Zero-Copy Host Memory These notes will introduce: These materials comes from Chapter 11 of “CUDA by Example” by Jason Sanders and Edwards Kandrot. ITCS 6/8010 CUDA Programming, UNC-Charlotte, B. Wilkinson, Jan 15, 2011 1 Zero-copy memory Zero-copy refers to the GPU accessing the host memory without explicitly copying the data from the host memory to the GPU memory i.e. zero copying Depending upon the hardware structure the data may get copied though! • Integrated GPUs that are part of the system chipset and share system memory do not. --- example MacBook Pro • Discrete GPU cards with their own device memory do. 2 CUDA routines for zero-copy memory Use page-locked memory. Allocate with: cudaHostAlloc (void ** ptr, size_t size, unsigned int flags) Allocates page-locked memory and accessible to the device. One flag: cudaHostAllocMapped -- Maps allocation into CUDA address space. Device pointer to memory obtained by calling cudaHostGetDevicePointer() – needed to account for different address spaces. Reference: NVIDIA CUDA library http://developer.download.nvidia.com/compute/cuda/3_0/toolkit/docs/online/ 3 Flags continued cudaHostAllocWriteCombined -- Allocates memory as “write-combined”, which can be transferred more quickly across PCIe bus on some system configurations, but cannot be read efficiently by most CPUs. Use for memory written by CPU and read by device via mapped pinned memory. Combining flags: cudaHostAllocMapped || cudaHostAllocWriteCombined Reference: NVIDIA CUDA library http://developer.download.nvidia.com/compute/cuda/3_0/toolkit/docs/online/ 4 Code to allocate memory and get pointer for device int *a; size = … ; // host pointer // number of bytes to allocate cudaHostAlloc( (void**)&a, size, cudaHostAllocMapped || cudaHostAllocWriteCombined ); cudaHostGetDevicePointer(&dev_a, a, 0); If needed Always zero currently (for future extensions) 5 Using pointer to host memory Simply use in kernel call where one would otherwise have used a device memory pointer: MyKernel<<< B,T>>> (*a, … ); without needing to modify the kernel code at all! 6 #define N 32 // size of vectors Example Vector addition without hostint main(int argc, char *argv[]) { int T = 32, B = 1; // threads per block and blocks per grid device transfers int *a,*b,*c; // host pointers __global__ void add(int *a,int *b, int *c) { int tid = blockIdx.x * blockDim.x + threadIdx.x; if(tid < N) c[tid] = a[tid]+b[tid]; } int *dev_a, *dev_b, *dev_c; // device pointers to host memory cudaEvent_t start, stop; float elapsed_time_ms; // to measure time cudaHostAlloc( (void**)&a, size, cudaHostAllocMapped || cudaHostAllocWriteCombined ); cudaHostAlloc( (void**)&b, size, cudaHostAllocMapped || cudaHostAllocWriteCombined ); cudaHostAlloc( (void**)&c, size, cudaHostAllocMapped ); … // load arrays with some numbers Note flags cudaHostGetDevicePointer(&dev_a, a, 0); // mem. copy to device not need now, but ptrs needed instead cudaHostGetDevicePointer(&dev_b, b, 0); cudaHostGetDevicePointer(&dev_c ,c, 0); … // start time add<<<B,T>>>(dev_a,dev_b,dev_c); cudaThreadSynchronize(); … … // copy back not needed but now need thread synchronization // end time // print results printf("Time to calculate results: %f ms.\n", elapsed_time_ms); // print out execution time cudaFreeHost(a); // clean up cudaFreeHost(b); cudaFreeHost(c); cudaEventDestroy(start); cudaEventDestroy(stop); book seems to miss out this special routine 7 when using cudaHostAlloc Code to determine whether GPU has the capability of features being used Look at device properties: cudaDeviceProp prop; int myDevice; cudaGetDevice(&myDevice); Returns device executing thread Returns a structure, see next cudaGetDeviceProperties(&prop, myDevice); If (prop.property_name != 1) printf(“Feature not available\n”); Various property names, see next 8 Properties struct cudaDeviceProp { char name[256]; size_t totalGlobalMem; size_t sharedMemPerBlock; int regsPerBlock; int warpSize; size_t memPitch; int maxThreadsPerBlock; int maxThreadsDim[3]; int maxGridSize[3]; size_t totalConstMem; int major; int minor; int clockRate; size_t textureAlignment; int deviceOverlap; int multiProcessorCount; int kernelExecTimeoutEnabled; int integrated; int canMapHostMemory; int computeMode; int concurrentKernels; 9 Checking can map page-locked host memory into device address space … cudaDeviceProp prop; int myDevice; cudaGetDevice(&myDevice); Very likely as only needs compute capability > 1.0 cudaGetDeviceProperties(&prop, myDevice); If (prop.canMapHostMemory != 1) { printf(“Feature not available\n”); return 0; } … 10 Integrated GPU systems Example: My 13” MacBook Pro, 2010 Zero-copy memory particularly interesting with integrated GPU systems where system memory is shared between CPU and GPU. Increased performance will always result when using zero-copy memory (according to the course textbook) CPU GPU 2.4 GHz Intel Core 2 Duo NVIDIA GeForce 320M Shared between CPU and GPU 256 MB DDR3 SDRAM DDR3 SDRAM 4 GB Main memory Intel Graphics Media Accelerator (GMA ) shared bus on 15/17” models 11 Using multiple GPU on one system Each GPU needs to be controlled by a separate thread: Code Thread 1 Thread 2 GPU 1 GPU 2 So need to write a multi-threaded program using thread APIs/tools such as Pthreads, WinThreads, OpenMP, … . 12 … #if _WIN32 //Windows threads. #include <windows.h> Textbook utility routines for multi-threading Found in ../common/book.h typedef HANDLE CUTThread; typedef unsigned (WINAPI *CUT_THREADROUTINE)(void *); #define CUT_THREADPROC unsigned WINAPI #define CUT_THREADEND return 0 #else //POSIX threads. #include <pthread.h> typedef pthread_t CUTThread; typedef void *(*CUT_THREADROUTINE)(void *); #define CUT_THREADPROC void #define CUT_THREADEND #endif //Create thread. CUTThread start_thread( CUT_THREADROUTINE, void *data ); //Wait for thread to finish. void end_thread( CUTThread thread ); Provides for Win32 Threads for Windows or Pthreads for Linux thread = start_thread(funct,ptr) Used to start a new thread Takes as arguments: void* funct (void*) void* ptr Returns CUTThread type thread identifier //Destroy thread. void destroy_thread( CUTThread thread ); //Wait for multiple threads. void wait_for_threads( const CUTThread *threads, int num ); #if _WIN32 //Create thread CUTThread start_thread(CUT_THREADROUTINE func, void *data){ return CreateThread(NULL, 0, (LPTHREAD_START_ROUTINE)func, data, 0, NULL); } //Wait for thread to finish void end_thread(CUTThread thread){ WaitForSingleObject(thread, INFINITE); CloseHandle(thread); } //Destroy thread void destroy_thread( CUTThread thread ){ TerminateThread(thread, 0); CloseHandle(thread); } //Wait for multiple threads void wait_for_threads(const CUTThread * threads, int num){ WaitForMultipleObjects(num, threads, true, INFINITE); for(int i = 0; i < num; i++) CloseHandle(threads[i]); } #else //Create thread CUTThread start_thread(CUT_THREADROUTINE func, void * data){ pthread_t thread; pthread_create(&thread, NULL, func, data); return thread; } //Wait for thread to finish void end_thread(CUTThread thread){ pthread_join(thread, NULL); } To terminate thread (join to main thread): end_thread(thread) //Destroy thread void destroy_thread( CUTThread thread ){ pthread_cancel(thread); } //Wait for multiple threads void wait_for_threads(const CUTThread * threads, int num){ for(int i = 0; i < num; i++) end_thread( threads[i] ); } #endif … 13 Book code Two structures containing data to be sent to each thread: Elements of structure previously defined DataStruct data[2]; data[0].a = a; data[0].b = b; … data[1].a = a + N/2; data[1].b = b+N/2; … … CUTThread thread = start_thread(routine, &(data[0]); routine (&(data[1])); … end_thread(thread); Two threads, each executing routine() 14 Pinned memory on multiple GPUs Pinned memory only pinned by thread allocating the pinned memory Other threads see it as pageable and access slower. These threads cannot use cudaMemcpyAsync, which requires pinned memory “Portable” pinned memory Memory allowed to move between host threads and any thread to see it as pinned memory Use cudaHostAlloc and include cudaAllocPortable flag 15 Questions