PxCudaContextManager

Defined in include/cudamanager/PxCudaContextManager.h

class PxCudaContextManager

Manages thread locks, and task scheduling for a CUDA context.

A PxCudaContextManager manages access to a single CUDA context, allowing it to be shared between multiple scenes. The context must be acquired from the manager before using any CUDA APIs unless stated differently.

The PxCudaContextManager is based on the CUDA driver API and explicitly does not support the CUDA runtime API (aka, CUDART).

Public Functions

template<typename T> inline void clearDeviceBufferAsync(T *deviceBuffer, PxU32 numElements, CUstream stream, PxI32 value = 0)

Schedules clear operation for a device memory buffer on the specified stream.

The cuda context will get acquired automatically

template<typename T> inline void copyDToH(T *hostBuffer, const T *deviceBuffer, PxU32 numElements)

Copies a device buffer to the host.

The cuda context will get acquired automatically

template<typename T> inline void copyHToD(T *deviceBuffer, const T *hostBuffer, PxU32 numElements)

Copies a host buffer to the device.

The cuda context will get acquired automatically

template<typename T> inline void copyDToHAsync(T *hostBuffer, const T *deviceBuffer, PxU32 numElements, CUstream stream)

Schedules device to host copy operation on the specified stream.

The cuda context will get acquired automatically

template<typename T> inline void copyHToDAsync(T *deviceBuffer, const T *hostBuffer, PxU32 numElements, CUstream stream)

Schedules host to device copy operation on the specified stream.

The cuda context will get acquired automatically

template<typename T> inline void copyDToDAsync(T *dstDeviceBuffer, const T *srcDeviceBuffer, PxU32 numElements, CUstream stream)

Schedules device to device copy operation on the specified stream.

The cuda context will get acquired automatically

template<typename T> inline void memsetAsync(T *dstDeviceBuffer, const T &value, PxU32 numElements, CUstream stream)

Schedules a memset operation on the device on the specified stream.

Only supported for 1 byte or 4 byte data types.

The cuda context will get acquired automatically

template<typename T> inline void allocDeviceBuffer(T *&deviceBuffer, PxU32 numElements, const char *filename = __FILE__, PxI32 line = __LINE__)

Allocates a device buffer.

The cuda context will get acquired automatically

template<typename T> inline T *allocDeviceBuffer(PxU32 numElements, const char *filename = __FILE__, PxI32 line = __LINE__)

Allocates a device buffer and returns the pointer to the memory.

The cuda context will get acquired automatically

template<typename T> inline void freeDeviceBuffer(T *&deviceBuffer)

Frees a device buffer.

The cuda context will get acquired automatically

template<typename T> inline void allocPinnedHostBuffer(T *&pinnedHostBuffer, PxU32 numElements, const char *filename = __FILE__, PxI32 line = __LINE__)

Allocates a pinned host buffer.

A pinned host buffer can be used on the gpu after getting a mapped device pointer from the pinned host buffer pointer, see getMappedDevicePtr The cuda context will get acquired automatically

See also

getMappedDevicePtr

template<typename T> inline void freePinnedHostBuffer(T *&pinnedHostBuffer)

Frees a pinned host buffer.

The cuda context will get acquired automatically

virtual CUdeviceptr getMappedDevicePtr(void *pinnedHostBuffer) = 0

Gets a mapped pointer from a pinned host buffer that can be used in cuda kernels directly.

Data access performance with a mapped pinned host pointer will be slower than using a device pointer directly but the changes done in the kernel will be available on the host immediately. The cuda context will get acquired automatically

virtual void acquireContext() = 0

Acquire the CUDA context for the current thread.

Acquisitions are allowed to be recursive within a single thread. You can acquire the context multiple times so long as you release it the same count.

The context must be acquired before using most CUDA functions.

virtual void releaseContext() = 0

Release the CUDA context from the current thread.

The CUDA context should be released as soon as practically possible, to allow other CPU threads to work efficiently.

virtual CUcontext getContext() = 0: Return the CUcontext.

virtual PxCudaContext *getCudaContext() = 0: Return the CudaContext.

virtual bool contextIsValid() const = 0

Context manager has a valid CUDA context.

This method should be called after creating a PxCudaContextManager, especially if the manager was responsible for allocating its own CUDA context (desc.ctx == NULL).

virtual bool supportsArchSM10() const = 0: G80.

virtual bool supportsArchSM11() const = 0: G92.

virtual bool supportsArchSM12() const = 0: GT200.

virtual bool supportsArchSM13() const = 0: GT260.

virtual bool supportsArchSM20() const = 0: GF100.

virtual bool supportsArchSM30() const = 0: GK100.

virtual bool supportsArchSM35() const = 0: GK110.

virtual bool supportsArchSM50() const = 0: GM100.

virtual bool supportsArchSM52() const = 0: GM200.

virtual bool supportsArchSM60() const = 0: GP100.

virtual bool isIntegrated() const = 0: true if GPU is an integrated (MCP) part

virtual bool canMapHostMemory() const = 0: true if GPU map host memory to GPU (0-copy)

virtual int getDriverVersion() const = 0: returns cached value of cuGetDriverVersion()

virtual size_t getDeviceTotalMemBytes() const = 0: returns cached value of device memory size

virtual int getMultiprocessorCount() const = 0: returns cache value of SM unit count

virtual unsigned int getClockRate() const = 0: returns cached value of SM clock frequency

virtual int getSharedMemPerBlock() const = 0: returns total amount of shared memory available per block in bytes

virtual int getSharedMemPerMultiprocessor() const = 0: returns total amount of shared memory available per multiprocessor in bytes

virtual unsigned int getMaxThreadsPerBlock() const = 0: returns the maximum number of threads per block

virtual const char *getDeviceName() const = 0: returns device name retrieved from driver

virtual CUdevice getDevice() const = 0: returns device handle retrieved from driver

virtual void setUsingConcurrentStreams(bool) = 0: turn on/off using concurrent streams for GPU work

virtual bool getUsingConcurrentStreams() const = 0: true if GPU work can run in concurrent streams

virtual void getDeviceMemoryInfo(size_t &free, size_t &total) const = 0: get currently available and total memory

virtual int usingDedicatedGPU() const = 0

Determine if the user has configured a dedicated PhysX GPU in the NV Control Panel.

Note

If using CUDA Interop, this will always return false

Returns: 1 if there is a dedicated GPU 0 if there is NOT a dedicated GPU -1 if the routine is not implemented

virtual CUmodule *getCuModules() = 0

Get the cuda modules that have been loaded into this context on construction.

Returns: Pointer to the cuda modules

virtual void release() = 0

Release the PxCudaContextManager.

If the PxCudaContextManager created the CUDA context it was responsible for, it also frees that context.

Do not release the PxCudaContextManager if there are any scenes using it. Those scenes must be released first.

Protected Functions

inline virtual ~PxCudaContextManager(): protected destructor, use release() method

virtual void *allocDeviceBufferInternal(PxU32 numBytes, const char *filename = NULL, PxI32 line = -1) = 0

virtual void *allocPinnedHostBufferInternal(PxU32 numBytes, const char *filename = NULL, PxI32 line = -1) = 0

virtual void freeDeviceBufferInternal(void *deviceBuffer) = 0

virtual void freePinnedHostBufferInternal(void *pinnedHostBuffer) = 0

virtual void clearDeviceBufferAsyncInternal(void *deviceBuffer, PxU32 numBytes, CUstream stream, PxI32 value) = 0

virtual void copyDToHAsyncInternal(void *hostBuffer, const void *deviceBuffer, PxU32 numBytes, CUstream stream) = 0

virtual void copyHToDAsyncInternal(void *deviceBuffer, const void *hostBuffer, PxU32 numBytes, CUstream stream) = 0

virtual void copyDToDAsyncInternal(void *dstDeviceBuffer, const void *srcDeviceBuffer, PxU32 numBytes, CUstream stream) = 0

virtual void copyDToHInternal(void *hostBuffer, const void *deviceBuffer, PxU32 numBytes) = 0

virtual void copyHToDInternal(void *deviceBuffer, const void *hostBuffer, PxU32 numBytes) = 0

virtual void memsetD8AsyncInternal(void *dstDeviceBuffer, const PxU8 &value, PxU32 numBytes, CUstream stream) = 0

virtual void memsetD32AsyncInternal(void *dstDeviceBuffer, const PxU32 &value, PxU32 numIntegers, CUstream stream) = 0