From 0f7e584913fe015013db51a3e29b22d733f2e664 Mon Sep 17 00:00:00 2001
From: Jorrit Schaap <schaap@astron.nl>
Date: Fri, 15 Mar 2019 07:50:49 +0000
Subject: [PATCH] COB-60: moved reamining cuda files up one level to the
one-and-only implementation dir.
---
.../CobaltTest/test/tManyPartTABOutput.cc | 2 +-
.../CobaltTest/test/tMultiPartTABOutput.cc | 2 +-
RTCP/Cobalt/GPUProc/src/CMakeLists.txt | 14 +-
.../GPUProc/src/{cuda => }/KernelFactory.cc | 0
RTCP/Cobalt/GPUProc/src/KernelFactory.h | 133 +++-
.../GPUProc/src/MultiDimArrayHostBuffer.h | 53 +-
.../src/{cuda => }/PerformanceCounter.cc | 0
RTCP/Cobalt/GPUProc/src/PerformanceCounter.h | 56 +-
RTCP/Cobalt/GPUProc/src/cuda/CMakeLists.txt | 7 -
RTCP/Cobalt/GPUProc/src/cuda/KernelFactory.h | 140 ----
.../src/cuda/MultiDimArrayHostBuffer.h | 62 --
.../GPUProc/src/cuda/PerformanceCounter.h | 65 --
RTCP/Cobalt/GPUProc/src/cuda/gpu_incl.h | 33 -
RTCP/Cobalt/GPUProc/src/cuda/gpu_wrapper.h | 627 ------------------
.../GPUProc/src/{cuda => }/cuda_config.h.in | 0
RTCP/Cobalt/GPUProc/src/gpu_incl.h | 12 +-
.../GPUProc/src/{cuda => }/gpu_utils.cc | 0
.../GPUProc/src/{cuda => }/gpu_wrapper.cc | 0
RTCP/Cobalt/GPUProc/src/gpu_wrapper.h | 601 ++++++++++++++++-
.../GPUProc/src/{cuda => }/gpu_wrapper.tcc | 0
.../test/SubbandProcs/tCorrelatorStep.cc | 2 +-
RTCP/Cobalt/GPUProc/test/cuda/tFFT_leakage.cc | 2 +-
RTCP/Cobalt/GPUProc/test/cuda/tGPUWrapper.cc | 2 +-
23 files changed, 794 insertions(+), 1019 deletions(-)
rename RTCP/Cobalt/GPUProc/src/{cuda => }/KernelFactory.cc (100%)
rename RTCP/Cobalt/GPUProc/src/{cuda => }/PerformanceCounter.cc (100%)
delete mode 100644 RTCP/Cobalt/GPUProc/src/cuda/CMakeLists.txt
delete mode 100644 RTCP/Cobalt/GPUProc/src/cuda/KernelFactory.h
delete mode 100644 RTCP/Cobalt/GPUProc/src/cuda/MultiDimArrayHostBuffer.h
delete mode 100644 RTCP/Cobalt/GPUProc/src/cuda/PerformanceCounter.h
delete mode 100644 RTCP/Cobalt/GPUProc/src/cuda/gpu_incl.h
delete mode 100644 RTCP/Cobalt/GPUProc/src/cuda/gpu_wrapper.h
rename RTCP/Cobalt/GPUProc/src/{cuda => }/cuda_config.h.in (100%)
rename RTCP/Cobalt/GPUProc/src/{cuda => }/gpu_utils.cc (100%)
rename RTCP/Cobalt/GPUProc/src/{cuda => }/gpu_wrapper.cc (100%)
rename RTCP/Cobalt/GPUProc/src/{cuda => }/gpu_wrapper.tcc (100%)
diff --git a/RTCP/Cobalt/CobaltTest/test/tManyPartTABOutput.cc b/RTCP/Cobalt/CobaltTest/test/tManyPartTABOutput.cc
index d310005bdb5..3cd86fb6639 100644
--- a/RTCP/Cobalt/CobaltTest/test/tManyPartTABOutput.cc
+++ b/RTCP/Cobalt/CobaltTest/test/tManyPartTABOutput.cc
@@ -30,7 +30,7 @@
#include <Common/LofarLogger.h>
#include <CoInterface/Parset.h>
#include <GPUProc/Pipelines/Pipeline.h>
-#include <GPUProc/cuda/SubbandProcs/SubbandProcOutputData.h>
+#include <GPUProc/SubbandProcs/SubbandProcOutputData.h>
#include <GPUProc/Station/StationInput.h>
#include <GPUProc/Storage/StorageProcesses.h>
diff --git a/RTCP/Cobalt/CobaltTest/test/tMultiPartTABOutput.cc b/RTCP/Cobalt/CobaltTest/test/tMultiPartTABOutput.cc
index f7cb9a1464c..9177c52d59d 100644
--- a/RTCP/Cobalt/CobaltTest/test/tMultiPartTABOutput.cc
+++ b/RTCP/Cobalt/CobaltTest/test/tMultiPartTABOutput.cc
@@ -30,7 +30,7 @@
#include <Common/LofarLogger.h>
#include <CoInterface/Parset.h>
#include <GPUProc/Pipelines/Pipeline.h>
-#include <GPUProc/cuda/SubbandProcs/SubbandProcOutputData.h>
+#include <GPUProc/SubbandProcs/SubbandProcOutputData.h>
#include <GPUProc/Station/StationInput.h>
#include <GPUProc/Storage/StorageProcesses.h>
diff --git a/RTCP/Cobalt/GPUProc/src/CMakeLists.txt b/RTCP/Cobalt/GPUProc/src/CMakeLists.txt
index 8b52611b0b5..a73efffb36a 100644
--- a/RTCP/Cobalt/GPUProc/src/CMakeLists.txt
+++ b/RTCP/Cobalt/GPUProc/src/CMakeLists.txt
@@ -7,6 +7,11 @@ execute_process(COMMAND ${CMAKE_COMMAND} -E create_symlink
${CMAKE_CURRENT_SOURCE_DIR}
${CMAKE_BINARY_DIR}/include/${PACKAGE_NAME})
+configure_file(
+ "${CMAKE_CURRENT_SOURCE_DIR}/cuda_config.h.in"
+ "${CMAKE_BINARY_DIR}/include/cuda_config.h" # internal, no need to install
+)
+
set(_gpuproc_sources
#Package__Version.cc
BandPass.cc
@@ -26,10 +31,10 @@ set(_gpuproc_sources
)
list(APPEND _gpuproc_sources
- cuda/gpu_wrapper.cc
- cuda/gpu_utils.cc
- cuda/KernelFactory.cc
- cuda/PerformanceCounter.cc
+ gpu_wrapper.cc
+ gpu_utils.cc
+ KernelFactory.cc
+ PerformanceCounter.cc
Kernels/Kernel.cc
Kernels/BeamFormerKernel.cc
Kernels/BeamFormerTransposeKernel.cc
@@ -67,7 +72,6 @@ list(APPEND _gpuproc_sources
# SubbandProcs/UHEP_SubbandProc.cc
)
-add_subdirectory(cuda)
lofar_add_library(gpuproc ${_gpuproc_sources})
if(CUDA_cufft_LIBRARY)
diff --git a/RTCP/Cobalt/GPUProc/src/cuda/KernelFactory.cc b/RTCP/Cobalt/GPUProc/src/KernelFactory.cc
similarity index 100%
rename from RTCP/Cobalt/GPUProc/src/cuda/KernelFactory.cc
rename to RTCP/Cobalt/GPUProc/src/KernelFactory.cc
diff --git a/RTCP/Cobalt/GPUProc/src/KernelFactory.h b/RTCP/Cobalt/GPUProc/src/KernelFactory.h
index 8a3bcef789d..9fc75474dd1 100644
--- a/RTCP/Cobalt/GPUProc/src/KernelFactory.h
+++ b/RTCP/Cobalt/GPUProc/src/KernelFactory.h
@@ -1,6 +1,6 @@
-//# KernelFactory.h: Factory for Kernel objects.
+//# KernelFactory.h
//#
-//# Copyright (C) 2013 ASTRON (Netherlands Institute for Radio Astronomy)
+//# Copyright (C) 2012-2013 ASTRON (Netherlands Institute for Radio Astronomy)
//# P.O. Box 2, 7990 AA Dwingeloo, The Netherlands
//#
//# This file is part of the LOFAR software suite.
@@ -19,23 +19,122 @@
//#
//# $Id$
-// \file
-// Factory for Kernel objects.
+#ifndef LOFAR_GPUPROC_CUDA_KERNELFACTORY_H
+#define LOFAR_GPUPROC_CUDA_KERNELFACTORY_H
-#ifndef LOFAR_GPUPROC_KERNELFACTORY_H
-#define LOFAR_GPUPROC_KERNELFACTORY_H
+#include <string>
+#include <CoInterface/Parset.h>
+#include <GPUProc/Kernels/Kernel.h>
+#include <GPUProc/gpu_wrapper.h>
+#include <GPUProc/gpu_utils.h>
-#if defined (USE_CUDA) && defined (USE_OPENCL)
-# error "Either CUDA or OpenCL must be enabled, not both"
-#endif
+namespace LOFAR
+{
+ namespace Cobalt
+ {
+ // Abstract base class of the templated KernelFactory class.
+ class KernelFactoryBase
+ {
+ public:
+ // Pure virtual destructor, because this is an abstract base class.
+ virtual ~KernelFactoryBase() = 0;
-#if defined (USE_CUDA)
-# include "cuda/KernelFactory.h"
-#elif defined (USE_OPENCL)
-# include "opencl/KernelFactory.h"
-#else
-# error "Either CUDA or OpenCL must be enabled, not neither"
-#endif
+ protected:
+ // Return compile definitions to use when creating PTX code for any
+ // Kernel.
+ CompileDefinitions
+ compileDefinitions(const Kernel::Parameters& param) const;
-#endif
+ // Return compile flags to use when creating PTX code for any Kernel.
+ CompileFlags
+ compileFlags(const Kernel::Parameters& param) const;
+ };
+
+ // Declaration of a generic factory class. For each concrete Kernel class
+ // (e.g. FIR_FilterKernel), a specialization must exist of the constructor
+ // and of the bufferSize() method.
+ template<typename T> class KernelFactory : public KernelFactoryBase
+ {
+ public:
+ // typedef typename T::Parameters Parameters;
+ typedef typename T::BufferType BufferType;
+ typedef typename T::Buffers Buffers;
+
+ // Construct a factory for creating Kernel objects of type \c T, using the
+ // settings provided by \a params.
+ KernelFactory(const typename T::Parameters ¶ms) :
+ itsParameters(params),
+ itsPTX(_createPTX())
+ {
+ }
+
+ // Create a new Kernel object of type \c T.
+ T* create(const gpu::Stream& stream,
+ gpu::DeviceMemory &inputBuffer,
+ gpu::DeviceMemory &outputBuffer) const
+ {
+ const typename T::Buffers buffers(inputBuffer, outputBuffer);
+
+ return create(stream, buffers);
+ }
+
+ // Return required buffer size for \a bufferType
+ size_t bufferSize(BufferType bufferType) const
+ {
+ return itsParameters.bufferSize(bufferType);
+ }
+
+ private:
+ // Used by the constructors to construct the PTX from the other
+ // members.
+ std::string _createPTX() const {
+ return createPTX(T::theirSourceFile,
+ compileDefinitions(),
+ compileFlags());
+ }
+ // Create a new Kernel object of type \c T.
+ T* create(const gpu::Stream& stream,
+ const typename T::Buffers& buffers) const
+ {
+ // Since we use overlapping input/output buffers, their size
+ // could be larger than we need.
+ ASSERTSTR(buffers.input.size() >= bufferSize(T::INPUT_DATA),
+ "Require " << bufferSize(T::INPUT_DATA) << " bytes for input, "
+ "but buffer is only " << buffers.input.size() << " bytes.");
+ ASSERTSTR(buffers.output.size() >= bufferSize(T::OUTPUT_DATA),
+ "Require " << bufferSize(T::OUTPUT_DATA) << " bytes for output, "
+ "but buffer is only " << buffers.output.size() << " bytes.");
+
+ return new T(
+ stream, createModule(stream.getContext(),
+ T::theirSourceFile,
+ itsPTX),
+ buffers, itsParameters);
+ }
+
+ // Return compile definitions to use when creating PTX code for kernels of
+ // type \c T, using the parameters stored in \c itsParameters.
+ CompileDefinitions compileDefinitions() const {
+ return KernelFactoryBase::compileDefinitions(itsParameters);
+ }
+
+ // Return compile flags to use when creating PTX code for kernels of type
+ // \c T.
+ CompileFlags compileFlags() const {
+ return KernelFactoryBase::compileFlags(itsParameters);
+ }
+
+ // Additional parameters needed to create a Kernel object of type \c T.
+ typename T::Parameters itsParameters;
+
+ // PTX code, generated for kernels of type \c T, using information in the
+ // Parset that was passed to the constructor.
+ std::string itsPTX;
+ };
+
+ } // namespace Cobalt
+
+} // namespace LOFAR
+
+#endif
diff --git a/RTCP/Cobalt/GPUProc/src/MultiDimArrayHostBuffer.h b/RTCP/Cobalt/GPUProc/src/MultiDimArrayHostBuffer.h
index dee99d41fa5..dafb5ba201a 100644
--- a/RTCP/Cobalt/GPUProc/src/MultiDimArrayHostBuffer.h
+++ b/RTCP/Cobalt/GPUProc/src/MultiDimArrayHostBuffer.h
@@ -1,6 +1,5 @@
//# MultiDimArrayHostBuffer.h
-//#
-//# Copyright (C) 2013 ASTRON (Netherlands Institute for Radio Astronomy)
+//# Copyright (C) 2012-2013 ASTRON (Netherlands Institute for Radio Astronomy)
//# P.O. Box 2, 7990 AA Dwingeloo, The Netherlands
//#
//# This file is part of the LOFAR software suite.
@@ -19,23 +18,45 @@
//#
//# $Id$
-// \file
-// Support for our multi-dim array-ed GPU host buffer.
+#ifndef LOFAR_GPUPROC_CUDA_MULTI_DIM_ARRAY_HOST_BUFFER_H
+#define LOFAR_GPUPROC_CUDA_MULTI_DIM_ARRAY_HOST_BUFFER_H
-#ifndef LOFAR_GPUPROC_MULTI_DIM_ARRAY_HOST_BUFFER_H
-#define LOFAR_GPUPROC_MULTI_DIM_ARRAY_HOST_BUFFER_H
+#include <CoInterface/MultiDimArray.h>
-#if defined (USE_CUDA) && defined (USE_OPENCL)
-# error "Either CUDA or OpenCL must be enabled, not both"
-#endif
+#include "gpu_wrapper.h"
-#if defined (USE_CUDA)
-# include "cuda/MultiDimArrayHostBuffer.h"
-#elif defined (USE_OPENCL)
-# include "opencl/MultiDimArrayHostBuffer.h"
-#else
-# error "Either CUDA or OpenCL must be enabled, not neither"
-#endif
+namespace LOFAR
+{
+ namespace Cobalt
+ {
+
+ // A MultiDimArray allocated as a HostBuffer
+ // Note: Elements are not constructed/destructed.
+ template <typename T, unsigned DIM>
+ class MultiDimArrayHostBuffer : public gpu::HostMemory,
+ public MultiDimArray<T, DIM>
+ {
+ public:
+ template <typename ExtentList>
+ MultiDimArrayHostBuffer(const ExtentList &extents, const gpu::Context &context,
+ unsigned int flags = 0)
+ :
+ HostMemory(context, MultiDimArray<T, DIM>::nrElements(extents) * sizeof(T), flags),
+ MultiDimArray<T, DIM>(extents, gpu::HostMemory::get<T>(), false)
+ {
+ }
+
+ using HostMemory::size;
+
+ private:
+ MultiDimArrayHostBuffer(); // don't use
+ MultiDimArrayHostBuffer(const MultiDimArrayHostBuffer<T, DIM> &rhs); // don't use
+ MultiDimArrayHostBuffer<T, DIM> &operator=(const MultiDimArrayHostBuffer<T, DIM> &rhs); // don't use
+ using MultiDimArray<T, DIM>::resize; // don't use
+ };
+
+ } // namespace Cobalt
+} // namespace LOFAR
#endif
diff --git a/RTCP/Cobalt/GPUProc/src/cuda/PerformanceCounter.cc b/RTCP/Cobalt/GPUProc/src/PerformanceCounter.cc
similarity index 100%
rename from RTCP/Cobalt/GPUProc/src/cuda/PerformanceCounter.cc
rename to RTCP/Cobalt/GPUProc/src/PerformanceCounter.cc
diff --git a/RTCP/Cobalt/GPUProc/src/PerformanceCounter.h b/RTCP/Cobalt/GPUProc/src/PerformanceCounter.h
index d7bca1ca404..5748b5f4d19 100644
--- a/RTCP/Cobalt/GPUProc/src/PerformanceCounter.h
+++ b/RTCP/Cobalt/GPUProc/src/PerformanceCounter.h
@@ -1,6 +1,5 @@
//# PerformanceCounter.h
-//#
-//# Copyright (C) 2013 ASTRON (Netherlands Institute for Radio Astronomy)
+//# Copyright (C) 2012-2013 ASTRON (Netherlands Institute for Radio Astronomy)
//# P.O. Box 2, 7990 AA Dwingeloo, The Netherlands
//#
//# This file is part of the LOFAR software suite.
@@ -19,23 +18,48 @@
//#
//# $Id$
-// \file
-// Support GPU kernel performance timing.
+#ifndef LOFAR_GPUPROC_CUDA_PERFORMANCECOUNTER_H
+#define LOFAR_GPUPROC_CUDA_PERFORMANCECOUNTER_H
-#ifndef LOFAR_PERFORMANCE_COUNTER_H
-#define LOFAR_PERFORMANCE_COUNTER_H
-#if defined (USE_CUDA) && defined (USE_OPENCL)
-# error "Either CUDA or OpenCL must be enabled, not both"
-#endif
+#include <GPUProc/gpu_wrapper.h>
+#include <CoInterface/RunningStatistics.h>
-#if defined (USE_CUDA)
-# include "cuda/PerformanceCounter.h"
-#elif defined (USE_OPENCL)
-# include "opencl/PerformanceCounter.h"
-#else
-# error "Either CUDA or OpenCL must be enabled, not neither"
-#endif
+namespace LOFAR
+{
+ namespace Cobalt
+ {
+ class PerformanceCounter
+ {
+ public:
+ PerformanceCounter(const gpu::Context &context, const std::string &name);
+ ~PerformanceCounter();
+
+ void recordStart(const gpu::Stream &stream);
+ void recordStop(const gpu::Stream &stream);
+
+ // Warning: user must make sure that the counter is not running!
+ RunningStatistics getStats() { logTime(); return stats; }
+
+ private:
+ const std::string name;
+
+ // Public event: it needs to be inserted into a stream.
+ // @{
+ gpu::Event start;
+ gpu::Event stop;
+ // @}
+
+ // Whether we have posted events that still need to be
+ // processed in logTime()
+ bool recording;
+
+ RunningStatistics stats;
+
+ void logTime();
+ };
+ }
+}
#endif
diff --git a/RTCP/Cobalt/GPUProc/src/cuda/CMakeLists.txt b/RTCP/Cobalt/GPUProc/src/cuda/CMakeLists.txt
deleted file mode 100644
index 17eac6967b6..00000000000
--- a/RTCP/Cobalt/GPUProc/src/cuda/CMakeLists.txt
+++ /dev/null
@@ -1,7 +0,0 @@
-# $Id$
-
-configure_file(
- "${CMAKE_CURRENT_SOURCE_DIR}/cuda_config.h.in"
- "${CMAKE_BINARY_DIR}/include/cuda_config.h" # internal, no need to install
-)
-
diff --git a/RTCP/Cobalt/GPUProc/src/cuda/KernelFactory.h b/RTCP/Cobalt/GPUProc/src/cuda/KernelFactory.h
deleted file mode 100644
index 9fc75474dd1..00000000000
--- a/RTCP/Cobalt/GPUProc/src/cuda/KernelFactory.h
+++ /dev/null
@@ -1,140 +0,0 @@
-//# KernelFactory.h
-//#
-//# Copyright (C) 2012-2013 ASTRON (Netherlands Institute for Radio Astronomy)
-//# P.O. Box 2, 7990 AA Dwingeloo, The Netherlands
-//#
-//# This file is part of the LOFAR software suite.
-//# The LOFAR software suite is free software: you can redistribute it and/or
-//# modify it under the terms of the GNU General Public License as published
-//# by the Free Software Foundation, either version 3 of the License, or
-//# (at your option) any later version.
-//#
-//# The LOFAR software suite is distributed in the hope that it will be useful,
-//# but WITHOUT ANY WARRANTY; without even the implied warranty of
-//# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
-//# GNU General Public License for more details.
-//#
-//# You should have received a copy of the GNU General Public License along
-//# with the LOFAR software suite. If not, see <http://www.gnu.org/licenses/>.
-//#
-//# $Id$
-
-#ifndef LOFAR_GPUPROC_CUDA_KERNELFACTORY_H
-#define LOFAR_GPUPROC_CUDA_KERNELFACTORY_H
-
-#include <string>
-#include <CoInterface/Parset.h>
-#include <GPUProc/Kernels/Kernel.h>
-#include <GPUProc/gpu_wrapper.h>
-#include <GPUProc/gpu_utils.h>
-
-namespace LOFAR
-{
- namespace Cobalt
- {
- // Abstract base class of the templated KernelFactory class.
- class KernelFactoryBase
- {
- public:
- // Pure virtual destructor, because this is an abstract base class.
- virtual ~KernelFactoryBase() = 0;
-
- protected:
- // Return compile definitions to use when creating PTX code for any
- // Kernel.
- CompileDefinitions
- compileDefinitions(const Kernel::Parameters& param) const;
-
- // Return compile flags to use when creating PTX code for any Kernel.
- CompileFlags
- compileFlags(const Kernel::Parameters& param) const;
- };
-
- // Declaration of a generic factory class. For each concrete Kernel class
- // (e.g. FIR_FilterKernel), a specialization must exist of the constructor
- // and of the bufferSize() method.
- template<typename T> class KernelFactory : public KernelFactoryBase
- {
- public:
- // typedef typename T::Parameters Parameters;
- typedef typename T::BufferType BufferType;
- typedef typename T::Buffers Buffers;
-
- // Construct a factory for creating Kernel objects of type \c T, using the
- // settings provided by \a params.
- KernelFactory(const typename T::Parameters ¶ms) :
- itsParameters(params),
- itsPTX(_createPTX())
- {
- }
-
- // Create a new Kernel object of type \c T.
- T* create(const gpu::Stream& stream,
- gpu::DeviceMemory &inputBuffer,
- gpu::DeviceMemory &outputBuffer) const
- {
- const typename T::Buffers buffers(inputBuffer, outputBuffer);
-
- return create(stream, buffers);
- }
-
- // Return required buffer size for \a bufferType
- size_t bufferSize(BufferType bufferType) const
- {
- return itsParameters.bufferSize(bufferType);
- }
-
- private:
- // Used by the constructors to construct the PTX from the other
- // members.
- std::string _createPTX() const {
- return createPTX(T::theirSourceFile,
- compileDefinitions(),
- compileFlags());
- }
-
- // Create a new Kernel object of type \c T.
- T* create(const gpu::Stream& stream,
- const typename T::Buffers& buffers) const
- {
- // Since we use overlapping input/output buffers, their size
- // could be larger than we need.
- ASSERTSTR(buffers.input.size() >= bufferSize(T::INPUT_DATA),
- "Require " << bufferSize(T::INPUT_DATA) << " bytes for input, "
- "but buffer is only " << buffers.input.size() << " bytes.");
- ASSERTSTR(buffers.output.size() >= bufferSize(T::OUTPUT_DATA),
- "Require " << bufferSize(T::OUTPUT_DATA) << " bytes for output, "
- "but buffer is only " << buffers.output.size() << " bytes.");
-
- return new T(
- stream, createModule(stream.getContext(),
- T::theirSourceFile,
- itsPTX),
- buffers, itsParameters);
- }
-
- // Return compile definitions to use when creating PTX code for kernels of
- // type \c T, using the parameters stored in \c itsParameters.
- CompileDefinitions compileDefinitions() const {
- return KernelFactoryBase::compileDefinitions(itsParameters);
- }
-
- // Return compile flags to use when creating PTX code for kernels of type
- // \c T.
- CompileFlags compileFlags() const {
- return KernelFactoryBase::compileFlags(itsParameters);
- }
-
- // Additional parameters needed to create a Kernel object of type \c T.
- typename T::Parameters itsParameters;
-
- // PTX code, generated for kernels of type \c T, using information in the
- // Parset that was passed to the constructor.
- std::string itsPTX;
- };
-
- } // namespace Cobalt
-
-} // namespace LOFAR
-
-#endif
diff --git a/RTCP/Cobalt/GPUProc/src/cuda/MultiDimArrayHostBuffer.h b/RTCP/Cobalt/GPUProc/src/cuda/MultiDimArrayHostBuffer.h
deleted file mode 100644
index dafb5ba201a..00000000000
--- a/RTCP/Cobalt/GPUProc/src/cuda/MultiDimArrayHostBuffer.h
+++ /dev/null
@@ -1,62 +0,0 @@
-//# MultiDimArrayHostBuffer.h
-//# Copyright (C) 2012-2013 ASTRON (Netherlands Institute for Radio Astronomy)
-//# P.O. Box 2, 7990 AA Dwingeloo, The Netherlands
-//#
-//# This file is part of the LOFAR software suite.
-//# The LOFAR software suite is free software: you can redistribute it and/or
-//# modify it under the terms of the GNU General Public License as published
-//# by the Free Software Foundation, either version 3 of the License, or
-//# (at your option) any later version.
-//#
-//# The LOFAR software suite is distributed in the hope that it will be useful,
-//# but WITHOUT ANY WARRANTY; without even the implied warranty of
-//# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
-//# GNU General Public License for more details.
-//#
-//# You should have received a copy of the GNU General Public License along
-//# with the LOFAR software suite. If not, see <http://www.gnu.org/licenses/>.
-//#
-//# $Id$
-
-#ifndef LOFAR_GPUPROC_CUDA_MULTI_DIM_ARRAY_HOST_BUFFER_H
-#define LOFAR_GPUPROC_CUDA_MULTI_DIM_ARRAY_HOST_BUFFER_H
-
-#include <CoInterface/MultiDimArray.h>
-
-#include "gpu_wrapper.h"
-
-namespace LOFAR
-{
- namespace Cobalt
- {
-
- // A MultiDimArray allocated as a HostBuffer
- // Note: Elements are not constructed/destructed.
- template <typename T, unsigned DIM>
- class MultiDimArrayHostBuffer : public gpu::HostMemory,
- public MultiDimArray<T, DIM>
- {
- public:
- template <typename ExtentList>
- MultiDimArrayHostBuffer(const ExtentList &extents, const gpu::Context &context,
- unsigned int flags = 0)
- :
- HostMemory(context, MultiDimArray<T, DIM>::nrElements(extents) * sizeof(T), flags),
- MultiDimArray<T, DIM>(extents, gpu::HostMemory::get<T>(), false)
- {
- }
-
- using HostMemory::size;
-
- private:
- MultiDimArrayHostBuffer(); // don't use
- MultiDimArrayHostBuffer(const MultiDimArrayHostBuffer<T, DIM> &rhs); // don't use
- MultiDimArrayHostBuffer<T, DIM> &operator=(const MultiDimArrayHostBuffer<T, DIM> &rhs); // don't use
- using MultiDimArray<T, DIM>::resize; // don't use
- };
-
- } // namespace Cobalt
-} // namespace LOFAR
-
-#endif
-
diff --git a/RTCP/Cobalt/GPUProc/src/cuda/PerformanceCounter.h b/RTCP/Cobalt/GPUProc/src/cuda/PerformanceCounter.h
deleted file mode 100644
index 5748b5f4d19..00000000000
--- a/RTCP/Cobalt/GPUProc/src/cuda/PerformanceCounter.h
+++ /dev/null
@@ -1,65 +0,0 @@
-//# PerformanceCounter.h
-//# Copyright (C) 2012-2013 ASTRON (Netherlands Institute for Radio Astronomy)
-//# P.O. Box 2, 7990 AA Dwingeloo, The Netherlands
-//#
-//# This file is part of the LOFAR software suite.
-//# The LOFAR software suite is free software: you can redistribute it and/or
-//# modify it under the terms of the GNU General Public License as published
-//# by the Free Software Foundation, either version 3 of the License, or
-//# (at your option) any later version.
-//#
-//# The LOFAR software suite is distributed in the hope that it will be useful,
-//# but WITHOUT ANY WARRANTY; without even the implied warranty of
-//# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
-//# GNU General Public License for more details.
-//#
-//# You should have received a copy of the GNU General Public License along
-//# with the LOFAR software suite. If not, see <http://www.gnu.org/licenses/>.
-//#
-//# $Id$
-
-#ifndef LOFAR_GPUPROC_CUDA_PERFORMANCECOUNTER_H
-#define LOFAR_GPUPROC_CUDA_PERFORMANCECOUNTER_H
-
-
-#include <GPUProc/gpu_wrapper.h>
-#include <CoInterface/RunningStatistics.h>
-
-namespace LOFAR
-{
- namespace Cobalt
- {
- class PerformanceCounter
- {
- public:
- PerformanceCounter(const gpu::Context &context, const std::string &name);
- ~PerformanceCounter();
-
- void recordStart(const gpu::Stream &stream);
- void recordStop(const gpu::Stream &stream);
-
- // Warning: user must make sure that the counter is not running!
- RunningStatistics getStats() { logTime(); return stats; }
-
- private:
- const std::string name;
-
- // Public event: it needs to be inserted into a stream.
- // @{
- gpu::Event start;
- gpu::Event stop;
- // @}
-
- // Whether we have posted events that still need to be
- // processed in logTime()
- bool recording;
-
- RunningStatistics stats;
-
- void logTime();
- };
- }
-}
-
-#endif
-
diff --git a/RTCP/Cobalt/GPUProc/src/cuda/gpu_incl.h b/RTCP/Cobalt/GPUProc/src/cuda/gpu_incl.h
deleted file mode 100644
index 7e56db7b5b4..00000000000
--- a/RTCP/Cobalt/GPUProc/src/cuda/gpu_incl.h
+++ /dev/null
@@ -1,33 +0,0 @@
-//# gpu_incl.h: portable CUDA header to mirror OpenCL sources
-//# Copyright (C) 2012-2013 ASTRON (Netherlands Institute for Radio Astronomy)
-//# P.O. Box 2, 7990 AA Dwingeloo, The Netherlands
-//#
-//# This file is part of the LOFAR software suite.
-//# The LOFAR software suite is free software: you can redistribute it and/or
-//# modify it under the terms of the GNU General Public License as published
-//# by the Free Software Foundation, either version 3 of the License, or
-//# (at your option) any later version.
-//#
-//# The LOFAR software suite is distributed in the hope that it will be useful,
-//# but WITHOUT ANY WARRANTY; without even the implied warranty of
-//# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
-//# GNU General Public License for more details.
-//#
-//# You should have received a copy of the GNU General Public License along
-//# with the LOFAR software suite. If not, see <http://www.gnu.org/licenses/>.
-//#
-//# $Id$
-
-#ifndef LOFAR_GPUPROC_CUDA_GPU_INCL_H
-#define LOFAR_GPUPROC_CUDA_GPU_INCL_H
-
-// Pointless in itself; to mirror the OpenCL sources
-// Note: nvcc automatically includes cuda.h, but for most code we don't need it.
-
-// CUDA include option(s)
-//<none>
-
-#include <cuda.h>
-
-#endif
-
diff --git a/RTCP/Cobalt/GPUProc/src/cuda/gpu_wrapper.h b/RTCP/Cobalt/GPUProc/src/cuda/gpu_wrapper.h
deleted file mode 100644
index 831a5e05993..00000000000
--- a/RTCP/Cobalt/GPUProc/src/cuda/gpu_wrapper.h
+++ /dev/null
@@ -1,627 +0,0 @@
-//# gpu_wrapper.h: CUDA-specific wrapper classes for GPU types.
-//#
-//# Copyright (C) 2013 ASTRON (Netherlands Institute for Radio Astronomy)
-//# P.O. Box 2, 7990 AA Dwingeloo, The Netherlands
-//#
-//# This file is part of the LOFAR software suite.
-//# The LOFAR software suite is free software: you can redistribute it and/or
-//# modify it under the terms of the GNU General Public License as published
-//# by the Free Software Foundation, either version 3 of the License, or
-//# (at your option) any later version.
-//#
-//# The LOFAR software suite is distributed in the hope that it will be useful,
-//# but WITHOUT ANY WARRANTY; without even the implied warranty of
-//# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
-//# GNU General Public License for more details.
-//#
-//# You should have received a copy of the GNU General Public License along
-//# with the LOFAR software suite. If not, see <http://www.gnu.org/licenses/>.
-//#
-//# $Id$
-
-#ifndef LOFAR_GPUPROC_CUDA_GPU_WRAPPER_H
-#define LOFAR_GPUPROC_CUDA_GPU_WRAPPER_H
-
-// \file cuda/gpu_wrapper.h
-// C++ wrappers for CUDA akin the OpenCL C++ wrappers.
-// Uses the "Pimpl" idiom for resource managing classes (i.e. that need to
-// control copying having a non-trivial destructor. For more info on Pimpl, see
-// http://www.boost.org/doc/libs/release/libs/smart_ptr/sp_techniques.html#pimpl
-// Not Pimpl-ed are class Platform, Device, and Function.
-// These are also passed by value.
-
-#include <cstddef>
-#include <string>
-#include <vector>
-#include <map>
-#include <iosfwd>
-
-#include <boost/shared_ptr.hpp>
-#include "gpu_incl.h" // ideally, this goes into the .cc, but too much leakage
-#include <cufft.h>
-
-#include <GPUProc/gpu_wrapper.h> // GPUException
-
-#if CUDA_VERSION < 4020
-typedef int CUsharedconfig;
-#endif
-
-namespace LOFAR
-{
- namespace Cobalt
- {
- class PerformanceCounter;
- namespace gpu
- {
-
- // Exception class for CUDA errors.
- EXCEPTION_CLASS(CUDAException, GPUException);
-
- // Return the cuFFT error string associated with \a errcode.
- std::string cufftErrorMessage(cufftResult errcode);
-
- // Return the CUDA error string associated with \a errcode.
- std::string errorMessage(CUresult errcode);
-
-
- // Struct representing a CUDA Grid, which is similar to the @c dim3 type
- // in the CUDA Runtime API.
- struct Grid
- {
- Grid(unsigned int x_ = 1, unsigned int y_ = 1, unsigned int z_ = 1);
- unsigned int x;
- unsigned int y;
- unsigned int z;
- friend std::ostream& operator<<(std::ostream& os, const Grid& grid);
- };
-
- // Struct representing a CUDA Block, which is similar to the @c dim3 type
- // in the CUDA Runtime API.
- //
- // @invariant x > 0, y > 0, z > 0
- struct Block
- {
- Block(unsigned int x_ = 1, unsigned int y_ = 1, unsigned int z_ = 1);
- unsigned int x;
- unsigned int y;
- unsigned int z;
- friend std::ostream& operator<<(std::ostream& os, const Block& block);
- };
-
- // Struct containing kernel launch configuration.
- struct ExecConfig
- {
- ExecConfig(Grid gr = Grid(), Block bl = Block(), size_t dynShMem = 0);
- Grid grid;
- Block block;
- size_t dynSharedMemSize;
- friend std::ostream& operator<<(std::ostream& os,
- const ExecConfig& execConfig);
- };
-
-
- // Forward declaration needed by Platform::devices.
- class Device;
-
- // This class is not strictly needed, because in CUDA there's only one
- // platform, but it hides the CUDA calls and makes it similar to OpenCL.
- class Platform
- {
- public:
- // Initialize the CUDA platform.
- // \param flags must be 0 (at least up till CUDA 5.0).
- Platform(unsigned int flags = 0);
-
- // The CUDA version (e.g. 5.0 -> 5000).
- int version() const;
-
- // Returns the number of devices in the CUDA platform.
- size_t size() const;
-
- // Returns a vector of all devices in the CUDA platform.
- std::vector<Device> devices() const;
-
- // Returns the name of the CUDA platform. (currently, "NVIDIA CUDA")
- std::string getName() const;
-
- // Return the maximum number of threads per block, that
- // is supported by all devices on the platform.
- //
- // Hardware dependent.
- // - Returns at least 512 (except for ancient hardware)
- // - Returns 1024 for K10 (= Cobalt hardware)
- unsigned getMaxThreadsPerBlock() const;
- };
-
- // Wrap a CUDA Device.
- class Device
- {
- public:
- // Create a device.
- // \param ordinal is the device number;
- // valid range: [0, Platform.size()-1]
- Device(int ordinal = 0);
-
- // Order Devices by PCI ID (used in std::sort)
- bool operator<(const Device &other) const;
-
- // Return the name of the device in human readable form.
- std::string getName() const;
-
- // Return the compute capability (major)
- unsigned getComputeCapabilityMajor() const;
-
- // Return the compute capability (minor)
- unsigned getComputeCapabilityMinor() const;
-
- // Return the total amount of global memory, in bytes
- size_t getTotalGlobalMem() const;
-
- // Return the maximum amount of shared memory per block
- size_t getBlockSharedMem() const;
-
- // Return the total amount of constant memory
- size_t getTotalConstMem() const;
-
- // Return the PCI ID (bus:device) of this GPU
- std::string pciId() const;
-
- // Return the maximum number of threads per block
- //
- // Hardware dependent.
- // - Returns at least 512 (except for ancient hardware)
- // - Returns 1024 for K10 (= Cobalt hardware)
- unsigned getMaxThreadsPerBlock() const;
-
- // Return the maximum dimensions of a block of threads.
- struct Block getMaxBlockDims() const;
-
- // Return the maximum dimensions of a grid of blocks.
- struct Grid getMaxGridDims() const;
-
- // Return the number of multi-processors.
- unsigned getMultiProcessorCount() const;
-
- // Return the maximum number of threads that can be
- // resident on a multi-processor.
- unsigned getMaxThreadsPerMultiProcessor() const;
-
- // Return information on a specific \a attribute.
- // \param attribute CUDA device attribute
- int getAttribute(CUdevice_attribute attribute) const;
-
- private:
- // Context needs access to our \c _device to create a context.
- friend class Context;
-
- // The CUDA device.
- CUdevice _device;
- };
-
-
- // Wrap a CUDA Context. Since this class manages a resource (a CUDA
- // context), it uses the pimpl idiom in combination with a reference
- // counted pointer to make it copyable.
- //
- // We do not tie any context to any thread by default -- all contexts
- // are `floating', and are to be tied to a thread only by pushing them
- // as the current context, performing operation(s), and popping them
- // from the current context stack. The pushing and popping is automated
- // in the ScopedCurrentContext class.
- class Context
- {
- public:
- // Create a new CUDA context and associate it with the calling thread.
- // In other words, \c setCurrent() is implied.
- //
- // Flags:
-
- // CU_CTX_SCHED_AUTO:
- // The default value if the flags parameter is zero, uses a
- // heuristic based on the number of active CUDA contexts in the
- // process C and the number of logical processors in the system P.
- // If C > P, then CUDA will yield to other OS threads when waiting
- // for the GPU, otherwise CUDA will not yield while waiting for
- // results and actively spin on the processor.
- // CU_CTX_SCHED_SPIN:
- // Instruct CUDA to actively spin when waiting for results from the GPU.
- // This can decrease latency when waiting for the GPU, but may lower
- // the performance of CPU threads if they are performing work in parallel
- // with the CUDA thread.
- // CU_CTX_SCHED_YIELD:
- // Instruct CUDA to yield its thread when waiting for results from the GPU.
- // This can increase latency when waiting for the GPU, but can increase
- // the performance of CPU threads performing work in parallel with the GPU.
- // CU_CTX_SCHED_BLOCKING_SYNC:
- // Instruct CUDA to block the CPU thread on a synchronization primitive
- // when waiting for the GPU to finish work.
- Context(const Device &device, unsigned int flags = CU_CTX_SCHED_YIELD);
-
- // Returns the device associated to this context.
- Device getDevice() const;
-
- // Set the cache configuration for kernel launches in this context.
- void setCacheConfig(CUfunc_cache config) const;
-
- // Set the shared memory configuration for kernel launches in this context.
- void setSharedMemConfig(CUsharedconfig config) const;
-
- private:
- // Non-copyable implementation class.
- class Impl;
-
- // Reference counted pointer to the implementation class.
- boost::shared_ptr<Impl> _impl;
-
- friend class ScopedCurrentContext;
- };
-
-
- // Make a certain context the current one for a certain scope.
- class ScopedCurrentContext
- {
- public:
- ScopedCurrentContext( const Context &context );
- ~ScopedCurrentContext();
-
- private:
- const Context &_context;
- };
-
-
- // Wrap CUDA Host Memory. This is the equivalent of a OpenCL Buffer. CUDA
- // distinguishes between between host- and device memory, OpenCL does not.
- class HostMemory
- {
- public:
- // Allocate \a size bytes of host memory.
- // \param context CUDA context associated with this HostMemory object.
- // \param size number of bytes to allocate
- // \param flags affect allocation
- // \note To create pinned memory, we need to set
- // \code
- // flags = CU_MEMHOSTALLOC_PORTABLE
- // \endcode
- // \note For input buffers we may consider setting
- // \code
- // flags = CU_MEMHOSTALLOC_PORTABLE | CU_MEMHOSTALLOC_WRITECOMBINED
- // \endcode
- // Please refer to the documentation of the function \c cuMemHostAlloc()
- // in the CUDA Driver API for details.
- HostMemory(const Context &context, size_t size, unsigned int flags = 0);
-
- // Return a pointer to the actual memory.
- // \warning The returned pointer shall not have a lifetime beyond the
- // lifetime of this object (actually the last copy).
- template <typename T>
- T *get() const;
-
- // Return the size of this memory block.
- size_t size() const;
-
- private:
- // Get a void pointer to the actual memory from our Impl class. This
- // method is only used by our templated get() method.
- void* getPtr() const;
-
- // Non-copyable implementation class.
- class Impl;
-
- // Reference counted pointer to the implementation class.
- boost::shared_ptr<Impl> _impl;
- };
-
-
- // Wrap CUDA Device Memory. This is the equivalent of an OpenCL
- // Buffer. CUDA distinguishes between between host- and device memory,
- // OpenCL does not.
- class DeviceMemory
- {
- public:
- // Allocate \a size bytes of device memory.
- DeviceMemory(const Context &context, size_t size);
-
- // Return a device pointer as a handle to the memory.
- void *get() const;
-
- // Fill the first \a n bytes of memory with the constant byte \a uc.
- // \param uc Constant byte value to put into memory
- // \param n Number of bytes to set. Defaults to the complete block.
- // If \a n is larger than the current memory block size, then
- // the complete block will be set to \a uc.
- void set(unsigned char uc, size_t n = (size_t)-1) const;
-
- // Return the size of this memory block.
- size_t size() const;
-
- // Fetch the contents of this buffer in a new HostMemory buffer.
- HostMemory fetch() const;
-
- private:
- // Function needs access to our device ptr location to set this as a kernel arg.
- friend class Function;
-
- // Non-copyable implementation class.
- class Impl;
-
- // Reference counted pointer to the implementation class.
- boost::shared_ptr<Impl> _impl;
- };
-
-
- // Wrap a CUDA Module. This is the equivalent of a OpenCL Program.
- class Module
- {
- public:
- typedef std::map<CUjit_option, void*> optionmap_t;
-
- Module(); // TODO: tmp, as long as CorrelatorPipelinePrograms needs a default init
-
- // Load the module in the file \a fname into the given \a context. The
- // file should be a \e cubin file or a \e ptx file as output by \c nvcc.
- // \param context CUDA context associated with this Module object.
- // \param fname name of a module file
- // \note For details, please refer to the documentation of \c
- // cuModuleLoad in the CUDA Driver API.
- Module(const Context &context, const std::string &fname);
-
- // Load the module pointed to by \a image into the given \a context. The
- // pointer may point to a null-terminated string containing \e cubin or
- // \e ptx code.
- // \param context CUDA context associated with this Module object.
- // \param image pointer to a module image in memory
- // \note For details, please refer to the documentation of \c
- // cuModuleLoadData in the CUDA Driver API.
- Module(const Context &context, const void *image);
-
- // Load the module pointed to by \a image into the given \a context. The
- // pointer may point to a null-terminated string containing \e cubin or
- // \e ptx code.
- // \param context CUDA context associated with this Module object.
- // \param image pointer to a module image in memory
- // \param options map of \c CUjit_option items, with their associated
- // values.
- // \note All values are cast to void*, so if an option requires
- // an unsigned int as value, the unsigned int's value itself is cast to void*!
- // \note For details, please refer to the documentation of \c
- // cuModuleLoadDataEx in the CUDA Driver API.
- Module(const Context &context, const void *image, optionmap_t &options);
-
- // Return the Context in which this Module was created.
- Context getContext() const;
-
- private:
- // Function needs access to our module to create a function.
- friend class Function;
-
- // Non-copyable implementation class.
- class Impl;
-
- // Reference counted pointer to the implementation class.
- boost::shared_ptr<Impl> _impl;
- };
-
- // Wrap a CUDA Device Function. This is the equivalent of an OpenCL
- // Program.
- class Function
- {
- public:
- // Construct a function object by looking up the function \a name in the
- // module \a module.
- Function(const Module &module, const std::string &name);
-
- // Return the name of the function.
- std::string name() const;
-
- // Set kernel immediate argument number \a index to \a val.
- // \a val must outlive kernel execution.
- // Not for device memory objects (be it as DeviceMemory or as void *).
- template <typename T>
- void setArg(size_t index, const T &val);
-
- // Set kernel DeviceMemory object argument number \a index to \a mem.
- // \a mem must outlive kernel execution.
- void setArg(size_t index, const DeviceMemory &mem);
-
- // Set pointer to kernel device memory object (as void *) number \a index
- // to \a val. \a *val must outlive kernel execution.
- // Note: Prefer to use setArg() passing a DeviceMemory ref over this overload.
- void setArg(size_t index, const void **val);
-
- // Return information about a function.
- // \note For details on valid values for \a attribute, please refer to
- // the documentation of cuFuncGetAttribute in the CUDA Driver API.
- int getAttribute(CUfunction_attribute attribute) const;
-
- // Set the shared memory configuration for a device function.
- // \note For details on valid values for \a config, please refer to the
- // documentation of cuFuncSetSharedMemConfig in the CUDA Driver API.
- void setSharedMemConfig(CUsharedconfig config) const;
-
- protected:
- const Context _context;
-
- private:
- // Keep the Module alive, because Function actually wraps a pointer
- // to a function within the Module.
- const Module _module;
-
- // The name of the function, for error reporting purposes
- const std::string _name;
-
- // Stream needs access to our CUDA function to launch a kernel.
- friend class Stream;
-
- // CUDA function.
- CUfunction _function;
-
- // Function arguments as set.
- std::vector<const void *> _kernelArgs;
-
- // Helper function to modify _kernelArgs.
- void doSetArg(size_t index, const void *argp);
-
- // Do not use. To guard against passing pointers.
- // Note that even device void * cannot be passed, because we need its
- // address with a life time longer than this formal parameter.
- template<typename T>
- void setArg(size_t index, const T *&); // intentionally not impl.
-
- // Do not use. To guard against passing HostMemory references to kernels.
- void setArg(size_t index, const HostMemory &); // intentionally not impl.
-
- // Do not use. To guard against passing HostMemory pointers to kernels.
- void setArg(size_t index, const HostMemory *); // intentionally not impl.
- };
-
- // Wrap a CUDA Event. This is the equivalent of an OpenCL Event.
- class Event
- {
- public:
- // Construct a CUDA event. This class manages a resource (a CUDA event)
- // and is therefore implemented using the pimpl idiom, using a reference
- // counted pointer to a non-copyable implementation class.
- // \note For details on valid values for \a flags, please refer to the
- // documentation of cuEventCreate in the CUDA Driver API.
- Event(const Context &context, unsigned int flags = CU_EVENT_DEFAULT);
-
- // Return the elapsed time in milliseconds between this event and the \a
- // second event.
- float elapsedTime(Event &second) const;
-
- // Wait until all work preceding this event in the same stream has
- // completed.
- void wait();
-
- private:
- // Stream needs access to our CUDA event to wait for and record events.
- friend class Stream;
-
- // Non-copyable implementation class.
- class Impl;
-
- // Reference counted pointer to the implementation class.
- boost::shared_ptr<Impl> _impl;
- };
-
-
- // Wrap a CUDA Stream. This is the equivalent of an OpenCL
- // CommandQueue. This class manages a resource (a CUDA stream) and is
- // therefore implemented using the pimpl idiom, using a reference counted
- // pointer to a non-copyable implementation class.
- class Stream
- {
- public:
- // Create a stream.
- // \param flags must be 0 for CUDA < 5.0
- // \param context CUDA context associated with this Stream object.
- // \note For details on valid values for \a flags, please refer to the
- // documentation of \c cuStreamCreate in the CUDA Driver API.
- explicit Stream(const Context &context, unsigned int flags = 0); // named CU_STREAM_DEFAULT (0) since CUDA 5.0
-
- // Transfer data from host memory \a hostMem to device memory \a devMem.
- // \param devMem Device memory that will be copied to.
- // \param hostMem Host memory that will be copied from.
- // \param synchronous Indicates whether the transfer must be done
- // synchronously or asynchronously.
- void writeBuffer(const DeviceMemory &devMem, const HostMemory &hostMem,
- bool synchronous = false) const;
-
- // Transfer data from host memory \a hostMem to device memory \a devMem.
- // When gpuProfiling is enabled this transfer is synchronous
- // \param devMem Device memory that will be copied to.
- // \param hostMem Host memory that will be copied from.
- // \param counter PerformanceCounter that will receive transfer duration
- // if gpuProfiling is enabled
- // \param synchronous Indicates whether the transfer must be done
- // synchronously or asynchronously. Default == false
- void writeBuffer(const DeviceMemory &devMem, const HostMemory &hostMem,
- PerformanceCounter &counter, bool synchronous = false) const;
-
- // Transfer data from device memory \a devMem to host memory \a hostMem.
- // \param hostMem Host memory that will be copied to.
- // \param devMem Device memory that will be copied from.
- // \param synchronous Indicates whether the transfer must be done
- // synchronously or asynchronously.
- void readBuffer(const HostMemory &hostMem, const DeviceMemory &devMem,
- bool synchronous = false) const;
-
- // Transfer data from device memory \a devMem to host memory \a hostMem.
- // When gpuProfiling is enabled this transfer is synchronous
- // \param hostMem Host memory that will be copied to.
- // \param devMem Device memory that will be copied from.
- // \param counter PerformanceCounter that will receive transfer duration
- // if gpuProfiling is enabled
- // \param synchronous Indicates whether the transfer must be done
- // synchronously or asynchronously. Default == false
- void readBuffer(const HostMemory &hostMem, const DeviceMemory &devMem,
- PerformanceCounter &counter, bool synchronous = false) const;
-
- // Transfer data from device memory \a devSource to device memory \a devTarget.
- // \param devTarget Device memory that will be copied to.
- // \param devSource Device memory that will be copied from.
- // \param synchronous Indicates whether the transfer must be done
- // synchronously or asynchronously.
- void copyBuffer(const DeviceMemory &devTarget, const DeviceMemory &devSource,
- bool synchronous = false) const;
-
- // Transfer data from device memory \a devSource to device memory \a devTarget.
- // When gpuProfiling is enabled this transfer is synchronous
- // \param devTarget Device memory that will be copied to.
- // \param devSource Device memory that will be copied from.
- // \param counter PerformanceCounter that will receive transfer duration
- // if gpuProfiling is enabled
- // \param synchronous Indicates whether the transfer must be done
- // synchronously or asynchronously. Defaults to \c false
- // (asynchronously).
- void copyBuffer(const DeviceMemory &devTarget, const DeviceMemory &devSource,
- PerformanceCounter &counter, bool synchronous = false) const;
-
- // Launch a CUDA function.
- // \param function object containing the function to launch
- // \param grid Grid size (in terms of blocks (not threads (OpenCL)))
- // \param block Block (thread group) size
- void launchKernel(const Function &function,
- const Grid &grid, const Block &block) const;
-
- // Check if all operations on this stream have completed.
- // \return true if all completed, or false otherwise.
- bool query() const;
-
- // Wait until a this stream's tasks are completed.
- void synchronize() const;
-
- // Let this stream wait on the event \a event.
- void waitEvent(const Event &event) const;
-
- // Record the event \a event for this stream.
- void recordEvent(const Event &event) const;
-
- // Return the underlying CUDA stream. TODO: try to get rid of CUstream here: FFT thing to here or make it friend
- CUstream get() const;
-
- // Returns the context associated with the underlying CUDA stream.
- Context getContext() const; // TODO: consider using this in the SubbandProcs (now has Stream and Context stored)
-
- // Return whether this stream mandates synchronous behaviour
- bool isSynchronous() const;
-
- private:
- // Non-copyable implementation class.
- class Impl;
-
- // Reference counted pointer to the implementation class.
- boost::shared_ptr<Impl> _impl;
-
- // Force synchronous transfers and kernel launches
- bool force_synchronous;
- };
-
- } // namespace gpu
- } // namespace Cobalt
-} // namespace LOFAR
-
-#include "gpu_wrapper.tcc"
-
-#endif
-
diff --git a/RTCP/Cobalt/GPUProc/src/cuda/cuda_config.h.in b/RTCP/Cobalt/GPUProc/src/cuda_config.h.in
similarity index 100%
rename from RTCP/Cobalt/GPUProc/src/cuda/cuda_config.h.in
rename to RTCP/Cobalt/GPUProc/src/cuda_config.h.in
diff --git a/RTCP/Cobalt/GPUProc/src/gpu_incl.h b/RTCP/Cobalt/GPUProc/src/gpu_incl.h
index 4884cbd3c55..efb6c8185bb 100644
--- a/RTCP/Cobalt/GPUProc/src/gpu_incl.h
+++ b/RTCP/Cobalt/GPUProc/src/gpu_incl.h
@@ -25,17 +25,7 @@
#ifndef LOFAR_GPUPROC_GPU_INCL_H
#define LOFAR_GPUPROC_GPU_INCL_H
-#if defined (USE_CUDA) && defined (USE_OPENCL)
-# error "Either CUDA or OpenCL must be enabled, not both"
-#endif
-
-#if defined (USE_CUDA)
-# include "cuda/gpu_incl.h"
-#elif defined (USE_OPENCL)
-# include "opencl/gpu_incl.h"
-#else
-# error "Either CUDA or OpenCL must be enabled, not neither"
-#endif
+#include <cuda.h>
#endif
diff --git a/RTCP/Cobalt/GPUProc/src/cuda/gpu_utils.cc b/RTCP/Cobalt/GPUProc/src/gpu_utils.cc
similarity index 100%
rename from RTCP/Cobalt/GPUProc/src/cuda/gpu_utils.cc
rename to RTCP/Cobalt/GPUProc/src/gpu_utils.cc
diff --git a/RTCP/Cobalt/GPUProc/src/cuda/gpu_wrapper.cc b/RTCP/Cobalt/GPUProc/src/gpu_wrapper.cc
similarity index 100%
rename from RTCP/Cobalt/GPUProc/src/cuda/gpu_wrapper.cc
rename to RTCP/Cobalt/GPUProc/src/gpu_wrapper.cc
diff --git a/RTCP/Cobalt/GPUProc/src/gpu_wrapper.h b/RTCP/Cobalt/GPUProc/src/gpu_wrapper.h
index 640e89ccbc5..6f19a4f10ee 100644
--- a/RTCP/Cobalt/GPUProc/src/gpu_wrapper.h
+++ b/RTCP/Cobalt/GPUProc/src/gpu_wrapper.h
@@ -19,41 +19,612 @@
//#
//# $Id$
-// \file
-// Wrapper classes for GPU types.
+// \file cuda/gpu_wrapper.h
+// C++ wrappers for CUDA akin the OpenCL C++ wrappers.
+// Uses the "Pimpl" idiom for resource managing classes (i.e. that need to
+// control copying having a non-trivial destructor. For more info on Pimpl, see
+// http://www.boost.org/doc/libs/release/libs/smart_ptr/sp_techniques.html#pimpl
+// Not Pimpl-ed are class Platform, Device, and Function.
+// These are also passed by value.
#ifndef LOFAR_GPUPROC_GPU_WRAPPER_H
#define LOFAR_GPUPROC_GPU_WRAPPER_H
-#if defined (USE_CUDA) && defined (USE_OPENCL)
-# error "Either CUDA or OpenCL must be enabled, not both"
-#endif
+#include <cstddef>
+#include <string>
+#include <vector>
+#include <map>
+#include <iosfwd>
+
+#include <boost/shared_ptr.hpp>
+#include "gpu_incl.h" // ideally, this goes into the .cc, but too much leakage
+#include <cufft.h>
#include <Common/Exception.h>
+#include <GPUProc/gpu_wrapper.h> // GPUException
+
+#if CUDA_VERSION < 4020
+typedef int CUsharedconfig;
+#endif
namespace LOFAR
{
namespace Cobalt
{
+ class PerformanceCounter;
+
namespace gpu
{
// Exception class for GPU errors.
EXCEPTION_CLASS(GPUException, LOFAR::Exception);
- } // namespace gpu
+ // Exception class for CUDA errors.
+ EXCEPTION_CLASS(CUDAException, GPUException);
- } // namespace Cobalt
+ // Return the cuFFT error string associated with \a errcode.
+ std::string cufftErrorMessage(cufftResult errcode);
-} // namespace LOFAR
+ // Return the CUDA error string associated with \a errcode.
+ std::string errorMessage(CUresult errcode);
+ // Struct representing a CUDA Grid, which is similar to the @c dim3 type
+ // in the CUDA Runtime API.
+ struct Grid
+ {
+ Grid(unsigned int x_ = 1, unsigned int y_ = 1, unsigned int z_ = 1);
+ unsigned int x;
+ unsigned int y;
+ unsigned int z;
+ friend std::ostream& operator<<(std::ostream& os, const Grid& grid);
+ };
-#if defined (USE_CUDA)
-# include "cuda/gpu_wrapper.h"
-#elif defined (USE_OPENCL)
-# include "opencl/gpu_wrapper.h"
-#else
-# error "Either CUDA or OpenCL must be enabled, not neither"
-#endif
+ // Struct representing a CUDA Block, which is similar to the @c dim3 type
+ // in the CUDA Runtime API.
+ //
+ // @invariant x > 0, y > 0, z > 0
+ struct Block
+ {
+ Block(unsigned int x_ = 1, unsigned int y_ = 1, unsigned int z_ = 1);
+ unsigned int x;
+ unsigned int y;
+ unsigned int z;
+ friend std::ostream& operator<<(std::ostream& os, const Block& block);
+ };
+
+ // Struct containing kernel launch configuration.
+ struct ExecConfig
+ {
+ ExecConfig(Grid gr = Grid(), Block bl = Block(), size_t dynShMem = 0);
+ Grid grid;
+ Block block;
+ size_t dynSharedMemSize;
+ friend std::ostream& operator<<(std::ostream& os,
+ const ExecConfig& execConfig);
+ };
+
+
+ // Forward declaration needed by Platform::devices.
+ class Device;
+
+ // This class is not strictly needed, because in CUDA there's only one
+ // platform, but it hides the CUDA calls and makes it similar to OpenCL.
+ class Platform
+ {
+ public:
+ // Initialize the CUDA platform.
+ // \param flags must be 0 (at least up till CUDA 5.0).
+ Platform(unsigned int flags = 0);
+
+ // The CUDA version (e.g. 5.0 -> 5000).
+ int version() const;
+
+ // Returns the number of devices in the CUDA platform.
+ size_t size() const;
+
+ // Returns a vector of all devices in the CUDA platform.
+ std::vector<Device> devices() const;
+
+ // Returns the name of the CUDA platform. (currently, "NVIDIA CUDA")
+ std::string getName() const;
+
+ // Return the maximum number of threads per block, that
+ // is supported by all devices on the platform.
+ //
+ // Hardware dependent.
+ // - Returns at least 512 (except for ancient hardware)
+ // - Returns 1024 for K10 (= Cobalt hardware)
+ unsigned getMaxThreadsPerBlock() const;
+ };
+
+ // Wrap a CUDA Device.
+ class Device
+ {
+ public:
+ // Create a device.
+ // \param ordinal is the device number;
+ // valid range: [0, Platform.size()-1]
+ Device(int ordinal = 0);
+
+ // Order Devices by PCI ID (used in std::sort)
+ bool operator<(const Device &other) const;
+
+ // Return the name of the device in human readable form.
+ std::string getName() const;
+
+ // Return the compute capability (major)
+ unsigned getComputeCapabilityMajor() const;
+
+ // Return the compute capability (minor)
+ unsigned getComputeCapabilityMinor() const;
+
+ // Return the total amount of global memory, in bytes
+ size_t getTotalGlobalMem() const;
+
+ // Return the maximum amount of shared memory per block
+ size_t getBlockSharedMem() const;
+
+ // Return the total amount of constant memory
+ size_t getTotalConstMem() const;
+
+ // Return the PCI ID (bus:device) of this GPU
+ std::string pciId() const;
+
+ // Return the maximum number of threads per block
+ //
+ // Hardware dependent.
+ // - Returns at least 512 (except for ancient hardware)
+ // - Returns 1024 for K10 (= Cobalt hardware)
+ unsigned getMaxThreadsPerBlock() const;
+
+ // Return the maximum dimensions of a block of threads.
+ struct Block getMaxBlockDims() const;
+
+ // Return the maximum dimensions of a grid of blocks.
+ struct Grid getMaxGridDims() const;
+
+ // Return the number of multi-processors.
+ unsigned getMultiProcessorCount() const;
+
+ // Return the maximum number of threads that can be
+ // resident on a multi-processor.
+ unsigned getMaxThreadsPerMultiProcessor() const;
+
+ // Return information on a specific \a attribute.
+ // \param attribute CUDA device attribute
+ int getAttribute(CUdevice_attribute attribute) const;
+
+ private:
+ // Context needs access to our \c _device to create a context.
+ friend class Context;
+
+ // The CUDA device.
+ CUdevice _device;
+ };
+
+
+ // Wrap a CUDA Context. Since this class manages a resource (a CUDA
+ // context), it uses the pimpl idiom in combination with a reference
+ // counted pointer to make it copyable.
+ //
+ // We do not tie any context to any thread by default -- all contexts
+ // are `floating', and are to be tied to a thread only by pushing them
+ // as the current context, performing operation(s), and popping them
+ // from the current context stack. The pushing and popping is automated
+ // in the ScopedCurrentContext class.
+ class Context
+ {
+ public:
+ // Create a new CUDA context and associate it with the calling thread.
+ // In other words, \c setCurrent() is implied.
+ //
+ // Flags:
+
+ // CU_CTX_SCHED_AUTO:
+ // The default value if the flags parameter is zero, uses a
+ // heuristic based on the number of active CUDA contexts in the
+ // process C and the number of logical processors in the system P.
+ // If C > P, then CUDA will yield to other OS threads when waiting
+ // for the GPU, otherwise CUDA will not yield while waiting for
+ // results and actively spin on the processor.
+ // CU_CTX_SCHED_SPIN:
+ // Instruct CUDA to actively spin when waiting for results from the GPU.
+ // This can decrease latency when waiting for the GPU, but may lower
+ // the performance of CPU threads if they are performing work in parallel
+ // with the CUDA thread.
+ // CU_CTX_SCHED_YIELD:
+ // Instruct CUDA to yield its thread when waiting for results from the GPU.
+ // This can increase latency when waiting for the GPU, but can increase
+ // the performance of CPU threads performing work in parallel with the GPU.
+ // CU_CTX_SCHED_BLOCKING_SYNC:
+ // Instruct CUDA to block the CPU thread on a synchronization primitive
+ // when waiting for the GPU to finish work.
+ Context(const Device &device, unsigned int flags = CU_CTX_SCHED_YIELD);
+
+ // Returns the device associated to this context.
+ Device getDevice() const;
+
+ // Set the cache configuration for kernel launches in this context.
+ void setCacheConfig(CUfunc_cache config) const;
+
+ // Set the shared memory configuration for kernel launches in this context.
+ void setSharedMemConfig(CUsharedconfig config) const;
+
+ private:
+ // Non-copyable implementation class.
+ class Impl;
+
+ // Reference counted pointer to the implementation class.
+ boost::shared_ptr<Impl> _impl;
+
+ friend class ScopedCurrentContext;
+ };
+
+
+ // Make a certain context the current one for a certain scope.
+ class ScopedCurrentContext
+ {
+ public:
+ ScopedCurrentContext( const Context &context );
+ ~ScopedCurrentContext();
+
+ private:
+ const Context &_context;
+ };
+
+
+ // Wrap CUDA Host Memory. This is the equivalent of a OpenCL Buffer. CUDA
+ // distinguishes between between host- and device memory, OpenCL does not.
+ class HostMemory
+ {
+ public:
+ // Allocate \a size bytes of host memory.
+ // \param context CUDA context associated with this HostMemory object.
+ // \param size number of bytes to allocate
+ // \param flags affect allocation
+ // \note To create pinned memory, we need to set
+ // \code
+ // flags = CU_MEMHOSTALLOC_PORTABLE
+ // \endcode
+ // \note For input buffers we may consider setting
+ // \code
+ // flags = CU_MEMHOSTALLOC_PORTABLE | CU_MEMHOSTALLOC_WRITECOMBINED
+ // \endcode
+ // Please refer to the documentation of the function \c cuMemHostAlloc()
+ // in the CUDA Driver API for details.
+ HostMemory(const Context &context, size_t size, unsigned int flags = 0);
+
+ // Return a pointer to the actual memory.
+ // \warning The returned pointer shall not have a lifetime beyond the
+ // lifetime of this object (actually the last copy).
+ template <typename T>
+ T *get() const;
+
+ // Return the size of this memory block.
+ size_t size() const;
+
+ private:
+ // Get a void pointer to the actual memory from our Impl class. This
+ // method is only used by our templated get() method.
+ void* getPtr() const;
+
+ // Non-copyable implementation class.
+ class Impl;
+
+ // Reference counted pointer to the implementation class.
+ boost::shared_ptr<Impl> _impl;
+ };
+
+
+ // Wrap CUDA Device Memory. This is the equivalent of an OpenCL
+ // Buffer. CUDA distinguishes between between host- and device memory,
+ // OpenCL does not.
+ class DeviceMemory
+ {
+ public:
+ // Allocate \a size bytes of device memory.
+ DeviceMemory(const Context &context, size_t size);
+
+ // Return a device pointer as a handle to the memory.
+ void *get() const;
+
+ // Fill the first \a n bytes of memory with the constant byte \a uc.
+ // \param uc Constant byte value to put into memory
+ // \param n Number of bytes to set. Defaults to the complete block.
+ // If \a n is larger than the current memory block size, then
+ // the complete block will be set to \a uc.
+ void set(unsigned char uc, size_t n = (size_t)-1) const;
+
+ // Return the size of this memory block.
+ size_t size() const;
+
+ // Fetch the contents of this buffer in a new HostMemory buffer.
+ HostMemory fetch() const;
+
+ private:
+ // Function needs access to our device ptr location to set this as a kernel arg.
+ friend class Function;
+
+ // Non-copyable implementation class.
+ class Impl;
+
+ // Reference counted pointer to the implementation class.
+ boost::shared_ptr<Impl> _impl;
+ };
+
+
+ // Wrap a CUDA Module. This is the equivalent of a OpenCL Program.
+ class Module
+ {
+ public:
+ typedef std::map<CUjit_option, void*> optionmap_t;
+
+ Module(); // TODO: tmp, as long as CorrelatorPipelinePrograms needs a default init
+
+ // Load the module in the file \a fname into the given \a context. The
+ // file should be a \e cubin file or a \e ptx file as output by \c nvcc.
+ // \param context CUDA context associated with this Module object.
+ // \param fname name of a module file
+ // \note For details, please refer to the documentation of \c
+ // cuModuleLoad in the CUDA Driver API.
+ Module(const Context &context, const std::string &fname);
+
+ // Load the module pointed to by \a image into the given \a context. The
+ // pointer may point to a null-terminated string containing \e cubin or
+ // \e ptx code.
+ // \param context CUDA context associated with this Module object.
+ // \param image pointer to a module image in memory
+ // \note For details, please refer to the documentation of \c
+ // cuModuleLoadData in the CUDA Driver API.
+ Module(const Context &context, const void *image);
+
+ // Load the module pointed to by \a image into the given \a context. The
+ // pointer may point to a null-terminated string containing \e cubin or
+ // \e ptx code.
+ // \param context CUDA context associated with this Module object.
+ // \param image pointer to a module image in memory
+ // \param options map of \c CUjit_option items, with their associated
+ // values.
+ // \note All values are cast to void*, so if an option requires
+ // an unsigned int as value, the unsigned int's value itself is cast to void*!
+ // \note For details, please refer to the documentation of \c
+ // cuModuleLoadDataEx in the CUDA Driver API.
+ Module(const Context &context, const void *image, optionmap_t &options);
+
+ // Return the Context in which this Module was created.
+ Context getContext() const;
+
+ private:
+ // Function needs access to our module to create a function.
+ friend class Function;
+
+ // Non-copyable implementation class.
+ class Impl;
+
+ // Reference counted pointer to the implementation class.
+ boost::shared_ptr<Impl> _impl;
+ };
+
+ // Wrap a CUDA Device Function. This is the equivalent of an OpenCL
+ // Program.
+ class Function
+ {
+ public:
+ // Construct a function object by looking up the function \a name in the
+ // module \a module.
+ Function(const Module &module, const std::string &name);
+
+ // Return the name of the function.
+ std::string name() const;
+
+ // Set kernel immediate argument number \a index to \a val.
+ // \a val must outlive kernel execution.
+ // Not for device memory objects (be it as DeviceMemory or as void *).
+ template <typename T>
+ void setArg(size_t index, const T &val);
+
+ // Set kernel DeviceMemory object argument number \a index to \a mem.
+ // \a mem must outlive kernel execution.
+ void setArg(size_t index, const DeviceMemory &mem);
+
+ // Set pointer to kernel device memory object (as void *) number \a index
+ // to \a val. \a *val must outlive kernel execution.
+ // Note: Prefer to use setArg() passing a DeviceMemory ref over this overload.
+ void setArg(size_t index, const void **val);
+
+ // Return information about a function.
+ // \note For details on valid values for \a attribute, please refer to
+ // the documentation of cuFuncGetAttribute in the CUDA Driver API.
+ int getAttribute(CUfunction_attribute attribute) const;
+
+ // Set the shared memory configuration for a device function.
+ // \note For details on valid values for \a config, please refer to the
+ // documentation of cuFuncSetSharedMemConfig in the CUDA Driver API.
+ void setSharedMemConfig(CUsharedconfig config) const;
+
+ protected:
+ const Context _context;
+
+ private:
+ // Keep the Module alive, because Function actually wraps a pointer
+ // to a function within the Module.
+ const Module _module;
+
+ // The name of the function, for error reporting purposes
+ const std::string _name;
+
+ // Stream needs access to our CUDA function to launch a kernel.
+ friend class Stream;
+
+ // CUDA function.
+ CUfunction _function;
+
+ // Function arguments as set.
+ std::vector<const void *> _kernelArgs;
+
+ // Helper function to modify _kernelArgs.
+ void doSetArg(size_t index, const void *argp);
+
+ // Do not use. To guard against passing pointers.
+ // Note that even device void * cannot be passed, because we need its
+ // address with a life time longer than this formal parameter.
+ template<typename T>
+ void setArg(size_t index, const T *&); // intentionally not impl.
+
+ // Do not use. To guard against passing HostMemory references to kernels.
+ void setArg(size_t index, const HostMemory &); // intentionally not impl.
+
+ // Do not use. To guard against passing HostMemory pointers to kernels.
+ void setArg(size_t index, const HostMemory *); // intentionally not impl.
+ };
+
+ // Wrap a CUDA Event. This is the equivalent of an OpenCL Event.
+ class Event
+ {
+ public:
+ // Construct a CUDA event. This class manages a resource (a CUDA event)
+ // and is therefore implemented using the pimpl idiom, using a reference
+ // counted pointer to a non-copyable implementation class.
+ // \note For details on valid values for \a flags, please refer to the
+ // documentation of cuEventCreate in the CUDA Driver API.
+ Event(const Context &context, unsigned int flags = CU_EVENT_DEFAULT);
+
+ // Return the elapsed time in milliseconds between this event and the \a
+ // second event.
+ float elapsedTime(Event &second) const;
+
+ // Wait until all work preceding this event in the same stream has
+ // completed.
+ void wait();
+
+ private:
+ // Stream needs access to our CUDA event to wait for and record events.
+ friend class Stream;
+
+ // Non-copyable implementation class.
+ class Impl;
+
+ // Reference counted pointer to the implementation class.
+ boost::shared_ptr<Impl> _impl;
+ };
+
+
+ // Wrap a CUDA Stream. This is the equivalent of an OpenCL
+ // CommandQueue. This class manages a resource (a CUDA stream) and is
+ // therefore implemented using the pimpl idiom, using a reference counted
+ // pointer to a non-copyable implementation class.
+ class Stream
+ {
+ public:
+ // Create a stream.
+ // \param flags must be 0 for CUDA < 5.0
+ // \param context CUDA context associated with this Stream object.
+ // \note For details on valid values for \a flags, please refer to the
+ // documentation of \c cuStreamCreate in the CUDA Driver API.
+ explicit Stream(const Context &context, unsigned int flags = 0); // named CU_STREAM_DEFAULT (0) since CUDA 5.0
+
+ // Transfer data from host memory \a hostMem to device memory \a devMem.
+ // \param devMem Device memory that will be copied to.
+ // \param hostMem Host memory that will be copied from.
+ // \param synchronous Indicates whether the transfer must be done
+ // synchronously or asynchronously.
+ void writeBuffer(const DeviceMemory &devMem, const HostMemory &hostMem,
+ bool synchronous = false) const;
+
+ // Transfer data from host memory \a hostMem to device memory \a devMem.
+ // When gpuProfiling is enabled this transfer is synchronous
+ // \param devMem Device memory that will be copied to.
+ // \param hostMem Host memory that will be copied from.
+ // \param counter PerformanceCounter that will receive transfer duration
+ // if gpuProfiling is enabled
+ // \param synchronous Indicates whether the transfer must be done
+ // synchronously or asynchronously. Default == false
+ void writeBuffer(const DeviceMemory &devMem, const HostMemory &hostMem,
+ PerformanceCounter &counter, bool synchronous = false) const;
+
+ // Transfer data from device memory \a devMem to host memory \a hostMem.
+ // \param hostMem Host memory that will be copied to.
+ // \param devMem Device memory that will be copied from.
+ // \param synchronous Indicates whether the transfer must be done
+ // synchronously or asynchronously.
+ void readBuffer(const HostMemory &hostMem, const DeviceMemory &devMem,
+ bool synchronous = false) const;
+
+ // Transfer data from device memory \a devMem to host memory \a hostMem.
+ // When gpuProfiling is enabled this transfer is synchronous
+ // \param hostMem Host memory that will be copied to.
+ // \param devMem Device memory that will be copied from.
+ // \param counter PerformanceCounter that will receive transfer duration
+ // if gpuProfiling is enabled
+ // \param synchronous Indicates whether the transfer must be done
+ // synchronously or asynchronously. Default == false
+ void readBuffer(const HostMemory &hostMem, const DeviceMemory &devMem,
+ PerformanceCounter &counter, bool synchronous = false) const;
+
+ // Transfer data from device memory \a devSource to device memory \a devTarget.
+ // \param devTarget Device memory that will be copied to.
+ // \param devSource Device memory that will be copied from.
+ // \param synchronous Indicates whether the transfer must be done
+ // synchronously or asynchronously.
+ void copyBuffer(const DeviceMemory &devTarget, const DeviceMemory &devSource,
+ bool synchronous = false) const;
+
+ // Transfer data from device memory \a devSource to device memory \a devTarget.
+ // When gpuProfiling is enabled this transfer is synchronous
+ // \param devTarget Device memory that will be copied to.
+ // \param devSource Device memory that will be copied from.
+ // \param counter PerformanceCounter that will receive transfer duration
+ // if gpuProfiling is enabled
+ // \param synchronous Indicates whether the transfer must be done
+ // synchronously or asynchronously. Defaults to \c false
+ // (asynchronously).
+ void copyBuffer(const DeviceMemory &devTarget, const DeviceMemory &devSource,
+ PerformanceCounter &counter, bool synchronous = false) const;
+
+ // Launch a CUDA function.
+ // \param function object containing the function to launch
+ // \param grid Grid size (in terms of blocks (not threads (OpenCL)))
+ // \param block Block (thread group) size
+ void launchKernel(const Function &function,
+ const Grid &grid, const Block &block) const;
+
+ // Check if all operations on this stream have completed.
+ // \return true if all completed, or false otherwise.
+ bool query() const;
+
+ // Wait until a this stream's tasks are completed.
+ void synchronize() const;
+
+ // Let this stream wait on the event \a event.
+ void waitEvent(const Event &event) const;
+
+ // Record the event \a event for this stream.
+ void recordEvent(const Event &event) const;
+
+ // Return the underlying CUDA stream. TODO: try to get rid of CUstream here: FFT thing to here or make it friend
+ CUstream get() const;
+
+ // Returns the context associated with the underlying CUDA stream.
+ Context getContext() const; // TODO: consider using this in the SubbandProcs (now has Stream and Context stored)
+
+ // Return whether this stream mandates synchronous behaviour
+ bool isSynchronous() const;
+
+ private:
+ // Non-copyable implementation class.
+ class Impl;
+
+ // Reference counted pointer to the implementation class.
+ boost::shared_ptr<Impl> _impl;
+
+ // Force synchronous transfers and kernel launches
+ bool force_synchronous;
+ };
+
+ } // namespace gpu
+ } // namespace Cobalt
+} // namespace LOFAR
+
+#include "gpu_wrapper.tcc"
#endif
diff --git a/RTCP/Cobalt/GPUProc/src/cuda/gpu_wrapper.tcc b/RTCP/Cobalt/GPUProc/src/gpu_wrapper.tcc
similarity index 100%
rename from RTCP/Cobalt/GPUProc/src/cuda/gpu_wrapper.tcc
rename to RTCP/Cobalt/GPUProc/src/gpu_wrapper.tcc
diff --git a/RTCP/Cobalt/GPUProc/test/SubbandProcs/tCorrelatorStep.cc b/RTCP/Cobalt/GPUProc/test/SubbandProcs/tCorrelatorStep.cc
index 6f78be3eb99..eb1850b2e75 100644
--- a/RTCP/Cobalt/GPUProc/test/SubbandProcs/tCorrelatorStep.cc
+++ b/RTCP/Cobalt/GPUProc/test/SubbandProcs/tCorrelatorStep.cc
@@ -22,7 +22,7 @@
#include <CoInterface/BudgetTimer.h>
#include <GPUProc/SubbandProcs/CorrelatorStep.h>
-#include <GPUProc/cuda/gpu_wrapper.h>
+#include <GPUProc/gpu_wrapper.h>
#include <UnitTest++.h>
#include <iostream>
diff --git a/RTCP/Cobalt/GPUProc/test/cuda/tFFT_leakage.cc b/RTCP/Cobalt/GPUProc/test/cuda/tFFT_leakage.cc
index 9571fe3438d..319ab116ac1 100644
--- a/RTCP/Cobalt/GPUProc/test/cuda/tFFT_leakage.cc
+++ b/RTCP/Cobalt/GPUProc/test/cuda/tFFT_leakage.cc
@@ -37,7 +37,7 @@
#include <CoInterface/Parset.h>
#include <GPUProc/FilterBank.h>
#include <GPUProc/SubbandProcs/CorrelatorSubbandProc.h>
-#include <GPUProc/cuda/Pipelines/Pipeline.h>
+#include <GPUProc/Pipelines/Pipeline.h>
#include <GPUProc/gpu_utils.h>
#include <GPUProc/gpu_wrapper.h>
diff --git a/RTCP/Cobalt/GPUProc/test/cuda/tGPUWrapper.cc b/RTCP/Cobalt/GPUProc/test/cuda/tGPUWrapper.cc
index e0d1b940506..1e1bb0f2ca3 100644
--- a/RTCP/Cobalt/GPUProc/test/cuda/tGPUWrapper.cc
+++ b/RTCP/Cobalt/GPUProc/test/cuda/tGPUWrapper.cc
@@ -28,7 +28,7 @@
#include <Common/LofarLogger.h>
#include <GPUProc/gpu_wrapper.h>
-#include <GPUProc/cuda/PerformanceCounter.h>
+#include <GPUProc/PerformanceCounter.h>
#include <UnitTest++.h>
using namespace std;
--
GitLab