0% found this document useful (0 votes)

566 views

PG-02829-001_v5.0|x

Chapter 1. INTRODUCTION

1.1From Graphics Processing to General Purpose Parallel Computing

Driven by the insatiable market demand for realtime, high-definition 3D graphics, the programmable Graphic Processor Unit or GPU has evolved into a highly parallel, multithreaded, manycore processor with tremendous computational horsepower and very high memory bandwidth, as illustrated by Figure 1 Floating-Point Operations per Second for the CPU and GPU and Figure 2 Memory Bandwidth for the CPU and GPU.

Introduction

Figure4CUDA Is Designed to Support Various Languages and Application Programming Interfaces

1.3A Scalable Programming Model

The advent of multicore CPUs and manycore GPUs means that mainstream processor chips are now parallel systems. Furthermore, their parallelism continues to scale with Moore's law. The challenge is to develop application software that transparently scales its parallelism to leverage the increasing number of processor cores, much as 3D graphics applications transparently scale their parallelism to manycore GPUs with widely varying numbers of cores. The CUDA parallel programming model is designed to overcome this challenge while maintaining a low learning curve for programmers familiar with standard programming languages such as C. At its core are three key abstractions a hierarchy of thread groups, shared memories, and barrier synchronization that are simply exposed to the programmer as a minimal set of language extensions. These abstractions provide fine-grained data parallelism and thread parallelism, nested within coarse-grained data parallelism and task parallelism. They guide the programmer to partition the problem into coarse sub-problems that can be solved independently in parallel by blocks of threads, and each sub-problem into finer pieces that can be solved cooperatively in parallel by all threads within the block.

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|4

Introduction

This decomposition preserves language expressivity by allowing threads to cooperate when solving each sub-problem, and at the same time enables automatic scalability. Indeed, each block of threads can be scheduled on any of the available multiprocessors within a GPU, in any order, concurrently or sequentially, so that a compiled CUDA program can execute on any number of multiprocessors as illustrated by Figure 5 Automatic Scalability, and only the runtime system needs to know the physical multiprocessor count. This scalable programming model allows the GPU architecture to span a wide market range by simply scaling the number of multiprocessors and memory partitions: from the high-performance enthusiast GeForce GPUs and professional Quadro and Tesla computing products to a variety of inexpensive, mainstream GeForce GPUs (see CUDAEnabled GPUs for a list of all CUDA-enabled GPUs).

A GPU is built around an array of Streaming Multiprocessors (SMs) (see Hardware Implementation for more details). A multithreaded program is partitioned into blocks of threads that execute independently from each other, so that a GPU with more multiprocessors will automatically execute the program in less time than a GPU with fewer multiprocessors.

Figure5Automatic Scalability

1.4Document Structure
This document is organized into the following chapters: Introduction is a general introduction to CUDA. Programming Model outlines the CUDA programming model Programming Interface describes the programming interface

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|5

Introduction

Hardware Implementation describes the hardware implementation Performance Guidelines gives some guidance on how to achieve maximum performance CUDA-Enabled GPUs lists all CUDA-enabled devices C Language Extensions is a detailed description of all extensions to the C language Mathematical Functions lists the mathematical functions supported in CUDA C/C++ Language Support lists the C++ features supported in device code Texture Fetching gives more details on texture fetching Compute Capabilities gives the technical specifications of various devices, as well as more architectural details Driver API introduces the low-level driver API

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|6

Chapter 2. PROGRAMMING MODEL

This chapter introduces the main concepts behind the CUDA programming model by outlining how they are exposed in C. An extensive description of CUDA C is given in Programming Interface. Full code for the vector addition example used in this chapter and the next can be found in the vectorAdd SDK code sample.

2.1Kernels
CUDA C extends C by allowing the programmer to define C functions, called kernels, that, when called, are executed N times in parallel by N different CUDA threads, as opposed to only once like regular C functions. A kernel is defined using the __global__ declaration specifier and the number of CUDA threads that execute that kernel for a given kernel call is specified using a new <<<>>> execution configuration syntax (see C Language Extensions). Each thread that executes the kernel is given a unique thread ID that is accessible within the kernel through the built-in threadIdx variable. As an illustration, the following sample code adds two vectors A and B of size N and stores the result into vector C:
// Kernel definition __global__ void VecAdd(float* A, float* B, float* C) { int i = threadIdx.x; C[i] = A[i] + B[i]; } int main() { ... // Kernel invocation with N threads VecAdd<<<1, N>>>(A, B, C); ... }

Here, each of the N threads that execute VecAdd() performs one pair-wise addition.

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|7

Programming Model

2.2Thread Hierarchy
For convenience, threadIdx is a 3-component vector, so that threads can be identified using a one-dimensional, two-dimensional, or three-dimensional thread index, forming a one-dimensional, two-dimensional, or three-dimensional thread block. This provides a natural way to invoke computation across the elements in a domain such as a vector, matrix, or volume. The index of a thread and its thread ID relate to each other in a straightforward way: For a one-dimensional block, they are the same; for a two-dimensional block of size (Dx, Dy),the thread ID of a thread of index (x, y) is (x + y Dx); for a three-dimensional block of size (Dx, Dy, Dz), the thread ID of a thread of index (x, y, z) is (x + y Dx + z Dx Dy). As an example, the following code adds two matrices A and B of size NxN and stores the result into matrix C:
// Kernel definition __global__ void MatAdd(float A[N][N], float B[N][N], float C[N][N]) { int i = threadIdx.x; int j = threadIdx.y; C[i][j] = A[i][j] + B[i][j]; } int main() { ... // Kernel invocation with one block of N * N * 1 threads int numBlocks = 1; dim3 threadsPerBlock(N, N); MatAdd<<<numBlocks, threadsPerBlock>>>(A, B, C); ... }

Figure8Heterogeneous Programming

2.5Compute Capability
The compute capability of a device is defined by a major revision number and a minor revision number. Devices with the same major revision number are of the same core architecture. The major revision number is 3 for devices based on the Kepler architecture, 2 for devices based on the Fermi architecture, and 1 for devices based on the Tesla architecture.

PG-02829-001_v5.0|15

Programming Interface

CUDA_CACHE_PATH specifies the folder where the compute cache files are stored; the default values are: on Windows, %APPDATA%\NVIDIA\ComputeCache, on MacOS, $HOME/Library/Application\ Support/NVIDIA/ ComputeCache, on Linux, ~/.nv/ComputeCache Setting CUDA_FORCE_PTX_JIT to 1 forces the device driver to ignore any binary code embedded in an application (see Application Compatibility) and to just-in-time compile embedded PTX code instead; if a kernel does not have embedded PTX code, it will fail to load; this environment variable can be used to validate that PTX code is embedded in an application and that its just-in-time compilation works as expected to guarantee application forward compatibility with future architectures.

3.1.2Binary Compatibility
Binary code is architecture-specific. A cubin object is generated using the compiler option -code that specifies the targeted architecture: For example, compiling with -code=sm_13 produces binary code for devices of compute capability 1.3 (see Compute Capability for a description of the compute capability). Binary compatibility is guaranteed from one minor revision to the next one, but not from one minor revision to the previous one or across major revisions. In other words, a cubin object generated for compute capability X.y is only guaranteed to execute on devices of compute capability X.z where zy.

3.1.3PTX Compatibility
Some PTX instructions are only supported on devices of higher compute capabilities. For example, atomic instructions on global memory are only supported on devices of compute capability 1.1 and above; double-precision instructions are only supported on devices of compute capability 1.3 and above. The -arch compiler option specifies the compute capability that is assumed when compiling C to PTX code. So, code that contains double-precision arithmetic, for example, must be compiled with arch=sm_13 (or higher compute capability), otherwise double-precision arithmetic will get demoted to single-precision arithmetic. PTX code produced for some specific compute capability can always be compiled to binary code of greater or equal compute capability.

3.1.4Application Compatibility
To execute code on devices of specific compute capability, an application must load binary or PTX code that is compatible with this compute capability as described in Binary Compatibility and PTX Compatibility. In particular, to be able to execute code on future architectures with higher compute capability for which no binary code can be generated yet , an application must load PTX code that will be just-in-time compiled for these devices (see Just-in-Time Compilation).

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|16

Programming Interface

Which PTX and binary code gets embedded in a CUDA C application is controlled by the -arch and -code compiler options or the -gencode compiler option as detailed in the nvcc user manual. For example,
nvcc x.cu -gencode arch=compute_10,code=sm_10 -gencode arch=compute_11,code=\compute_11,sm_11\

embeds binary code compatible with compute capability 1.0 (first -gencode option) and PTX and binary code compatible with compute capability 1.1 (second -gencode option). Host code is generated to automatically select at runtime the most appropriate code to load and execute, which, in the above example, will be: 1.0 binary code for devices with compute capability 1.0, 1.1 binary code for devices with compute capability 1.1, 1.2, 1.3, binary code obtained by compiling 1.1 PTX code for devices with compute capabilities 2.0 and higher. x.cu can have an optimized code path that uses atomic operations, for example, which are only supported in devices of compute capability 1.1 and higher. The __CUDA_ARCH__ macro can be used to differentiate various code paths based on compute capability. It is only defined for device code. When compiling with arch=compute_11 for example, __CUDA_ARCH__ is equal to 110. Applications using the driver API must compile code to separate files and explicitly load and execute the most appropriate file at runtime. The nvcc user manual lists various shorthands for the -arch, -code, and gencode compiler options. For example, -arch=sm_13 is a shorthand for arch=compute_13 -code=compute_13,sm_13 (which is the same as -gencode arch=compute_13,code=\compute_13,sm_13\).

3.1.5C/C++ Compatibility
The front end of the compiler processes CUDA source files according to C++ syntax rules. Full C++ is supported for the host code. However, only a subset of C++ is fully supported for the device code as described in C/C++ Language Support.

3.1.664-Bit Compatibility
The 64-bit version of nvcc compiles device code in 64-bit mode (i.e., pointers are 64-bit). Device code compiled in 64-bit mode is only supported with host code compiled in 64bit mode. Similarly, the 32-bit version of nvcc compiles device code in 32-bit mode and device code compiled in 32-bit mode is only supported with host code compiled in 32-bit mode. The 32-bit version of nvcc can compile device code in 64-bit mode also using the -m64 compiler option. The 64-bit version of nvcc can compile device code in 32-bit mode also using the -m32 compiler option.

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|17

Programming Interface

3.2CUDA C Runtime
The runtime is implemented in the cudart dynamic library which is typically included in the application installation package. All its entry points are prefixed with cuda. As mentioned in Heterogeneous Programming, the CUDA programming model assumes a system composed of a host and a device, each with their own separate memory. Device Memory gives an overview of the runtime functions used to manage device memory. Shared Memory illustrates the use of shared memory, introduced in Thread Hierarchy, to maximize performance. Page-Locked Host Memory introduces page-locked host memory that is required to overlap kernel execution with data transfers between host and device memory. Asynchronous Concurrent Execution describes the concepts and API used to enable asynchronous concurrent execution at various levels in the system. Multi-Device System shows how the programming model extends to a system with multiple devices attached to the same host. Error Checking describes how to properly check the errors generated by the runtime. Call Stack mentions the runtime functions used to manage the CUDA C call stack. Texture and Surface Memory presents the texture and surface memory spaces that provide another way to access device memory; they also expose a subset of the GPU texturing hardware. Graphics Interoperability introduces the various functions the runtime provides to interoperate with the two main graphics APIs, OpenGL and Direct3D.

3.2.1Initialization
There is no explicit initialization function for the runtime; it initializes the first time a runtime function is called (more specifically any function other than functions from the device and version management sections of the reference manual). One needs to keep this in mind when timing runtime function calls and when interpreting the error code from the first call into the runtime. During initialization, the runtime creates a CUDA context for each device in the system (see Context for more details on CUDA contexts). This context is the primary context for this device and it is shared among all the host threads of the application. This all happens under the hood and the runtime does not expose the primary context to the application. When a host thread calls cudaDeviceReset(), this destroys the primary context of the device the host thread currently operates on (i.e., the current device as defined in Device Selection). The next runtime function call made by any host thread that has this device as current will create a new primary context for this device.

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|18

Programming Interface

3.2.2Device Memory
As mentioned in Heterogeneous Programming, the CUDA programming model assumes a system composed of a host and a device, each with their own separate memory. Kernels operate out of device memory, so the runtime provides functions to allocate, deallocate, and copy device memory, as well as transfer data between host memory and device memory. Device memory can be allocated either as linear memory or as CUDA arrays. CUDA arrays are opaque memory layouts optimized for texture fetching. They are described in Texture and Surface Memory. Linear memory exists on the device in a 32-bit address space for devices of compute capability 1.x and 40-bit address space of devices of higher compute capability, so separately allocated entities can reference one another via pointers, for example, in a binary tree. Linear memory is typically allocated using cudaMalloc() and freed using cudaFree() and data transfer between host memory and device memory are typically done using cudaMemcpy(). In the vector addition code sample of Kernels, the vectors need to be copied from host memory to device memory:
// Device code __global__ void VecAdd(float* A, float* B, float* C, int N) { int i = blockDim.x * blockIdx.x + threadIdx.x; if (i < N) C[i] = A[i] + B[i]; } // Host code int main() { int N = ...; size_t size = N * sizeof(float); // Allocate input vectors h_A and h_B in host memory float* h_A = (float*)malloc(size); float* h_B = (float*)malloc(size); // Initialize input vectors ... // Allocate vectors in device memory float* d_A; cudaMalloc(&d_A, size); float* d_B; cudaMalloc(&d_B, size); float* d_C; cudaMalloc(&d_C, size); // Copy vectors from host memory to device memory cudaMemcpy(d_A, h_A, size, cudaMemcpyHostToDevice); cudaMemcpy(d_B, h_B, size, cudaMemcpyHostToDevice); // Invoke kernel int threadsPerBlock = 256; int blocksPerGrid = (N + threadsPerBlock 1) / threadsPerBlock; VecAdd<<<blocksPerGrid, threadsPerBlock>>>(d_A, d_B, d_C, N);

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|19

Programming Interface

// Copy result from device memory to host memory // h_C contains the result in host memory cudaMemcpy(h_C, d_C, size, cudaMemcpyDeviceToHost); // Free device memory cudaFree(d_A); cudaFree(d_B); cudaFree(d_C); // Free host memory ...

Linear memory can also be allocated through cudaMallocPitch() and cudaMalloc3D(). These functions are recommended for allocations of 2D or 3D arrays as it makes sure that the allocation is appropriately padded to meet the alignment requirements described in Device Memory Accesses, therefore ensuring best performance when accessing the row addresses or performing copies between 2D arrays and other regions of device memory (using the cudaMemcpy2D() and cudaMemcpy3D() functions). The returned pitch (or stride) must be used to access array elements. The following code sample allocates a widthheight 2D array of floating-point values and shows how to loop over the array elements in device code:
// Host code int width = 64, height = 64; float* devPtr; size_t pitch; cudaMallocPitch(&devPtr, &pitch, width * sizeof(float), height); MyKernel<<<100, 512>>>(devPtr, pitch, width, height); // Device code __global__ void MyKernel(float* devPtr, size_t pitch, int width, int height) { for (int r = 0; r < height; ++r) { float* row = (float*)((char*)devPtr + r * pitch); for (int c = 0; c > width; ++c) { float element = row[c]; } } }

The following code sample allocates a widthheightdepth 3D array of floatingpoint values and shows how to loop over the array elements in device code:
// Host code int width = 64, height = 64, depth = 64; cudaExtent extent = make_cudaExtent(width * sizeof(float), height, depth); cudaPitchedPtr devPitchedPtr; cudaMalloc3D(&devPitchedPtr, extent); MyKernel<<<100, 512>>>(devPitchedPtr, width, height, depth); // Device code __global__ void MyKernel(cudaPitchedPtr devPitchedPtr, int width, int height, int depth) { char* devPtr = devPitchedPtr.ptr; size_t pitch = devPitchedPtr.pitch; size_t slicePitch = pitch * height; for (int z = 0; z < depth; ++z) { char* slice = devPtr + z * slicePitch; for (int y = 0; y < height; ++y) {

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|20

Programming Interface

float* row = (float*)(slice + y * pitch); for (int x = 0; x < width; ++x) { float element = row[x]; }

The reference manual lists all the various functions used to copy memory between linear memory allocated with cudaMalloc(), linear memory allocated with cudaMallocPitch() or cudaMalloc3D(), CUDA arrays, and memory allocated for variables declared in global or constant memory space. The following code sample illustrates various ways of accessing global variables via the runtime API:
__constant__ float constData[256]; float data[256]; cudaMemcpyToSymbol(constData, data, sizeof(data)); cudaMemcpyFromSymbol(data, constData, sizeof(data)); __device__ float devData; float value = 3.14f; cudaMemcpyToSymbol(devData, &value, sizeof(float)); __device__ float* devPointer; float* ptr; cudaMalloc(&ptr, 256 * sizeof(float)); cudaMemcpyToSymbol(devPointer, &ptr, sizeof(ptr));

cudaGetSymbolAddress() is used to retrieve the address pointing to the memory allocated for a variable declared in global memory space. The size of the allocated memory is obtained through cudaGetSymbolSize().

3.2.3Shared Memory
As detailed in Variable Type Qualifiers shared memory is allocated using the __shared__ qualifier. Shared memory is expected to be much faster than global memory as mentioned in Thread Hierarchy and detailed in Shared Memory. Any opportunity to replace global memory accesses by shared memory accesses should therefore be exploited as illustrated by the following matrix multiplication example. The following code sample is a straightforward implementation of matrix multiplication that does not take advantage of shared memory. Each thread reads one row of A and one column of B and computes the corresponding element of C as illustrated in Figure 9 Matrix Multiplication without Shared Memory. A is therefore read B.width times from global memory and B is read A.height times.
// Matrices are stored in row-major order: // M(row, col) = *(M.elements + row * M.width + col) typedef struct { int width; int height; float* elements; } Matrix; // Thread block size #define BLOCK_SIZE 16

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|21

Programming Interface

// Forward declaration of the matrix multiplication kernel __global__ void MatMulKernel(const Matrix, const Matrix, Matrix); // Matrix multiplication - Host code // Matrix dimensions are assumed to be multiples of BLOCK_SIZE void MatMul(const Matrix A, const Matrix B, Matrix C) { // Load A and B to device memory Matrix d_A; d_A.width = A.width; d_A.height = A.height; size_t size = A.width * A.height * sizeof(float); cudaMalloc(&d_A.elements, size); cudaMemcpy(d_A.elements, A.elements, size, cudaMemcpyHostToDevice); Matrix d_B; d_B.width = B.width; d_B.height = B.height; size = B.width * B.height * sizeof(float); cudaMalloc(&d_B.elements, size); cudaMemcpy(d_B.elements, B.elements, size, cudaMemcpyHostToDevice); // Allocate C in device memory Matrix d_C; d_C.width = C.width; d_C.height = C.height; size = C.width * C.height * sizeof(float); cudaMalloc(&d_C.elements, size); // Invoke kernel dim3 dimBlock(BLOCK_SIZE, BLOCK_SIZE); dim3 dimGrid(B.width / dimBlock.x, A.height / dimBlock.y); MatMulKernel<<<dimGrid, dimBlock>>>(d_A, d_B, d_C); // Read C from device memory cudaMemcpy(C.elements, Cd.elements, size, cudaMemcpyDeviceToHost); // Free device memory cudaFree(d_A.elements); cudaFree(d_B.elements); cudaFree(d_C.elements);

// Matrix multiplication kernel called by MatMul() __global__ void MatMulKernel(Matrix A, Matrix B, Matrix C) { // Each thread computes one element of C // by accumulating results into Cvalue float Cvalue = 0; int row = blockIdx.y * blockDim.y + threadIdx.y; int col = blockIdx.x * blockDim.x + threadIdx.x; for (int e = 0; e < A.width; ++e) Cvalue += A.elements[row * A.width + e] * B.elements[e * B.width + col]; C.elements[row * C.width + col] = Cvalue; }

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|22

Programming Interface

Figure9Matrix Multiplication without Shared Memory

The following code sample is an implementation of matrix multiplication that does take advantage of shared memory. In this implementation, each thread block is responsible for computing one square sub-matrix Csub of C and each thread within the block is responsible for computing one element of Csub. As illustrated in Figure 10 Matrix Multiplication with Shared Memory, Csub is equal to the product of two rectangular matrices: the sub-matrix of A of dimension (A.width, block_size) that has the same row indices as Csub, and the sub-matrix of B of dimension (block_size, A.width )that has the same column indices as Csub. In order to fit into the devices resources, these two rectangular matrices are divided into as many square matrices of dimension block_size as necessary and Csub is computed as the sum of the products of these square matrices. Each of these products is performed by first loading the two corresponding square

www.nvidia.com

PG-02829-001_v5.0|26

Programming Interface

Figure10Matrix Multiplication with Shared Memory

3.2.4Page-Locked Host Memory

The runtime provides functions to allow the use of page-locked (also known as pinned) host memory (as opposed to regular pageable host memory allocated by malloc()): cudaHostAlloc() and cudaFreeHost() allocate and free page-locked host memory; cudaHostRegister() page-locks a range of memory allocated by malloc() (see reference manual for limitations). Using page-locked host memory has several benefits: Copies between page-locked host memory and device memory can be performed concurrently with kernel execution for some devices as mentioned in Asynchronous Concurrent Execution.

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|27

Programming Interface

On some devices, page-locked host memory can be mapped into the address space of the device, eliminating the need to copy it to or from device memory as detailed in Mapped Memory. On systems with a front-side bus, bandwidth between host memory and device memory is higher if host memory is allocated as page-locked and even higher if in addition it is allocated as write-combining as described in Write-Combining Memory. Page-locked host memory is a scarce resource however, so allocations in page-locked memory will start failing long before allocations in pageable memory. In addition, by reducing the amount of physical memory available to the operating system for paging, consuming too much page-locked memory reduces overall system performance. The simple zero-copy SDK sample comes with a detailed document on the page-locked memory APIs.

3.2.4.1Portable Memory
A block of page-locked memory can be used in conjunction with any device in the system (see Multi-Device System for more details on multi-device systems), but by default, the benefits of using page-locked memory described above are only available in conjunction with the device that was current when the block was allocated (and with all devices sharing the same unified address space, if any, as described in Unified Virtual Address Space). To make these advantages available to all devices, the block needs to be allocated by passing the flag cudaHostAllocPortable to cudaHostAlloc() or page-locked by passing the flag cudaHostRegisterPortable to cudaHostRegister().

3.2.4.2Write-Combining Memory
By default page-locked host memory is allocated as cacheable. It can optionally be allocated as write-combining instead by passing flag cudaHostAllocWriteCombined to cudaHostAlloc(). Write-combining memory frees up the hosts L1 and L2 cache resources, making more cache available to the rest of the application. In addition, writecombining memory is not snooped during transfers across the PCI Express bus, which can improve transfer performance by up to 40%. Reading from write-combining memory from the host is prohibitively slow, so writecombining memory should in general be used for memory that the host only writes to.

3.2.4.3Mapped Memory
On devices of compute capability greater than 1.0, a block of page-locked host memory can also be mapped into the address space of the device by passing flag cudaHostAllocMapped to cudaHostAlloc() or by passing flag cudaHostRegisterMapped to cudaHostRegister(). Such a block has therefore in general two addresses: one in host memory that is returned by cudaHostAlloc() or malloc(), and one in device memory that can be retrieved using cudaHostGetDevicePointer() and then used to access the block from within a kernel. The only exception is for pointers allocated with cudaHostAlloc() and when a unified address space is used for the host and the device as mentioned in Unified Virtual Address Space.

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|28

Programming Interface

Accessing host memory directly from within a kernel has several advantages: There is no need to allocate a block in device memory and copy data between this block and the block in host memory; data transfers are implicitly performed as needed by the kernel; There is no need to use streams (see Concurrent Data Transfers) to overlap data transfers with kernel execution; the kernel-originated data transfers automatically overlap with kernel execution. Since mapped page-locked memory is shared between host and device however, the application must synchronize memory accesses using streams or events (see Asynchronous Concurrent Execution) to avoid any potential read-after-write, writeafter-read, or write-after-write hazards. To be able to retrieve the device pointer to any mapped page-locked memory, pagelocked memory mapping must be enabled by calling cudaSetDeviceFlags() with the cudaDeviceMapHost flag before any other CUDA call is performed. Otherwise, cudaHostGetDevicePointer() will return an error. cudaHostGetDevicePointer() also returns an error if the device does not support mapped page-locked host memory. Applications may query this capability by checking the canMapHostMemory device property (see Device Enumeration), which is equal to 1 for devices that support mapped page-locked host memory. Note that atomic functions (see Atomic Functions) operating on mapped page-locked memory are not atomic from the point of view of the host or other devices.

3.2.5Asynchronous Concurrent Execution

3.2.5.1Concurrent Execution between Host and Device
In order to facilitate concurrent execution between host and device, some function calls are asynchronous: Control is returned to the host thread before the device has completed the requested task. These are: Kernel launches; Memory copies between two addresses to the same device memory; Memory copies from host to device of a memory block of 64 KB or less; Memory copies performed by functions that are suffixed with Async; Memory set function calls.

Programmers can globally disable asynchronous kernel launches for all CUDA applications running on a system by setting the CUDA_LAUNCH_BLOCKING environment variable to 1. This feature is provided for debugging purposes only and should never be used as a way to make production software run reliably. Kernel launches are synchronous in the following cases: The application is run via a debugger or memory checker (cuda-gdb, cudamemcheck, Nsight) on a device of compute capability 1.x; Hardware counters are collected via a profiler (Nsight, Visual Profiler).

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|29

Programming Interface

3.2.5.2Overlap of Data Transfer and Kernel Execution

Some devices of compute capability 1.1 and higher can perform copies between page-locked host memory and device memory concurrently with kernel execution. Applications may query this capability by checking the asyncEngineCount device property (see Device Enumeration), which is greater than zero for devices that support it. For devices of compute capability 1.x, this capability is only supported for memory copies that do not involve CUDA arrays or 2D arrays allocated through cudaMallocPitch() (see Device Memory).

3.2.5.3Concurrent Kernel Execution

Some devices of compute capability 2.x and higher can execute multiple kernels concurrently. Applications may query this capability by checking the concurrentKernels device property (see Device Enumeration), which is equal to 1 for devices that support it. The maximum number of kernel launches that a device can execute concurrently is 32 on devices of compute capability 3.5 and 16 on devices of lower compute capabiliy. A kernel from one CUDA context cannot execute concurrently with a kernel from another CUDA context. Kernels that use many textures or a large amount of local memory are less likely to execute concurrently with other kernels.

3.2.5.4Concurrent Data Transfers

Some devices of compute capability 2.x and higher can perform a copy from page-locked host memory to device memory concurrently with a copy from device memory to pagelocked host memory. Applications may query this capability by checking theasyncEngineCount device property (see Device Enumeration), which is equal to 2 for devices that support it.

3.2.5.5Streams
Applications manage concurrency through streams. A stream is a sequence of commands (possibly issued by different host threads) that execute in order. Different streams, on the other hand, may execute their commands out of order with respect to one another or concurrently; this behavior is not guaranteed and should therefore not be relied upon for correctness (e.g., inter-kernel communication is undefined).

3.2.5.5.1Creation and Destruction

A stream is defined by creating a stream object and specifying it as the stream parameter to a sequence of kernel launches and host <-> device memory copies. The following code sample creates two streams and allocates an array hostPtr of float in pagelocked memory.
cudaStream_t stream[2]; for (int i = 0; i < 2; ++i) cudaStreamCreate(&stream[i]);

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|30

Programming Interface

PG-02829-001_v5.0|33

Programming Interface

Blocking callback must not make CUDA API calls (directly or indirectly), as it might end up waiting on itself if it makes such a call leading to a deadlock.

3.2.5.6Events
The runtime also provides a way to closely monitor the devices progress, as well as perform accurate timing, by letting the application asynchronously record events at any point in the program and query when these events are completed. An event has completed when all tasks or optionally, all commands in a given stream preceding the event have completed. Events in stream zero are completed after all preceding tasks and commands in all streams are completed.

3.2.5.6.1Creation and Destruction

The following code sample creates two events:
cudaEvent_t start, stop; cudaEventCreate(&start); cudaEventCreate(&stop);

They are destroyed this way:

cudaEventDestroy(start); cudaEventDestroy(stop);

3.2.5.6.2Elapsed Time
The events created in Creation and Destruction can be used to time the code sample of Creation and Destruction the following way:
cudaEventRecord(start, 0); for (int i = 0; i < 2; ++i) { cudaMemcpyAsync(inputDev + i * size, inputHost + i * size, size, cudaMemcpyHostToDevice, stream[i]); MyKernel<<<100, 512, 0, stream[i]>>> (outputDev + i * size, inputDev + i * size, size); cudaMemcpyAsync(outputHost + i * size, outputDev + i * size, size, cudaMemcpyDeviceToHost, stream[i]); } cudaEventRecord(stop, 0); cudaEventSynchronize(stop); float elapsedTime; cudaEventElapsedTime(&elapsedTime, start, stop);

3.2.5.7Synchronous Calls
When a synchronous function is called, control is not returned to the host thread before the device has completed the requested task. Whether the host thread will then yield, block, or spin can be specified by calling cudaSetDeviceFlags()with some specific flags (see reference manual for details) before any other CUDA call is performed by the host thread.

3.2.6Multi-Device System

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|34

Programming Interface

3.2.6.1Device Enumeration
A host system can have multiple devices. The following code sample shows how to enumerate these devices, query their properties, and determine the number of CUDAenabled devices.
int deviceCount; cudaGetDeviceCount(&deviceCount); int device; for (device = 0; device < deviceCount; ++device) { cudaDeviceProp deviceProp; cudaGetDeviceProperties(&deviceProp, device); printf("Device %d has compute capability %d.%d.\n", device, deviceProp.major, deviceProp.minor); }

3.2.6.2Device Selection
A host thread can set the device it operates on at any time by calling cudaSetDevice(). Device memory allocations and kernel launches are made on the currently set device; streams and events are created in association with the currently set device. If no call to cudaSetDevice() is made, the current device is device 0. The following code sample illustrates how setting the current device affects memory allocation and kernel execution.
size_t size = 1024 * sizeof(float); cudaSetDevice(0); // Set device 0 as current float* p0; cudaMalloc(&p0, size); // Allocate memory on device 0 MyKernel<<<1000, 128>>>(p0); // Launch kernel on device 0 cudaSetDevice(1); // Set device 1 as current float* p1; cudaMalloc(&p1, size); // Allocate memory on device 1 MyKernel<<<1000, 128>>>(p1); // Launch kernel on device 1

3.2.6.3Stream and Event Behavior

A kernel launch or memory copy will fail if it is issued to a stream that is not associated to the current device as illustrated in the following code sample.
cudaSetDevice(0); // Set device 0 as current cudaStream_t s0; cudaStreamCreate(&s0); // Create stream s0 on device 0 MyKernel<<<100, 64, 0, s0>>>(); // Launch kernel on device 0 in s0 cudaSetDevice(1); // Set device 1 as current cudaStream_t s1; cudaStreamCreate(&s1); // Create stream s1 on device 1 MyKernel<<<100, 64, 0, s1>>>(); // Launch kernel on device 1 in s1 // This kernel launch will fail: MyKernel<<<100, 64, 0, s0>>>(); // Launch kernel on device 1 in s0

cudaEventRecord() will fail if the input event and input stream are associated to different devices. cudaEventElapsedTime() will fail if the two input events are associated to different devices. cudaEventSynchronize() and cudaEventQuery() will succeed even if the input event is associated to a device that is different from the current device.

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|35

Programming Interface

cudaStreamWaitEvent() will succeed even if the input stream and input event are associated to different devices. cudaStreamWaitEvent() can therefore be used to synchronize multiple devices with each other. Each device has its own default stream (see Default Stream), so commands issued to the default stream of a device may execute out of order or concurrently with respect to commands issued to the default stream of any other device.

3.2.6.4Peer-to-Peer Memory Access

When the application is run as a 64-bit process on Windows Vista/7 in TCC mode (see Tesla Compute Cluster Mode for Windows), on Windows XP, or on Linux, devices of compute capability 2.0 and higher from the Tesla series may address each others memory (i.e., a kernel executing on one device can dereference a pointer to the memory of the other device). This peer-to-peer memory access feature is supported between two devices if cudaDeviceCanAccessPeer() returns true for these two devices. Peer-to-peer memory access must be enabled between two devices by calling cudaDeviceEnablePeerAccess() as illustrated in the following code sample. A unified address space is used for both devices (see Unified Virtual Address Space), so the same pointer can be used to address memory from both devices as shown in the code sample below.
cudaSetDevice(0); float* p0; size_t size = 1024 * sizeof(float); cudaMalloc(&p0, size); MyKernel<<<1000, 128>>>(p0); cudaSetDevice(1); cudaDeviceEnablePeerAccess(0, 0); // Set device 0 as current // // // // // Allocate memory on device 0 Launch kernel on device 0 Set device 1 as current Enable peer-to-peer access with device 0

// Launch kernel on device 1 // This kernel launch can access memory on device 0 at address p0 MyKernel<<<1000, 128>>>(p0);

3.2.6.5Peer-to-Peer Memory Copy

Memory copies can be performed between the memories of two different devices. When a unified address space is used for both devices (see Unified Virtual Address Space), this is done using the regular memory copy functions mentioned in Device Memory. Otherwise, this is done using cudaMemcpyPeer(), cudaMemcpyPeerAsync(), cudaMemcpy3DPeer(), or cudaMemcpy3DPeerAsync() as illustrated in the following code sample.
cudaSetDevice(0); float* p0; size_t size = 1024 * sizeof(float); cudaMalloc(&p0, size); cudaSetDevice(1); float* p1; cudaMalloc(&p1, size); cudaSetDevice(0); MyKernel<<<1000, 128>>>(p0); cudaSetDevice(1); cudaMemcpyPeer(p1, 1, p0, 0, size); // Set device 0 as current // Allocate memory on device 0 // Set device 1 as current // // // // // Allocate memory on device 1 Set device 0 as current Launch kernel on device 0 Set device 1 as current Copy p0 to p1

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|36

A texture object is created using cudaCreateTextureObject() from a resource description of type struct cudaResourceDesc, which specifies the texture, and from a texture description defined as such:
struct cudaTextureDesc { enum cudaTextureAddressMode enum cudaTextureFilterMode enum cudaTextureReadMode int int unsigned int enum cudaTextureFilterMode float float float }; addressMode[3]; filterMode; readMode; sRGB; normalizedCoords; maxAnisotropy; mipmapFilterMode; mipmapLevelBias; minMipmapLevelClamp; maxMipmapLevelClamp;

addressMode specifies the addressing mode; filterMode specifies the filter mode; readMode specifies the read mode; normalizedCoords specifies whether texture coordinates are normalized or not;

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|40

Programming Interface

See reference manual for sRGB, maxAnisotropy, mipmapFilterMode, mipmapLevelBias, minMipmapLevelClamp, and maxMipmapLevelClamp. The following code sample applies some simple transformation kernel to a texture.
// Simple transformation kernel __global__ void transformKernel(float* output, cudaTextureObject_t texObj, int width, int height, float theta) { // Calculate normalized texture coordinates unsigned int x = blockIdx.x * blockDim.x + threadIdx.x; unsigned int y = blockIdx.y * blockDim.y + threadIdx.y; float u = x / (float)width; float v = y / (float)height; // Transform coordinates u -= 0.5f; v -= 0.5f; float tu = u * cosf(theta) v * sinf(theta) + 0.5f; float tv = v * cosf(theta) + u * sinf(theta) + 0.5f; // Read from texture and write to global memory output[y * width + x] = tex2D<float>(texObj, tu, tv);

// Host code int main() { // Allocate CUDA array in device memory cudaChannelFormatDesc channelDesc = cudaCreateChannelDesc(32, 0, 0, 0, cudaChannelFormatKindFloat); cudaArray* cuArray; cudaMallocArray(&cuArray, &channelDesc, width, height); // Copy to device memory some data located at address h_data // in host memory cudaMemcpyToArray(cuArray, 0, 0, h_data, size, cudaMemcpyHostToDevice); // Specify texture struct cudaResourceDesc resDesc; memset(&resDesc, 0, sizeof(resDesc)); resDesc.resType = cudaResourceTypeArray; resDesc.res.array.array = cuArray; // Specify texture object parameters struct cudaTextureDesc texDesc; memset(&texDesc, 0, sizeof(texDesc)); texDesc.addressMode[0] = cudaAddressModeWrap; texDesc.addressMode[1] = cudaAddressModeWrap; texDesc.filterMode = cudaFilterModeLinear; texDesc.readMode = cudaReadModeElementType; texDesc.normalizedCoords = 1; // Create texture object cudaTextureObject_t texObj = 0; cudaCreateTextureObject(&texObj, &resDesc, &texDesc, NULL); // Allocate result of transformation in device memory float* output; cudaMalloc(&output, width * height * sizeof(float)); // Invoke kernel

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|41

Programming Interface

dim3 dimBlock(16, 16); dim3 dimGrid((width + dimBlock.x 1) / dimBlock.x, (height + dimBlock.y 1) / dimBlock.y); transformKernel<<<dimGrid, dimBlock>>>(output, texObj, width, height, angle); // Destroy texture object cudaDestroyTextureObject(texObject); // Free device memory cudaFreeArray(cuArray); cudaFree(output); } return 0;

3.2.10.1.2Texture Reference API

Some of the attributes of a texture reference are immutable and must be known at compile time; they are specified when declaring the texture reference. A texture reference is declared at file scope as a variable of type texture:
texture<DataType, Type, ReadMode> texRef;

where: DataType specifies the type of the texel; Type specifies the type of the texture reference and is equal to cudaTextureType1D, cudaTextureType2D, or cudaTextureType3D, for a one-dimensional, two-dimensional, or three-dimensional texture, respectively, or cudaTextureType1DLayered or cudaTextureType2DLayered for a onedimensional or two-dimensional layered texture respectively; Type is an optional argument which defaults to cudaTextureType1D; ReadMode specifies the read mode; it is an optional argument which defaults to cudaReadModeElementType. A texture reference can only be declared as a static global variable and cannot be passed as an argument to a function. The other attributes of a texture reference are mutable and can be changed at runtime through the host runtime. As explained in the reference manual, the runtime API has a low-level C-style interface and a high-level C++-style interface. The texture type is defined in the high-level API as a structure publicly derived from the textureReference type defined in the low-level API as such:
struct textureReference { int enum cudaTextureFilterMode enum cudaTextureAddressMode struct cudaChannelFormatDesc int unsigned int enum cudaTextureFilterMode float float float } normalized; filterMode; addressMode[3]; channelDesc; sRGB; maxAnisotropy; mipmapFilterMode; mipmapLevelBias; minMipmapLevelClamp; maxMipmapLevelClamp;

normalized specifies whether texture coordinates are normalized or not; filterMode specifies the filtering mode;
www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|42

Programming Interface

addressMode specifies the addressing mode; channelDesc describes the format of the texel; it must match the DataType argument of the texture reference declaration; channelDesc is of the following type:
struct cudaChannelFormatDesc { int x, y, z, w; enum cudaChannelFormatKind f; };

where x, y, z, and w are equal to the number of bits of each component of the returned value and f is: cudaChannelFormatKindSigned if these components are of signed integer type, cudaChannelFormatKindUnsigned if they are of unsigned integer type, cudaChannelFormatKindFloat if they are of floating point type. See reference manual for sRGB, maxAnisotropy, mipmapFilterMode, mipmapLevelBias, minMipmapLevelClamp, and maxMipmapLevelClamp. normalized, addressMode, and filterMode may be directly modified in host code. Before a kernel can use a texture reference to read from texture memory, the texture reference must be bound to a texture using cudaBindTexture() or cudaBindTexture2D() for linear memory, or cudaBindTextureToArray() for CUDA arrays. cudaUnbindTexture() is used to unbind a texture reference. It is recommended to allocate two-dimensional textures in linear memory using cudaMallocPitch() and use the pitch returned by cudaMallocPitch() as input parameter to cudaBindTexture2D(). The following code samples bind a texture reference to linear memory pointed to by devPtr: Using the low-level API:
texture<float, cudaTextureType2D, cudaReadModeElementType> texRef; textureReference* texRefPtr; cudaGetTextureReference(&texRefPtr, texRef); cudaChannelFormatDesc channelDesc = cudaCreateChannelDesc<float>(); size_t offset; cudaBindTexture2D(&offset, texRefPtr, devPtr, &channelDesc, width, height, pitch); texture<float, cudaTextureType2D, cudaReadModeElementType> texRef; cudaChannelFormatDesc channelDesc = cudaCreateChannelDesc<float>(); size_t offset; cudaBindTexture2D(&offset, texRef, devPtr, channelDesc, width, height, pitch);

Using the high-level API:

The following code samples bind a texture reference to a CUDA array cuArray: Using the low-level API:
texture<float, cudaTextureType2D, cudaReadModeElementType> texRef; textureReference* texRefPtr; cudaGetTextureReference(&texRefPtr, texRef); cudaChannelFormatDesc channelDesc;

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|43

Programming Interface

cudaGetChannelDesc(&channelDesc, cuArray); cudaBindTextureToArray(texRef, cuArray, &channelDesc); ? Using the high-level API: texture<float, cudaTextureType2D, cudaReadModeElementType> texRef; cudaBindTextureToArray(texRef, cuArray);

The format specified when binding a texture to a texture reference must match the parameters specified when declaring the texture reference; otherwise, the results of texture fetches are undefined. There is a limit to the number of textures that can be bound to a kernel as specified in Table 10 Technical Specifications per Compute Capability. The following code sample applies some simple transformation kernel to a texture.
// 2D float texture texture<float, cudaTextureType2D, cudaReadModeElementType> texRef; // Simple transformation kernel __global__ void transformKernel(float* output, int width, int height, float theta) { // Calculate normalized texture coordinates unsigned int x = blockIdx.x * blockDim.x + threadIdx.x; unsigned int y = blockIdx.y * blockDim.y + threadIdx.y; float u = x / (float)width; float v = y / (float)height; // Transform coordinates u -= 0.5f; v -= 0.5f; float tu = u * cosf(theta) - v * sinf(theta) + 0.5f; float tv = v * cosf(theta) + u * sinf(theta) + 0.5f; // Read from texture and write to global memory output[y * width + x] = tex2D(texRef, tu, tv);

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|44

Programming Interface

float* output; cudaMalloc(&output, width * height * sizeof(float)); // Invoke kernel dim3 dimBlock(16, 16); dim3 dimGrid((width + dimBlock.x - 1) / dimBlock.x, (height + dimBlock.y - 1) / dimBlock.y); transformKernel<<<dimGrid, dimBlock>>>(output, width, height, angle); // Free device memory cudaFreeArray(cuArray); cudaFree(output); } return 0;

3.2.10.1.316-Bit Floating-Point Textures

The 16-bit floating-point or half format supported by CUDA arrays is the same as the IEEE 754-2008 binary2 format. CUDA C does not support a matching data type, but provides intrinsic functions to convert to and from the 32-bit floating-point format via the unsigned short type: __float2half_rn(float) and __half2float(unsigned short). These functions are only supported in device code. Equivalent functions for the host code can be found in the OpenEXR library, for example. 16-bit floating-point components are promoted to 32 bit float during texture fetching before any filtering is performed. A channel description for the 16-bit floating-point format can be created by calling one of the cudaCreateChannelDescHalf*() functions.

3.2.10.1.4Layered Textures
A one-dimensional or two-dimensional layered texture (also know as texture array in Direct3D and array texture in OpenGL) is a texture made up of a sequence of layers, all of which are regular textures of same dimensionality, size, and data type. A one-dimensional layered texture is addressed using an integer index and a floatingpoint texture coordinate; the index denotes a layer within the sequence and the coordinate addresses a texel within that layer. A two-dimensional layered texture is addressed using an integer index and two floating-point texture coordinates; the index denotes a layer within the sequence and the coordinates address a texel within that layer. A layered texture can only be a CUDA array by calling cudaMalloc3DArray() with the cudaArrayLayered flag (and a height of zero for one-dimensional layered texture). Layered textures are fetched using the device functions described in tex1Dlayered() and tex2Dlayered(). Texture filtering (see Texture Fetching) is done only within a layer, not across layers. Layered textures are only supported on devices of compute capability 2.0 and higher.

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|45

Programming Interface

3.2.10.1.5Cubemap Textures
A cubemap texture is a special type of two-dimensional layered texture that has six layers representing the faces of a cube: The width of a layer is equal to its height. The cubemap is addressed using three texture coordinates x, y, and z that are interpreted as a direction vector emanating from the center of the cube and pointing to one face of the cube and a texel within the layer corresponding to that face. More specifically, the face is selected by the coordinate with largest magnitude m and the corresponding layer is addressed using coordinates (s/m+1)/2 and (t/m+1)/2 where s and t are defined in Table 1 Cubemap Fetch.

Table1Cubemap Fetch
face
|x| > |y| and |x| > |z| x>0 x<0 y>0 y<0 z>0 z<0 0 1 2 3 4 5

m
x -x y -y z -z

s
-z z x x x -x

t
-y -y z -z -y -y

|y| > |x| and |y| > |z|

|z| > |x| and |z| > |y|

A layered texture can only be a CUDA array by calling cudaMalloc3DArray() with the cudaArrayCubemap flag. Cubemap textures are fetched using the device function described in texCubemap(). Cubemap textures are only supported on devices of compute capability 2.0 and higher.

3.2.10.1.6Cubemap Layered Textures

A cubemap layered texture is a layered texture whose layers are cubemaps of same dimension. A cubemap layered texture is addressed using an integer index and three floatingpoint texture coordinates; the index denotes a cubemap within the sequence and the coordinates address a texel within that cubemap. A layered texture can only be a CUDA array by calling cudaMalloc3DArray() with the cudaArrayLayered and cudaArrayCubemap flags. Cubemap layered textures are fetched using the device function described in texCubemapLayered(). Texture filtering (see Texture Fetching) is done only within a layer, not across layers. Cubemap layered textures are only supported on devices of compute capability 2.0 and higher.

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|46

Programming Interface

3.2.10.1.7Texture Gather
Texture gather is a special texture fetch that is available for two-dimensional textures only. It is performed by the tex2Dgather() function, which has the same parameters as tex2D(), plus an additional comp parameter equal to 0, 1, 2, or 3 (see tex2Dgather()). It returns four 32-bit numbers that correspond to the value of the component comp of each of the four texels that would have been used for bilinear filtering during a regular texture fetch. For example, if these texels are of values (253, 20, 31, 255), (250, 25, 29, 254), (249, 16, 37, 253), (251, 22, 30, 250), and comp is 2, tex2Dgather() returns (31, 29, 37, 30). Texture gather is only supported for CUDA arrays created with the cudaArrayTextureGather flag and of width and height less than the maximum specified in Table 10 Technical Specifications per Compute Capability for texture gather, which is smaller than for regular texture fetch. Texture gather is only supported on devices of compute capability 2.0 and higher.

3.2.10.2Surface Memory
For devices of compute capability 2.0 and higher, a CUDA array (described in Cubemap Surfaces), created with the cudaArraySurfaceLoadStore flag, can be read and written via a surface object or surface reference using the functions described in Surface Functions. Table 10 Technical Specifications per Compute Capability lists the maximum surface width, height, and depth depending on the compute capability of the device.

3.2.10.2.1Surface Object API

A surface object is created using cudaCreateSurfaceObject() from a resource description of type struct cudaResourceDesc. The following code sample applies some simple transformation kernel to a texture.
// Simple copy kernel __global__ void copyKernel(cudaSurfaceObject inputSurfObj, cudaSurfaceObject outputSurfObj, int width, int height) { // Calculate surface coordinates unsigned int x = blockIdx.x * blockDim.x + threadIdx.x; unsigned int y = blockIdx.y * blockDim.y + threadIdx.y; if (x < width && y < height) { uchar4 data; // Read from input surface surf2Dread(&data, inputSurfObj, x * 4, y); // Write to output surface surf2Dwrite(data, outputSurfObj, x * 4, y); } } // Host code int main() { // Allocate CUDA arrays in device memory cudaChannelFormatDesc channelDesc = cudaCreateChannelDesc(8, 8, 8, 8,

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|47

Programming Interface

cudaChannelFormatKindUnsigned); cudaArray* cuInputArray; cudaMallocArray(&cuInputArray, &channelDesc, width, height, cudaArraySurfaceLoadStore); cudaArray* cuOutputArray; cudaMallocArray(&cuOutputArray, &channelDesc, width, height, cudaArraySurfaceLoadStore); // Copy to device memory some data located at address h_data // in host memory cudaMemcpyToArray(cuInputArray, 0, 0, h_data, size, cudaMemcpyHostToDevice); // Specify surface struct cudaResourceDesc resDesc; memset(&resDesc, 0, sizeof(resDesc)); resDesc.resType = cudaResourceTypeArray; // Create the surface objects resDesc.res.array.array = cuInputArray; cudaSurfaceObject inputSurfObj = 0; cudaCreateSurfaceObject(&inputSurfObj, &resDesc); resDesc.res.array.array = cuOutputArray; cudaSurfaceObject outputSurfObj = 0; cudaCreateSurfaceObject(&outputSurfObj, &resDesc); // Invoke kernel dim3 dimBlock(16, 16); dim3 dimGrid((width + dimBlock.x - 1) / dimBlock.x, (height + dimBlock.y - 1) / dimBlock.y); copyKernel<<<dimGrid, dimBlock>>>(inputSurfObj, outputSurfObj, width, height); // Destroy surface objects cudaDestroySurfaceObject(inputSurfObj); cudaDestroySurfaceObject(outputSurfObj); // Free device memory cudaFreeArray(cuInputArray); cudaFreeArray(cuOutputArray); } return 0;

3.2.10.2.2Surface Reference API

A surface reference is declared at file scope as a variable of type surface:
surface<void, Type> surfRef;

where Type specifies the type of the surface reference and is equal to cudaSurfaceType1D, cudaSurfaceType2D, cudaSurfaceType3D, cudaSurfaceTypeCubemap, cudaSurfaceType1DLayered, cudaSurfaceType2DLayered, or cudaSurfaceTypeCubemapLayered; Type is an optional argument which defaults to cudaSurfaceType1D. A surface reference can only be declared as a static global variable and cannot be passed as an argument to a function. Before a kernel can use a surface reference to access a CUDA array, the surface reference must be bound to the CUDA array using cudaBindSurfaceToArray(). The following code samples bind a surface reference to a CUDA array cuArray:

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|48

Programming Interface

Using the low-level API:

surface<void, cudaSurfaceType2D> surfRef; surfaceReference* surfRefPtr; cudaGetSurfaceReference(&surfRefPtr, "surfRef"); cudaChannelFormatDesc channelDesc; cudaGetChannelDesc(&channelDesc, cuArray); cudaBindSurfaceToArray(surfRef, cuArray, &channelDesc); surface<void, cudaSurfaceType2D> surfRef; cudaBindSurfaceToArray(surfRef, cuArray);

Using the high-level API:

A CUDA array must be read and written using surface functions of matching dimensionality and type and via a surface reference of matching dimensionality; otherwise, the results of reading and writing the CUDA array are undefined. Unlike texture memory, surface memory uses byte addressing. This means that the xcoordinate used to access a texture element via texture functions needs to be multiplied by the byte size of the element to access the same element via a surface function. For example, the element at texture coordinate x of a one-dimensional floating-point CUDA array bound to a texture reference texRef and a surface reference surfRef is read using tex1d(texRef, x) via texRef, but surf1Dread(surfRef, 4*x) via surfRef. Similarly, the element at texture coordinate x and y of a twodimensional floating-point CUDA array bound to a texture reference texRef and a surface reference surfRef is accessed using tex2d(texRef, x, y) via texRef, but surf2Dread(surfRef, 4*x, y) via surfRef (the byte offset of the y-coordinate is internally calculated from the underlying line pitch of the CUDA array). The following code sample applies some simple transformation kernel to a texture.
// 2D surfaces surface<void, 2> inputSurfRef; surface<void, 2> outputSurfRef; // Simple copy kernel __global__ void copyKernel(int width, int height) { // Calculate surface coordinates unsigned int x = blockIdx.x * blockDim.x + threadIdx.x; unsigned int y = blockIdx.y * blockDim.y + threadIdx.y; if (x < width && y < height) { uchar4 data; // Read from input surface surf2Dread(&data, inputSurfRef, x * 4, y); // Write to output surface surf2Dwrite(data, outputSurfRef, x * 4, y); } } // Host code int main() { // Allocate CUDA arrays in device memory cudaChannelFormatDesc channelDesc = cudaCreateChannelDesc(8, 8, 8, 8, cudaChannelFormatKindUnsigned); cudaArray* cuInputArray; cudaMallocArray(&cuInputArray, &channelDesc, width, height, cudaArraySurfaceLoadStore); cudaArray* cuOutputArray; cudaMallocArray(&cuOutputArray, &channelDesc, width, height, cudaArraySurfaceLoadStore);

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|49

Programming Interface

// Copy to device memory some data located at address h_data // in host memory cudaMemcpyToArray(cuInputArray, 0, 0, h_data, size, cudaMemcpyHostToDevice); // Bind the arrays to the surface references cudaBindSurfaceToArray(inputSurfRef, cuInputArray); cudaBindSurfaceToArray(outputSurfRef, cuOutputArray); // Invoke kernel dim3 dimBlock(16, 16); dim3 dimGrid((width + dimBlock.x - 1) / dimBlock.x, (height + dimBlock.y - 1) / dimBlock.y); copyKernel<<<dimGrid, dimBlock>>>(width, height); // Free device memory cudaFreeArray(cuInputArray); cudaFreeArray(cuOutputArray); } return 0;

3.2.10.2.3Cubemap Surfaces
Cubemap surfaces are accessed usingsurfCubemapread() and surfCubemapwrite() (surfCubemapread() and surfCubemapwrite()) as a twodimensional layered surface, i.e., using an integer index denoting a face and two floating-point texture coordinates addressing a texel within the layer corresponding to this face. Faces are ordered as indicated in Table 1 Cubemap Fetch.

3.2.10.2.4Cubemap Layered Surfaces

Cubemap layered surfaces are accessed using surfCubemapLayeredread() and surfCubemapLayeredwrite() (surfCubemapLayeredread() and surfCubemapLayeredwrite()) as a two-dimensional layered surface, i.e., using an integer index denoting a face of one of the cubemaps and two floating-point texture coordinates addressing a texel within the layer corresponding to this face. Faces are ordered as indicated in Table 1 Cubemap Fetch, so index ((2 * 6) + 3), for example, accesses the fourth face of the third cubemap.

3.2.10.3CUDA Arrays
CUDA arrays are opaque memory layouts optimized for texture fetching. They are one dimensional, two dimensional, or three-dimensional and composed of elements, each of which has 1, 2 or 4 components that may be signed or unsigned 8 , 16 or 32 bit integers, 16 bit floats, or 32 bit floats. CUDA arrays are only accessible by kernels through texture fetching as described in Texture Memory or surface reading and writing as described in Surface Memory .

3.2.10.4Read/Write Coherency
The texture and surface memory is cached (see Device Memory Accesses) and within the same kernel call, the cache is not kept coherent with respect to global memory writes and surface memory writes, so any texture fetch or surface read to an address

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|50

Programming Interface

that has been written to via a global write or a surface write in the same kernel call returns undefined data. In other words, a thread can safely read some texture or surface memory location only if this memory location has been updated by a previous kernel call or memory copy, but not if it has been previously updated by the same thread or another thread from the same kernel call.

3.2.11Graphics Interoperability
Some resources from OpenGL and Direct3D may be mapped into the address space of CUDA, either to enable CUDA to read data written by OpenGL or Direct3D, or to enable CUDA to write data for consumption by OpenGL or Direct3D. A resource must be registered to CUDA before it can be mapped using the functions mentioned in OpenGL Interoperability and Direct3D Interoperability. These functions return a pointer to a CUDA graphics resource of type struct cudaGraphicsResource. Registering a resource is potentially high-overhead and therefore typically called only once per resource. A CUDA graphics resource is unregistered using cudaGraphicsUnregisterResource(). Once a resource is registered to CUDA, it can be mapped and unmapped as many times as necessary using cudaGraphicsMapResources() and cudaGraphicsUnmapResources(). cudaGraphicsResourceSetMapFlags() can be called to specify usage hints (write-only, read-only) that the CUDA driver can use to optimize resource management. A mapped resource can be read from or written to by kernels using the device memory address returned by cudaGraphicsResourceGetMappedPointer() for buffers and cudaGraphicsSubResourceGetMappedArray() for CUDA arrays. Accessing a resource through OpenGL or Direct3D while it is mapped to CUDA produces undefined results. OpenGL Interoperability and Direct3D Interoperability give specifics for each graphics API and some code samples. SLI Interoperability gives specifics for when the system is in SLI mode.

3.2.11.1OpenGL Interoperability
Interoperability with OpenGL requires that the CUDA device be specified by cudaGLSetGLDevice() before any other runtime calls. Note that cudaSetDevice() and cudaGLSetGLDevice() are mutually exclusive. The OpenGL resources that may be mapped into the address space of CUDA are OpenGL buffer, texture, and renderbuffer objects. A buffer object is registered using cudaGraphicsGLRegisterBuffer(). In CUDA, it appears as a device pointer and can therefore be read and written by kernels or via cudaMemcpy() calls. A texture or renderbuffer object is registered using cudaGraphicsGLRegisterImage(). In CUDA, it appears as a CUDA array. Kernels can read from the array by binding it to a texture or surface reference. They can also write to it via the surface write functions if the resource has been registered with the cudaGraphicsRegisterFlagsSurfaceLoadStore flag. The array can also be

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|51

Programming Interface

read and written via cudaMemcpy2D() calls. cudaGraphicsGLRegisterImage() supports all texture formats with 1, 2, or 4 components and an internal type of float (e.g., GL_RGBA_FLOAT32), normalized integer (e.g., GL_RGBA8, GL_INTENSITY16), and unnormalized integer (e.g., GL_RGBA8UI) (please note that since unnormalized integer formats require OpenGL 3.0, they can only be written by shaders, not the fixed function pipeline). The OpenGL context whose resources are being shared has to be current to the host thread making any OpenGL interoperability API calls. The following code sample uses a kernel to dynamically modify a 2D width x height grid of vertices stored in a vertex buffer object:
GLuint positionsVBO; struct cudaGraphicsResource* positionsVBO_CUDA; int main() { // Initialize OpenGL and GLUT for device 0 // and make the OpenGL context current ... glutDisplayFunc(display); // Explicitly set device 0 cudaGLSetGLDevice(0); // Create buffer object and register it with CUDA glGenBuffers(1, positionsVBO); glBindBuffer(GL_ARRAY_BUFFER, &positionsVBO); unsigned int size = width * height * 4 * sizeof(float); glBufferData(GL_ARRAY_BUFFER, size, 0, GL_DYNAMIC_DRAW); glBindBuffer(GL_ARRAY_BUFFER, 0); cudaGraphicsGLRegisterBuffer(&positionsVBO_CUDA, positionsVBO, cudaGraphicsMapFlagsWriteDiscard); // Launch rendering loop glutMainLoop(); } ...

void display() { // Map buffer object for writing from CUDA float4* positions; cudaGraphicsMapResources(1, &positionsVBO_CUDA, 0); size_t num_bytes; cudaGraphicsResourceGetMappedPointer((void**)&positions, &num_bytes, positionsVBO_CUDA)); // Execute kernel dim3 dimBlock(16, 16, 1); dim3 dimGrid(width / dimBlock.x, height / dimBlock.y, 1); createVertices<<<dimGrid, dimBlock>>>(positions, time, width, height); // Unmap buffer object cudaGraphicsUnmapResources(1, &positionsVBO_CUDA, 0); // Render from buffer object glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT); glBindBuffer(GL_ARRAY_BUFFER, positionsVBO); glVertexPointer(4, GL_FLOAT, 0, 0);

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|52

Programming Interface

glEnableClientState(GL_VERTEX_ARRAY); glDrawArrays(GL_POINTS, 0, width * height); glDisableClientState(GL_VERTEX_ARRAY); // Swap buffers glutSwapBuffers(); glutPostRedisplay();

void deleteVBO() { cudaGraphicsUnregisterResource(positionsVBO_CUDA); glDeleteBuffers(1, &positionsVBO); } __global__ void createVertices(float4* positions, float time, unsigned int width, unsigned int height) { unsigned int x = blockIdx.x * blockDim.x + threadIdx.x; unsigned int y = blockIdx.y * blockDim.y + threadIdx.y; // Calculate uv coordinates float u = x / (float)width; float v = y / (float)height; u = u * 2.0f - 1.0f; v = v * 2.0f - 1.0f; // calculate simple sine wave pattern float freq = 4.0f; float w = sinf(u * freq + time) * cosf(v * freq + time) * 0.5f; // Write positions positions[y * width + x] = make_float4(u, w, v, 1.0f);

On Windows and for Quadro GPUs, cudaWGLGetDevice() can be used to retrieve the CUDA device associated to the handle returned by wglEnumGpusNV(). Quadro GPUs offer higher performance OpenGL interoperability than GeForce and Tesla GPUs in a multi-GPU configuration where OpenGL rendering is performed on the Quadro GPU and CUDA computations are performed on other GPUs in the system.

3.2.11.2Direct3D Interoperability
Direct3D interoperability is supported for Direct3D 9, Direct3D 10, and Direct3D 11. A CUDA context may interoperate with only one Direct3D device at a time and the CUDA context and Direct3D device must be created on the same GPU. In addition the following considerations must be taken when creating the device: Direct3D 9 devices must be created with DeviceType set to D3DDEVTYPE_HAL and BehaviorFlags with the D3DCREATE_HARDWARE_VERTEXPROCESSING flag. Direct3D 10 and Direct3D 11 devices must be created with DriverType set to D3D_DRIVER_TYPE_HARDWARE. Interoperability with Direct3D requires that the Direct3D device be specified by cudaD3D9SetDirect3DDevice(), cudaD3D10SetDirect3DDevice() and cudaD3D11SetDirect3DDevice(), before any other runtime calls. cudaD3D9GetDevice(), cudaD3D10GetDevice(), and cudaD3D11GetDevice() can be used to retrieve the CUDA device associated to some adapter. A set of calls is also available to allow the creation of CUDA contexts with interoperability with Direct3D devices that use NVIDIA SLI in AFR (Alternate

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|53

Programming Interface

Frame Rendering) mode: cudaD3D[9|10|11]GetDevices(). A call to cudaD3D[9|10|11]GetDevices()can be used to obtain a list of CUDA device handles that can be passed as the (optional) last parameter to cudaD3D[9|10| 11]SetDirect3DDevice(). The application has the choice to either create multiple CPU threads, each using a different CUDA context, or a single CPU thread using multiple CUDA context. If using separate CPU threads for each GPU each of the CUDA contexts would be created by the CUDA runtime by calling in a separate CPU thread cudaD3D[9|10| 11]SetDirect3DDevice() using one of the CUDA device handles returned by cudaD3D[9|10|11]GetDevices(). If using a single CPU thread the CUDA contexts would have to be created using the CUDA driver API context creation functions for interoperability with Direct3D devices that use NVIDIA SLI (cuD3D[9|10|11]CtxCreateOnDevice()). The application relies on the interoperability between CUDA driver and runtime APIs (Interoperablility between Runtime and Driver APIs), which allows it to call cuCtxPushCurrent() and cuCtxPopCurrent() to change the CUDA context active at a given time. The Direct3D resources that may be mapped into the address space of CUDA are Direct3D buffers, textures, and surfaces. These resources are registered using cudaGraphicsD3D9RegisterResource(), cudaGraphicsD3D10RegisterResource(), and cudaGraphicsD3D11RegisterResource(). The following code sample uses a kernel to dynamically modify a 2D width x height grid of vertices stored in a vertex buffer object.

3.2.11.2.1Direct3D 9 Version
IDirect3D9* D3D; IDirect3DDevice9* device; struct CUSTOMVERTEX { FLOAT x, y, z; DWORD color; }; IDirect3DVertexBuffer9* positionsVB; struct cudaGraphicsResource* positionsVB_CUDA; int main() { // Initialize Direct3D D3D = Direct3DCreate9(D3D_SDK_VERSION); // Get a CUDA-enabled adapter unsigned int adapter = 0; for (; adapter < g_pD3D->GetAdapterCount(); adapter++) { D3DADAPTER_IDENTIFIER9 adapterId; g_pD3D->GetAdapterIdentifier(adapter, 0, &adapterId); int dev; if (cudaD3D9GetDevice(&dev, adapterId.DeviceName) == cudaSuccess) break; } // Create device ... D3D->CreateDevice(adapter, D3DDEVTYPE_HAL, hWnd, D3DCREATE_HARDWARE_VERTEXPROCESSING, &params, &device);

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|54

Programming Interface

// Register device with CUDA cudaD3D9SetDirect3DDevice(device); // Create vertex buffer and register it with CUDA unsigned int size = width * height * sizeof(CUSTOMVERTEX); device->CreateVertexBuffer(size, 0, D3DFVF_CUSTOMVERTEX, D3DPOOL_DEFAULT, &positionsVB, 0); cudaGraphicsD3D9RegisterResource(&positionsVB_CUDA, positionsVB, cudaGraphicsRegisterFlagsNone); cudaGraphicsResourceSetMapFlags(positionsVB_CUDA, cudaGraphicsMapFlagsWriteDiscard); // Launch rendering loop while (...) { ... Render(); ... } ...

void Render() { // Map vertex buffer for writing from CUDA float4* positions; cudaGraphicsMapResources(1, &positionsVB_CUDA, 0); size_t num_bytes; cudaGraphicsResourceGetMappedPointer((void**)&positions, &num_bytes, positionsVB_CUDA)); // Execute kernel dim3 dimBlock(16, 16, 1); dim3 dimGrid(width / dimBlock.x, height / dimBlock.y, 1); createVertices<<<dimGrid, dimBlock>>>(positions, time, width, height); // Unmap vertex buffer cudaGraphicsUnmapResources(1, &positionsVB_CUDA, 0); // Draw and present ...

void releaseVB() { cudaGraphicsUnregisterResource(positionsVB_CUDA); positionsVB->Release(); } __global__ void createVertices(float4* positions, float time, unsigned int width, unsigned int height) { unsigned int x = blockIdx.x * blockDim.x + threadIdx.x; unsigned int y = blockIdx.y * blockDim.y + threadIdx.y; // Calculate uv coordinates float u = x / (float)width; float v = y / (float)height; u = u * 2.0f - 1.0f; v = v * 2.0f - 1.0f; // Calculate simple sine wave pattern float freq = 4.0f; float w = sinf(u * freq + time)

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|55

Programming Interface

* cosf(v * freq + time) * 0.5f; // Write positions positions[y * width + x] = make_float4(u, w, v, __int_as_float(0xff00ff00));

3.2.11.2.2Direct3D 10 Version
ID3D10Device* device; struct CUSTOMVERTEX { FLOAT x, y, z; DWORD color; }; ID3D10Buffer* positionsVB; struct cudaGraphicsResource* positionsVB_CUDA; int main() { // Get a CUDA-enabled adapter IDXGIFactory* factory; CreateDXGIFactory(__uuidof(IDXGIFactory), (void**)&factory); IDXGIAdapter* adapter = 0; for (unsigned int i = 0; !adapter; ++i) { if (FAILED(factory->EnumAdapters(i, &adapter)) break; int dev; if (cudaD3D10GetDevice(&dev, adapter) == cudaSuccess) break; adapter->Release(); } factory->Release(); // Create swap chain and device ... D3D10CreateDeviceAndSwapChain(adapter, D3D10_DRIVER_TYPE_HARDWARE, 0, D3D10_CREATE_DEVICE_DEBUG, D3D10_SDK_VERSION, &swapChainDesc, &swapChain, &device); adapter->Release(); // Register device with CUDA cudaD3D10SetDirect3DDevice(device); // Create vertex buffer and register it with CUDA unsigned int size = width * height * sizeof(CUSTOMVERTEX); D3D10_BUFFER_DESC bufferDesc; bufferDesc.Usage = D3D10_USAGE_DEFAULT; bufferDesc.ByteWidth = size; bufferDesc.BindFlags = D3D10_BIND_VERTEX_BUFFER; bufferDesc.CPUAccessFlags = 0; bufferDesc.MiscFlags = 0; device->CreateBuffer(&bufferDesc, 0, &positionsVB); cudaGraphicsD3D10RegisterResource(&positionsVB_CUDA, positionsVB, cudaGraphicsRegisterFlagsNone); cudaGraphicsResourceSetMapFlags(positionsVB_CUDA, cudaGraphicsMapFlagsWriteDiscard); // Launch rendering loop while (...) { ... Render(); ...

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|56

Programming Interface

} ...

3.2.11.2.3Direct3D 11 Version
ID3D11Device* device; struct CUSTOMVERTEX { FLOAT x, y, z; DWORD color; }; ID3D11Buffer* positionsVB; struct cudaGraphicsResource* positionsVB_CUDA; int main() {

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|57

Programming Interface

// Get a CUDA-enabled adapter IDXGIFactory* factory; CreateDXGIFactory(__uuidof(IDXGIFactory), (void**)&factory); IDXGIAdapter* adapter = 0; for (unsigned int i = 0; !adapter; ++i) { if (FAILED(factory->EnumAdapters(i, &adapter)) break; int dev; if (cudaD3D11GetDevice(&dev, adapter) == cudaSuccess) break; adapter->Release(); } factory->Release(); // Create swap chain and device ... sFnPtr_D3D11CreateDeviceAndSwapChain(adapter, D3D11_DRIVER_TYPE_HARDWARE, 0, D3D11_CREATE_DEVICE_DEBUG, featureLevels, 3, D3D11_SDK_VERSION, &swapChainDesc, &swapChain, &device, &featureLevel, &deviceContext); adapter->Release(); // Register device with CUDA cudaD3D11SetDirect3DDevice(device); // Create vertex buffer and register it with CUDA unsigned int size = width * height * sizeof(CUSTOMVERTEX); D3D11_BUFFER_DESC bufferDesc; bufferDesc.Usage = D3D11_USAGE_DEFAULT; bufferDesc.ByteWidth = size; bufferDesc.BindFlags = D3D11_BIND_VERTEX_BUFFER; bufferDesc.CPUAccessFlags = 0; bufferDesc.MiscFlags = 0; device->CreateBuffer(&bufferDesc, 0, &positionsVB); cudaGraphicsD3D11RegisterResource(&positionsVB_CUDA, positionsVB, cudaGraphicsRegisterFlagsNone); cudaGraphicsResourceSetMapFlags(positionsVB_CUDA, cudaGraphicsMapFlagsWriteDiscard); // Launch rendering loop while (...) { ... Render(); ... } ...

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|58

Programming Interface

dim3 dimGrid(width / dimBlock.x, height / dimBlock.y, 1); createVertices<<<dimGrid, dimBlock>>>(positions, time, width, height); // Unmap vertex buffer cudaGraphicsUnmapResources(1, &positionsVB_CUDA, 0); // Draw and present ...

// Calculate uv coordinates float u = x / (float)width; float v = y / (float)height; u = u * 2.0f - 1.0f; v = v * 2.0f - 1.0f; // Calculate simple sine wave pattern float freq = 4.0f; float w = sinf(u * freq + time) * cosf(v * freq + time) * 0.5f; // Write positions positions[y * width + x] = make_float4(u, w, v, __int_as_float(0xff00ff00));

3.2.11.3SLI Interoperability
In a system with multiple GPUs, all CUDA-enabled GPUs are accessible via the CUDA driver and runtime as separate devices. There are however special considerations as described below when the system is in SLI mode. First, an allocation in one CUDA device on one GPU will consume memory on other GPUs that are part of the SLI configuration of the Direct3D or OpenGL device. Because of this, allocations may fail earlier than otherwise expected. Second, applications have to create multiple CUDA contexts, one for each GPU in the SLI configuration and deal with the fact that a different GPU is used for rendering by the Direct3D or OpenGL device at every frame. The application can use the cudaD3D[9|10|11]GetDevices() for Direct3D and cudaGLGetDevices() for OpenGL set of calls to identify the CUDA device handle(s) for the device(s) that are performing the rendering in the current and next frame. Given this information the application will typically map Direct3D or OpenGL resources to the CUDA context corresponding to the CUDA device returned by cudaD3D[9|10|11]GetDevices() or cudaGLGetDevices() when the deviceList parameter is set to CU_D3D10_DEVICE_LIST_CURRENT_FRAME or cudaGLDeviceListCurrentFrame.

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|59

Programming Interface

See Direct3D Interoperability and OpenGL Interoperability for details on how the CUDA runtime interoperate with Direct3D and OpenGL, respectively.

3.3Versioning and Compatibility

There are two version numbers that developers should care about when developing a CUDA application: The compute capability that describes the general specifications and features of the compute device (see Compute Capability) and the version of the CUDA driver API that describes the features supported by the driver API and runtime. The version of the driver API is defined in the driver header file as CUDA_VERSION. It allows developers to check whether their application requires a newer device driver than the one currently installed. This is important, because the driver API is backward compatible, meaning that applications, plug-ins, and libraries (including the C runtime) compiled against a particular version of the driver API will continue to work on subsequent device driver releases as illustrated in Figure 11 The Driver API Is Backward, but Not Forward Compatible. The driver API is not forward compatible, which means that applications, plug-ins, and libraries (including the C runtime) compiled against a particular version of the driver API will not work on previous versions of the device driver. It is important to note that mixing and matching versions is not supported; specifically: All applications, plug-ins, and libraries on a system must use the same version of the CUDA driver API, since only one version of the CUDA device driver can be installed on a system. All plug-ins and libraries used by an application must use the same version of the runtime. All plug-ins and libraries used by an application must use the same version of any libraries that use the runtime (such as CUFFT, CUBLAS, ).

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|60

Programming Interface

Figure11The Driver API Is Backward, but Not Forward Compatible

3.4Compute Modes
On Tesla solutions running Windows Server 2008 and later or Linux, one can set any device in a system in one of the three following modes using NVIDIAs System Management Interface (nvidia-smi), which is a tool distributed as part of the driver: Default compute mode: Multiple host threads can use the device (by calling cudaSetDevice() on this device, when using the runtime API, or by making current a context associated to the device, when using the driver API) at the same time. Exclusive-process compute mode: Only one CUDA context may be created on the device across all processes in the system and that context may be current to as many threads as desired within the process that created that context. Exclusive-process-and-thread compute mode: Only one CUDA context may be created on the device across all processes in the system and that context may only be current to one thread at a time. Prohibited compute mode: No CUDA context can be created on the device. This means, in particular, that a host thread using the runtime API without explicitly calling cudaSetDevice() might be associated with a device other than device 0 if device 0 turns out to be in the exclusive-process mode and used by another process, or in the exclusive-process-and-thread mode and used by another thread, or in prohibited mode. cudaSetValidDevices() can be used to set a device from a prioritized list of devices. Applications may query the compute mode of a device by checking the computeMode device property (see Device Enumeration).

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|61

Programming Interface

3.5Mode Switches
GPUs that have a display output dedicate some DRAM memory to the so-called primary surface, which is used to refresh the display device whose output is viewed by the user. When users initiate a mode switch of the display by changing the resolution or bit depth of the display (using NVIDIA control panel or the Display control panel on Windows), the amount of memory needed for the primary surface changes. For example, if the user changes the display resolution from 1280x1024x32-bit to 1600x1200x32-bit, the system must dedicate 7.68 MB to the primary surface rather than 5.24 MB. (Full-screen graphics applications running with anti-aliasing enabled may require much more display memory for the primary surface.) On Windows, other events that may initiate display mode switches include launching a full-screen DirectX application, hitting Alt +Tab to task switch away from a full-screen DirectX application, or hitting Ctrl+Alt+Del to lock the computer. If a mode switch increases the amount of memory needed for the primary surface, the system may have to cannibalize memory allocations dedicated to CUDA applications. Therefore, a mode switch results in any call to the CUDA runtime to fail and return an invalid context error.

3.6Tesla Compute Cluster Mode for Windows

Performance Guidelines

For the parallel workloads, at points in the algorithm where parallelism is broken because some threads need to synchronize in order to share data with each other, there are two cases: Either these threads belong to the same block, in which case they should use __syncthreads() and share data through shared memory within the same kernel invocation, or they belong to different blocks, in which case they must share data through global memory using two separate kernel invocations, one for writing to and one for reading from global memory. The second case is much less optimal since it adds the overhead of extra kernel invocations and global memory traffic. Its occurrence should therefore be minimized by mapping the algorithm to the CUDA programming model in such a way that the computations that require inter-thread communication are performed within a single thread block as much as possible.

5.2.2Device Level
At a lower level, the application should maximize parallel execution between the multiprocessors of a device. For devices of compute capability 1.x, only one kernel can execute on a device at one time, so the kernel should be launched with at least as many thread blocks as there are multiprocessors in the device. For devices of compute capability 2.x and higher, multiple kernels can execute concurrently on a device, so maximum utilization can also be achieved by using streams to enable enough kernels to execute concurrently as described in Asynchronous Concurrent Execution.

5.2.3Multiprocessor Level
At an even lower level, the application should maximize parallel execution between the various functional units within a multiprocessor. As described in Hardware Multithreading, a GPU multiprocessor relies on threadlevel parallelism to maximize utilization of its functional units. Utilization is therefore directly linked to the number of resident warps. At every instruction issue time, a warp scheduler selects a warp that is ready to execute its next instruction, if any, and issues the instruction to the active threads of the warp. The number of clock cycles it takes for a warp to be ready to execute its next instruction is called the latency, and full utilization is achieved when all warp schedulers always have some instruction to issue for some warp at every clock cycle during that latency period, or in other words, when latency is completely hidden. The number of instructions required to hide a latency of L clock cycles depends on the respective throughputs of these instructions (see Arithmetic Instructions for the throughputs of various arithmetic instructions); assuming maximum throughput for all instructions, it is: L/4 (rounded up to nearest integer) for devices of compute capability 1.x since a multiprocessor issues one instruction per warp over four clock cycles, as mentioned in Compute Capability 1.x, L for devices of compute capability 2.0 since a multiprocessor issues one instruction per warp over two clock cycles for two warps at a time, as mentioned in Compute Capability 2.x,

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|67

Performance Guidelines

2L for devices of compute capability 2.1 since a multiprocessor issues a pair of instructions per warp over two clock cycles for two warps at a time, as mentioned in Compute Capability 2.x, 8L for devices of compute capability 3.x since a multiprocessor issues a pair of instructions per warp over one clock cycle for four warps at a time, as mentioned in Compute Capability 3.x. For devices of compute capability 2.0, the two instructions issued every other cycle are for two different warps. For devices of compute capability 2.1, the four instructions issued every other cycle are two pairs for two different warps, each pair being for the same warp. For devices of compute capability 3.x, the eight instructions issued every cycle are four pairs for four different warps, each pair being for the same warp. The most common reason a warp is not ready to execute its next instruction is that the instructions input operands are not available yet. If all input operands are registers, latency is caused by register dependencies, i.e., some of the input operands are written by some previous instruction(s) whose execution has not completed yet. In the case of a back-to-back register dependency (i.e., some input operand is written by the previous instruction), the latency is equal to the execution time of the previous instruction and the warp schedulers must schedule instructions for different warps during that time. Execution time varies depending on the instruction, but it is typically about 22 clock cycles for devices of compute capability 1.x and 2.x and about 11 clock cycles for devices of compute capability 3.x, which translates to 6 warps for devices of compute capability 1.x, 22 warps for devices of compute capability 2.x, and 44 warps for devices of compute capability 3.x and higher (still assuming that warps execute instructions with maximum throughput, otherwise fewer warps are needed). For devices of compute capability 2.1 and higher, this is also assuming enough instruction-level parallelism so that schedulers are always able to issue pairs of instructions for each warp. If some input operand resides in off-chip memory, the latency is much higher: 400 to 800 clock cycles for devices of compute capability 1.x and 2.x and about 200 to 400 clock cycles for devices of compute capability 3.x. The number of warps required to keep the warp schedulers busy during such high latency periods depends on the kernel code and its degree of instruction-level parallelism. In general, more warps are required if the ratio of the number of instructions with no off-chip memory operands (i.e., arithmetic instructions most of the time) to the number of instructions with off-chip memory operands is low (this ratio is commonly called the arithmetic intensity of the program). For example, assume this ratio is 30, also assume the latencies are 600 cycles on devices of compute capability 1.x and 2.x and 300 cycles on devices of compute capability 3.x. Then about 5 warps are required for devices of compute capability 1.x, about 20 for devices of compute capability 2.x and about 40 for devices of compute capability 3.x (with the same assumptions as in the previous paragraph). Another reason a warp is not ready to execute its next instruction is that it is waiting at some memory fence (Memory Fence Functions) or synchronization point (Memory Fence Functions). A synchronization point can force the multiprocessor to idle as more and more warps wait for other warps in the same block to complete execution of

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|68

Performance Guidelines

instructions prior to the synchronization point. Having multiple resident blocks per multiprocessor can help reduce idling in this case, as warps from different blocks do not need to wait for each other at synchronization points. The number of blocks and warps residing on each multiprocessor for a given kernel call depends on the execution configuration of the call (Execution Configuration), the memory resources of the multiprocessor, and the resource requirements of the kernel as described in Hardware Multithreading. To assist programmers in choosing thread block size based on register and shared memory requirements, the CUDA Software Development Kit provides a spreadsheet, called the CUDA Occupancy Calculator, where occupancy is defined as the ratio of the number of resident warps to the maximum number of resident warps (given in Compute Capabilities for various compute capabilities). Register, local, shared, and constant memory usages are reported by the compiler when compiling with the -ptxas-options=-v option. The total amount of shared memory required for a block is equal to the sum of the amount of statically allocated shared memory, the amount of dynamically allocated shared memory, and for devices of compute capability 1.x, the amount of shared memory used to pass the kernels arguments (see __noinline__ and __forceinline__). The number of registers used by a kernel can have a significant impact on the number of resident warps. For example, for devices of compute capability 1.2, if a kernel uses 16 registers and each block has 512 threads and requires very little shared memory, then two blocks (i.e., 32 warps) can reside on the multiprocessor since they require 2x512x16 registers, which exactly matches the number of registers available on the multiprocessor. But as soon as the kernel uses one more register, only one block (i.e., 16 warps) can be resident since two blocks would require 2x512x17 registers, which are more registers than are available on the multiprocessor. Therefore, the compiler attempts to minimize register usage while keeping register spilling (see Device Memory Accesses) and the number of instructions to a minimum. Register usage can be controlled using the maxrregcount compiler option or launch bounds as described in Launch Bounds. Each double variable (on devices that supports native double precision, i.e., devices of compute capability 1.2 and higher) and each long long variable uses two registers. However, devices of compute capability 1.2 and higher have at least twice as many registers per multiprocessor as devices with lower compute capability. The effect of execution configuration on performance for a given kernel call generally depends on the kernel code. Experimentation is therefore recommended. Applications can also parameterize execution configurations based on register file size and shared memory size, which depends on the compute capability of the device, as well as on the number of multiprocessors and memory bandwidth of the device, all of which can be queried using the runtime (see reference manual). The number of threads per block should be chosen as a multiple of the warp size to avoid wasting computing resources with under-populated warps as much as possible.

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|69

Performance Guidelines

5.3Maximize Memory Throughput

The first step in maximizing overall memory throughput for the application is to minimize data transfers with low bandwidth. That means minimizing data transfers between the host and the device, as detailed in Data Transfer between Host and Device, since these have much lower bandwidth than data transfers between global memory and the device. That also means minimizing data transfers between global memory and the device by maximizing use of on-chip memory: shared memory and caches (i.e., L1/L2 caches available on devices of compute capability 2.x and higher, texture cache and constant cache available on all devices). Shared memory is equivalent to a user-managed cache: The application explicitly allocates and accesses it. As illustrated in CUDA C Runtime, a typical programming pattern is to stage data coming from device memory into shared memory; in other words, to have each thread of a block: Load data from device memory to shared memory, Synchronize with all the other threads of the block so that each thread can safely read shared memory locations that were populated by different threads, Process the data in shared memory, Synchronize again if necessary to make sure that shared memory has been updated with the results, Write the results back to device memory. For some applications (e.g., for which global memory access patterns are datadependent), a traditional hardware-managed cache is more appropriate to exploit data locality. As mentioned in Compute Capability 2.x, for devices of compute capability 2.x and higher, the same on-chip memory is used for both L1 and shared memory, and how much of it is dedicated to L1 versus shared memory is configurable for each kernel call. The throughput of memory accesses by a kernel can vary by an order of magnitude depending on access pattern for each type of memory. The next step in maximizing memory throughput is therefore to organize memory accesses as optimally as possible based on the optimal memory access patterns described in Device Memory Accesses. This optimization is especially important for global memory accesses as global memory bandwidth is low, so non-optimal global memory accesses have a higher impact on performance.

5.3.1Data Transfer between Host and Device

Applications should strive to minimize data transfer between the host and the device. One way to accomplish this is to move more code from the host to the device, even if that means running kernels with low parallelism computations. Intermediate data structures may be created in device memory, operated on by the device, and destroyed without ever being mapped by the host or copied to host memory.

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|70

Performance Guidelines

Also, because of the overhead associated with each transfer, batching many small transfers into a single large transfer always performs better than making each transfer separately. On systems with a front-side bus, higher performance for data transfers between host and device is achieved by using page-locked host memory as described in Page-Locked Host Memory. In addition, when using mapped page-locked memory (Mapped Memory), there is no need to allocate any device memory and explicitly copy data between device and host memory. Data transfers are implicitly performed each time the kernel accesses the mapped memory. For maximum performance, these memory accesses must be coalesced as with accesses to global memory (see Device Memory Accesses). Assuming that they are and that the mapped memory is read or written only once, using mapped pagelocked memory instead of explicit copies between device and host memory can be a win for performance. On integrated systems where device memory and host memory are physically the same, any copy between host and device memory is superfluous and mapped page-locked memory should be used instead. Applications may query a device is integrated by checking that the integrated device property (see Device Enumeration) is equal to 1.

5.3.2Device Memory Accesses

An instruction that accesses addressable memory (i.e., global, local, shared, constant, or texture memory) might need to be re-issued multiple times depending on the distribution of the memory addresses across the threads within the warp. How the distribution affects the instruction throughput this way is specific to each type of memory and described in the following sections. For example, for global memory, as a general rule, the more scattered the addresses are, the more reduced the throughput is. Global Memory Global memory resides in device memory and device memory is accessed via 32-, 64-, or 128-byte memory transactions. These memory transactions must be naturally aligned: Only the 32-, 64-, or 128-byte segments of device memory that are aligned to their size (i.e., whose first address is a multiple of their size) can be read or written by memory transactions. When a warp executes an instruction that accesses global memory, it coalesces the memory accesses of the threads within the warp into one or more of these memory transactions depending on the size of the word accessed by each thread and the distribution of the memory addresses across the threads. In general, the more transactions are necessary, the more unused words are transferred in addition to the words accessed by the threads, reducing the instruction throughput accordingly. For example, if a 32-byte memory transaction is generated for each thread's 4-byte access, throughput is divided by 8. How many transactions are necessary and how much throughput is ultimately affected varies with the compute capability of the device. For devices of compute capability 1.0

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|71

16(*) 32 32 16 32 16

4 48 48 16 48 16

8 160 160 32 160 32

64 160 160 64 160 32

Multiple Multiple instructions instructions 8 8

Multiple Multiple Multiple Multiple instructions instructions instructions instructions

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|76

Performance Guidelines

Compute Capability 1.0 1.1 1.2

Type conversions from 8-bit and 16-bit integer to 32-bit types Type conversions from and to 64-bit types All other type conversions 8 Multiple instructions 8 8 16 16 128 128

1.3

2.0

2.1

3.0

3.5

1 8

16(*) 16

4 16

8 32

32 32

(*) Throughput is lower for GeForce GPUs.

Other instructions and functions are implemented on top of the native instructions. The implementation may be different for devices of different compute capabilities, and the number of native instructions after compilation may fluctuate with every compiler version. For complicated functions, there can be multiple code paths depending on input. cuobjdump can be used to inspect a particular implementation in a cubin object. The implementation of some functions are readily available on the CUDA header files (math_functions.h, device_functions.h, ). In general, code compiled with -ftz=true (denormalized numbers are flushed to zero) tends to have higher performance than code compiled with -ftz=false. Similarly, code compiled with -prec div=false (less precise division) tends to have higher performance code than code compiled with -prec div=true, and code compiled with -prec-sqrt=false (less precise square root) tends to have higher performance than code compiled with -prec-sqrt=true. The nvcc user manual describes these compilation flags in more details. Single-Precision Floating-Point Addition and Multiplication Intrinsics fadd_r[d,u], __fmul_r[d,u], and __fmaf_r[n,z,d,u] (see Intrinsic Functions) compile to tens of instructions for devices of compute capability 1.x, but map to a single native instruction for devices of compute capability 2.x and higher. Single-Precision Floating-Point Division __fdividef(x, y) (see Intrinsic Functions) provides faster single-precision floatingpoint division than the division operator. Single-Precision Floating-Point Reciprocal Square Root To preserve IEEE-754 semantics the compiler can optimize 1.0/sqrtf() into rsqrtf() only when both reciprocal and square root are approximate, (i.e., with prec-div=false and -prec-sqrt=false). It is therefore recommended to invoke rsqrtf() directly where desired.

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|77

Performance Guidelines

Single-Precision Floating-Point Square Root Single-precision floating-point square root is implemented as a reciprocal square root followed by a reciprocal instead of a reciprocal square root followed by a multiplication so that it gives correct results for 0 and infinity. Sine and Cosine sinf(x), cosf(x), tanf(x), sincosf(x), and corresponding double-precision instructions are much more expensive and even more so if the argument x is large in magnitude. More precisely, the argument reduction code (see Mathematical Functions for implementation) comprises two code paths referred to as the fast path and the slow path, respectively. The fast path is used for arguments sufficiently small in magnitude and essentially consists of a few multiply-add operations. The slow path is used for arguments large in magnitude and consists of lengthy computations required to achieve correct results over the entire argument range. At present, the argument reduction code for the trigonometric functions selects the fast path for arguments whose magnitude is less than 48039.0f for the single-precision functions, and less than 2147483648.0 for the double-precision functions. As the slow path requires more registers than the fast path, an attempt has been made to reduce register pressure in the slow path by storing some intermediate variables in local memory, which may affect performance because of local memory high latency and bandwidth (see Device Memory Accesses). At present, 28 bytes of local memory are used by single-precision functions, and 44 bytes are used by double-precision functions. However, the exact amount is subject to change. Due to the lengthy computations and use of local memory in the slow path, the throughput of these trigonometric functions is lower by one order of magnitude when the slow path reduction is required as opposed to the fast path reduction. Integer Arithmetic On devices of compute capability 1.x, 32-bit integer multiplication is implemented using multiple instructions as it is not natively supported. 24-bit integer multiplication is natively supported however via the __[u]mul24 intrinsic. Using __[u]mul24 instead of the 32-bit multiplication operator whenever possible usually improves performance for instruction bound kernels. It can have the opposite effect however in cases where the use of __[u]mul24 inhibits compiler optimizations. On devices of compute capability 2.x and beyond, 32-bit integer multiplication is natively supported, but 24-bit integer multiplication is not. __[u]mul24 is therefore implemented using multiple instructions and should not be used.

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|78

Performance Guidelines

Integer division and modulo operation are costly: tens of instructions on devices of compute capability 1.x, below 20 instructions on devices of compute capability 2.x and higher. They can be replaced with bitwise operations in some cases: If n is a power of 2, (i/n) is equivalent to (i>>log2(n)) and (i%n) is equivalent to (i&(n-1)); the compiler will perform these conversions if n is literal. __brev, __brevll, __popc, and __popcll compile to tens of instructions for devices of compute capability 1.x, but __brev and __popc map to a single instruction for devices of compute capability 2.x and higher and __brevll and __popcll to just a few. __clz, __clzll, __ffs, and __ffsll compile to fewer instructions for devices of compute capability 2.x and higher than for devices of compute capability 1.x. Type Conversion Sometimes, the compiler must insert conversion instructions, introducing additional execution cycles. This is the case for: Functions operating on variables of type char or short whose operands generally need to be converted to int, Double-precision floating-point constants (i.e., those constants defined without any type suffix) used as input to single-precision floating-point computations (as mandated by C/C++ standards). This last case can be avoided by using single-precision floating-point constants, defined with an f suffix such as 3.141592653589793f, 1.0f, 0.5f.

5.4.2Control Flow Instructions

Any flow control instruction (if, switch, do, for, while) can significantly impact the effective instruction throughput by causing threads of the same warp to diverge (i.e., to follow different execution paths). If this happens, the different executions paths have to be serialized, increasing the total number of instructions executed for this warp. When all the different execution paths have completed, the threads converge back to the same execution path. To obtain best performance in cases where the control flow depends on the thread ID, the controlling condition should be written so as to minimize the number of divergent warps. This is possible because the distribution of the warps across the block is deterministic as mentioned in SIMT Architecture. A trivial example is when the controlling condition only depends on (threadIdx / warpSize) where warpSize is the warp size. In this case, no warp diverges since the controlling condition is perfectly aligned with the warps. Sometimes, the compiler may unroll loops or it may optimize out if or switch statements by using branch predication instead, as detailed below. In these cases, no warp can ever diverge. The programmer can also control loop unrolling using the #pragma unroll directive (see #pragma unroll).

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|79

Performance Guidelines

When using branch predication none of the instructions whose execution depends on the controlling condition gets skipped. Instead, each of them is associated with a perthread condition code or predicate that is set to true or false based on the controlling condition and although each of these instructions gets scheduled for execution, only the instructions with a true predicate are actually executed. Instructions with a false predicate do not write results, and also do not evaluate addresses or read operands. The compiler replaces a branch instruction with predicated instructions only if the number of instructions controlled by the branch condition is less or equal to a certain threshold: If the compiler determines that the condition is likely to produce many divergent warps, this threshold is 7, otherwise it is 4.

5.4.3Synchronization Instruction
Throughput for __syncthreads() is 8 operations per clock cycle for devices of compute capability 1.x, 16 operations per clock cycle for devices of compute capability 2.x, and 128 operations per clock cycle for devices of compute capability 3.x. Note that __syncthreads() can impact performance by forcing the multiprocessor to idle as detailed in Device Memory Accesses.

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|80

Appendix A. CUDA-ENABLED GPUS

http://developer.nvidia.com/cuda-gpus lists all CUDA-enabled devices with their compute capability. The compute capability, number of multiprocessors, clock frequency, total amount of device memory, and other properties can be queried using the runtime (see reference manual).

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|81

Appendix B. C LANGUAGE EXTENSIONS

B.1Function Type Qualifiers

Function type qualifiers specify whether a function executes on the host or on the device and whether it is callable from the host or from the device.

B.1.1__device__
The __device__ qualifier declares a function that is: Executed on the device, Callable from the device only.

B.1.2__global__
The __global__ qualifier declares a function as being a kernel. Such a function is: Executed on the device, Callable from the host, Callable from the device for devices of compute capability 3.x (see the CUDA Dynamic Parallelism programming guide for more details). __global__ functions must have void return type. Any call to a __global__ function must specify its execution configuration as described in Execution Configuration. A call to a __global__ function is asynchronous, meaning it returns before the device has completed its execution.

B.1.3__host__
The __host__ qualifier declares a function that is:

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|82

C Language Extensions

Executed on the host, Callable from the host only. It is equivalent to declare a function with only the __host__ qualifier or to declare it without any of the __host__, __device__, or __global__ qualifier; in either case the function is compiled for the host only. The __global__ and __host__ qualifiers cannot be used together. The __device__ and __host__ qualifiers can be used together however, in which case the function is compiled for both the host and the device. The __CUDA_ARCH__ macro introduced in Application Compatibility can be used to differentiate code paths between host and device:
__host__ __device__ func() { #if __CUDA_ARCH__ == 100 // Device code path for compute capability 1.0 #elif __CUDA_ARCH__ == 200 // Device code path for compute capability 2.0 #elif __CUDA_ARCH__ == 300 // Device code path for compute capability 3.0 #elif !defined(__CUDA_ARCH__) // Host code path #endif }

B.1.4noinline and forceinline

The effects here are a reduced number of memory accesses and reduced number of computations. This is balanced by an increase in register pressure due to "cached" loads and common sub-expressions. Since register pressure is a critical issue in many CUDA codes, use of restricted pointers can have negative performance impact on CUDA code, due to reduced occupancy.

B.3Built-in Vector Types

B.3.1char, short, int, long, longlong, float, double
These are vector types derived from the basic integer and floating-point types. They are structures and the 1st, 2nd, 3rd, and 4th components are accessible through the fields x, y, z, and w, respectively. They all come with a constructor function of the form make_<type name>; for example,
int2 make_int2(int x, int y);

which creates a vector of type int2 with value(x, y). In host code, the alignment requirement of a vector type is equal to the alignment requirement of its base type. This is not always the case in device code as detailed in Table 3 Alignment Requirements in Device Code.

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|86

C Language Extensions

Table3Alignment Requirements in Device Code

Type
char1, uchar1 char2, uchar2 char3, uchar3 char4, uchar4 short1, ushort1 short2, ushort2 short3, ushort3 short4, ushort4 int1, uint1 int2, uint2 int3, uint3 int4, uint4 long1, ulong1 long2, ulong2 long3, ulong3 long4, ulong4 longlong1, ulonglong1 longlong2, ulonglong2 float1 float2 float3 float4 double1 double2

Alignment
1 2 1 4 2 4 2 8 4 8 4 16 4 if sizeof(long) is equal to sizeof(int) 8, otherwise 8 if sizeof(long) is equal to sizeof(int), 16, otherwise 4 if sizeof(long) is equal to sizeof(int), 8, otherwise 16 8 16 4 8 4 16 8 16

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|87

C Language Extensions

B.3.2dim3
This type is an integer vector type based on uint3 that is used to specify dimensions. When defining a variable of type dim3, any component left unspecified is initialized to 1.

B.4Built-in Variables
Built-in variables specify the grid and block dimensions and the block and thread indices. They are only valid within functions that are executed on the device.

B.4.1gridDim
This variable is of type dim3 (see dim3) and contains the dimensions of the grid.

B.4.2blockIdx
This variable is of type uint3 (see char, short, int, long, longlong, float, double) and contains the block index within the grid.

B.4.3blockDim
This variable is of type dim3 (see dim3) and contains the dimensions of the block.

B.4.4threadIdx
This variable is of type uint3 (see char, short, int, long, longlong, float, double ) and contains the thread index within the block.

B.4.5warpSize
This variable is of type int and contains the warp size in threads (see SIMT Architecture for the definition of a warp).

B.5Memory Fence Functions

void __threadfence_block();

waits until all global and shared memory accesses made by the calling thread prior to __threadfence_block() are visible to all threads in the thread block.
void __threadfence();

waits until all global and shared memory accesses made by the calling thread prior to __threadfence() are visible to:

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|88

C Language Extensions

All threads in the thread block for shared memory accesses, All threads in the device for global memory accesses.
void __threadfence_system();

waits until all global and shared memory accesses made by the calling thread prior to __threadfence_system() are visible to: All threads in the thread block for shared memory accesses, All threads in the device for global memory accesses, Host threads for page-locked host memory accesses (see CUDA C Runtime.4.3). __threadfence_system() is only supported by devices of compute capability 2.x and higher. In general, when a thread issues a series of writes to memory in a particular order, other threads may see the effects of these memory writes in a different order. __threadfence_block(), __threadfence(), and __threadfence_system() can be used to enforce some ordering. One use case is when threads consume some data produced by other threads as illustrated by the following code sample of a kernel that computes the sum of an array of N numbers in one call. Each block first sums a subset of the array and stores the result in global memory. When all blocks are done, the last block done reads each of these partial sums from global memory and sums them to obtain the final result. In order to determine which block is finished last, each block atomically increments a counter to signal that it is done with computing and storing its partial sum (see Atomic Functions about atomic functions). The last block is the one that receives the counter value equal to gridDim.x-1. If no fence is placed between storing the partial sum and incrementing the counter, the counter might increment before the partial sum is stored and therefore, might reach gridDim.x-1 and let the last block start reading partial sums before they have been actually updated in memory.
__device__ unsigned int count = 0; __shared__ bool isLastBlockDone; __global__ void sum(const float* array, unsigned int N, float* result) { // Each block sums a subset of the input array float partialSum = calculatePartialSum(array, N); if (threadIdx.x == 0) { // Thread 0 of each block stores the partial sum // to global memory result[blockIdx.x] = partialSum; // Thread 0 makes sure its result is visible to // all other threads __threadfence(); // Thread 0 of each block signals that it is done unsigned int value = atomicInc(&count, gridDim.x); // Thread 0 of each block determines if its block is // the last block to be done

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|89

C Language Extensions

isLastBlockDone = (value == (gridDim.x - 1));

// Synchronize to make sure that each thread reads // the correct value of isLastBlockDone __syncthreads(); if (isLastBlockDone) { // The last block sums the partial sums // stored in result[0 .. gridDim.x-1] float totalSum = calculateTotalSum(result); if (threadIdx.x == 0) { // Thread 0 of last block stores total sum // to global memory and resets count so that // next kernel call works properly result[0] = totalSum; count = 0;

B.6Synchronization Functions
void __syncthreads();

waits until all threads in the thread block have reached this point and all global and shared memory accesses made by these threads prior to __syncthreads() are visible to all threads in the block. __syncthreads() is used to coordinate communication between the threads of the same block. When some threads within a block access the same addresses in shared or global memory, there are potential read-after-write, write-after-read, or write-afterwrite hazards for some of these memory accesses. These data hazards can be avoided by synchronizing threads in-between these accesses. __syncthreads() is allowed in conditional code but only if the conditional evaluates identically across the entire thread block, otherwise the code execution is likely to hang or produce unintended side effects. Devices of compute capability 2.x and higher support three variations of __syncthreads() described below.
int __syncthreads_count(int predicate);

is identical to __syncthreads() with the additional feature that it evaluates predicate for all threads of the block and returns the number of threads for which predicate evaluates to non-zero.
int __syncthreads_and(int predicate);

is identical to __syncthreads() with the additional feature that it evaluates predicate for all threads of the block and returns non-zero if and only if predicate evaluates to nonzero for all of them.
int __syncthreads_or(int predicate);

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|90

C Language Extensions

B.7Mathematical Functions
The reference manual lists all C/C++ standard library mathematical functions that are supported in device code and all intrinsic functions that are only supported in device code. Mathematical Functions provides accuracy information for some of these functions when relevant.

B.8Texture Functions
Texture objects are described in Texture Object API Texture references are described in Texture Reference API Texture fetching is described in Texture Fetching.

B.8.1Texture Object API

B.8.1.1tex1Dfetch()
template<class T> T tex1Dfetch(cudaTextureObject_t texObj, int x);

fetches the region of linear memory specified by the one-dimensional texture object texObj using integer texture coordinate x. tex1Dfetch() only works with nonnormalized coordinates, so only the border and clamp addressing modes are supported. It does not perform any texture filtering. For integer types, it may optionally promote the integer to single-precision floating point.

B.8.1.2tex1D()
template<class DataType, enum cudaTextureReadMode readMode> Type tex1D(texture<DataType, cudaTextureType1D, readMode> texRef, float x);

fetches the CUDA array bound to the one-dimensional texture reference texRef using texture coordinate x.

B.8.1.3tex2D()
template<class DataType, enum cudaTextureReadMode readMode> Type tex2D(texture<DataType, cudaTextureType2D, readMode> texRef, float x, float y);

fetches the CUDA array or the region of linear memory bound to the two-dimensional texture reference texRef using texture coordinates x and y.

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|91

C Language Extensions

B.8.1.4tex3D()
template<class DataType, enum cudaTextureReadMode readMode> Type tex3D(texture<DataType, cudaTextureType3D, readMode> texRef, float x, float y, float z);

fetches the CUDA array bound to the three-dimensional texture reference texRef using texture coordinates x, y, and z.

B.8.1.5tex1Dlayered()
template<class DataType, enum cudaTextureReadMode readMode> Type tex1DLayered( texture<DataType, cudaTextureType1DLayered, readMode> texRef, float x, int layer);

fetches the CUDA array bound to the one-dimensional layered texture reference texRef using texture coordinate x and index layer, as described in Layered Textures.

B.8.1.6tex2Dlayered()
template<class DataType, enum cudaTextureReadMode readMode> Type tex2DLayered( texture<DataType, cudaTextureType2DLayered, readMode> texRef, float x, float y, int layer);

fetches the CUDA array bound to the two-dimensional layered texture reference texRef using texture coordinates x and y, and index layer, as described in Texture Memory

B.8.1.7texCubemap()
template<class DataType, enum cudaTextureReadMode readMode> Type texCubemap( texture<DataType, cudaTextureTypeCubemap, readMode> texRef, float x, float y, float z);

fetches the CUDA array bound to the cubemap texture reference texRef using texture coordinates x, y, and z, as described in Cubemap Textures.

B.8.1.8texCubemapLayered()
template<class DataType, enum cudaTextureReadMode readMode> Type texCubemapLayered( texture<DataType, cudaTextureTypeCubemapLayered, readMode> texRef, float x, float y, float z, int layer);

fetches the CUDA array bound to the cubemap layered texture reference texRef using texture coordinates x, y, and z, and index layer, as described in Cubemap Layered Textures.

B.8.1.9tex2Dgather()
template<class T> T tex2Dgather(cudaTextureObject_t texObj, float x, float y, int comp = 0); ;

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|92

C Language Extensions

fetches the CUDA array specified by the cubemap texture object texObj using texture coordinates x and y, as described in CUDA C Runtime.

B.8.2Texture Reference API

B.8.2.1tex1Dfetch()
template<class DataType> Type tex1Dfetch( texture<DataType, cudaTextureType1D, cudaReadModeElementType> texRef, int x); float tex1Dfetch( texture<unsigned char, cudaTextureType1D, cudaReadModeNormalizedFloat> texRef, int x); float tex1Dfetch( texture<signed char, cudaTextureType1D, cudaReadModeNormalizedFloat> texRef, int x); float tex1Dfetch( texture<unsigned short, cudaTextureType1D, cudaReadModeNormalizedFloat> texRef, int x); float tex1Dfetch( texture<signed short, cudaTextureType1D, cudaReadModeNormalizedFloat> texRef, int x);

fetches the region of linear memory bound to the one-dimensional texture reference texRef using integer texture coordinate x. tex1Dfetch() only works with nonnormalized coordinates, so only the border and clamp addressing modes are supported. It does not perform any texture filtering. For integer types, it may optionally promote the integer to single-precision floating point. Besides the functions shown above, 2-, and 4-tuples are supported; for example:
float4 tex1Dfetch( texture<uchar4, cudaTextureType1D, cudaReadModeNormalizedFloat> texRef, int x);

fetches the region of linear memory bound to texture reference texRef using texture coordinate x.

B.8.2.2tex1D()
template<class DataType, enum cudaTextureReadMode readMode> Type tex1D(texture<DataType, cudaTextureType1D, readMode> texRef, float x);

fetches the CUDA array bound to the one-dimensional texture reference texRef using texture coordinate x.

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|93

C Language Extensions

B.8.2.3tex2D()
template<class DataType, enum cudaTextureReadMode readMode> Type tex2D(texture<DataType, cudaTextureType2D, readMode> texRef, float x, float y);

fetches the CUDA array or the region of linear memory bound to the two-dimensional texture reference texRef using texture coordinates x and y.

B.8.2.4tex3D()
template<class DataType, enum cudaTextureReadMode readMode> Type tex3D(texture<DataType, cudaTextureType3D, readMode> texRef, float x, float y, float z);

fetches the CUDA array bound to the three-dimensional texture reference texRef using texture coordinates x, y, and z.

B.8.2.5tex1Dlayered()
template<class DataType, enum cudaTextureReadMode readMode> Type tex1DLayered( texture<DataType, cudaTextureType1DLayered, readMode> texRef, float x, int layer);

fetches the CUDA array bound to the one-dimensional layered texture reference texRef using texture coordinate x and index layer, as described in Layered Textures.

B.8.2.6tex2Dlayered()
template<class DataType, enum cudaTextureReadMode readMode> Type tex2DLayered( texture<DataType, cudaTextureType2DLayered, readMode> texRef, float x, float y, int layer);

fetches the CUDA array bound to the two-dimensional layered texture reference texRef using texture coordinates x and y, and index layer, as described in Texture Memory

B.8.2.7texCubemap()
template<class DataType, enum cudaTextureReadMode readMode> Type texCubemap( texture<DataType, cudaTextureTypeCubemap, readMode> texRef, float x, float y, float z);

fetches the CUDA array bound to the cubemap texture reference texRef using texture coordinates x, y, and z, as described in Cubemap Textures.

B.8.2.8texCubemapLayered()
template<class DataType, enum cudaTextureReadMode readMode> Type texCubemapLayered( texture<DataType, cudaTextureTypeCubemapLayered, readMode> texRef, float x, float y, float z, int layer);

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|94

PG-02829-001_v5.0|97

C Language Extensions

B.9.1.12surfCubemapwrite()
template<class T> void surfCubemapwrite(T data, cudaSurfaceObject surfObj, int x, int y, int face, boundaryMode = cudaBoundaryModeTrap);

writes value data to the CUDA array specified by the cubemap object surfObj at coordinate x and y, and face index face.

B.9.1.13surfCubemapLayeredread()
template<class T> T surfCubemapLayeredread( cudaSurfaceObject surfObj, int x, int y, int layerFace, boundaryMode = cudaBoundaryModeTrap); template<class T> void surfCubemapLayeredread(T data, cudaSurfaceObject surfObj, int x, int y, int layerFace, boundaryMode = cudaBoundaryModeTrap);

reads the CUDA array specified by the cubemap layered surface object surfObj using coordinate x and y, and index layerFace.

B.9.1.14surfCubemapLayeredwrite()
template<class T> void surfCubemapLayeredwrite(T data, cudaSurfaceObject surfObj, int x, int y, int layerFace, boundaryMode = cudaBoundaryModeTrap);

writes value data to the CUDA array specified by the cubemap layered object surfObj at coordinate x and y, and index layerFace.

B.9.2Surface Reference API

A surface reference is declared at file scope as a variable of type surface:
surface<void, Type> surfRef;

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|98

C Language Extensions

Using the high-level API:

surfaceReference* surfRefPtr; cudaGetSurfaceReference(&surfRefPtr, "surfRef"); cudaChannelFormatDesc channelDesc; cudaGetChannelDesc(&channelDesc, cuArray); cudaBindSurfaceToArray(surfRef, cuArray, &channelDesc); surface<void, cudaSurfaceType2D> surfRef; cudaBindSurfaceToArray(surfRef, cuArray);

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|99

C Language Extensions

cudaMemcpyToArray(cuInputArray, 0, 0, h_data, size, cudaMemcpyHostToDevice); // Bind the arrays to the surface references cudaBindSurfaceToArray(inputSurfRef, cuInputArray); cudaBindSurfaceToArray(outputSurfRef, cuOutputArray); // Invoke kernel dim3 dimBlock(16, 16); dim3 dimGrid((width + dimBlock.x - 1) / dimBlock.x, (height + dimBlock.y - 1) / dimBlock.y); copyKernel<<<dimGrid, dimBlock>>>(width, height); // Free device memory cudaFreeArray(cuInputArray); cudaFreeArray(cuOutputArray); } return 0;

B.9.2.1surf1Dread()
template<class Type> Type surf1Dread(surface<void, cudaSurfaceType1D> surfRef, int x, boundaryMode = cudaBoundaryModeTrap); template<class Type> void surf1Dread(Type data, surface<void, cudaSurfaceType1D> surfRef, int x, boundaryMode = cudaBoundaryModeTrap);

reads the CUDA array bound to the one-dimensional surface reference surfRef using coordinate x.

B.9.2.2surf1Dwrite
template<class Type> void surf1Dwrite(Type data, surface<void, cudaSurfaceType1D> surfRef, int x, boundaryMode = cudaBoundaryModeTrap);

writes value data to the CUDA array bound to the one-dimensional surface reference surfRef at coordinate x.

B.9.2.3surf2Dread()
template<class Type> Type surf2Dread(surface<void, cudaSurfaceType2D> surfRef, int x, int y, boundaryMode = cudaBoundaryModeTrap); template<class Type> void surf2Dread(Type* data, surface<void, cudaSurfaceType2D> surfRef, int x, int y, boundaryMode = cudaBoundaryModeTrap);

reads the CUDA array bound to the two-dimensional surface reference surfRef using coordinates x and y.

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|100

C Language Extensions

B.9.2.4surf2Dwrite()
template<class Type> void surf3Dwrite(Type data, surface<void, cudaSurfaceType3D> surfRef, int x, int y, int z, boundaryMode = cudaBoundaryModeTrap);

writes value data to the CUDA array bound to the two-dimensional surface reference surfRef at coordinate x and y.

B.9.2.5surf3Dread()
template<class Type> Type surf3Dread(surface<void, cudaSurfaceType3D> surfRef, int x, int y, int z, boundaryMode = cudaBoundaryModeTrap); template<class Type> void surf3Dread(Type* data, surface<void, cudaSurfaceType3D> surfRef, int x, int y, int z, boundaryMode = cudaBoundaryModeTrap);

reads the CUDA array bound to the three-dimensional surface reference surfRef using coordinates x, y, and z.

B.9.2.6surf3Dwrite()
template<class Type> void surf3Dwrite(Type data, surface<void, cudaSurfaceType3D> surfRef, int x, int y, int z, boundaryMode = cudaBoundaryModeTrap);

writes value data to the CUDA array bound to the three-dimensional surface reference surfRef at coordinate x, y, and z.

B.9.2.7surf1DLayeredread()
template<class Type> Type surf1DLayeredread( surface<void, cudaSurfaceType1DLayered> surfRef, int x, int layer, boundaryMode = cudaBoundaryModeTrap); template<class Type> void surf1DLayeredread(Type data, surface<void, cudaSurfaceType1DLayered> surfRef, int x, int layer, boundaryMode = cudaBoundaryModeTrap);

reads the CUDA array bound to the one-dimensional layered surface reference surfRef using coordinate x and index layer.

B.9.2.8surf1DLayeredwrite()
template<class Type> void surf1DLayeredwrite(Type data, surface<void, cudaSurfaceType1DLayered> surfRef, int x, int layer,

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|101

C Language Extensions

boundaryMode = cudaBoundaryModeTrap);

writes value data to the CUDA array bound to the two-dimensional layered surface reference surfRef at coordinate x and index layer.

B.9.2.9surf2DLayeredread()
template<class Type> Type surf2DLayeredread( surface<void, cudaSurfaceType2DLayered> surfRef, int x, int y, int layer, boundaryMode = cudaBoundaryModeTrap); template<class Type> void surf2DLayeredread(Type data, surface<void, cudaSurfaceType2DLayered> surfRef, int x, int y, int layer, boundaryMode = cudaBoundaryModeTrap);

reads the CUDA array bound to the two-dimensional layered surface reference surfRef using coordinate x and y, and index layer.

B.9.2.10surf2DLayeredwrite()
template<class Type> void surf2DLayeredwrite(Type data, surface<void, cudaSurfaceType2DLayered> surfRef, int x, int y, int layer, boundaryMode = cudaBoundaryModeTrap);

writes value data to the CUDA array bound to the one-dimensional layered surface reference surfRef at coordinate x and y, and index layer.

B.9.2.11surfCubemapread()
template<class Type> Type surfCubemapread( surface<void, cudaSurfaceTypeCubemap> surfRef, int x, int y, int face, boundaryMode = cudaBoundaryModeTrap); template<class Type> void surfCubemapread(Type data, surface<void, cudaSurfaceTypeCubemap> surfRef, int x, int y, int face, boundaryMode = cudaBoundaryModeTrap);

reads the CUDA array bound to the cubemap surface reference surfRef using coordinate x and y, and face index face.

B.9.2.12surfCubemapwrite()
template<class Type> void surfCubemapwrite(Type data, surface<void, cudaSurfaceTypeCubemap> surfRef, int x, int y, int face, boundaryMode = cudaBoundaryModeTrap);

writes value data to the CUDA array bound to the cubemap reference surfRef at coordinate x and y, and face index face.

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|102

C Language Extensions

B.9.2.13surfCubemapLayeredread()
template<class Type> Type surfCubemapLayeredread( surface<void, cudaSurfaceTypeCubemapLayered> surfRef, int x, int y, int layerFace, boundaryMode = cudaBoundaryModeTrap); template<class Type> void surfCubemapLayeredread(Type data, surface<void, cudaSurfaceTypeCubemapLayered> surfRef, int x, int y, int layerFace, boundaryMode = cudaBoundaryModeTrap);

reads the CUDA array bound to the cubemap layered surface reference surfRef using coordinate x and y, and index layerFace.

B.9.2.14surfCubemapLayeredwrite()
template<class Type> void surfCubemapLayeredwrite(Type data, surface<void, cudaSurfaceTypeCubemapLayered> surfRef, int x, int y, int layerFace, boundaryMode = cudaBoundaryModeTrap);

PG-02829-001_v5.0|106

C Language Extensions

The 64-bit version of atomicAnd() is only supported by devices of compute capability 3.5 and higher.

B.11.2.2atomicOr()
int atomicOr(int* address, int val); unsigned int atomicOr(unsigned int* address, unsigned int val); unsigned long long int atomicOr(unsigned long long int* address, unsigned long long int val);

reads the 32-bit or 64-bit word old located at the address address in global or shared memory, computes (old | val), and stores the result back to memory at the same address. These three operations are performed in one atomic transaction. The function returns old. The 64-bit version of atomicOr() is only supported by devices of compute capability 3.5 and higher.

B.11.2.3atomicXor()
int atomicXor(int* address, int val); unsigned int atomicXor(unsigned int* address, unsigned int val); unsigned long long int atomicXor(unsigned long long int* address, unsigned long long int val);

reads the 32-bit or 64-bit word old located at the address address in global or shared memory, computes (old ^ val), and stores the result back to memory at the same address. These three operations are performed in one atomic transaction. The function returns old. The 64-bit version of atomicXor() is only supported by devices of compute capability 3.5 and higher.

B.12Warp Vote Functions

Warp vote functions are only supported by devices of compute capability 1.2 and higher (see SIMT Architecture for the definition of a warp).
int __all(int predicate);

evaluates predicate for all active threads of the warp and returns non-zero if and only if predicate evaluates to non-zero for all of them.
int __any(int predicate);

evaluates predicate for all active threads of the warp and returns non-zero if and only if predicate evaluates to non-zero for any of them.
unsigned int __ballot(int predicate);

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|107

C Language Extensions

evaluates predicate for all active threads of the warp and returns an integer whose Nth bit is set if and only if predicate evaluates to non-zero for the Nth thread of the warp. This function is only supported by devices of compute capability 2.x and higher.

B.13Warp Shuffle Functions

__shfl, __shfl_up, __shfl_down, __shfl_xor exchange a variable between threads within a warp. They are only supported by devices of compute capability 3.x (see SIMT Architecture for the definition of a warp).

B.13.1Synopsis
int int int int __shfl(int var, int srcLane, int width=warpSize); __shfl_up(int var, unsigned int delta, int width=warpSize); __shfl_down(int var, unsigned int delta, int width=warpSize); __shfl_xor(int var, int laneMask, int width=warpSize);

float __shfl(float var, int srcLane, int width=warpSize); float __shfl_up(float var, unsigned int delta, int width=warpSize); float __shfl_down(float var, unsigned int delta, int width=warpSize); float __shfl_xor(float var, int laneMask, int width=warpSize);

B.13.2Description
The __shfl() intrinsics permit exchanging of a variable between threads within a warp without use of shared memory. The exchange occurs simultaneously for all active threads within the warp, moving 4 bytes of data per thread. Exchange of 8-byte quantities must be broken into two separate invocations of __shfl(). Threads within a warp are referred to as lanes, and for devices of compute capability 3.x may have an index between 0 and warpSize-1 (inclusive). Four source-lane addressing modes are supported: __shfl() Direct copy from indexed lane __shfl_up() Copy from a lane with lower ID relative to caller __shfl_down() Copy from a lane with higher ID relative to caller __shfl_xor() Copy from a lane based on bitwise XOR of own lane ID Threads may only read data from another thread which is actively participating in the __shfl() command. If the target thread is inactive, the retrieved value is undefined. All the __shfl() intrinsics take an optional width parameter which permits subdivision of the warp into segments for example to exchange data between 4 groups of

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|108

C Language Extensions

8 lanes in a SIMD manner. If width is less than warpSize then each subsection of the warp behaves as a separate entity with a starting logical lane ID of 0. A thread may only exchange data with others in its own subsection. width must have a value which is a power of 2 so that the warp can be subdivided equally; results are undefined if width is not a power of 2, or is a number greater than warpSize. __shfl() returns the value of var held by the thread whose ID is given by srcLane. If srcLane is outside the range [0:width-1], then the thread's own value of var is returned. __shfl_up() calculates a source lane ID by subtracting delta from the caller's lane ID. The value of var held by the resulting lane ID is returned: in effect, var is shifted up the warp by delta lanes. The source lane index will not wrap around the value of width, so effectively the lower delta lanes will be unchanged. __shfl_down() calculates a source lane ID by adding delta to the caller's lane ID. The value of var held by the resulting lane ID is returned: this has the effect of shifting var down the warp by delta lanes. As for __shfl_up(), the ID number of the source lane will not wrap around the value of width and so the upper delta lanes will remain unchanged. __shfl_xor() calculates a source line ID by performing a bitwise XOR of the caller's lane ID with laneMask: the value of var held by the resulting lane ID is returned. If the resulting lane ID falls outside the range permitted by width, the thread's own value of var is returned. This mode implements a butterfly addressing pattern such as is used in tree reduction and broadcast.

B.13.3Return Value
All __shfl() intrinsics return the 4-byte word referenced by var from the source lane ID as an unsigned integer. If the source lane ID is out of range or the source thread has exited, the calling threads own var is returned.

B.13.4Notes
All __shfl() intrinsics share the same semantics with respect to code motion as the vote intrinsics __any() and __all(). Threads may only read data from another thread which is actively participating in the __shfl() command. If the target thread is inactive, the retrieved value is undefined. width must be a power-of-2 (i.e. 2, 4, 8, 16 or 32). Results are unspecified for other values. Types other than int or float must first be cast in order to use the __shfl() intrinsics.

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|109

C Language Extensions

B.13.5Examples
B.13.5.1Broadcast of a single value across a warp
__global__ void bcast(int arg) { int laneId = threadIdx.x & 0x1f; int value; if (laneId == 0) // Note unused variable for value = arg; // all threads except lane 0 value = __shfl(value, 0); // Get value from lane 0 if (value != arg) printf(Thread %d failed.\n, threadIdx.x);

void main() { bcast<<< 1, 32 >>>(1234); cudaDeviceSynchronize(); }

B.13.5.2Inclusive plus-scan across sub-partitions of 8 threads

__global__ void scan4() { // Seed sample starting value (inverse of lane ID) int value = 31 laneId; // Loop to accumulate scan within my partition. // Scan requires log2(n) == 3 steps for 8 threads // It works by an accumulated sum up the warp // by 1, 2, 4, 8 etc. steps. for (int i=1; i<=4; i*=2) { // Note: shfl requires all threads being // accessed to be active. Therefore we do // the __shfl unconditionally so that we // can read even from threads which wont do a // sum, and then conditionally assign the result. int n = __shfl_up(value, i, 8); if (laneId >= i) value += n; } } printf(Thread %d final value = %d\n, threadIdx.x, value);

void main() { scan4<<< 1, 32 >>>(); cudaDeviceSynchronize(); }

B.13.5.3Reduction across a warp

__global__ void warpReduce() { // Seed starting value as inverse lane ID int value = 31 laneId; // Use XOR mode to perform butterfly reduction for (int i=16; i>=1; i/=2) value += __shfl_xor(value, i, 32); // value now contains the sum across all threads printf(Thread %d final value = %d\n, threadIdx.x, value);

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|110

C Language Extensions

} void main() { warpReduce<<< 1, 32 >>>(); cudaDeviceSynchronize(); }

B.14Profiler Counter Function

Each multiprocessor has a set of sixteen hardware counters that an application can increment with a single instruction by calling the __prof_trigger() function.
[void __prof_trigger(int counter);

increments by one per warp the per-multiprocessor hardware counter of index counter. Counters 8 to 15 are reserved and should not be used by applications. The value of counters 0, 1, , 7 for the first multiprocessor can be obtained via the CUDA profiler by listing prof_trigger_00, prof_trigger_01, , prof_trigger_07 ,etc. in the profiler.conf file (see the profiler manual for more details). All counters are reset before each kernel launch (note that when collecting counters, kernel launches are synchronous as mentioned in Concurrent Execution between Host and Device).

B.15Assertion
Assertion is only supported by devices of compute capability 2.x and higher.
void assert(int expression);

stops the kernel execution if expression is equal to zero. If the program is run within a debugger, this triggers a breakpoint and the debugger can be used to inspect the current state of the device. Otherwise, each thread for which expression is equal to zero prints a message to stderr after synchronization with the host via cudaDeviceSynchronize(), cudaStreamSynchronize(), or cudaEventSynchronize(). The format of this message is as follows:
<filename>:<line number>:<function>: block: [blockId.x,blockId.x,blockIdx.z], thread: [threadIdx.x,threadIdx.y,threadIdx.z] Assertion `<expression>` failed.

Any subsequent host-side synchronization calls made for the same device will return cudaErrorAssert. No more commands can be sent to this device until cudaDeviceReset() is called to reinitialize the device. If expression is different from zero, the kernel execution is unaffected. For example, the following program from source file test.cu
#include <assert.h> // assert() is only supported // for devices of compute capability 2.0 and higher

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|111

C Language Extensions

#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ < 200) #undef assert #define assert(arg) #endif __global__ void testAssert(void) { int is_one = 1; int should_be_one = 0; // This will have no effect assert(is_one); // This will halt kernel execution assert(should_be_one);

int main(int argc, char* argv[]) { testAssert<<<1,1>>>(); cudaDeviceSynchronize(); return 0; }

will output:
test.cu:19: void testAssert(): block: [0,0,0], thread: [0,0,0] Assertion `should_be_one` failed.

Assertions are for debugging purposes. They can affect performance and it is therefore recommended to disable them in production code. They can be disabled at compile time by defining the NDEBUG preprocessor macro before including assert.h. Note that expression should not be an expression with side effects (something like (++i > 0), for example), otherwise disabling the assertion will affect the functionality of the code.

B.16Formatted Output
Formatted output is only supported by devices of compute capability 2.x and higher.
int printf(const char *format[, arg, ...]);

prints formatted output from a kernel to a host-side output stream. The in-kernel printf() function behaves in a similar way to the standard C-library printf() function, and the user is referred to the host systems manual pages for a complete description of printf() behavior. In essence, the string passed in as format is output to a stream on the host, with substitutions made from the argument list wherever a format specifier is encountered. Supported format specifiers are listed below. The printf() command is executed as any other device-side function: per-thread, and in the context of the calling thread. From a multi-threaded kernel, this means that a straightforward call to printf() will be executed by every thread, using that threads data as specified. Multiple versions of the output string will then appear at the host stream, once for each thread which encountered the printf().

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|112

C Language Extensions

It is up to the programmer to limit the output to a single thread if only a single output string is desired (see Examples for an illustrative example). Unlike the C-standard printf(), which returns the number of characters printed, CUDAs printf() returns the number of arguments parsed. If no arguments follow the format string, 0 is returned. If the format string is NULL, -1 is returned. If an internal error occurs, -2 is returned.

B.16.1Format Specifiers
As for standard printf(), format specifiers take the form: %[flags][width] [.precision][size]type The following fields are supported (see widely-available documentation for a complete description of all behaviors): Flags: # 0 + - Width: * 0-9 Precision: 0-9 Size: h l ll Type: %cdiouxXpeEfgGaAs

Note that CUDAs printf()will accept any combination of flag, width, precision, size and type, whether or not overall they form a valid format specifier. In other words, %hd will be accepted and printf will expect a double-precision variable in the corresponding location in the argument list.

B.16.2Limitations
Final formatting of the printf() output takes place on the host system. This means that the format string must be understood by the host-systems compiler and C library. Every effort has been made to ensure that the format specifiers supported by CUDAs printf function form a universal subset from the most common host compilers, but exact behavior will be host-O/S-dependent. As described in Format Specifiers, printf() will accept all combinations of valid flags and types. This is because it cannot determine what will and will not be valid on the host system where the final output is formatted. The effect of this is that output may be undefined if the program emits a format string which contains invalid combinations. The printf() command can accept at most 32 arguments in addition to the format string. Additional arguments beyond this will be ignored, and the format specifier output as-is. Owing to the differing size of the long type on 64-bit Windows platforms (four bytes on 64-bit Windows platforms, eight bytes on other 64-bit platforms), a kernel which is compiled on a non-Windows 64-bit machine but then run on a win64 machine will see

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|113

C Language Extensions

corrupted output for all format strings which include %ld. It is recommended that the compilation platform matches the execution platform to ensure safety. The output buffer for printf() is set to a fixed size before kernel launch (see Associated Host-Side API). It is circular and if more output is produced during kernel execution than can fit in the buffer, older output is overwritten. It is flushed only when one of these actions is performed: Kernel launch via <<<>>> or cuLaunchKernel() (at the start of the launch, and if the CUDA_LAUNCH_BLOCKING environment variable is set to 1, at the end of the launch as well), Synchronization via cudaDeviceSynchronize(), cuCtxSynchronize(), cudaStreamSynchronize(), cuStreamSynchronize(), cudaEventSynchronize(), or cuEventSynchronize(), Memory copies via any blocking version of cudaMemcpy*() or cuMemcpy*(), Module loading/unloading via cuModuleLoad() or cuModuleUnload(), Context destruction via cudaDeviceReset() or cuCtxDestroy(). Note that the buffer is not flushed automatically when the program exits. The user must call cudaDeviceReset() or cuCtxDestroy() explicitly, as shown in the examples below.

B.16.3Associated Host-Side API

The following API functions get and set the size of the buffer used to transfer the printf() arguments and internal metadata to the host (default is 1 megabyte): cudaDeviceGetLimit(size_t* size,cudaLimitPrintfFifoSize) cudaDeviceSetLimit(cudaLimitPrintfFifoSize, size_t size)

B.16.4Examples
The following code sample:
#include "stdio.h" // printf() is only supported // for devices of compute capability 2.0 and higher #if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ < 200) #define printf(f, ...) ((void)(f, __VA_ARGS__),0) #endif __global__ void helloCUDA(float f) { printf("Hello thread %d, f=%f\n", threadIdx.x, f); } int main() { helloCUDA<<<1, 5>>>(1.2345f); cudaDeviceSynchronize(); return 0; }

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|114

C Language Extensions

will output: Notice how each thread encounters the printf() command, so there are as many lines of output as there were threads launched in the grid. As expected, global values (i.e., float f) are common between all threads, and local values (i.e., threadIdx.x) are distinct per-thread. The following code sample:
#include "stdio.h" // printf() is only supported // for devices of compute capability 2.0 and higher #if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ < 200) #define printf(f, ...) ((void)(f, __VA_ARGS__),0) #endif __global__ void helloCUDA(float f) { if (threadIdx.x == 0) printf("Hello thread %d, f=%f\n", threadIdx.x, f) ; } int main() { helloCUDA<<<1, 5>>>(1.2345f); cudaDeviceSynchronize(); return 0; }

will output: Self-evidently, the if() statement limits which threads will call printf, so that only a single line of output is seen.

B.17Dynamic Global Memory Allocation and Operations

Dynamic global memory allocation and operations are only supported by devices of compute capability 2.x and higher.
void* malloc(size_t size); void free(void* ptr);

allocate and free memory dynamically from a fixed-size heap in global memory.
void* memcpy(void* dest, const void* src, size_t size);

copy size bytes from the memory location pointed by src to the memory location pointed by dest.
void* memset(void* ptr, int value, size_t size);

set size bytes of memory block pointed by ptr to value (interpreted as an unsigned char). The CUDA in-kernel malloc() function allocates at least size bytes from the device heap and returns a pointer to the allocated memory or NULL if insufficient memory

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|115

C Language Extensions

exists to fulfill the request. The returned pointer is guaranteed to be aligned to a 16-byte boundary. The CUDA in-kernel free() function deallocates the memory pointed to by ptr, which must have been returned by a previous call to malloc(). If ptr is NULL, the call to free() is ignored. Repeated calls to free() with the same ptr has undefined behavior. The memory allocated by a given CUDA thread via malloc() remains allocated for the lifetime of the CUDA context, or until it is explicitly released by a call to free(). It can be used by any other CUDA threads even from subsequent kernel launches. Any CUDA thread may free memory allocated by another thread, but care should be taken to ensure that the same pointer is not freed more than once.

B.17.1Heap Memory Allocation

The device memory heap has a fixed size that must be specified before any program using malloc() or free() is loaded into the context. A default heap of eight megabytes is allocated if any program uses malloc() without explicitly specifying the heap size. The following API functions get and set the heap size: cudaDeviceGetLimit(size_t* size, cudaLimitMallocHeapSize) cudaDeviceSetLimit(cudaLimitMallocHeapSize, size_t size) The heap size granted will be at least size bytes. cuCtxGetLimit()and cudaDeviceGetLimit() return the currently requested heap size. The actual memory allocation for the heap occurs when a module is loaded into the context, either explicitly via the CUDA driver API (see Module), or implicitly via the CUDA runtime API (see CUDA C Runtime). If the memory allocation fails, the module load will generate a CUDA_ERROR_SHARED_OBJECT_INIT_FAILED error. Heap size cannot be changed once a module load has occurred and it does not resize dynamically according to need. Memory reserved for the device heap is in addition to memory allocated through hostside CUDA API calls such as cudaMalloc().

B.17.2Interoperability with Host Memory API

Memory allocated via malloc() cannot be freed using the runtime (i.e., by calling any of the free memory functions from Device Memory). Similarly, memory allocated via the runtime (i.e., by calling any of the memory allocation functions from Device Memory) cannot be freed via free().

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|116

C Language Extensions

B.17.3Examples
B.17.3.1Per Thread Allocation
The following code sample:
#include <stdlib.h> #include <stdio.h> __global__ void mallocTest() { size_t size = 123; char* ptr = (char*)malloc(size); memset(ptr, 0, size); printf("Thread %d got pointer: %p\n", threadIdx.x, ptr); free(ptr); } int main() { // Set a heap size of 128 megabytes. Note that this must // be done before any kernel is launched. cudaDeviceSetLimit(cudaLimitMallocHeapSize, 128*1024*1024); mallocTest<<<1, 5>>>(); cudaDeviceSynchronize(); return 0; }

will output:
Thread Thread Thread Thread Thread 0 1 2 3 4 got got got got got pointer: pointer: pointer: pointer: pointer: 00057020 0005708c 000570f8 00057164 000571d0

Notice how each thread encounters the malloc() and memset() commands and so receives and initializes its own allocation. (Exact pointer values will vary: these are illustrative.)

B.17.3.2Per Thread Block Allocation

#include <stdlib.h> __global__ void mallocTest() { __shared__ int* data; // // // // if The first thread in the block does the allocation and initialization and then shares the pointer with all other threads through shared memory, so that access can easily be coalesced. 64 bytes per thread are allocated. (threadIdx.x == 0) { size_t size = blockDim.x * 64; data = (int*)malloc(size); memset(data, 0, size);

} __syncthreads();

// Check for failure

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|117

C Language Extensions

if (data == NULL) return; // Threads index into the memory, ensuring coalescence int* ptr = data; for (int i = 0; i < 64; ++i) ptr[i * blockDim.x + threadIdx.x] = threadIdx.x; // Ensure all threads complete before freeing __syncthreads(); // Only one thread may free the memory! if (threadIdx.x == 0) free(data);

int main() { cudaDeviceSetLimit(cudaLimitMallocHeapSize, 12810241024); mallocTest<<<10, 128>>>(); cudaDeviceSynchronize(); return 0; }

B.17.3.3Allocation Persisting Between Kernel Launches

#include <stdlib.h> #include <stdio.h> #define NUM_BLOCKS 20 __device__ int* dataptr[NUM_BLOCKS]; // Per-block pointer __global__ void allocmem() { // Only the first thread in the block does the allocation // since we want only one allocation per block. if (threadIdx.x == 0) dataptr[blockIdx.x] = (int*)malloc(blockDim.x * 4); __syncthreads(); // Check for failure if (dataptr[blockIdx.x] == NULL) return; // Zero the data with all threads in parallel dataptr[blockIdx.x][threadIdx.x] = 0;

// Simple example: store thread ID into each element __global__ void usemem() { int* ptr = dataptr[blockIdx.x]; if (ptr != NULL) ptr[threadIdx.x] += threadIdx.x; } // Print the content of the buffer before freeing it __global__ void freemem() { int* ptr = dataptr[blockIdx.x]; if (ptr != NULL) printf("Block %d, Thread %d: final value = %d\n", blockIdx.x, threadIdx.x, ptr[threadIdx.x]); // Only free from one thread! if (threadIdx.x == 0)

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|118

C Language Extensions

CUDA C Programming Guide

PG-02829-001_v5.0|121

C Language Extensions

(deviceProp.major >= 2 ? 2 * THREADS_PER_BLOCK : THREADS_PER_BLOCK); MyKernel<<<blocksPerGrid, threadsPerBlock>>>(...);

Register usage is reported by the --ptxas options=-v compiler option. The number of resident blocks can be derived from the occupancy reported by the CUDA profiler (see Device Memory Accessesfor a definition of occupancy). Register usage can also be controlled for all __global__ functions in a file using the maxrregcount compiler option. The value of maxrregcount is ignored for functions with launch bounds.

B.20#pragma unroll
By default, the compiler unrolls small loops with a known trip count. The #pragma unroll directive however can be used to control unrolling of any given loop. It must be placed immediately before the loop and only applies to that loop. It is optionally followed by a number that specifies how many times the loop must be unrolled. For example, in this code sample:
#pragma unroll 5 for (int i = 0; i < n; ++i)

the loop will be unrolled 5 times. The compiler will also insert code to ensure correctness (in the example above, to ensure that there will only be n iterations if n is less than 5, for example). It is up to the programmer to make sure that the specified unroll number gives the best performance. #pragma unroll 1 will prevent the compiler from ever unrolling a loop. If no number is specified after #pragma unroll, the loop is completely unrolled if its trip count is constant, otherwise it is not unrolled at all.

B.21SIMD Video Instructions

PTX ISA version 3.0 includes SIMD (Single Instruction, Multiple Data) video instructions which operate on pairs of 16-bit values and quads of 8-bit values. These are available on devices of compute capability 3.0. The SIMD video instructions are: vadd2, vadd4 vsub2, vsub4 vavrg2, vavrg4 vabsdiff2, vabsdiff4 vmin2, vmin4 vmax2, vmax4

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|122

C Language Extensions

vset2, vset4 PTX instructions, such as the SIMD video instructions, can be included in CUDA programs by way of the assembler, asm(), statement. The basic syntax of an asm() statement is:
asm("template-string" : "constraint"(output) : "constraint"(input)"));

An example of using the vabsdiff4 PTX instruction is:

asm("vabsdiff4.u32.u32.u32.add" " %0, %1, %2, %3;": "=r" (result):"r" (A), "r" (B), "r" (C));

This uses the vabsdiff4 instruction to compute an integer quad byte SIMD sum of absolute differences. The absolute difference value is computed for each byte of the unsigned integers A and B in SIMD fashion. The optional accumulate operation (.add) is specified to sum these differences. Refer to the document "Using Inline PTX Assembly in CUDA" for details on using the assembly statement in your code. Refer to the PTX ISA documentation ("Parallel Thread Execution ISA Version 3.0" for example) for details on the PTX instructions for the version of PTX that you are using.

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|123

Appendix C. MATHEMATICAL FUNCTIONS

The reference manual lists, along with their description, all the functions of the C/C++ standard library mathematical functions that are supported in device code, as well as all intrinsic functions (that are only supported in device code). This appendix provides accuracy information for some of these functions when applicable.

C.1Standard Functions
The functions from this section can be used in both host and device code. This section specifies the error bounds of each function when executed on the device and also when executed on the host in the case where the host does not supply the function. The error bounds are generated from extensive but not exhaustive tests, so they are not guaranteed bounds. Single-Precision Floating-Point Functions Addition and multiplication are IEEE-compliant, so have a maximum error of 0.5 ulp. However, on the device, the compiler often combines them into a single multiplyadd instruction (FMAD) and for devices of compute capability 1.x, FMAD truncates the intermediate result of the multiplication as mentioned in Floating-Point Standard. This combination can be avoided by using the __fadd_[rn,rz,ru,rd]() and __fmul_[rn,rz,ru,rd]() intrinsic functions (see Intrinsic Functions). The recommended way to round a single-precision floating-point operand to an integer, with the result being a single-precision floating-point number is rintf(), not roundf(). The reason is that roundf() maps to an 8-instruction sequence on the device, whereas rintf() maps to a single instruction. truncf(), ceilf(), and floorf() each map to a single instruction as well.

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|124

Mathematical Functions

Table4Mathematical Standard Library Functions with Maximum ULP Error

The maximum error is stated as the absolute value of the difference in ulps between a correctly rounded single-precision result and the result returned by the CUDA library function.
Function Maximum ulp error
0 (IEEE-754 round-to-nearest-even) (except for devices of compute capability 1.x when addition is merged into an FMAD)

x+y

x*y

0 (IEEE-754 round-to-nearest-even) (except for devices of compute capability 1.x when multiplication is merged into an FMAD)

x/y

0 for compute capability > 2 when compiled with -precdiv=true 2 (full range), otherwise

1/x

0 for compute capability > 2 when compiled with -precdiv=true 1 (full range), otherwise

rsqrtf(x) 1/sqrtf(x) sqrtf(x)

2 (full range)

1/sqrtf(x) only when it is converted to rsqrtf(x) by the compiler.

Applies to 0 for compute capability > 2 when compiled with -precsqrt=true 3 (full range), otherwise

cbrtf(x) rcbrtf(x) hypotf(x,y) expf(x) exp2f(x) exp10f(x) expm1f(x) logf(x)

1 (full range) 2 (full range) 3 (full range) 2 (full range) 2 (full range) 2 (full range) 1 (full range) 1 (full range)

www.nvidia.com

Maximum ulp error

0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range)

nearbyintf(x) ceilf(x) floorf(x) lrintf(x) lroundf(x) llrintf(x) llroundf(x)

Double-Precision Floating-Point Functions The errors listed below only apply when compiling for devices with native doubleprecision support. When compiling for devices without such support, such as devices of compute capability 1.2 and lower, the double type gets demoted to float by default and the double-precision math functions are mapped to their single-precision equivalents. The recommended way to round a double-precision floating-point operand to an integer, with the result being a double-precision floating-point number is rint(), not round(). The reason is that round() maps to an 8-instruction sequence on the device, whereas rint() maps to a single instruction. trunc(), ceil(), and floor() each map to a single instruction as well.

Table5Mathematical Standard Library Functions with Maximum ULP Error

The maximum error is stated as the absolute value of the difference in ulps between a correctly rounded double-precision result and the result returned by the CUDA library function.
Function Maximum ulp error
0 (IEEE-754 round-to-nearest-even) 0 (IEEE-754 round-to-nearest-even) 0 (IEEE-754 round-to-nearest-even) 0(IEEE-754 round-to-nearest-even) 0 (IEEE-754 round-to-nearest-even) 1 (full range)

x+y x*y x/y 1/x sqrt(x) rsqrt(x)

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|128

Mathematical Functions

Function

Maximum ulp error

1 (full range) 1 (full range) 2 (full range) 1 (full range) 1 (full range) 1 (full range) 1 (full range) 1 (full range) 1 (full range) 1 (full range) 1 (full range) 1 (full range) 1 (full range) 2 (full range) 1 (full range) 1 (full range) 1 (full range) 1 (full range) 2 (full range) 2 (full range) 2 (full range) 2 (full range) 1 (full range) 1 (full range) 1 (full range) 2 (full range) 2 (full range) 2 (full range)

cbrt(x) rcbrt(x) hypot(x,y) exp(x) exp2(x) exp10(x) expm1(x) log(x) log2(x) log10(x) log1p(x) sin(x) cos(x) tan(x) sincos(x,sptr,cptr) sinpi(x) cospi(x) sincospi(x,sptr,cptr) asin(x) acos(x) atan(x) atan2(y,x) sinh(x) cosh(x) tanh(x) asinh(x) acosh(x) atanh(x)

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|129

Mathematical Functions

Function

Maximum ulp error

2 (full range) 2 (full range) 4 (full range) 5 (full range) 8 (full range) 3 (full range) 5 (full range) 8 (full range) 4 (outside interval -11.0001 ... -2.2637; larger inside) 8 (full range) 0 (IEEE-754 round-to-nearest-even) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 7 for |x| < 8 otherwise, the maximum absolute error is 5 x 10
-12

pow(x,y) erf(x) erfc(x) erfinv(x) erfcinv(x) erfcx(x) normcdf(x) normcdfinv(x) lgamma(x) tgamma(x) fma(x,y,z) frexp(x,exp) ldexp(x,exp) scalbn(x,n) scalbln(x,l) logb(x) ilogb(x) j0(x)

Maximum ulp error

otherwise, the maximum absolute error is 5 x 10
-12

yn(x) fmod(x,y) remainder(x,y) remquo(x,y,iptr) mod(x,iptr) fdim(x,y) trunc(x) round(x) rint(x) nearbyint(x) ceil(x) floor(x) lrint(x) lround(x) llrint(x) llround(x)

For |x| > 1.5n, the maximum absolute error is 5 x 10

-12

0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range) 0 (full range)

C.2Intrinsic Functions
The functions from this section can only be used in device code. Among these functions are the less accurate, but faster versions of some of the functions of Standard Functions. They have the same name prefixed with __ (such as __sinf(x)). They are faster as they map to fewer native instructions. The compiler has an option (-use_fast_math) that forces each function in Table 6 Functions Affected by -use_fast_math to compile to its intrinsic counterpart. In addition to reducing the accuracy of the affected functions, it may also cause some differences in special case handling. A more robust approach is to selectively replace mathematical function calls by calls to intrinsic functions only where it is merited by the performance gains and where changed properties such as reduced accuracy and different special case handling can be tolerated.

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|131

Mathematical Functions

Table6Functions Affected by -use_fast_math

Operator/Function Device Function

x/y sinf(x) cosf(x) tanf(x) logf(x) sincosf(x,sptr,cptr) log2f(x) expf(x)

fdividef(x,y) sinf(x) cosf(x) tanf(x) logf(x) sincosf(x,sptr,cptr) log2f(x) expf(x)

log10f(x) exp10f(x) powf(x,y)

log10f(x) exp10f(x) __powf(x,y)

Functions suffixed with _rn operate using the round to nearest even rounding mode. Functions suffixed with _rz operate using the round towards zero rounding mode. Functions suffixed with _ru operate using the round up (to positive infinity) rounding mode. Functions suffixed with _rd operate using the round down (to negative infinity) rounding mode. Single-Precision Floating-Point Functions __fadd_[rn,rz,ru,rd]() and __fmul_[rn,rz,ru,rd]() map to addition and multiplication operations that the compiler never merges into FMADs. By contrast, additions and multiplications generated from the '*' and '+' operators will frequently be combined into FMADs. The accuracy of floating-point division varies depending on the compute capability of the device and whether the code is compiled with -prec-div=false or -precdiv=true. For devices of compute capability 1.x or for devices of compute capability 2.x and higher when the code is compiled with -prec-div=false, both the regular division / operator and __fdividef(x,y) have the same accuracy, but for 2126 < y < 2128, __fdividef(x,y) delivers a result of zero, whereas the / operator delivers the correct result to within the accuracy stated in Table 7 Single-Precision Floating-Point Intrinsic Functions. Also, for 2126 < y < 2128, if x is infinity, __fdividef(x,y) delivers a NaN (as a result of multiplying infinity by zero), while the / operator returns infinity. On the other hand, the / operator is IEEE-compliant on devices of compute capability 2.x and higher when the code is compiled with -prec-div=true or without any -precdiv option at all since its default value is true.

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|132

Mathematical Functions

Table7Single-Precision Floating-Point Intrinsic Functions

(Supported by the CUDA Runtime Library with Respective Error Bounds)
Function Error bounds
IEEE-compliant. IEEE-compliant. IEEE-compliant. IEEE-compliant. IEEE-compliant. IEEE-compliant. IEEE-compliant. For

__fadd_[rn,rz,ru,rd](x,y) __fmul_[rn,rz,ru,rd](x,y) __fmaf_[rn,rz,ru,rd](x,y,z) __frcp_[rn,rz,ru,rd](x) __fsqrt_[rn,rz,ru,rd](x) __frsqrt_rn(x) __fdiv_[rn,rz,ru,rd](x,y) __fdividef(x,y) __expf(x) __exp10f(x) __logf(x) __log2f(x) __log10f(x) __sinf(x) __cosf(x) __sincosf(x,sptr,cptr) __tanf(x) __powf(x, y)

y in [2-126, 2126], the maximum ulp error is 2.

2 + floor(abs(1.16 * x)).
The maximum ulp error is

2 + floor(abs(2.95 * x)).
The maximum ulp error is For 2
-21.41

x in [0.5, 2], the maximum absolute error is

, otherwise, the maximum ulp error is 3.

For 2
-22

, otherwise, the maximum ulp error is 2.

x in [0.5, 2], the maximum absolute error is x in [0.5, 2], the maximum absolute error is x in [-,], the maximum absolute error is
, and larger otherwise.

For 2
-24

, otherwise, the maximum ulp error is 3.

For 2

-21.41

For 2

-21.19

x in [-,], the maximum absolute error is

, and larger otherwise.

Same as

sinf(x) and cosf(x). __sinf(x) exp2f(y *

Derived from its implementation as

* (1/__cosf(x)).
Derived from its implementation as

__log2f(x)).

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|133

Mathematical Functions

PG-02829-001_v5.0|137

C/C++ Language Support

In the following example, the instantiation kern<int> only occurs when __CUDA_ARCH__ is not defined. This is not supported.
__device__ int result; template <typename T> __global__ void kern(T in) { result = in; } __host__ __device__ void foo(void) { #if !defined(__CUDA_ARCH__) kern<<<1,1>>>(1); // instantiation "kern<int>" #endif /* !defined(__CUDA_ARCH__) */ } int main(void) { foo(); cudaDeviceSynchronize(); return 0; }

D.2.2Qualifiers
D.2.2.1Device Memory Qualifiers
The __device__, __shared__ and __constant__ qualifiers are not allowed on: class, struct, and union data members, formal parameters, local variables within a function that executes on the host. __shared__ and __constant__ variables have implied static storage. __device__ and __constant__ variable definitions are only allowed in namespace scope (including global namespace scope). __device__, __constant__ and __shared__ variables defined in namespace scope, that are of class type, cannot have a non-empty constructor or a non-empty destructor. A constructor for a class type is considered empty at a point in the translation unit, if it is either a trivial constructor or it satisfies all of the following conditions: The constructor function has been defined. The constructor function has no parameters, the initializer list is empty and the function body is an empty compound statement. Its class has no virtual functions and no virtual base classes. The default constructors of all base classes of its class can be considered empty. For all the nonstatic data members of its class that are of class type (or array thereof), the default constructors can be considered empty. A destructor for a class is considered empty at a point in the translation unit, if it is either a trivial destructor or it satisfies all of the following conditions: The destructor function has been defined. The destructor function body is an empty compound statement.

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|138

C/C++ Language Support

Its class has no virtual functions and no virtual base classes. The destructors of all base classes of its class can be considered empty. For all the nonstatic data members of its class that are of class type (or array thereof), the destructor can be considered empty. When compiling in the whole program compilation mode (see the nvcc user manual for a description of this mode), __device__, __shared__, and __constant__ variables cannot be defined as external using the extern keyword. The only exception is for dynamically allocated __shared__ variables as described in __shared__. When compiling in the separate compilation mode (see the nvcc user manual for a description of this mode), __device__, __shared__, and __constant__ variables can be defined as external using the extern keyword. nvlink will generate an error when it cannot find a definition for an external variable (unless it is a dynamically allocated __shared__ variable).

D.2.2.2Volatile Qualifier
Only after the execution of a __threadfence_block(), __threadfence(), or __syncthreads() (Memory Fence Functions and Synchronization Functions) are prior writes to global or shared memory guaranteed to be visible by other threads. As long as this requirement is met, the compiler is free to optimize reads and writes to global or shared memory. This behavior can be changed using the volatile keyword: If a variable located in global or shared memory is declared as volatile, the compiler assumes that its value can be changed or used at any time by another thread and therefore any reference to this variable compiles to an actual memory read or write instruction.

D.2.3Pointers
For devices of compute capability 1.x, pointers in code that is executed on the device are supported as long as the compiler is able to resolve whether they point to either the shared memory space, the global memory space, or the local memory space, otherwise they are restricted to only point to memory allocated or declared in the global memory space. For devices of compute capability 2.x and higher, pointers are supported without any restriction. Dereferencing a pointer either to global or shared memory in code that is executed on the host, or to host memory in code that is executed on the device results in an undefined behavior, most often in a segmentation fault and application termination. The address obtained by taking the address of a __device__, __shared__ or __constant__ variable can only be used in device code. The address of a __device__ or __constant__ variable obtained through cudaGetSymbolAddress() as described in Device Memory can only be used in host code. As a consequence of the use of C++ syntax rules, void pointers (e.g., returned by malloc()) cannot be assigned to non-void pointers without a typecast.

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|139

C/C++ Language Support

D.2.4Operators
D.2.4.1Assignment Operator
__constant__ variables can only be assigned from the host code through runtime functions (Device Memory); they cannot be assigned from the device code. __shared__ variables cannot have an initialization as part of their declaration. It is not allowed to assign values to any of the built-in variables defined in Built-in Variables.

D.2.4.2Address Operator
It is not allowed to take the address of any of the built-in variables defined in Built-in Variables.

D.2.5Functions
D.2.5.1Compiler generated functions
The execution space qualifiers (__host__, __device__) for a compiler generated function are the union of the execution space qualifiers of all the functions that invoke it (note that a __global__ caller will be treated as a __device__ caller for this analysis). For example:
class Base { int x; public: __host__ __device__ Base(void) : x(10) {} }; class Derived : public Base { int y; }; class Other: public Base { int z; }; __device__ void foo(void) { Derived D1; Other D2; } __host__ void bar(void) { Other D3; }

Here, the compiler generated constructor function "Derived::Derived" will be treated as a __device__ function, since it is invoked only from the __device__ function "foo". The compiler generated constructor function "Other::Other" will be treated as

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|140

C/C++ Language Support

a __host__ __device__ function, since it is invoked both from a __device__ function "foo" and a __host__ function "bar".

D.2.5.2Function Parameters
__global__ function parameters are passed to the device: via shared memory and are limited to 256 bytes on devices of compute capability 1.x, via constant memory and are limited to 4 KB on devices of compute capability 2.x and higher. __device__ and __global__ functions cannot have a variable number of arguments.

D.2.5.3Static Variables within Function

Static variables cannot be declared within the body of __device__ and __global__ functions.

D.2.5.4Function Pointers
Function pointers to __global__ functions are supported in host code, but not in device code. Function pointers to __device__ functions are only supported in device code compiled for devices of compute capability 2.x and higher. It is not allowed to take the address of a __device__ function in host code.

D.2.5.5Function Recursion
__global__ functions do not support recursion. __device__ functions only support recursion in device code compiled for devices of compute capability 2.x and higher.

D.2.6Classes
D.2.6.1Data Members
Static data members are not supported. The layout of bit-fields in device code may currently not match the layout in host code on Windows.

D.2.6.2Function Members
Static member functions cannot be __global__ functions.

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|141

C/C++ Language Support

D.2.6.3Virtual Functions
When a function in a derived class overrides a virtual function in a base class, the execution space qualifiers (i.e., __host__, __device__) on the overridden and overriding functions must match. It is not allowed to pass as an argument to a __global__ function an object of a class with virtual functions. The virtual function table is placed in global or constant memory by the compiler.

D.2.6.4Virtual Base Classes

It is not allowed to pass as an argument to a __global__ function an object of a class derived from virtual base classes.

Figure12Nearest-Point Sampling of a One-Dimensional Texture of Four Texels

E.2Linear Filtering
In this filtering mode, which is only available for floating-point textures, the value returned by the texture fetch is tex(x)=(1)T[i]+T[i+1] for a one-dimensional texture, tex(x,y)=(1)(1)T[i,j]+(1)T[i+1,j]+(1)T[i,j+1]+T[i+1,j+1] for a twodimensional texture, tex(x,y,z) = (1)(1)(1)T[i,j,k]+(1)(1)T[i+1,j,k]+ (1)(1)T[i,j+1,k]+(1)T[i+1,j+1,k]+ (1)(1)T[i,j,k+1]+(1)T[i+1,j,k+1]+ (1)T[i,j+1,k+1]+T[i+1,j+1,k+1] for a three-dimensional texture, where: i=floor(xB), =frac(xB), xB=x-0.5, j=floor(yB), =frac(yB), yB=y-0.5, k=floor(zB), =frac(zB), zB= z-0.5, , , and are stored in 9-bit fixed point format with 8 bits of fractional value (so 1.0 is exactly represented).

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|145

Texture Fetching

Figure 13 Linear Filtering of a One-Dimensional Texture of Four Texels in Clamp Addressing Mode illustrates nearest-point sampling for a one-dimensional texture with N=4.

Figure13Linear Filtering of a One-Dimensional Texture of Four Texels in Clamp Addressing Mode

E.3Table Lookup
A table lookup TL(x) where x spans the interval [0,R] can be implemented as TL(x)=tex((N1)/R)x+0.5) in order to ensure that TL(0)=T[0] and TL(R)=T[N1]. Figure 14 One-Dimensional Table Lookup Using Linear Filtering illustrates the use of texture filtering to implement a table lookup with R=4 or R=1 from a one-dimensional texture with N=4.

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|146

Texture Fetching

Figure14One-Dimensional Table Lookup Using Linear Filtering

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|147

Appendix F. COMPUTE CAPABILITIES

The general specifications and features of a compute device depend on its compute capability (see Compute Capability). Table 9 Feature Support per Compute Capability gives the features and technical specifications associated to each compute capability. Floating-Point Standard reviews the compliance with the IEEE floating-point standard. Section Compute Capability 1.x, Compute Capability 2.x, and Compute Capability 3.x give more details on the architecture of devices of compute capability 1.x, 2.x, and 3.x, respectively.

F.1Features and Technical Specifications

Table9Feature Support per Compute Capability
Feature Support (Unlisted features are supported for all compute capabilities)
Atomic functions operating on 32-bit integer values in global memory (Atomic Functions) atomicExch() operating on 32-bit floating point values in global memory (atomicExch()) Atomic functions operating on 32-bit integer values in shared memory (Atomic Functions) atomicExch() operating on 32-bit floating point values in shared memory (atomicExch()) Atomic functions operating on 64-bit integer values in global memory (Atomic Functions) Warp vote functions (Warp Vote Functions) Double-precision floating-point numbers No Yes No Yes

Compute Capability 1.0 1.1 1.2 1.3 2.x, 3.0 3.5

Yes

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|148

Compute Capabilities

Feature Support (Unlisted features are supported for all compute capabilities)
Atomic functions operating on 64-bit integer values in shared memory (Atomic Functions) Atomic addition operating on 32-bit floating point values in global and shared memory (atomicAdd()) __ballot() (Warp Vote Functions) __threadfence_system() (Memory Fence Functions) __syncthreads_count(), __syncthreads_and(), __syncthreads_or() (Synchronization Functions) Surface functions (Surface Functions) 3D grid of thread blocks Funnel shift (see reference manual)

Compute Capability 1.0 1.1 1.2 1.3 2.x, 3.0 3.5

Yes

Table10Technical Specifications per Compute Capability

Compute Capability Technical Specifications
Maximum dimensionality of grid of thread blocks Maximum x-dimension of a grid of thread blocks Maximum y- or z-dimension of a grid of thread blocks Maximum dimensionality of thread block Maximum x- or y-dimension of a block Maximum z-dimension of a block Maximum number of threads per block Warp size Maximum number of resident blocks per multiprocessor Maximum number of resident warps per multiprocessor Maximum number of resident threads per multiprocessor Number of 32-bit registers per multiprocessor 24 768 8K 8 32 1024 16 K 48 1536 32 K 512 32 16 64 2048 64 K 512 64 1024

1.0

1.1
2

1.2

1.3

2.x

3.0
3

3.5

65535 65535 3

2 -1

1024

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|149

Compute Capabilities

Compute Capability Technical Specifications

Maximum number of 32-bit registers per thread Maximum amount of shared memory per multiprocessor Number of shared memory banks Amount of local memory per thread Constant memory size Cache working set per multiprocessor for constant memory Cache working set per multiprocessor for texture memory Maximum width for a 1D texture reference bound to a CUDA array Maximum width for a 1D texture reference bound to linear memory Maximum width and number of layers for a 1D layered texture reference Maximum width and height for a 2D texture reference bound to a CUDA array Maximum width and height for a 2D texture reference bound to linear memory Maximum width and height for a 2D texture reference bound to a CUDA array supporting texture gather Maximum width, height, and number of layers for a 2D layered texture reference Maximum width, height, and depth for a 3D texture reference bound to a CUDA array Maximum width (and height) for a cubemap texture reference Maximum width (and height) and number of layers for a cubemap layered texture reference Maximum number of textures that can be bound to a kernel Maximum width for a 1D surface reference bound to a CUDA array Maximum width and number of layers for a 1D layered surface reference Maximum width and height for a 2D surface reference bound to a CUDA array N/A 8192 x 512 65536 x 32768 65000 x 65000

1.0

1.1
128

1.2

1.3

2.x
63

3.0

3.5
255

16 KB 16 16 KB 64 KB 8 KB

48 KB 32 512 KB

Device dependent, between 6 KB and 8 KB 8192 2

65536

16384 x 2048 65536 x 65535 65000 x 65000

N/A

16384 x 16384

8192 x 8192 x 512

16384 x 16384 x 2048 4096 x 4096 x 4096 16384

2048 x 2048 x 2048

N/A

16384 x 2046

128

256 65536 65536 x 2048 65536 x 32768

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|150

Compute Capabilities

Compute Capability Technical Specifications

Maximum width, height, and number of layers for a 2D layered surface reference Maximum width, height, and depth for a 3D surface reference bound to a CUDA array Maximum width (and height) for a cubemap surface reference bound to a CUDA array Maximum width (and height) and number of layers for a cubemap layered surface reference Maximum number of surfaces that can be bound to a kernel Maximum number of instructions per kernel 2 million 8

1.0

1.1

1.2

1.3

2.x

3.0

3.5

65536 x 32768 x 2048

32768

32768 x 2046

16 512 million

F.2Floating-Point Standard
All compute devices follow the IEEE 754-2008 standard for binary floating-point arithmetic with the following deviations: There is no dynamically configurable rounding mode; however, most of the operations support multiple IEEE rounding modes, exposed via device intrinsics; There is no mechanism for detecting that a floating-point exception has occurred and all operations behave as if the IEEE-754 exceptions are always masked, and deliver the masked response as defined by IEEE-754 if there is an exceptional event; for the same reason, while SNaN encodings are supported, they are not signaling and are handled as quiet; The result of a single-precision floating-point operation involving one or more input NaNs is the quiet NaN of bit pattern 0x7fffffff; Double-precision floating-point absolute value and negation are not compliant with IEEE-754 with respect to NaNs; these are passed through unchanged; For single-precision floating-point numbers on devices of compute capability 1.x: Denormalized numbers are not supported; floating-point arithmetic and comparison instructions convert denormalized operands to zero prior to the floating-point operation; Underflowed results are flushed to zero; Some instructions are not IEEE-compliant: Addition and multiplication are often combined into a single multiply-add instruction (FMAD), which truncates (i.e., without rounding) the intermediate mantissa of the multiplication; Division is implemented via the reciprocal in a non-standard-compliant way;

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|151

Compute Capabilities

Square root is implemented via the reciprocal square root in a non-standardcompliant way; For addition and multiplication, only round-to-nearest-even and roundtowards-zero are supported via static rounding modes; directed rounding towards +/- infinity is not supported. To mitigate the impact of these restrictions, IEEE-compliant software (and therefore slower) implementations are provided through the following intrinsics (c.f. Intrinsic Functions): __fmaf_r{n,z,u,d}(float, float, float): single-precision fused multiply-add with IEEE rounding modes, __frcp_r[n,z,u,d](float): single-precision reciprocal with IEEE rounding modes, __fdiv_r[n,z,u,d](float, float): single-precision division with IEEE rounding modes, __fsqrt_r[n,z,u,d](float): single-precision square root with IEEE rounding modes, __fadd_r[u,d](float, float): single-precision addition with IEEE directed rounding, __fmul_r[u,d](float, float): single-precision multiplication with IEEE directed rounding; For double-precision floating-point numbers on devices of compute capability 1.x: Round-to-nearest-even is the only supported IEEE rounding mode for reciprocal, division, and square root. When compiling for devices without native double-precision floating-point support, i.e., devices of compute capability 1.2 and lower, each double variable is converted to single-precision floating-point format (but retains its size of 64 bits) and doubleprecision floating-point arithmetic gets demoted to single-precision floating-point arithmetic. For devices of compute capability 2.x and higher, code must be compiled with ftz=false, -prec-div=true, and -prec-sqrt=true to ensure IEEE compliance (this is the default setting; see the nvcc user manual for description of these compilation flags); code compiled with -ftz=true, -prec-div=false, and -prec-sqrt=false comes closest to the code generated for devices of compute capability 1.x. Addition and multiplication are often combined into a single multiply-add instruction: FMAD for single precision on devices of compute capability 1.x, FFMA for single precision on devices of compute capability 2.x and higher. As mentioned above, FMAD truncates the mantissa prior to use it in the addition. FFMA, on the other hand, is an IEEE-754(2008) compliant fused multiply-add instruction, so the full-width product is being used in the addition and a single rounding occurs during generation of the final result. While FFMA in general has superior numerical properties compared to FMAD, the switch from FMAD to FFMA can cause slight changes in numeric results and can in rare circumstances lead to slighty larger error in final results.

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|152

Compute Capabilities

In accordance to the IEEE-754R standard, if one of the input parameters to fminf(), fmin(), fmaxf(), or fmax() is NaN, but not the other, the result is the non-NaN parameter. The conversion of a floating-point value to an integer value in the case where the floating-point value falls outside the range of the integer format is left undefined by IEEE-754. For compute devices, the behavior is to clamp to the end of the supported range. This is unlike the x86 architecture behavior. The behavior of integer division by zero and integer overflow is left undefined by IEEE-754. For compute devices, there is no mechanism for detecting that such integer operation exceptions have occurred. Integer division by zero yields an unspecified, machine-specific value. http://developer.nvidia.com/content/precision-performance-floating-point-and-ieee-754compliance-nvidia-gpus includes more information on the floating point accuracy and compliance of NVIDIA GPUs.

F.3Compute Capability 1.x

F.3.1Architecture
For devices of compute capability 1.x, a multiprocessor consists of: 8 CUDA cores for arithmetic operations (see Arithmetic Instructions for throughputs of arithmetic operations), 1 double-precision floating-point unit for double-precision floating-point arithmetic operations (this is only for devices of compute capability 1.3 and above), 2 special function units for single-precision floating-point transcendental functions (these units can also handle single-precision floating-point multiplications), 1 warp scheduler. To execute an instruction for all threads of a warp, the warp scheduler must therefore issue the instruction over: 4 clock cycles for an integer or single-precision floating-point arithmetic instruction, 32 clock cycles for a double-precision floating-point arithmetic instruction (this is only for devices of compute capability 1.3 and above), 16 clock cycles for a single-precision floating-point transcendental instruction. A multiprocessor also has a read-only constant cache that is shared by all functional units and speeds up reads from the constant memory space, which resides in device memory. Multiprocessors are grouped into Texture Processor Clusters (TPCs). The number of multiprocessors per TPC is: 2 for devices of compute capabilities 1.0 and 1.1, 3 for devices of compute capabilities 1.2 and 1.3.

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|153

Compute Capabilities

Each TPC has a read-only texture cache that is shared by all multiprocessors and speeds up reads from the texture memory space, which resides in device memory. Each multiprocessor accesses the texture cache via a texture unit that implements the various addressing modes and data filtering mentioned in Texture and Surface Memory. The local and global memory spaces reside in device memory and are not cached.

F.3.2Global Memory
A global memory request for a warp is split into two memory requests, one for each half-warp, that are issued independently. Devices of Compute Capability 1.0 and 1.1 and Devices of Compute Capability 1.2 and 1.3 describe how the memory accesses of threads within a half-warp are coalesced into one or more memory transactions depending on the compute capability of the device. Figure 15 Examples of Global Memory Accesses shows some examples of global memory accesses and corresponding memory transactions based on compute capability. The resulting memory transactions are serviced at the throughput of device memory. Devices of Compute Capability 1.0 and 1.1 To coalesce, the memory request for a half-warp must satisfy the following conditions: The size of the words accessed by the threads must be 4, 8, or 16 bytes; If this size is: 4, all 16 words must lie in the same 64-byte segment, 8, all 16 words must lie in the same 128-byte segment, 16, the first 8 words must lie in the same 128-byte segment and the last 8 words in the following 128-byte segment; th Threads must access the words in sequence: The k thread in the half-warp must th access the k word. If the half-warp meets these requirements, a 64-byte memory transaction, a 128-byte memory transaction, or two 128-byte memory transactions are issued if the size of the words accessed by the threads is 4, 8, or 16, respectively. Coalescing is achieved even if the warp is divergent, i.e., there are some inactive threads that do not actually access memory. If the half-warp does not meet these requirements, 16 separate 32-byte memory transactions are issued. Devices of Compute Capability 1.2 and 1.3 Threads can access any words in any order, including the same words, and a single memory transaction for each segment addressed by the half-warp is issued. This is in contrast with devices of compute capabilities 1.0 and 1.1 where threads need to access words in sequence and coalescing only happens if the half-warp addresses a single segment.

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|154

Compute Capabilities

More precisely, the following protocol is used to determine the memory transactions necessary to service all threads in a half-warp: Find the memory segment that contains the address requested by the active thread with the lowest thread ID. The segment size depends on the size of the words accessed by the threads: 32 bytes for 1-byte words, 64 bytes for 2 byte words, 128 bytes for 4-, 8- and 16-byte words. Find all other active threads whose requested address lies in the same segment. Reduce the transaction size, if possible: If the transaction size is 128 bytes and only the lower or upper half is used, reduce the transaction size to 64 bytes; If the transaction size is 64 bytes (originally or after reduction from 128 bytes) and only the lower or upper half is used, reduce the transaction size to 32 bytes. Carry out the transaction and mark the serviced threads as inactive. Repeat until all threads in the half-warp are serviced.

F.3.3Shared Memory
Shared memory has 16 banks that are organized such that successive 32-bit words map to successive banks. Each bank has a bandwidth of 32 bits per two clock cycles. A shared memory request for a warp is split into two memory requests, one for each half-warp, that are issued independently. As a consequence, there can be no bank conflict between a thread belonging to the first half of a warp and a thread belonging to the second half of the same warp. If a non-atomic instruction executed by a warp writes to the same location in shared memory for more than one of the threads of the warp, only one thread per half-warp performs a write and which thread performs the final write is undefined. 32-Bit Strided Access A common access pattern is for each thread to access a 32-bit word from an array indexed by the thread ID tid and with some stride s:
extern __shared__ float shared[]; float data = shared[BaseIndex + s * tid];

In this case, threads tid and tid+n access the same bank whenever s*n is a multiple of the number of banks (i.e., 16) or, equivalently, whenever n is a multiple of 16/d where d is the greatest common divisor of 16 and s. As a consequence, there will be no bank conflict only if half the warp size (i.e., 16) is less than or equal to 16/d, i.e., only if d is equal to 1, i.e., s is odd. Figure 16 Examples of Strided Shared Memory Accesses for Devices of Compute Capability 3.x (in 32-bit mode) shows some examples of strided access for devices of

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|155

Compute Capabilities

compute capability 3.x. The same examples apply for devices of compute capability 1.x, but with 16 banks instead of 32. Also, the access pattern for the example in the middle generates 2-way bank conflicts for devices of compute capability 1.x. 32-Bit Broadcast Access Shared memory features a broadcast mechanism whereby a 32-bit word can be read and broadcast to several threads simultaneously when servicing one memory read request. This reduces the number of bank conflicts when several threads read from an address within the same 32-bit word. More precisely, a memory read request made of several addresses is serviced in several steps over time by servicing one conflict-free subset of these addresses per step until all addresses have been serviced; at each step, the subset is built from the remaining addresses that have yet to be serviced using the following procedure: Select one of the words pointed to by the remaining addresses as the broadcast word; Include in the subset: All addresses that are within the broadcast word, One address for each bank (other than the broadcasting bank) pointed to by the remaining addresses. Which word is selected as the broadcast word and which address is picked up for each bank at each cycle are unspecified. A common conflict-free case is when all threads of a half-warp read from an address within the same 32-bit word. Figure 17 Examples of Irregular Shared Memory Accesses for Devices of Compute Capability 3.x shows some examples of memory read accesses that involve the broadcast mechanism for devices of compute capability 3.x. The same examples apply for devices of compute capability 1.x, but with 16 banks instead of 32. Also, the access pattern for the example at the right generates 2-way bank conflicts for devices of compute capability 1.x. 8-Bit and 16-Bit Access 8-bit and 16-bit accesses typically generate bank conflicts. For example, there are bank conflicts if an array of char is accessed the following way:
extern __shared__ float shared[]; char data = shared[BaseIndex + tid];

because shared[0], shared[1], shared[2], and shared[3], for example, belong to the same bank. There are no bank conflicts however, if the same array is accessed the following way:
char data = shared[BaseIndex + 4 * tid];

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|156

Compute Capabilities

Larger Than 32-Bit Access Accesses that are larger than 32-bit per thread are split into 32-bit accesses that typically generate bank conflicts. For example, there are 2-way bank conflicts for arrays of doubles accessed as follows:
extern __shared__ float shared[]; double data = shared[BaseIndex + tid];

as the memory request is compiled into two separate 32-bit requests with a stride of two. One way to avoid bank conflicts in this case is two split the double operands like in the following sample code:
__shared__ int shared_lo[32]; __shared__ int shared_hi[32]; double dataIn; shared_lo[BaseIndex + tid] = __double2loint(dataIn); shared_hi[BaseIndex + tid] = __double2hiint(dataIn); double dataOut = __hiloint2double(shared_hi[BaseIndex + tid], shared_lo[BaseIndex + tid]);

This might not always improve performance however and does perform worse on devices of compute capabilities 2.x and higher. The same applies to structure assignments. The following code, for example:
extern __shared__ float shared[]; struct type data = shared[BaseIndex + tid];

results in: Three separate reads without bank conflicts if type is defined as
struct type { float x, y, z; };

since each member is accessed with an odd stride of three 32-bit words; Two separate reads with bank conflicts if type is defined as
struct type { float x, y; };

since each member is accessed with an even stride of two 32-bit words.

F.4Compute Capability 2.x

F.4.1Architecture
For devices of compute capability 2.x, a multiprocessor consists of: For devices of compute capability 2.0: 32 CUDA cores for arithmetic operations (see Arithmetic Instructions for throughputs of arithmetic operations),

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|157

Compute Capabilities

4 special function units for single-precision floating-point transcendental functions, For devices of compute capability 2.1: 48 CUDA cores for arithmetic operations (see Arithmetic Instructions for throughputs of arithmetic operations), 8 special function units for single-precision floating-point transcendental functions, 2 warp schedulers. At every instruction issue time, each scheduler issues: One instruction for devices of compute capability 2.0, Two independent instructions for devices of compute capability 2.1, for some warp that is ready to execute, if any. The first scheduler is in charge of the warps with an odd ID and the second scheduler is in charge of the warps with an even ID. Note that when a scheduler issues a double-precision floating-point instruction, the other scheduler cannot issue any instruction. A warp scheduler can issue an instruction to only half of the CUDA cores. To execute an instruction for all threads of a warp, a warp scheduler must therefore issue the instruction over two clock cycles for an integer or floating-point arithmetic instruction. A multiprocessor also has a read-only constant cache that is shared by all functional units and speeds up reads from the constant memory space, which resides in device memory. There is an L1 cache for each multiprocessor and an L2 cache shared by all multiprocessors, both of which are used to cache accesses to local or global memory, including temporary register spills. The cache behavior (e.g., whether reads are cached in both L1 and L2 or in L2 only) can be partially configured on a per-access basis using modifiers to the load or store instruction. The same on-chip memory is used for both L1 and shared memory: It can be configured as 48 KB of shared memory and 16 KB of L1 cache or as 16 KB of shared memory and 48 KB of L1 cache, using cudaFuncSetCacheConfig()/cuFuncSetCacheConfig():
// Device code __global__ void MyKernel(int* foo, int* bar, int a) { ... } // Host code // Runtime API // cudaFuncCachePreferShared: shared memory is 48 KB // cudaFuncCachePreferL1: shared memory is 16 KB // cudaFuncCachePreferNone: no preference cudaFuncSetCacheConfig(MyKernel, cudaFuncCachePreferShared) // Or via a function pointer: void (*funcPtr)(int*, int*, int); funcPtr = MyKernel; cudaFuncSetCacheConfig(*funcPtr, cudaFuncCachePreferShared);

www.nvidia.com

F.5.1Architecture
A multiprocessor consists of: 192 CUDA cores for arithmetic operations (see Arithmetic Instructions for throughputs of arithmetic operations), 32 special function units for single-precision floating-point transcendental functions, 4 warp schedulers.

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|161

Compute Capabilities

When a multiprocessor is given warps to execute, it first distributes them among the four schedulers. Then, at every instruction issue time, each scheduler issues two independent instructions for one of its assigned warps that is ready to execute, if any. A multiprocessor has a read-only constant cache that is shared by all functional units and speeds up reads from the constant memory space, which resides in device memory. There is an L1 cache for each multiprocessor and an L2 cache shared by all multiprocessors, both of which are used to cache accesses to local or global memory, including temporary register spills. The cache behavior (e.g., whether reads are cached in both L1 and L2 or in L2 only) can be partially configured on a per-access basis using modifiers to the load or store instruction. The same on-chip memory is used for both L1 and shared memory: It can be configured as 48 KB of shared memory and 16 KB of L1 cache or as 16 KB of shared memory and 48 KB of L1 cache or as 32 KB of shared memory and 32 KB of L1 cache, using cudaFuncSetCacheConfig()/cuFuncSetCacheConfig():
// Device code __global__ void MyKernel() { ... } // Host code // Runtime API // cudaFuncCachePreferShared: shared memory is 48 KB // cudaFuncCachePreferEqual: shared memory is 32 KB // cudaFuncCachePreferL1: shared memory is 16 KB // cudaFuncCachePreferNone: no preference cudaFuncSetCacheConfig(MyKernel, cudaFuncCachePreferShared)

The default cache configuration is "prefer none," meaning "no preference." If a kernel is configured to have no preference, then it will default to the preference of the current thread/context, which is set using cudaDeviceSetCacheConfig()/cuCtxSetCacheConfig() (see the reference manual for details). If the current thread/context also has no preference (which is again the default setting), then whichever cache configuration was most recently used for any kernel will be the one that is used, unless a different cache configuration is required to launch the kernel (e.g., due to shared memory requirements). The initial configuration is 48 KB of shared memory and 16 KB of L1 cache. Applications may query the L2 cache size by checking the l2CacheSize device property (see Device Enumeration). The maximum L2 cache size is 1.5 MB. Multiprocessors are grouped into Graphics Processor Clusters (GPCs). A GPC includes three multiprocessors. Each multiprocessor has a read-only data cache of 48 KB to speed up reads from device memory. It accesses this cache either directly (for devices of compute capability 3.5 only), or via a texture unit that implements the various addressing modes and data filtering mentioned in Texture and Surface Memory. When accessed via the texture unit, the read-only data cache is also referred to as texture cache.

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|162

Compute Capabilities

F.5.2Global Memory
Global memory accesses for devices of compute capability 3.x are cached in L2 and for devices of compute capability 3.5, may also be cached in the read-only data cache described in the previous section; they are not cached in L1. Caching in L2 behaves in the same way as for devices of compute capability 2.x (see Global Memory). The compiler determines whether a given global memory read is cached in the readonly data cache or not. A requirement for data to be cached in the read-only cache is that it must be read-only. In order to allow the compiler to detect that this condition is satisfied, pointers used for loading such data should be marked with both the const and __restrict__ qualifiers. Figure 15 Examples of Global Memory Accesses shows some examples of global memory accesses and corresponding memory transactions based on compute capability. Examples of Global Memory Accesses by a Warp, 4-Byte Word per Thread, and Associated Memory Transactions Based on Compute Capability

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|163

Figure17Examples of Irregular Shared Memory Accesses for Devices of Compute Capability 3.x

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|168

Appendix G. DRIVER API

This appendix assumes knowledge of the concepts described in CUDA C Runtime. The driver API is implemented in the nvcuda dynamic library which is copied on the system during the installation of the device driver. All its entry points are prefixed with cu. It is a handle-based, imperative API: Most objects are referenced by opaque handles that may be specified to functions to manipulate the objects. The objects available in the driver API are summarized in Table 11 Objects Available in the CUDA Driver API.

Table11Objects Available in the CUDA Driver API

Object
Device Context Module Function Heap memory CUDA array

Handle
CUdevice CUcontext CUmodule CUfunction CUdeviceptr CUarray

Description
CUDA-enabled device Roughly equivalent to a CPU process Roughly equivalent to a dynamic library Kernel Pointer to device memory Opaque container for one-dimensional or twodimensional data on the device, readable via texture or surface references Object that describes how to interpret texture memory data Object that describes how to read or write CUDA arrays Object that describes a CUDA event

Texture reference Surface reference Event

CUtexref CUsurfref CUevent

The driver API must be initialized with cuInit() before any function from the driver API is called. A CUDA context must then be created that is attached to a specific device and made current to the calling host thread as detailed in Context.

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|169

Driver API

Within a CUDA context, kernels are explicitly loaded as PTX or binary objects by the host code as described in Module. Kernels written in C must therefore be compiled separately into PTX or binary objects. Kernels are launched using API entry points as described in Kernel Execution. Any application that wants to run on future device architectures must load PTX, not binary code. This is because binary code is architecture-specific and therefore incompatible with future architectures, whereas PTX code is compiled to binary code at load time by the device driver. Here is the host code of the sample from Kernels written using the driver API:
int main() { int N = ...; size_t size = N * sizeof(float); // Allocate input vectors h_A and h_B in host memory float* h_A = (float*)malloc(size); float* h_B = (float*)malloc(size); // Initialize input vectors ... // Initialize cuInit(0); // Get number of devices supporting CUDA int deviceCount = 0; cuDeviceGetCount(&deviceCount); if (deviceCount == 0) { printf("There is no device supporting CUDA.\n"); exit (0); } // Get handle for device 0 CUdevice cuDevice; cuDeviceGet(&cuDevice, 0); // Create context CUcontext cuContext; cuCtxCreate(&cuContext, 0, cuDevice); // Create module from binary file CUmodule cuModule; cuModuleLoad(&cuModule, "VecAdd.ptx"); // Allocate vectors in device memory CUdeviceptr d_A; cuMemAlloc(&d_A, size); CUdeviceptr d_B; cuMemAlloc(&d_B, size); CUdeviceptr d_C; cuMemAlloc(&d_C, size); // Copy vectors from host memory to device memory cuMemcpyHtoD(d_A, h_A, size); cuMemcpyHtoD(d_B, h_B, size); // Get function handle from module CUfunction vecAdd; cuModuleGetFunction(&vecAdd, cuModule, "VecAdd"); // Invoke kernel int threadsPerBlock = 256;

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|170

Driver API

int blocksPerGrid = (N + threadsPerBlock 1) / threadsPerBlock; void* args[] = { &d_A, &d_B, &d_C, &N }; cuLaunchKernel(vecAdd, blocksPerGrid, 1, 1, threadsPerBlock, 1, 1, 0, 0, args, 0); } ...

Full code can be found in the vectorAddDrv SDK code sample.

G.1Context
A CUDA context is analogous to a CPU process. All resources and actions performed within the driver API are encapsulated inside a CUDA context, and the system automatically cleans up these resources when the context is destroyed. Besides objects such as modules and texture or surface references, each context has its own distinct address space. As a result, CUdeviceptr values from different contexts reference different memory locations. A host thread may have only one device context current at a time. When a context is created with cuCtxCreate(), it is made current to the calling host thread. CUDA functions that operate in a context (most functions that do not involve device enumeration or context management) will return CUDA_ERROR_INVALID_CONTEXT if a valid context is not current to the thread. Each host thread has a stack of current contexts. cuCtxCreate() pushes the new context onto the top of the stack. cuCtxPopCurrent() may be called to detach the context from the host thread. The context is then "floating" and may be pushed as the current context for any host thread. cuCtxPopCurrent() also restores the previous current context, if any. A usage count is also maintained for each context. cuCtxCreate() creates a context with a usage count of 1. cuCtxAttach() increments the usage count and cuCtxDetach() decrements it. A context is destroyed when the usage count goes to 0 when calling cuCtxDetach() or cuCtxDestroy(). Usage count facilitates interoperability between third party authored code operating in the same context. For example, if three libraries are loaded to use the same context, each library would call cuCtxAttach() to increment the usage count and cuCtxDetach() to decrement the usage count when the library is done using the context. For most libraries, it is expected that the application will have created a context before loading or initializing the library; that way, the application can create the context using its own heuristics, and the library simply operates on the context handed to it. Libraries that wish to create their own contexts unbeknownst to their API clients who may or may not have created contexts of their own would use cuCtxPushCurrent() and cuCtxPopCurrent() as illustrated in Figure 18 Library Context Management.

PG-02829-001_v5.0|173

Driver API

gridWidth, gridHeight, gridDepth, 0, 0, 0, extra);

The alignment requirement of a structure is equal to the maximum of the alignment requirements of its fields. The alignment requirement of a structure that contains built-in vector types, CUdeviceptr, or non-aligned double and long long, might therefore differ between device code and host code. Such a structure might also be padded differently. The following structure, for example, is not padded at all in host code, but it is padded in device code with 12 bytes after field f since the alignment requirement for field f4 is 16.
typedef struct { float f; float4 f4; } myStruct;

G.4Interoperablility between Runtime and Driver APIs

An application can mix runtime API code with driver API code. If a context is created and made current via the driver API, subsequent runtime calls will pick up this context instead of creating a new one. If the runtime is initialized (implicitly as mentioned in CUDA C Runtime), cuCtxGetCurrent() can be used to retrieve the context created during initialization. This context can be used by subsequent driver API calls. Device memory can be allocated and freed using either API. CUdeviceptr can be cast to regular pointers and vice-versa:
CUdeviceptr devPtr; float* d_data; // Allocation using driver API cuMemAlloc(&devPtr, size); d_data = (float*)devPtr; // Allocation using runtime API cudaMalloc(&d_data, size); devPtr = (CUdeviceptr)d_data;

In particular, this means that applications written using the driver API can invoke libraries written using the runtime API (such as CUFFT, CUBLAS, ). All functions from the device and version management sections of the reference manual can be used interchangeably.

www.nvidia.com

CUDA C Programming Guide

PG-02829-001_v5.0|174

Notice ALL NVIDIA DESIGN SPECIFICATIONS, REFERENCE BOARDS, FILES, DRAWINGS, DIAGNOSTICS, LISTS, AND OTHER DOCUMENTS (TOGETHER AND SEPARATELY, "MATERIALS") ARE BEING PROVIDED "AS IS." NVIDIA MAKES NO WARRANTIES, EXPRESSED, IMPLIED, STATUTORY, OR OTHERWISE WITH RESPECT TO THE MATERIALS, AND EXPRESSLY DISCLAIMS ALL IMPLIED WARRANTIES OF NONINFRINGEMENT, MERCHANTABILITY, AND FITNESS FOR A PARTICULAR PURPOSE. Information furnished is believed to be accurate and reliable. However, NVIDIA Corporation assumes no responsibility for the consequences of use of such information or for any infringement of patents or other rights of third parties that may result from its use. No license is granted by implication of otherwise under any patent rights of NVIDIA Corporation. Specifications mentioned in this publication are subject to change without notice. This publication supersedes and replaces all other information previously supplied. NVIDIA Corporation products are not authorized as critical components in life support devices or systems without express written approval of NVIDIA Corporation. Trademarks NVIDIA and the NVIDIA logo are trademarks or registered trademarks of NVIDIA Corporation in the U.S. and other countries. Other company and product names may be trademarks of the respective companies with which they are associated. Copyright
2007-2012 NVIDIA Corporation. All rights reserved.

www.nvidia.com