diff --git a/libtcc/Correlator.h b/libtcc/Correlator.h
index 723361f8109930ddf789c7e63a33d62a9c5d6d1e..7b33fa2f3fdffda90843ed22efdb1ab7c8b3638c 100644
--- a/libtcc/Correlator.h
+++ b/libtcc/Correlator.h
@@ -1,5 +1,5 @@
-#if !defined CORRELATOR_H
-#define CORRELATOR_H
+#if !defined TCC_CORRELATOR_H
+#define TCC_CORRELATOR_H
#include "libtcc/CorrelatorKernel.h"
#include "util/cu.h"
@@ -17,7 +17,7 @@ namespace tcc {
unsigned nrReceiversPerBlock = 64
); // throw (cu::Error, nvrtc::Error)
- void launchAsync(cu::Stream &, cu::DeviceMemory &isibilities, cu::DeviceMemory &samples); // throw (cu::Error)
+ void launchAsync(cu::Stream &, cu::DeviceMemory &visibilities, cu::DeviceMemory &samples); // throw (cu::Error)
void launchAsync(CUstream, CUdeviceptr visibilities, CUdeviceptr samples); // throw (cu::Error)
uint64_t FLOPS() const;
diff --git a/libtcc/CorrelatorKernel.h b/libtcc/CorrelatorKernel.h
index e28a2fdb197bcf08ff566331c6b2a1259dc7d60b..a0a92f1cd65e6859984b2a027bb54cfed2d6d1bf 100644
--- a/libtcc/CorrelatorKernel.h
+++ b/libtcc/CorrelatorKernel.h
@@ -1,5 +1,5 @@
-#if !defined CORRELATOR_KERNEL_H
-#define CORRELATOR_KERNEL_H
+#if !defined TCC_CORRELATOR_KERNEL_H
+#define TCC_CORRELATOR_KERNEL_H
#include "libtcc/Kernel.h"
diff --git a/libtcc/Kernel.h b/libtcc/Kernel.h
index f6c9ab410017c32c92fb3e2356a2546a94a76b49..63329ee32a91a50e6c261b5f5f48e97bfbe533df 100644
--- a/libtcc/Kernel.h
+++ b/libtcc/Kernel.h
@@ -1,5 +1,5 @@
-#if !defined KERNEL_H
-#define KERNEL_H
+#if !defined TCC_KERNEL_H
+#define TCC_KERNEL_H
#include "util/cu.h"
diff --git a/test/Common/Config.h b/test/Common/Config.h
index 0ecedb4b6185e0741431a3067fa137ffa684bae6..fcacde73d55f489feb1b820504ebbffcfa960305 100644
--- a/test/Common/Config.h
+++ b/test/Common/Config.h
@@ -1,5 +1,5 @@
-#if !defined CONFIG_H
-#define CONFIG_H
+#if !defined TCC_CONFIG_H
+#define TCC_CONFIG_H
#if defined __ARM_ARCH
#define UNIFIED_MEMORY // assume this is a Jetson Xavier
diff --git a/test/CorrelatorTest/CorrelatorTest.cc b/test/CorrelatorTest/CorrelatorTest.cc
index 23bc7c156344e3f78766f4422d49e53cb09b81f9..2735a36378a9545b678b78b089623571e107d66c 100644
--- a/test/CorrelatorTest/CorrelatorTest.cc
+++ b/test/CorrelatorTest/CorrelatorTest.cc
@@ -101,7 +101,7 @@ template <typename SampleType, typename VisibilityType> void CorrelatorTest::doT
hostToDeviceStream.memcpyHtoDAsync(deviceSamples, hostSamples, samplesRef.bytesize());
hostToDeviceRecordStop.enqueue(hostToDeviceStream);
- stream.waitEvent(hostToDeviceRecordStop.event);
+ stream.wait(hostToDeviceRecordStop.event);
#endif
computeRecordStart.enqueue(stream);
@@ -112,7 +112,7 @@ template <typename SampleType, typename VisibilityType> void CorrelatorTest::doT
computeRecordStop.enqueue(stream);
#if !defined UNIFIED_MEMORY
- deviceToHostStream.waitEvent(computeRecordStop.event);
+ deviceToHostStream.wait(computeRecordStop.event);
deviceToHostRecordStart.enqueue(deviceToHostStream);
deviceToHostStream.memcpyDtoHAsync(hostVisibilities, deviceVisibilities, visibilitiesRef.bytesize());
deviceToHostRecordStop.enqueue(deviceToHostStream);
diff --git a/test/CorrelatorTest/Options.h b/test/CorrelatorTest/Options.h
index 9e51307fcb5e2eefdabf08b11c04119bbab433e6..441e1fae0d0623f37cb5fd28c424a81fcb82c3cd 100644
--- a/test/CorrelatorTest/Options.h
+++ b/test/CorrelatorTest/Options.h
@@ -1,5 +1,5 @@
-#if !defined OPTIONS_H
-#define OPTIONS_H
+#if !defined TCC_OPTIONS_H
+#define TCC_OPTIONS_H
#include <exception>
#include <string>
diff --git a/util/cu.h b/util/cu.h
index 46dfdea4dce9cbb44ee4b26b6f45331cebc49175..72f6cf575a7512b097539202a623b0f089ed951a 100644
--- a/util/cu.h
+++ b/util/cu.h
@@ -5,6 +5,7 @@
#include <cuda_runtime_api.h>
#include <exception>
#include <fstream>
+#include <memory>
#include <string>
#include <vector>
@@ -61,21 +62,18 @@ namespace cu {
template <typename T> class Wrapper
{
public:
- // Wrapper<T>(Wrapper<T> &) = delete; // disallow copies
-
// conversion to C-style T
- operator const T & () const
+ operator T () const
{
return _obj;
}
- operator T & ()
+ operator T ()
{
return _obj;
}
-#if 0 // makes no sense if object is not copyable
bool operator == (const Wrapper<T> &other)
{
return _obj == other._obj;
@@ -85,24 +83,35 @@ namespace cu {
{
return _obj != other._obj;
}
-#endif
protected:
Wrapper<T>()
+ {
+ }
+
+ Wrapper<T>(const Wrapper<T> &other)
+ :
+ _obj(other._obj),
+ manager(other.manager)
+ {
+ }
+
+ Wrapper<T>(Wrapper<T> &&other)
:
- hasOwnership(true)
+ _obj(other._obj),
+ manager(std::move(other.manager))
{
+ other._obj = 0;
}
Wrapper<T>(T &obj)
:
- _obj(obj),
- hasOwnership(false)
+ _obj(obj)
{
}
- bool hasOwnership;
T _obj;
+ std::shared_ptr<T> manager;
};
class Device : public Wrapper<CUdevice>
@@ -185,6 +194,7 @@ namespace cu {
_primaryContext(false)
{
checkCudaCall(cuCtxCreate(&_obj, flags, device));
+ manager = std::shared_ptr<CUcontext>(new CUcontext(_obj), [] (CUcontext *ptr) { if (*ptr) cuCtxDestroy(*ptr); delete ptr; });
}
Context(CUcontext context)
@@ -194,14 +204,6 @@ namespace cu {
{
}
- ~Context()
- {
- if (hasOwnership)
- checkCudaCall(cuCtxDestroy(_obj));
- //else
- //checkCudaCall(cuDevicePrimaryCtxRelease(getDevice())); // FIXME
- }
-
unsigned getApiVersion() const
{
unsigned version;
@@ -225,7 +227,7 @@ namespace cu {
{
CUcontext context;
checkCudaCall(cuCtxGetCurrent(&context));
- return Context(context);
+ return std::move(Context(context));
}
void setCurrent() const
@@ -303,11 +305,7 @@ namespace cu {
HostMemory(size_t size, int flags = 0)
{
checkCudaCall(cuMemHostAlloc(&_obj, size, flags));
- }
-
- ~HostMemory()
- {
- checkCudaCall(cuMemFreeHost(_obj));
+ manager = std::shared_ptr<void *>(new (void *)(_obj), [] (void **ptr) { cuMemFreeHost(*ptr); delete ptr; });
}
template <typename T> operator T * ()
@@ -323,6 +321,7 @@ namespace cu {
DeviceMemory(size_t size)
{
checkCudaCall(cuMemAlloc(&_obj, size));
+ manager = std::shared_ptr<CUdeviceptr>(new CUdeviceptr(_obj), [] (CUdeviceptr *ptr) { cuMemFree(*ptr); delete ptr; });
}
DeviceMemory(CUdeviceptr ptr)
@@ -331,32 +330,15 @@ namespace cu {
{
}
- DeviceMemory(HostMemory &hostMemory)
+ DeviceMemory(const HostMemory &hostMemory)
{
- hasOwnership = false;
checkCudaCall(cuMemHostGetDevicePointer(&_obj, hostMemory, 0));
}
- ~DeviceMemory()
- {
- if (hasOwnership)
- checkCudaCall(cuMemFree(_obj));
- }
-
const void *parameter() const // used to construct parameter list for launchKernel();
{
return &_obj;
}
-
- template <typename T> operator T * () const
- {
- return (T *) _obj;
- }
-
- template <typename T> operator const T * () const
- {
- return (const T *) _obj;
- }
};
@@ -366,6 +348,7 @@ namespace cu {
Array(unsigned width, CUarray_format format, unsigned numChannels)
{
Array(width, 0, format, numChannels);
+ manager = std::shared_ptr<CUarray>(new CUarray(_obj), [] (CUarray *ptr) { cuArrayDestroy(*ptr); delete ptr; });
}
Array(unsigned width, unsigned height, CUarray_format format, unsigned numChannels)
@@ -376,6 +359,7 @@ namespace cu {
descriptor.Format = format;
descriptor.NumChannels = numChannels;
checkCudaCall(cuArrayCreate(&_obj, &descriptor));
+ manager = std::shared_ptr<CUarray>(new CUarray(_obj), [] (CUarray *ptr) { cuArrayDestroy(*ptr); delete ptr; });
}
Array(unsigned width, unsigned height, unsigned depth, CUarray_format format, unsigned numChannels)
@@ -388,6 +372,7 @@ namespace cu {
descriptor.NumChannels = numChannels;
descriptor.Flags = 0;
checkCudaCall(cuArray3DCreate(&_obj, &descriptor));
+ manager = std::shared_ptr<CUarray>(new CUarray(_obj), [] (CUarray *ptr) { cuArrayDestroy(*ptr); delete ptr; });
}
Array(CUarray &array)
@@ -395,79 +380,6 @@ namespace cu {
Wrapper(array)
{
}
-
- ~Array()
- {
- if (hasOwnership)
- checkCudaCall(cuArrayDestroy(_obj));
- }
- };
-
-
-#if 0
- class TexObject : public Wrapper<CUtexObject> {
- public:
- TexObject(onst CUDA_RESOURCE_DESC *pResDesc, const CUDA_TEXTURE_DESC *pTexDesc, const CUDA_RESOURCE_VIEW_DESC *pResViewDesc)
- {
- checkCudaCall(cuTexObjectCreate(&_obj, pResDesc, pTexDesc, pResViewDesc));
- }
-
- TexObject(CUtexObject &obj)
- :
- Wrapper<CUtexObject>(obj)
- {
- }
-
- ~TexObject()
- {
- if (hasOwnership)
- checkCudaCall(cuTexObjectDestroy(_obj));
- }
-
-
- };
-#endif
-
-
- class TexRef : public Wrapper<CUtexref> {
- public:
- TexRef(CUtexref texref)
- :
- Wrapper<CUtexref>(texref)
- {
- }
-
-#if 0 // deprecated
- void setAddress(size_t &byte_offset, DeviceMemory &memory, size_t size)
- {
- checkCudaCall(cuTexRefSetAddress(&byte_offset, _obj, memory, size));
- }
-
- void setArray(Array &array, unsigned flags = CU_TRSA_OVERRIDE_FORMAT)
- {
- checkCudaCall(cuTexRefSetArray(_obj, array, flags));
- }
-
- void setAddressMode(int dim, CUaddress_mode am)
- {
- checkCudaCall(cuTexRefSetAddressMode(_obj, dim, am));
- }
-
- void setFilterMode(CUfilter_mode fm)
- {
- checkCudaCall(cuTexRefSetFilterMode(_obj, fm));
- }
-
- void setFlags(int flags)
- {
- checkCudaCall(cuTexRefSetFlags(_obj, flags));
- }
-
- void setFormat(CUarray_format fmt, int numPackedComponents)
- {
- checkCudaCall(cuTexRefSetFormat(_obj, fmt, numPackedComponents));
- }
-#endif
};
@@ -499,11 +411,13 @@ namespace cu {
#else
checkCudaCall(cuModuleLoad(&_obj, file_name));
#endif
+ manager = std::shared_ptr<CUmodule>(new CUmodule(_obj), [] (CUmodule *ptr) { cuModuleUnload(*ptr); delete ptr; });
}
Module(const void *data)
{
checkCudaCall(cuModuleLoadData(&_obj, data));
+ manager = std::shared_ptr<CUmodule>(new CUmodule(_obj), [] (CUmodule *ptr) { cuModuleUnload(*ptr); delete ptr; });
}
Module(CUmodule &module)
@@ -512,18 +426,14 @@ namespace cu {
{
}
- ~Module()
- {
- if (hasOwnership)
- checkCudaCall(cuModuleUnload(_obj));
- }
-
+#if 0
TexRef getTexRef(const char *name) const
{
CUtexref texref;
checkCudaCall(cuModuleGetTexRef(&texref, _obj, name));
return TexRef(texref);
}
+#endif
CUdeviceptr getGlobal(const char *name) const
{
@@ -537,7 +447,7 @@ namespace cu {
class Function : public Wrapper<CUfunction>
{
public:
- Function(Module &module, const char *name)
+ Function(const Module &module, const char *name)
{
checkCudaCall(cuModuleGetFunction(&_obj, module, name));
}
@@ -559,13 +469,6 @@ namespace cu {
{
checkCudaCall(cuFuncSetCacheConfig(_obj, config));
}
-
-#if 0
- void paramSetTexRef(TexRef &texref)
- {
- checkCudaCall(cuParamSetTexRef(_obj, CU_PARAM_TR_DEFAULT, texref));
- }
-#endif
};
@@ -575,6 +478,7 @@ namespace cu {
Event(int flags = CU_EVENT_DEFAULT)
{
checkCudaCall(cuEventCreate(&_obj, flags));
+ manager = std::shared_ptr<CUevent>(new CUevent(_obj), [] (CUevent *ptr) { cuEventDestroy(*ptr); delete ptr; });
}
Event(CUevent &event)
@@ -583,16 +487,10 @@ namespace cu {
{
}
- ~Event()
- {
- if (hasOwnership)
- checkCudaCall(cuEventDestroy(_obj));
- }
-
- float elapsedTime(Event &second) const
+ float elapsedTime(const Event &start) const
{
float ms;
- checkCudaCall(cuEventElapsedTime(&ms, second, _obj));
+ checkCudaCall(cuEventElapsedTime(&ms, start, _obj));
return ms;
}
@@ -623,6 +521,7 @@ namespace cu {
Stream(int flags = CU_STREAM_DEFAULT)
{
checkCudaCall(cuStreamCreate(&_obj, flags));
+ manager = std::shared_ptr<CUstream>(new CUstream(_obj), [] (CUstream *ptr) { cuStreamDestroy(*ptr); delete ptr; });
}
Stream(CUstream stream)
@@ -631,12 +530,6 @@ namespace cu {
{
}
- ~Stream()
- {
- if (hasOwnership)
- checkCudaCall(cuStreamDestroy(_obj));
- }
-
void memcpyHtoDAsync(CUdeviceptr devPtr, const void *hostPtr, size_t size)
{
checkCudaCall(cuMemcpyHtoDAsync(devPtr, hostPtr, size, _obj));
@@ -669,7 +562,7 @@ namespace cu {
checkCudaCall(cuStreamSynchronize(_obj));
}
- void waitEvent(Event &event)
+ void wait(Event &event)
{
checkCudaCall(cuStreamWaitEvent(_obj, event, 0));
}
@@ -700,7 +593,7 @@ namespace cu {
}
};
-#if 1
+#if 0
class Graph : public Wrapper<CUgraph>
{
public:
@@ -746,6 +639,7 @@ namespace cu {
Graph(unsigned flags = 0)
{
checkCudaCall(cuGraphCreate(&_obj, flags));
+ manager = std::shared_ptr<CUgraphNode>(new CUgraphNode(_obj), [] (CUgraphNode *ptr) { cuGraphDestroy(*ptr); delete ptr; });
}
Graph(CUgraph &graph)
@@ -754,12 +648,6 @@ namespace cu {
{
}
- ~Graph()
- {
- if (hasOwnership)
- checkCudaCall(cuGraphDestroy(_obj));
- }
-
ExecKernelNode addKernelNode(/* std::vector<GraphNode> dependencies, */ const KernelNodeParams &kernelArgs)
{
ExecKernelNode node;
diff --git a/util/multi_array.h b/util/multi_array.h
index 1db0ea5a62be9a2acc5d66d8b5128a14641db4e8..651061d2f563c40f32c9f9a50d781e4d9ee86077 100644
--- a/util/multi_array.h
+++ b/util/multi_array.h
@@ -1,5 +1,5 @@
-#if !defined MULTI_ARRAY_H
-#define MULTI_ARRAY_H
+#if !defined TCC_MULTI_ARRAY_H
+#define TCC_MULTI_ARRAY_H
// Experimental class that provides multi-dimensional arrays, similar to
// boost::multi_array[_ref], but much simpler.