+#if _POSIX_C_SOURCE >= 200112L || _XOPEN_SOURCE >= 600
+ if (posix_memalign(&ptr, ALIGNMENT, size * sizeof(T)) != 0)
+ throw std::bad_alloc();
+#else
+ if ((ptr = memalign(ALIGNMENT, size * sizeof(T))) == 0)
+ throw std::bad_alloc();
+#endif
+
+ return static_cast<pointer>(ptr);
+ }
+
+ void deallocate(pointer ptr, size_type /*size*/)
+ {
+ free(ptr);
+ }
+};
+
+} // namespace CS1
+} // namespace LOFAR
+
+#endif
diff --git a/Appl/CEP/CS1/CS1_Interface/include/CS1_Interface/Makefile.am b/Appl/CEP/CS1/CS1_Interface/include/CS1_Interface/Makefile.am
index 05e0f7ad9d7db5cff545931a208a6fd25a8c000e..a067287cd2ba74e65729682e8b18d73498bec379 100644
--- a/Appl/CEP/CS1/CS1_Interface/include/CS1_Interface/Makefile.am
+++ b/Appl/CEP/CS1/CS1_Interface/include/CS1_Interface/Makefile.am
@@ -1,4 +1,5 @@
pkginclude_HEADERS = Package__Version.h \
+ AlignedStdAllocator.h \
Allocator.h \
BGL_Command.h \
BGL_Configuration.h \