NVIDIA CUDA Programming Guide 1.1
User Manual:
Open the PDF directly: View PDF 
.
Page Count: 143
| Download | |
| Open PDF In Browser | View PDF |
NVIDIA CUDA
Compute Unified
Device Architecture
Programming Guide
Version 1.1
11/29/2007
ii
CUDA Programming Guide Version 1.1
Table of Contents
Chapter 1. Introduction to CUDA....................................................................... 1
1.1
The Graphics Processor Unit as a Data-Parallel Computing Device ...................1
1.2
CUDA: A New Architecture for Computing on the GPU ....................................3
1.3
Document’s Structure ...................................................................................6
Chapter 2. Programming Model......................................................................... 7
2.1
A Highly Multithreaded Coprocessor...............................................................7
2.2
Thread Batching...........................................................................................7
2.2.1
Thread Block .........................................................................................7
2.2.2
Grid of Thread Blocks.............................................................................8
2.3
Memory Model ........................................................................................... 10
Chapter 3. Hardware Implementation ............................................................ 13
3.1
A Set of SIMD Multiprocessors with On-Chip Shared Memory ........................ 13
3.2
Execution Model ......................................................................................... 14
3.3
Compute Capability .................................................................................... 15
3.4
Multiple Devices ......................................................................................... 16
3.5
Mode Switches ........................................................................................... 16
Chapter 4. Application Programming Interface .............................................. 17
4.1
An Extension to the C Programming Language ............................................. 17
4.2
Language Extensions .................................................................................. 17
4.2.1
Function Type Qualifiers....................................................................... 18
4.2.1.1
__device__................................................................................ 18
4.2.1.2
__global__................................................................................ 18
4.2.1.3
__host__.................................................................................... 18
4.2.1.4
Restrictions................................................................................... 18
4.2.2
Variable Type Qualifiers ....................................................................... 19
4.2.2.1
__device__................................................................................ 19
4.2.2.2
__constant__............................................................................ 19
4.2.2.3
__shared__................................................................................ 19
CUDA Programming Guide Version 1.1
iii
4.2.2.4
Restrictions................................................................................... 20
4.2.3
Execution Configuration ....................................................................... 21
4.2.4
Built-in Variables.................................................................................. 21
4.2.4.1
gridDim...................................................................................... 21
4.2.4.2
blockIdx.................................................................................... 22
4.2.4.3
blockDim.................................................................................... 22
4.2.4.4
threadIdx.................................................................................. 22
4.2.4.5
Restrictions................................................................................... 22
4.2.5
4.3
Compilation with NVCC ........................................................................ 22
4.2.5.1
__noinline__............................................................................ 22
4.2.5.2
#pragma unroll........................................................................ 23
Common Runtime Component..................................................................... 23
4.3.1
Built-in Vector Types............................................................................ 23
4.3.1.1
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, float1, float2, float3,
float4 23
4.3.1.2
4.3.2
Mathematical Functions........................................................................ 24
4.3.3
Time Function ..................................................................................... 24
4.3.4
Texture Type....................................................................................... 24
4.4
4.3.4.1
Texture Reference Declaration ....................................................... 24
4.3.4.2
Runtime Texture Reference Attributes ............................................ 25
4.3.4.3
Texturing from Linear Memory versus CUDA Arrays ........................ 25
Device Runtime Component ........................................................................ 26
4.4.1
Mathematical Functions........................................................................ 26
4.4.2
Synchronization Function ..................................................................... 26
4.4.3
Type Conversion Functions................................................................... 26
4.4.4
Type Casting Functions ........................................................................ 27
4.4.5
Texture Functions ................................................................................ 27
4.4.5.1
Texturing from Device Memory ...................................................... 27
4.4.5.2
Texturing from CUDA Arrays.......................................................... 28
4.4.6
iv
dim3 Type ................................................................................... 23
Atomic Functions ................................................................................. 28
CUDA Programming Guide Version 1.1
4.5
Host Runtime Component ........................................................................... 28
4.5.1
Common Concepts............................................................................... 29
4.5.1.1
Device .......................................................................................... 29
4.5.1.2
Memory........................................................................................ 29
4.5.1.3
OpenGL Interoperability ................................................................ 30
4.5.1.4
Direct3D Interoperability ............................................................... 30
4.5.1.5
Asynchronous Concurrent Execution............................................... 31
4.5.2
Runtime API ........................................................................................ 32
4.5.2.1
Initialization.................................................................................. 32
4.5.2.2
Device Management...................................................................... 32
4.5.2.3
Memory Management.................................................................... 32
4.5.2.4
Stream Management ..................................................................... 34
4.5.2.5
Event Management ....................................................................... 34
4.5.2.6
Texture Reference Management .................................................... 35
4.5.2.7
OpenGL Interoperability ................................................................ 36
4.5.2.8
Direct3D Interoperability ............................................................... 37
4.5.2.9
Debugging using the Device Emulation Mode.................................. 37
4.5.3
Driver API ........................................................................................... 39
4.5.3.1
Initialization.................................................................................. 39
4.5.3.2
Device Management...................................................................... 39
4.5.3.3
Context Management .................................................................... 39
4.5.3.4
Module Management ..................................................................... 40
4.5.3.5
Execution Control.......................................................................... 40
4.5.3.6
Memory Management.................................................................... 41
4.5.3.7
Stream Management ..................................................................... 42
4.5.3.8
Event Management ....................................................................... 43
4.5.3.9
Texture Reference Management .................................................... 43
4.5.3.10
OpenGL Interoperability ................................................................ 44
4.5.3.11
Direct3D Interoperability ............................................................... 44
Chapter 5. Performance Guidelines ................................................................. 47
5.1
Instruction Performance ............................................................................. 47
5.1.1
Instruction Throughput ........................................................................ 47
5.1.1.1
Arithmetic Instructions .................................................................. 47
CUDA Programming Guide Version 1.1
v
5.1.1.2
Control Flow Instructions............................................................... 48
5.1.1.3
Memory Instructions ..................................................................... 49
5.1.1.4
Synchronization Instruction ........................................................... 49
5.1.2
Memory Bandwidth .............................................................................. 49
5.1.2.1
Global Memory.............................................................................. 50
5.1.2.2
Constant Memory.......................................................................... 55
5.1.2.3
Texture Memory ........................................................................... 55
5.1.2.4
Shared Memory ............................................................................ 56
5.1.2.5
Registers ...................................................................................... 62
5.2
Number of Threads per Block...................................................................... 62
5.3
Data Transfer between Host and Device ...................................................... 63
5.4
Texture Fetch versus Global or Constant Memory Read................................. 63
5.5
Overall Performance Optimization Strategies ................................................ 64
Chapter 6. Example of Matrix Multiplication ................................................... 67
6.1
Overview ................................................................................................... 67
6.2
Source Code Listing .................................................................................... 69
6.3
Source Code Walkthrough........................................................................... 71
6.3.1
Mul() ................................................................................................ 71
6.3.2
Muld() .............................................................................................. 71
Appendix A. Technical Specifications .............................................................. 73
A.1
General Specifications................................................................................. 74
A.2
Floating-Point Standard .............................................................................. 74
Appendix B. Mathematical Functions .............................................................. 77
B.1
Common Runtime Component..................................................................... 77
B.2
Device Runtime Component ........................................................................ 80
Appendix C. Atomic Functions ......................................................................... 83
C.1
vi
Arithmetic Functions ................................................................................... 83
C.1.1
atomicAdd() .................................................................................... 83
C.1.2
atomicSub() .................................................................................... 83
C.1.3
atomicExch() .................................................................................. 83
C.1.4
atomicMin() .................................................................................... 84
C.1.5
atomicMax() .................................................................................... 84
C.1.6
atomicInc() .................................................................................... 84
CUDA Programming Guide Version 1.1
C.1.7
atomicDec() .................................................................................... 84
C.1.8
atomicCAS() .................................................................................... 84
C.2
Bitwise Functions........................................................................................ 85
C.2.1
atomicAnd() .................................................................................... 85
C.2.2
atomicOr() ...................................................................................... 85
C.2.3
atomicXor() .................................................................................... 85
Appendix D. Runtime API Reference ............................................................... 87
D.1
Device Management ................................................................................... 87
D.1.1
cudaGetDeviceCount() .................................................................. 87
D.1.2
cudaSetDevice() ............................................................................ 87
D.1.3
cudaGetDevice() ............................................................................ 87
D.1.4
cudaGetDeviceProperties() ........................................................ 88
D.1.5
cudaChooseDevice() ...................................................................... 89
D.2
Thread Management .................................................................................. 89
D.2.1
cudaThreadSynchronize() ............................................................ 89
D.2.2
cudaThreadExit() .......................................................................... 89
D.3
Stream Management .................................................................................. 89
D.3.1
cudaStreamCreate() ...................................................................... 89
D.3.2
cudaStreamQuery() ........................................................................ 89
D.3.3
cudaStreamSynchronize() ............................................................ 89
D.3.4
cudaStreamDestroy() .................................................................... 89
D.4
Event Management..................................................................................... 90
D.4.1
cudaEventCreate() ........................................................................ 90
D.4.2
cudaEventRecord() ........................................................................ 90
D.4.3
cudaEventQuery() .......................................................................... 90
D.4.4
cudaEventSynchronize() .............................................................. 90
D.4.5
cudaEventDestroy() ...................................................................... 90
D.4.6
cudaEventElapsedTime() .............................................................. 90
D.5
Memory Management ................................................................................. 91
D.5.1
cudaMalloc() .................................................................................. 91
D.5.2
cudaMallocPitch() ........................................................................ 91
D.5.3
cudaFree() ...................................................................................... 91
CUDA Programming Guide Version 1.1
vii
D.5.4
cudaMallocArray() ........................................................................ 92
D.5.5
cudaFreeArray() ............................................................................ 92
D.5.6
cudaMallocHost() .......................................................................... 92
D.5.7
cudaFreeHost() .............................................................................. 92
D.5.8
cudaMemset() .................................................................................. 92
D.5.9
cudaMemset2D() .............................................................................. 92
D.5.10
cudaMemcpy() ............................................................................... 93
D.5.11
cudaMemcpy2D() ........................................................................... 93
D.5.12
cudaMemcpyToArray() ................................................................. 94
D.5.13
cudaMemcpy2DToArray() ............................................................. 94
D.5.14
cudaMemcpyFromArray() ............................................................. 95
D.5.15
cudaMemcpy2DFromArray() ......................................................... 95
D.5.16
cudaMemcpyArrayToArray() ....................................................... 96
D.5.17
cudaMemcpy2DArrayToArray() ................................................... 96
D.5.18
cudaMemcpyToSymbol() ............................................................... 96
D.5.19
cudaMemcpyFromSymbol() ........................................................... 96
D.5.20
cudaGetSymbolAddress() ........................................................... 97
D.5.21
cudaGetSymbolSize() ................................................................. 97
D.6
Texture Reference Management.................................................................. 97
D.6.1
D.6.1.1
cudaCreateChannelDesc()...................................................... 97
D.6.1.2
cudaGetChannelDesc()............................................................ 97
D.6.1.3
cudaGetTextureReference().................................................. 97
D.6.1.4
cudaBindTexture().................................................................. 98
D.6.1.5
cudaBindTextureToArray().................................................... 98
D.6.1.6
cudaUnbindTexture().............................................................. 98
D.6.1.7
cudaGetTextureAlignmentOffset()...................................... 98
D.6.2
viii
Low-Level API ..................................................................................... 97
High-Level API..................................................................................... 98
D.6.2.1
cudaCreateChannelDesc()...................................................... 98
D.6.2.2
cudaBindTexture().................................................................. 99
D.6.2.3
cudaBindTextureToArray().................................................... 99
CUDA Programming Guide Version 1.1
D.6.2.4
D.7
cudaUnbindTexture().............................................................. 99
Execution Control ..................................................................................... 100
D.7.1
cudaConfigureCall() .................................................................. 100
D.7.2
cudaLaunch() ................................................................................ 100
D.7.3
cudaSetupArgument() .................................................................. 100
D.8
OpenGL Interoperability............................................................................ 100
D.8.1
cudaGLRegisterBufferObject() ................................................ 100
D.8.2
cudaGLMapBufferObject() .......................................................... 101
D.8.3
cudaGLUnmapBufferObject() ...................................................... 101
D.8.4
cudaGLUnregisterBufferObject() ............................................ 101
D.9
Direct3D Interoperability........................................................................... 101
D.9.1
cudaD3D9Begin() .......................................................................... 101
D.9.2
cudaD3D9End() .............................................................................. 101
D.9.3
cudaD3D9RegisterVertexBuffer() ............................................ 101
D.9.4
cudaD3D9MapVertexBuffer() ...................................................... 101
D.9.5
cudaD3D9UnmapVertexBuffer() .................................................. 102
D.9.6
cudaD3D9UnregisterVertexBuffer() ........................................ 102
D.9.7
cudaD3D9GetDevice() .................................................................. 102
D.10
Error Handling ...................................................................................... 102
D.10.1
cudaGetLastError() ................................................................. 102
D.10.2
cudaGetErrorString() ............................................................. 102
Appendix E. Driver API Reference ................................................................. 103
E.1
Initialization ............................................................................................. 103
E.1.1
E.2
cuInit() ........................................................................................ 103
Device Management ................................................................................. 103
E.2.1
cuDeviceGetCount() .................................................................... 103
E.2.2
cuDeviceGet() .............................................................................. 103
E.2.3
cuDeviceGetName() ...................................................................... 103
E.2.4
cuDeviceTotalMem() .................................................................... 104
E.2.5
cuDeviceComputeCapability() .................................................. 104
E.2.6
cuDeviceGetAttribute() ............................................................ 104
E.2.7
cuDeviceGetProperties() .......................................................... 105
CUDA Programming Guide Version 1.1
ix
E.3
E.3.1
cuCtxCreate() .............................................................................. 106
E.3.2
cuCtxAttach() .............................................................................. 106
E.3.3
cuCtxDetach() .............................................................................. 106
E.3.4
cuCtxGetDevice() ........................................................................ 106
E.3.5
cuCtxSynchronize() .................................................................... 106
E.4
Module Management ................................................................................ 106
E.4.1
cuModuleLoad() ............................................................................ 106
E.4.2
cuModuleLoadData() .................................................................... 107
E.4.3
cuModuleLoadFatBinary() .......................................................... 107
E.4.4
cuModuleUnload() ........................................................................ 107
E.4.5
cuModuleGetFunction() .............................................................. 107
E.4.6
cuModuleGetGlobal() .................................................................. 107
E.4.7
cuModuleGetTexRef() .................................................................. 108
E.5
Stream Management ................................................................................ 108
E.5.1
cuStreamCreate() ........................................................................ 108
E.5.2
cuStreamQuery() .......................................................................... 108
E.5.3
cuStreamSynchronize() .............................................................. 108
E.5.4
cuStreamDestroy() ...................................................................... 108
E.6
Event Management................................................................................... 108
E.6.1
cuEventCreate() .......................................................................... 108
E.6.2
cuEventRecord() .......................................................................... 108
E.6.3
cuEventQuery() ............................................................................ 109
E.6.4
cuEventSynchronize() ................................................................ 109
E.6.5
cuEventDestroy() ........................................................................ 109
E.6.6
cuEventElapsedTime() ................................................................ 109
E.7
x
Context Management................................................................................ 106
Execution Control ..................................................................................... 109
E.7.1
cuFuncSetBlockShape() .............................................................. 109
E.7.2
cuFuncSetSharedSize() .............................................................. 110
E.7.3
cuParamSetSize() ........................................................................ 110
E.7.4
cuParamSeti() .............................................................................. 110
E.7.5
cuParamSetf() .............................................................................. 110
CUDA Programming Guide Version 1.1
E.7.6
cuParamSetv() .............................................................................. 110
E.7.7
cuParamSetTexRef() .................................................................... 110
E.7.8
cuLaunch() .................................................................................... 110
E.7.9
cuLaunchGrid() ............................................................................ 111
E.8
Memory Management ............................................................................... 111
E.8.1
cuMemGetInfo() ............................................................................ 111
E.8.2
cuMemAlloc() ................................................................................ 111
E.8.3
cuMemAllocPitch() ...................................................................... 111
E.8.4
cuMemFree() .................................................................................. 112
E.8.5
cuMemAllocHost() ........................................................................ 112
E.8.6
cuMemFreeHost() .......................................................................... 112
E.8.7
cuMemGetAddressRange() ............................................................ 112
E.8.8
cuArrayCreate() .......................................................................... 113
E.8.9
cuArrayGetDescriptor() ............................................................ 114
E.8.10
cuArrayDestroy() ........................................................................ 114
E.8.11
cuMemset() .................................................................................... 114
E.8.12
cuMemset2D() ................................................................................ 114
E.8.13
cuMemcpyHtoD() ............................................................................ 115
E.8.14
cuMemcpyDtoH() ............................................................................ 115
E.8.15
cuMemcpyDtoD() ............................................................................ 115
E.8.16
cuMemcpyDtoA() ............................................................................ 116
E.8.17
cuMemcpyAtoD() ............................................................................ 116
E.8.18
cuMemcpyAtoH() ............................................................................ 116
E.8.19
cuMemcpyHtoA() ............................................................................ 116
E.8.20
cuMemcpyAtoA() ............................................................................ 117
E.8.21
cuMemcpy2D() ................................................................................ 117
E.9
Texture Reference Management................................................................ 119
E.9.1
cuTexRefCreate() ........................................................................ 119
E.9.2
cuTexRefDestroy() ...................................................................... 119
E.9.3
cuTexRefSetArray() .................................................................... 119
E.9.4
cuTexRefSetAddress() ................................................................ 120
CUDA Programming Guide Version 1.1
xi
E.9.5
cuTexRefSetFormat() .................................................................. 120
E.9.6
cuTexRefSetAddressMode() ........................................................ 120
E.9.7
cuTexRefSetFilterMode() .......................................................... 120
E.9.8
cuTexRefSetFlags() .................................................................... 121
E.9.9
cuTexRefGetAddress() ................................................................ 121
E.9.10
cuTexRefGetArray() .................................................................... 121
E.9.11
cuTexRefGetAddressMode() ........................................................ 121
E.9.12
cuTexRefGetFilterMode() .......................................................... 121
E.9.13
cuTexRefGetFormat() .................................................................. 122
E.9.14
cuTexRefGetFlags() .................................................................... 122
E.10
OpenGL Interoperability ........................................................................ 122
E.10.1
cuGLInit() .................................................................................... 122
E.10.2
cuGLRegisterBufferObject() .................................................... 122
E.10.3
cuGLMapBufferObject() .............................................................. 122
E.10.4
cuGLUnmapBufferObject() .......................................................... 122
E.10.5
cuGLUnregisterBufferObject() ................................................ 123
E.11
Direct3D Interoperability ....................................................................... 123
E.11.1
cuD3D9Begin() .............................................................................. 123
E.11.2
cuD3D9End() .................................................................................. 123
E.11.3
cuD3D9RegisterVertexBuffer() ................................................ 123
E.11.4
cuD3D9MapVertexBuffer() .......................................................... 123
E.11.5
cuD3D9UnmapVertexBuffer() ...................................................... 123
E.11.6
cuD3D9UnregisterVertexBuffer() ............................................ 123
E.11.7
cuD3D9GetDevice() ...................................................................... 124
Appendix F. Texture Fetching ........................................................................ 125
xii
F.1
Nearest-Point Sampling............................................................................. 126
F.2
Linear Filtering ......................................................................................... 127
F.3
Table Lookup ........................................................................................... 128
CUDA Programming Guide Version 1.1
List of Figures
Figure 1-1.
Floating-Point Operations per Second for the CPU and GPU.....................1
Figure 1-2.
The GPU Devotes More Transistors to Data Processing ............................2
Figure 1-3.
Compute Unified Device Architecture Software Stack ...............................3
Figure 1-4.
The Gather and Scatter Memory Operations ............................................4
Figure 1-5.
Shared Memory Brings Data Closer to the ALUs .......................................5
Figure 2-1.
Thread Batching ....................................................................................9
Figure 2-2.
Memory Model..................................................................................... 11
Figure 3-1.
Hardware Model .................................................................................. 14
Figure 5-1.
Examples of Coalesced Global Memory Access Patterns.......................... 52
Figure 5-2.
Examples of Non-Coalesced Global Memory Access Patterns................... 53
Figure 5-3.
Examples of Non-Coalesced Global Memory Access Patterns................... 54
Figure 5-4.
Examples of Shared Memory Access Patterns without Bank Conflicts ..... 58
Figure 5-5.
Example of a Shared Memory Access Pattern without Bank Conflicts...... 59
Figure 5-6.
Examples of Shared Memory Access Patterns with Bank Conflicts ........... 60
Figure 5-7.
Example of Shared Memory Read Access Patterns with Broadcast........... 61
Figure 6-1.
Matrix Multiplication ............................................................................. 68
CUDA Programming Guide Version 1.1
xiii
Chapter 1.
Introduction to CUDA
1.1
The Graphics Processor Unit as a
Data-Parallel Computing Device
In a matter of just a few years, the programmable graphics processor unit has
evolved into an absolute computing workhorse, as illustrated by Figure 1-1. With
multiple cores driven by very high memory bandwidth, today's GPUs offer
incredible resources for both graphics and non-graphics processing.
GFLOPS
G80GL = Quadro 5600 FX
G80GL
G80 = GeForce 8800 GTX
G71 = GeForce 7900 GTX
G70 = GeForce 7800 GTX
NV40 = GeForce 6800 Ultra
NV35 = GeForce FX 5950 Ultra
NV30 = GeForce FX 5800
Figure 1-1.
Floating-Point Operations per Second for the
CPU and GPU
The main reason behind such an evolution is that the GPU is specialized for
compute-intensive, highly parallel computation – exactly what graphics rendering is
about – and therefore is designed such that more transistors are devoted to data
processing rather than data caching and flow control, as schematically illustrated by
Figure 1-2.
CUDA Programming Guide Version 1.1
1
Chapter 1. Introduction to CUDA
Control
ALU
ALU
ALU
ALU
Cache
DRAM
DRAM
CPU
Figure 1-2.
GPU
The GPU Devotes More Transistors to Data
Processing
More specifically, the GPU is especially well-suited to address problems that can be
expressed as data-parallel computations – the same program is executed on many
data elements in parallel – with high arithmetic intensity – the ratio of arithmetic
operations to memory operations. Because the same program is executed for each
data element, there is a lower requirement for sophisticated flow control; and
because it is executed on many data elements and has high arithmetic intensity, the
memory access latency can be hidden with calculations instead of big data caches.
Data-parallel processing maps data elements to parallel processing threads. Many
applications that process large data sets such as arrays can use a data-parallel
programming model to speed up the computations. In 3D rendering large sets of
pixels and vertices are mapped to parallel threads. Similarly, image and media
processing applications such as post-processing of rendered images, video encoding
and decoding, image scaling, stereo vision, and pattern recognition can map image
blocks and pixels to parallel processing threads. In fact, many algorithms outside the
field of image rendering and processing are accelerated by data-parallel processing,
from general signal processing or physics simulation to computational finance or
computational biology.
Up until now, however, accessing all that computational power packed into the
GPU and efficiently leveraging it for non-graphics applications remained tricky:
The GPU could only be programmed through a graphics API, imposing a high
learning curve to the novice and the overhead of an inadequate API to the nongraphics application.
The GPU DRAM could be read in a general way – GPU programs can gather
data elements from any part of DRAM – but could not be written in a general
way – GPU programs cannot scatter information to any part of DRAM –,
removing a lot of the programming flexibility readily available on the CPU.
Some applications were bottlenecked by the DRAM memory bandwidth, underutilizing the GPU’s computational power.
This document describes a novel hardware and programming model that is a direct
answer to these problems and exposes the GPU as a truly generic data-parallel
computing device.
2
CUDA Programming Guide Version 1.1
Chapter 1. Introduction to CUDA
1.2
CUDA: A New Architecture for Computing on
the GPU
CUDA stands for Compute Unified Device Architecture and is a new hardware
and software architecture for issuing and managing computations on the GPU as a
data-parallel computing device without the need of mapping them to a graphics
API. It is available for the GeForce 8 Series, the Tesla solutions, and some
Quadro solutions (see Appendix A for details). The operating system’s multitasking
mechanism is responsible for managing the access to the GPU by several CUDA
and graphics applications running concurrently.
The CUDA software stack is composed of several layers as illustrated in Figure 1-3:
a hardware driver, an application programming interface (API) and its runtime, and
two higher-level mathematical libraries of common usage, CUFFT and CUBLAS
that are both described in separate documents. The hardware has been designed to
support lightweight driver and runtime layers, resulting in high performance.
CPU
Application
CUDA Libraries
CUDA Runtime
CUDA Driver
GPU
Figure 1-3.
Compute Unified Device Architecture Software
Stack
The CUDA API comprises an extension to the C programming language for a
minimum learning curve (see Chapter 4).
CUDA Programming Guide Version 1.1
3
Chapter 1. Introduction to CUDA
CUDA provides general DRAM memory addressing as illustrated in Figure 1-4 for
more programming flexibility: both scatter and gather memory operations. From a
programming perspective, this translates into the ability to read and write data at any
location in DRAM, just like on a CPU.
Control
ALU
ALU
ALU
...
Cache
Control
ALU
ALU
ALU
...
d4
d5
d6
d7
ALU
ALU
ALU
...
d4
d5
d6
d7
…
Cache
DRAM
d0
d1
d2
d3
…
Gather
Control
ALU
ALU
ALU
...
Cache
Control
…
Cache
DRAM
d0
d1
d2
d3
…
Scatter
Figure 1-4.
4
The Gather and Scatter Memory Operations
CUDA Programming Guide Version 1.1
Chapter 1. Introduction to CUDA
CUDA features a parallel data cache or on-chip shared memory with very fast
general read and write access, that threads use to share data with each other (see
Chapter 3). As illustrated in Figure 1-5, applications can take advantage of it by
minimizing overfetch and round-trips to DRAM and therefore becoming less
dependent on DRAM memory bandwidth.
Control
ALU
ALU
ALU
...
Cache
Control
ALU
ALU
ALU
...
d4
d5
d6
d7
…
Cache
DRAM
d0
d1
d2
d3
…
Without shared memory
Control
ALU
ALU
ALU
...
Control
Cache
Cache
Shared
memory
Shared
memory
ALU
ALU
ALU
...
d0
d1
d2
d3
d4
d5
d6
d7
d0
d1
d2
d3
d4
d5
d6
d7
DRAM
…
…
With shared memory
Figure 1-5.
Shared Memory Brings Data Closer to the ALUs
CUDA Programming Guide Version 1.1
5
Chapter 1. Introduction to CUDA
1.3
Document’s Structure
This document is organized into the following chapters:
6
Chapter 1 contains a general introduction to CUDA.
Chapter 2 outlines the programming model.
Chapter 3 describes its hardware implementation.
Chapter 4 describes the CUDA API and runtime.
Chapter 5 gives some guidance on how to achieve maximum performance.
Chapter 6 illustrates the previous chapters by walking through the code of some
simple example.
Appendix A gives the technical specifications of various devices.
Appendix B lists the mathematical functions supported in CUDA.
Appendix C lists the atomic functions supported in CUDA.
Appendix D is the CUDA runtime API reference.
Appendix E is the CUDA driver API reference.
Appendix F gives more details on texture fetching.
CUDA Programming Guide Version 1.1
Chapter 2.
Programming Model
2.1
A Highly Multithreaded Coprocessor
When programmed through CUDA, the GPU is viewed as a compute device capable of
executing a very high number of threads in parallel. It operates as a coprocessor to
the main CPU, or host: In other words, data-parallel, compute-intensive portions of
applications running on the host are off-loaded onto the device.
More precisely, a portion of an application that is executed many times, but
independently on different data, can be isolated into a function that is executed on
the device as many different threads. To that effect, such a function is compiled to
the instruction set of the device and the resulting program, called a kernel, is
downloaded to the device.
Both the host and the device maintain their own DRAM, referred to as host memory
and device memory, respectively. One can copy data from one DRAM to the other
through optimized API calls that utilize the device’s high-performance Direct
Memory Access (DMA) engines.
2.2
Thread Batching
The batch of threads that executes a kernel is organized as a grid of thread blocks as
described in Sections 2.2.1 and 2.2.2 and illustrated in Figure 2-1.
2.2.1
Thread Block
A thread block is a batch of threads that can cooperate together by efficiently
sharing data through some fast shared memory and synchronizing their execution to
coordinate memory accesses. More precisely, one can specify synchronization points
in the kernel, where threads in a block are suspended until they all reach the
synchronization point.
Each thread is identified by its thread ID, which is the thread number within the
block. To help with complex addressing based on the thread ID, an application can
also specify a block as a two- or three-dimensional array of arbitrary size and
identify each thread using a 2- or 3-component index instead. For a two-
CUDA Programming Guide Version 1.1
7
Chapter 2. Programming Model
dimensional block of size (Dx, Dy), the thread ID of a thread of index (x, y) is
(x + y Dx) and 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).
2.2.2
Grid of Thread Blocks
There is a limited maximum number of threads that a block can contain. However,
blocks of same dimensionality and size that execute the same kernel can be batched
together into a grid of blocks, so that the total number of threads that can be
launched in a single kernel invocation is much larger. This comes at the expense of
reduced thread cooperation, because threads in different thread blocks from the
same grid cannot communicate and synchronize with each other. This model allows
kernels to efficiently run without recompilation on various devices with different
parallel capabilities: A device may run all the blocks of a grid sequentially if it has
very few parallel capabilities, or in parallel if it has a lot of parallel capabilities, or
usually a combination of both.
Each block is identified by its block ID, which is the block number within the grid.
To help with complex addressing based on the block ID, an application can also
specify a grid as a two-dimensional array of arbitrary size and identify each block
using a 2-component index instead. For a two-dimensional block of size (Dx, Dy),
the block ID of a block of index (x, y) is (x + y Dx).
8
CUDA Programming Guide Version 1.1
Chapter 2. Programming Model
Host
Device
Grid 1
Kernel 1
Block
(0, 0)
Block
(1, 0)
Block
(2, 0)
Block
(0, 1)
Block
(1, 1)
Block
(2, 1)
Grid 2
Kernel 2
Block (1, 1)
Thread
(0, 0)
Thread
(1, 0)
Thread
(2, 0)
Thread
(3, 0)
Thread
(4, 0)
Thread
(0, 1)
Thread
(1, 1)
Thread
(2, 1)
Thread
(3, 1)
Thread
(4, 1)
Thread
(0, 2)
Thread
(1, 2)
Thread
(2, 2)
Thread
(3, 2)
Thread
(4, 2)
The host issues a succession of kernel invocations to the device. Each kernel is executed as a batch
of threads organized as a grid of thread blocks
Figure 2-1.
Thread Batching
CUDA Programming Guide Version 1.1
9
Chapter 2. Programming Model
2.3
Memory Model
A thread that executes on the device has only access to the device’s DRAM and
on-chip memory through the following memory spaces, as illustrated in Figure 2-2:
Read-write per-thread registers,
Read-write per-thread local memory,
Read-write per-block shared memory,
Read-write per-grid global memory,
Read-only per-grid constant memory,
Read-only per-grid texture memory.
The global, constant, and texture memory spaces can be read from or written to by
the host and are persistent across kernel launches by the same application.
The global, constant, and texture memory spaces are optimized for different
memory usages (see Sections 5.1.2.1, 5.1.2.2, and 5.1.2.3). Texture memory also
offers different addressing modes, as well as data filtering, for some specific data
formats (see Section 4.3.4).
10
CUDA Programming Guide Version 1.1
Chapter 2. Programming Model
Grid
Block (0, 0)
Block (1, 0)
Shared Memory
Shared Memory
Registers
Registers
Registers
Registers
Thread (0, 0)
Thread (1, 0)
Thread (0, 0)
Thread (1, 0)
Local
Memory
Local
Memory
Local
Memory
Local
Memory
Global
Memory
Constant
Memory
Texture
Memory
A thread has access to the device’s DRAM and on-chip memory through a set of
memory spaces of various scopes.
Figure 2-2.
Memory Model
CUDA Programming Guide Version 1.1
11
Chapter 3.
Hardware Implementation
3.1
A Set of SIMD Multiprocessors with On-Chip
Shared Memory
The device is implemented as a set of multiprocessors as illustrated in Figure 3-1. Each
multiprocessor has a Single Instruction, Multiple Data architecture (SIMD): At any
given clock cycle, each processor of the multiprocessor executes the same
instruction, but operates on different data.
Each multiprocessor has on-chip memory of the four following types:
One set of local 32-bit registers per processor,
A parallel data cache or shared memory that is shared by all the processors and
implements the shared memory space,
A read-only constant cache that is shared by all the processors and speeds up reads
from the constant memory space, which is implemented as a read-only region of
device memory,
A read-only texture cache that is shared by all the processors and speeds up reads
from the texture memory space, which is implemented as a read-only region of
device memory.
The local and global memory spaces are implemented as read-write regions of
device memory and are not cached.
Each multiprocessor accesses the texture cache via a texture unit that implements the
various addressing modes and data filtering mentioned in Section 2.3.
CUDA Programming Guide Version 1.1
13
Chapter 3. Hardware Implementation
Device
Multiprocessor N
Multiprocessor 2
Multiprocessor 1
Shared Memory
Registers
Processor 1
Registers
Processor 2
Registers
…
Instruction
Unit
Processor M
Constant
Cache
Texture
Cache
Device Memory
A set of SIMD multiprocessors with on-chip shared memory.
Figure 3-1.
3.2
Hardware Model
Execution Model
A grid of thread blocks is executed on the device by scheduling blocks for execution
on the multiprocessors. Each multiprocessor processes batches of blocks one batch
after the other. A block is processed by only one multiprocessor, so that the shared
memory space resides in the on-chip shared memory leading to very fast memory
accesses.
How many blocks each multiprocessor can process in one batch depends on how
many registers per thread and how much shared memory per block are required for
14
CUDA Programming Guide Version 1.1
Chapter 3: Hardware Implementation
a given kernel since the multiprocessor’s registers and shared memory are split
among all the threads of the batch of blocks. If there are not enough registers or
shared memory available per multiprocessor to process at least one block, the kernel
will fail to launch.
The blocks that are processed by one multiprocessor in one batch are referred to as
active. Each active block is split into SIMD groups of threads called warps: Each of
these warps contains the same number of threads, called the warp size, and is
executed by the multiprocessor in a SIMD fashion. Active warps – i.e. all the warps
from all active blocks – are time-sliced: A thread scheduler periodically switches from
one warp to another to maximize the use of the multiprocessor’s computational
resources. A half-warp is either the first or second half of a warp.
The way a block is split into warps is always the same; each warp contains threads of
consecutive, increasing thread IDs with the first warp containing thread 0.
Section 2.2.1 describes how thread IDs relate to thread indices in the block.
The issue order of the warps within a block is undefined, but their execution can be
synchronized, as mentioned in Section 2.2.1, to coordinate global or shared memory
accesses.
The issue order of the blocks within a grid of thread blocks is undefined and there is
no synchronization mechanism between blocks, so threads from two different
blocks of the same grid cannot safely communicate with each other through global
memory during the execution of the grid.
If a non-atomic instruction executed by a warp writes to the same location in global
or shared memory for more than one of the threads of the warp, the number of
serialized writes that occur to that location and the order in which they occur is
undefined, but one of the writes is guaranteed to succeed. If an atomic instruction
(see Section 4.4.6) executed by a warp reads, modifies, and writes to the same
location in global memory for more than one of the threads of the warp, each read,
modify, write to that location occurs and they are all serialized, but the order in
which they occur is undefined.
3.3
Compute 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
devices listed in Appendix A are all of compute capability 1.x (Their major revision
number is 1).
The minor revision number corresponds to an incremental improvement to the core
architecture, possibly including new features.
The technical specifications of the various compute capabilities are given in
Appendix A.
CUDA Programming Guide Version 1.1
15
Chapter 3. Hardware Implementation
3.4
Multiple Devices
The use of multiple GPUs as CUDA devices by an application running on a multiGPU system is only guaranteed to work if theses GPUs are of the same type. If the
system is in SLI mode however, only one GPU can be used as a CUDA device since
all the GPUs are fused at the lowest levels in the driver stack. SLI mode needs to be
turned off in the control panel for CUDA to be able to see each GPU as separate
devices.
3.5
Mode Switches
GPUs 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. (Fullscreen 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, resulting in a crash of these applications.
16
CUDA Programming Guide Version 1.1
Chapter 4.
Application Programming Interface
4.1
An Extension to the C Programming
Language
The goal of the CUDA programming interface is to provide a relatively simple path
for users familiar with the C programming language to easily write programs for
execution by the device.
It consists of:
A minimal set of extensions to the C language, described in Section 4.2, that
allow the programmer to target portions of the source code for execution on the
device;
A runtime library split into:
¾ A host component, described in Section 4.5, that runs on the host and
provides functions to control and access one or more compute devices
from the host;
¾ A device component, described in Section 4.4, that runs on the device and
provides device-specific functions;
¾ A common component, described in Section 4.3, that provides built-in
vector types and a subset of the C standard library that are supported in
both host and device code.
It should be emphasized that the only functions from the C standard library that are
supported to run on the device are the functions provided by the common runtime
component.
4.2
Language Extensions
The extensions to the C programming language are four-fold:
Function type qualifiers to specify whether a function executes on the host or on
the device and whether it is callable from the host or from the device
(Section 4.2.1);
Variable type qualifiers to specify the memory location on the device of a
variable (Section 4.2.2);
CUDA Programming Guide Version 1.1
17
Chapter 4. Application Programming Interface
A new directive to specify how a kernel is executed on the device from the host
(Section 4.2.3);
Four built-in variables that specify the grid and block dimensions and the block
and thread indices (Section 4.2.4).
Each source file containing these extensions must be compiled with the CUDA
compiler nvcc, as briefly described in Section 4.2.5. A detailed description of nvcc
can be found in a separate document.
Each of these extensions come with some restrictions described in each of the
sections below. nvcc will give an error or a warning on some violations of these
restrictions, but some of them cannot be detected.
4.2.1
Function Type Qualifiers
4.2.1.1
__device__
The __device__ qualifier declares a function that is:
Executed on the device
Callable from the device only.
4.2.1.2
__global__
The __global__ qualifier declares a function as being a kernel. Such a function is:
Executed on the device,
Callable from the host only.
4.2.1.3
__host__
The __host__ qualifier declares a function that is:
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.
However, the __host__ qualifier can also be used in combination with the
__device__ qualifier, in which case the function is compiled for both the host and
the device.
4.2.1.4
Restrictions
__device__ and __global__ functions do not support recursion.
__device__ and __global__ functions cannot declare static variables inside
their body.
__device__ and __global__ functions cannot have a variable number of
arguments.
__device__ functions cannot have their address taken; function pointers to
__global__ functions, on the other hand, are supported.
The __global__ and __host__ qualifiers cannot be used together.
18
CUDA Programming Guide Version 1.1
Chapter 4. Application Programming Interface
__global__ functions must have void return type.
Any call to a __global__ function must specify its execution configuration as
described in Section 4.2.3.
A call to a __global__ function is asynchronous, meaning it returns before the
device has completed its execution.
__global__ function parameters are currently passed via shared memory to the
device and limited to 256 bytes.
4.2.2
Variable Type Qualifiers
4.2.2.1
__device__
The __device__ qualifier declares a variable that resides on the device.
At most one of the other type qualifiers defined in the next three sections may be
used together with __device__ to further specify which memory space the
variable belongs to. If none of them is present, the variable:
Resides in global memory space,
Has the lifetime of an application,
Is accessible from all the threads within the grid and from the host through the
runtime library.
4.2.2.2
__constant__
The __constant__ qualifier, optionally used together with __device__,
declares a variable that:
Resides in constant memory space,
Has the lifetime of an application,
Is accessible from all the threads within the grid and from the host through the
runtime library.
4.2.2.3
__shared__
The __shared__ qualifier, optionally used together with __device__, declares a
variable that:
Resides in the shared memory space of a thread block,
Has the lifetime of the block,
Is only accessible from all the threads within the block.
There is full sequential consistency of shared variables within a thread, however
relaxed ordering across threads. Only after the execution of a __syncthreads()
(Section 4.4.2) are the writes from other threads guaranteed to be visible. Unless the
variable is declared as volatile, the compiler is free to optimize the reads and writes
to shared memory as long as the previous statement is met.
When declaring a variable in shared memory as an external array such as
extern __shared__ float shared[];
the size of the array is determined at launch time (see Section 4.2.3). All variables
declared in this fashion, start at the same address in memory, so that the layout of
CUDA Programming Guide Version 1.1
19
Chapter 4. Application Programming Interface
the variables in the array must be explicitly managed through offsets. For example, if
one wants the equivalent of
short array0[128];
float array1[64];
int
array2[256];
in dynamically allocated shared memory, one could declare and initialize the arrays
the following way:
extern __shared__ char array[];
__device__ void func()
// __device__ or __global__ function
{
short* array0 = (short*)array;
float* array1 = (float*)&array0[128];
int*
array2 =
(int*)&array1[64];
}
4.2.2.4
Restrictions
These qualifiers are not allowed on struct and union members, on formal
parameters and on local variables within a function that executes on the host.
__shared__ and __constant__ variables have implied static storage.
__device__, __shared__ and __constant__ variables cannot be defined as
external using the extern keyword.
__device__ and __constant__ variables are only allowed at file scope.
__constant__ variables cannot be assigned to from the device, only from the
host through host runtime functions (Sections 4.5.2.3 and 4.5.3.6).
__shared__ variables cannot have an initialization as part of their declaration.
An automatic variable declared in device code without any of these qualifiers
generally resides in a register. However in some cases the compiler might choose to
place it in local memory. This is often the case for large structures or arrays that
would consume too much register space, and arrays for which the compiler cannot
determine that they are indexed with constant quantities. Inspection of the ptx
assembly code (obtained by compiling with the –ptx or -keep option) will tell if a
variable has been placed in local memory during the first compilation phases as it
will be declared using the .local mnemonic and accessed using the ld.local
and st.local mnemonics. If it has not, subsequent compilation phases might still
decide otherwise though if they find it consumes too much register space for the
targeted architecture. This can be checked by compiling with the --ptxasoptions=-v option that reports local memory usage (lmem).
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 or the
global memory space, otherwise they are restricted to only point to memory
allocated or declared in the global memory space.
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
20
CUDA Programming Guide Version 1.1
Chapter 4. Application Programming Interface
cudaGetSymbolAddress() as described in Section 4.5.2.3 can only be used in
host code.
4.2.3
Execution Configuration
Any call to a __global__ function must specify the execution configuration for that
call.
The execution configuration defines the dimension of the grid and blocks that will
be used to execute the function on the device, as well as the associated stream (see
Section 4.5.1.5 for a description of streams). It is specified by inserting an
expression of the form <<< Dg, Db, Ns, S >>> between the function name
and the parenthesized argument list, where:
is of type dim3 (see Section 4.3.1.2) and specifies the dimension and size of
the grid, such that Dg.x * Dg.y equals the number of blocks being launched;
Dg.z is unused;
Db is of type dim3 (see Section 4.3.1.2) and specifies the dimension and size of
each block, such that Db.x * Db.y * Db.z equals the number of threads per
block;
Ns is of type size_t and specifies the number of bytes in shared memory that
is dynamically allocated per block for this call in addition to the statically
allocated memory; this dynamically allocated memory is used by any of the
variables declared as an external array as mentioned in Section 4.2.2.3; Ns is an
optional argument which defaults to 0;
S is of type cudaStream_t and specifies the associated stream; S is an optional
argument which defaults to 0.
As an example, a function declared as
Dg
__global__ void Func(float* parameter);
must be called like this:
Func<<< Dg, Db, Ns >>>(parameter);
The arguments to the execution configuration are evaluated before the actual
function arguments and like the function arguments, are currently passed via shared
memory to the device.
The function call will fail if Dg or Db are greater than the maximum sizes allowed
for the device as specified in Appendix A.1, or if Ns is greater than the maximum
amount of shared memory available on the device, minus the amount of shared
memory required for static allocation, functions arguments, and execution
configuration.
4.2.4
Built-in Variables
4.2.4.1
gridDim
This variable is of type dim3 (see Section 4.3.1.2) and contains the dimensions of
the grid.
CUDA Programming Guide Version 1.1
21
Chapter 4. Application Programming Interface
4.2.4.2
blockIdx
This variable is of type uint3 (see Section 4.3.1.1) and contains the block index
within the grid.
4.2.4.3
blockDim
This variable is of type dim3 (see Section 4.3.1.2) and contains the dimensions of
the block.
4.2.4.4
threadIdx
This variable is of type uint3 (see Section 4.3.1.1) and contains the thread index
within the block.
4.2.4.5
Restrictions
It is not allowed to take the address of any of the built-in variables.
It is not allowed to assign values to any of the built-in variables.
4.2.5
Compilation with NVCC
nvcc is a compiler driver that simplifies the process of compiling CUDA code: It
provides simple and familiar command line options and executes them by invoking
the collection of tools that implement the different compilation stages.
nvcc’s basic workflow consists in separating device code from host code and
compiling the device code into a binary form or cubin object. The generated host
code is output either as C code that is left to be compiled using another tool or as
object code directly by invoking the host compiler during the last compilation stage.
Applications can either ignore the generated host code and load and execute the
cubin object on the device using the CUDA driver API (see Section 4.5.3), or they
can link to the generated host code, which includes the cubin object as a global
initialized data array and contains a translation of the execution configuration syntax
described in Section 4.2.3 into the necessary CUDA runtime startup code to load
and launch each compiled kernel (see Section 4.5.2).
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 the C subset
of C++ is fully supported for the device code; C++ specific features such as classes,
inheritance, or declaration of variables within basic blocks are not. 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.
A detailed description of nvcc’s workflow and command options can be found in a
separate document.
nvcc introduces two compiler directives described in the following sections.
4.2.5.1
__noinline__
By default, a __device__ function is always inlined. The __noinline__
function qualifier however can be used as a hint for the compiler not to inline the
function if possible. The function body must still be in the same file where it is
called.
22
CUDA Programming Guide Version 1.1
Chapter 4. Application Programming Interface
The compiler will not honor the __noinline__ qualifier for functions with
pointer parameters and for functions with large parameter lists.
4.2.5.2
#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. It is up to the programmer to make sure that
unrolling will not affect the correctness of the program (which it might, in the above
example, if n is smaller than 5).
#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.
4.3
Common Runtime Component
The common runtime component can be used by both host and device functions.
4.3.1
Built-in Vector Types
4.3.1.1
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, float1, float2, float3, float4
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_; for example,
int2 make_int2(int x, int y);
which creates a vector of type int2 with value (x, y).
4.3.1.2
dim3 Type
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.
CUDA Programming Guide Version 1.1
23
Chapter 4. Application Programming Interface
4.3.2
Mathematical Functions
Table B-1 contains a comprehensive list of the C/C++ standard library
mathematical functions that are currently supported, along with their respective
error bounds when executed on the device.
When executed in host code, a given function uses the C runtime implementation if
available.
4.3.3
Time Function
clock_t clock();
returns the value of a counter that is incremented every clock cycle.
Sampling this counter at the beginning and at the end of a kernel, taking the
difference of the two samples, and recording the result per thread provides a
measure for each thread of the number of clock cycles taken by the device to
completely execute the thread, but not of the number of clock cycles the device
actually spent executing thread instructions. The former number is greater that the
latter since threads are time sliced.
4.3.4
Texture Type
CUDA supports a subset of the texturing hardware that the GPU uses for graphics
to access texture memory. Reading data from texture memory instead of global
memory can have several performance benefits as described in Section 5.4.
Texture memory is read from kernels using device functions called texture fetches,
described in Section 4.4.5. The first parameter of a texture fetch specifies an object
called a texture reference.
A texture reference defines which part of texture memory is fetched. It must be
bound through host runtime functions (Sections 4.5.2.6 and 4.5.3.9) to some region
of memory, called a texture, before it can be used by a kernel. Several distinct texture
references might be bound to the same texture or to textures that overlap in
memory.
A texture reference has several attributes. One of them is its dimensionality that
specifies whether the texture is addressed as a one-dimensional array using one
texture coordinate, or as a two-dimensional array using two texture coordinates.
Elements of the array are called texels, short for “texture elements.”
Other attributes define the input and output data types of the texture fetch, as well
as how the input coordinates are interpreted and what processing should be done.
4.3.4.1
Texture Reference Declaration
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 texRef;
where:
24
CUDA Programming Guide Version 1.1
Chapter 4. Application Programming Interface
specifies the type of data that is returned when fetching the texture; Type
is restricted to the basic integer and floating-point types and any of the 1-, 2-,
and 4-component vector types defined in Section 4.3.1.1;
Dim specifies the dimensionality of the texture reference and is equal to 1 or 2;
Dim is an optional argument which defaults to 1;
ReadMode is equal to cudaReadModeNormalizedFloat or
cudaReadModeElementType; if it is cudaReadModeNormalizedFloat
and Type is a 16-bit or 8-bit integer type, the value is actually returned as
floating-point type and the full range of the integer type is mapped to [0.0, 1.0]
for unsigned integer type and [-1.0, 1.0] for signed integer type; for example, an
unsigned 8-bit texture element with the value 0xff reads as 1; if it is
cudaReadModeElementType, no conversion is performed; ReadMode is an
optional argument which defaults to cudaReadModeElementType.
Type
4.3.4.2
Runtime Texture Reference Attributes
The other attributes of a texture reference are mutable and can be changed at
runtime through the host runtime (Section 4.5.2.6 for the runtime API and
Section 4.5.3.9 for the driver API). They specify whether texture coordinates are
normalized or not, the addressing mode, and texture filtering, as detailed below.
By default, textures are referenced using floating-point coordinates in the range
[0, N) where N is the size of the texture in the dimension corresponding to the
coordinate. For example, a texture that is 64×32 in size will be referenced with
coordinates in the range [0, 63] and [0, 31] for the x and y dimensions, respectively.
Normalized texture coordinates cause the coordinates to be specified in the range
[0.0, 1.0) instead of [0, N), so the same 64×32 texture would be addressed by
normalized coordinates in the range [0, 1) in both the x and y dimensions.
Normalized texture coordinates are a natural fit to some applications’ requirements,
if it is preferable for the texture coordinates to be independent of the texture size.
The addressing mode defines what happens when texture coordinates are out of
range. When using unnormalized texture coordinates, texture coordinates outside
the range [0, N) are clamped: Values below 0 are set to 0 and values greater or equal
to N are set to N-1. Clamping is also the default addressing mode when using
normalized texture coordinates: Values below 0.0 or above 1.0 are clamped to the
range [0.0, 1.0). For normalized coordinates, the “wrap” addressing mode also may
be specified. Wrap addressing is usually used when the texture contains a periodic
signal. It uses only the fractional part of the texture coordinate; for example, 1.25 is
treated the same as 0.25 and -1.25 is treated the same as 0.75.
Linear texture filtering may be done only for textures that are configured to return
floating-point data. It performs low-precision interpolation between neighboring
texels. When enabled, the texels surrounding a texture fetch location are read and
the return value of the texture fetch is interpolated based on where the texture
coordinates fell between the texels. Simple linear interpolation is performed for onedimensional textures and bilinear interpolation is performed for two-dimensional
textures.
Appendix F gives more details on texture fetching.
4.3.4.3
Texturing from Linear Memory versus CUDA Arrays
A texture can be any region of linear memory or a CUDA array (see Section 4.5.1.2).
CUDA Programming Guide Version 1.1
25
Chapter 4. Application Programming Interface
Textures allocated in linear memory:
Can only be of dimensionality equal to 1;
Do not support texture filtering;
Can only be addressed using a non-normalized integer texture coordinate;
Do not support the various addressing modes: Out-of-range texture accesses
return zero.
The hardware enforces an alignment requirement on texture base addresses. To
abstract this alignment requirement from programmers, the functions to bind
texture references onto device memory pass back a byte offset that must be applied
to texture fetches in order to read from the desired memory. The base pointers
returned by CUDA’s allocation routines conform to this alignment constraint, so
applications can avoid the offsets altogether by passing allocated pointers to
cudaBindTexture()/cuTexRefSetAddress().
4.4
Device Runtime Component
The device runtime component can only be used in device functions.
4.4.1
Mathematical Functions
For some of the functions of Table B-1, a less accurate, but faster version exists in
the device runtime component; it has the same name prefixed with __ (such as
__sin(x)). These intrinsic functions are listed in Table B-2, along with their
respective error bounds.
The compiler has an option (-use_fast_math) to force every function to compile
to its less accurate counterpart if it exists.
4.4.2
Synchronization Function
void __syncthreads();
synchronizes all threads in a block. Once all threads have reached this point,
execution resumes normally.
__syncthreads() is used to coordinate communication between the threads of a
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 writeafter-write 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.
4.4.3
Type Conversion Functions
The suffixes in the function below indicate IEEE-754 rounding modes:
rn
26
is round-to-nearest-even,
CUDA Programming Guide Version 1.1
Chapter 4. Application Programming Interface
is round-towards-zero,
ru is round-up (to positive infinity),
rd is round-down (to negative infinity).
rz
int __float2int_[rn,rz,ru,rd](float);
converts the floating-point argument to an integer, using the specified rounding
mode.
unsigned int __float2uint_[rn,rz,ru,rd](float);
converts the floating-point argument to an unsigned integer, using the specified
rounding mode.
float __int2float_[rn,rz,ru,rd](int);
converts the integer argument to a floating-point number, using the specified
rounding mode.
float __uint2float_[rn,rz,ru,rd](unsigned int);
converts the unsigned integer argument to a floating-point number, using the
specified rounding mode.
4.4.4
Type Casting Functions
float __int_as_float(int);
performs a floating-point type cast on the integer argument, leaving the value
unchanged. For example, __int_as_float(0xC0000000) is equal to -2.
int __float_as_int(float);
performs an integer type cast on the floating-point argument, leaving the value
unchanged. For example, __float_as_int(1.0f) is equal to 0x3f800000.
4.4.5
Texture Functions
4.4.5.1
Texturing from Device Memory
When texturing from device memory, the texture is accessed with the
tex1Dfetch() family of functions; for example:
template
Type tex1Dfetch(
texture texRef,
int x);
float tex1Dfetch(
texture texRef,
int x);
float tex1Dfetch(
texture texRef,
int x);
float tex1Dfetch(
texture texRef,
int x);
float tex1Dfetch(
CUDA Programming Guide Version 1.1
27
Chapter 4. Application Programming Interface
texture texRef,
int x);
These functions fetch the region of linear memory bound to texture reference
texRef using texture coordinate x. No texture filtering and addressing modes are
supported. For integer types, these functions may optionally promote the integer to
32-bit floating point.
Besides the functions shown above, 2-, and 4-tuples are supported; for example:
float4 tex1Dfetch(
texture texRef,
int x);
fetches the linear memory bound to texture reference texRef using texture
coordinate x.
4.4.5.2
Texturing from CUDA Arrays
When texturing from CUDA arrays, the texture is accessed with the tex1D() or
tex2D():
template
Type tex1D(texture texRef, float x);
template
Type tex2D(texture texRef, float x, float y);
These functions fetches the CUDA array bound to texture reference texRef using
texture coordinates x and y. A combination of the texture reference’s immutable
(compile-time) and mutable (runtime) attributes determine how the coordinates are
interpreted, what processing occurs during the texture fetch, and the return value
delivered by the texture fetch (see Sections 4.3.4.1 and 4.3.4.2).
4.4.6
Atomic Functions
Atomic functions are only available for devices of compute capability 1.1. They are
listed in Appendix C.
An atomic function performs a read-modify-write atomic operation on one 32-bit
word residing in global memory. For example, atomicAdd() reads a 32-bit word
at some address in global memory, adds an integer to it, and writes the result back to
the same address. The operation is atomic in the sense that it is guaranteed to be
performed without interference from other threads. In other words, no other thread
can access this address until the operation is complete.
Atomic operations only work with 32-bit signed and unsigned integers.
4.5
Host Runtime Component
The host runtime component can only be used by host functions.
It provides functions to handle:
Device management,
Context management,
Memory management,
28
CUDA Programming Guide Version 1.1
Chapter 4. Application Programming Interface
Code module management,
Execution control,
Texture reference management,
Interoperability with OpenGL and Direct3D.
It is composed of two APIs:
A low-level API called the CUDA driver API,
A higher-level API called the CUDA runtime API that is implemented on top of
the CUDA driver API.
These APIs are mutually exclusive: An application should use either one or the
other.
The CUDA runtime eases device code management by providing implicit
initialization, context management, and module management. The C host code
generated by nvcc is based on the CUDA runtime (see Section 4.2.5), so
applications that link to this code must use the CUDA runtime API.
In contrast, the CUDA driver API requires more code, is harder to program and
debug, but offers a better level of control and is language-independent since it only
deals with cubin objects (see Section 4.2.5). In particular, it is more difficult to
configure and launch kernels using the CUDA driver API, since the execution
configuration and kernel parameters must be specified with explicit function calls
instead of the execution configuration syntax described in Section 4.2.3. Also, device
emulation (see Section 4.5.2.7) does not work with the CUDA driver API.
The CUDA driver API is delivered through the cuda dynamic library and all its
entry points are prefixed with cu.
The CUDA runtime API is delivered through the cudart dynamic library and all
its entry points are prefixed with cuda.
4.5.1
Common Concepts
4.5.1.1
Device
Both APIs provide functions to enumerate the devices available on the system,
query their properties, and select one of them for kernel executions (see
Section 4.5.2.2 for the runtime API and Section 4.5.3.2 for the driver API).
Several host threads can execute device code on the same device, but by design, a
host thread can execute device code on only one device. As a consequence, multiple
host threads are required to execute device code on multiple devices. Also, any
CUDA resources created through the runtime in one host thread cannot be used by
the runtime from another host thread.
4.5.1.2
Memory
Device memory can be allocated either as linear memory or as CUDA arrays.
Linear memory exists on the device in a 32-bit address space, so separately allocated
entities can reference one another via pointers, for example, in a binary tree.
CUDA arrays are opaque memory layouts optimized for texture fetching (see
Section 4.3.4). They are one-dimensional or two-dimensional and composed of
elements, each of which has 1, 2 or 4 components that may be signed or unsigned
CUDA Programming Guide Version 1.1
29
Chapter 4. Application Programming Interface
8-, 16- or 32-bit integers, 16-bit floats (currently only supported through the driver
API), or 32-bit floats. CUDA arrays are only readable by kernels through texture
fetching and may only be bound to texture references with the same number of
packed components.
Both linear memory and CUDA arrays are readable and writable by the host
through the memory copy functions described in Sections 4.5.2.3 and 4.5.3.6.
The host runtime also provides functions to allocate and free page-locked host
memory – as opposed to regular pageable host memory allocated by malloc() (see
Sections D.5.6 and D.5.7 for the runtime API and Sections E.8.5 and Section E.8.6
for the driver API). One advantage of page-locked memory is that the bandwidth
between host memory and device memory is higher if host memory is allocated as
page-locked – only for data transfers performed by the host thread that allocated
host memory. Page-locked 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, allocating too much page-locked memory reduces overall system
performance.
4.5.1.3
OpenGL Interoperability
OpenGL buffer objects may be mapped into the address space of CUDA, either to
enable CUDA to read data written by OpenGL or to enable CUDA to write data
for consumption by OpenGL. Section 4.5.2.7 describes how this is done with the
runtime API and Section 4.5.3.10, with the driver API.
4.5.1.4
Direct3D Interoperability
Direct3D 9.0 vertex buffers may be mapped into the address space of CUDA, either
to enable CUDA to read data written by Direct3D or to enable CUDA to write data
for consumption by Direct3D. Section 4.5.2.8 describes how this is done with the
runtime API and Section 4.5.2.8, with the driver API.
A CUDA context may interoperate with only one Direct3D device at a time,
bracketed by calls to the begin/end functions described in Sections 4.5.2.8 and
4.5.3.11.
The CUDA context and Direct3D device must be created on the same GPU. This
can be ensured by querying the CUDA device corresponding to the adapter used by
Direct3D, using cudaD3D9GetDevice() for the runtime API (see Section D.9.7)
or cuD3D9GetDevice() for the driver API (see Section E.11.7).
The Direct3D device must also be created with the
D3DCREATE_HARDWARE_VERTEXPROCESSING flag.
CUDA does not yet support:
Versions other than Direct3D 9.0,
Direct3D objects other than vertex buffers.
Incidently, cudaD3D9GetDevice() or cuD3D9GetDevice() can also be used
to ensure that the Direct3D device and CUDA context are created on different
devices in the case where load balancing Direct3D and CUDA is preferred over
interoperability.
30
CUDA Programming Guide Version 1.1
Chapter 4. Application Programming Interface
4.5.1.5
Asynchronous Concurrent Execution
In order to facilitate concurrent execution between host and device, some runtime
functions are asynchronous: Control is returned to the application before the device
has completed the requested task. These are:
Kernel launches through __global__ functions or cuGridLaunch() and
cuGridLaunchAsync();
The functions that perform memory copies and are suffixed with Async;
The functions that perform device ↔ device memory copies;
The functions that set memory.
Some devices can also perform copies between page-locked host memory and
device memory concurrently with kernel execution. Applications may query this
capability by calling cuDeviceGetAttribute() with
CU_DEVICE_ATTRIBUTE_GPU_OVERLAP (see Sections E.2.6 respectively). This
capability is currently supported only for memory copies that do not involve CUDA
arrays or 2D arrays allocated through cudaMallocPitch() (see Section 4.5.2.3)
or cuMemAllocPitch() (see Section 4.5.3.6).
Applications manage concurrency through streams. A stream is a sequence of
operations that execute in order. Different streams, on the other hand, may execute
their operations out of order with respect to one another or concurrently.
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.
Section 4.5.2.4 describes how this is done with the runtime API and Section 4.5.3.7,
with the driver API.
Any kernel launch, memory set, or memory copy for which a zero stream parameter
has been specified begins only after all preceding operations are done, including
operations that are part of streams, and no subsequent operation may begin until it
is done.
cudaStreamQuery() for the runtime API and cuStreamQuery() for the
driver API (see Sections D.3.2 and E.5.2 respectively) provide applications with a
way to know if all preceding operations in a stream have completed.
cudaStreamSynchronize() for the runtime API and
cuStreamSynchronize() for the driver API (see Sections E.5.2 and E.5.3
respectively) provide a way to explicitly force the runtime to wait until all preceding
operations in a stream have completed.
Similarly, with cudaThreadSynchronize() for the runtime API and
cuCtxSynchronize() for the driver API (see Sections D.2.1 and E.3.5
respectively) applications can force the runtime to wait until all preceding device
tasks have completed. To avoid unnecessary slowdowns, these functions are best
used for timing purposes or to isolate a launch or memory copy that is failing.
The runtime also provides a way to closely monitor the device’s 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 actually recorded. An
event is recorded when all tasks – or optionally, all operations in a given stream –
preceding the event have completed. Section 4.5.2.5 describes how this is done with
the runtime API and Section 4.5.3.8, with the driver API.
CUDA Programming Guide Version 1.1
31
Chapter 4. Application Programming Interface
Two operations from different streams cannot run concurrently if either a pagelocked host memory allocation, a device memory allocation, a device memory set, a
device ↔ device memory copy, or an event recording occurs in-between them.
Programmers can globally disable asynchronous execution 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.
4.5.2
Runtime API
4.5.2.1
Initialization
There is no explicit initialization function for the runtime API; it initializes the first
time a runtime function is called. 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.
4.5.2.2
Device Management
The functions from Section D.1 are used to manage the devices present in the
system.
cudaGetDeviceCount() and cudaGetDeviceProperties() provide a way
to enumerate these devices and retrieve their properties:
int deviceCount;
cudaGetDeviceCount(&deviceCount);
int device;
for (device = 0; device < deviceCount; ++device) {
cudaDeviceProp deviceProp;
cudaGetDeviceProperties(&deviceProp, device);
}
cudaSetDevice() is used to select the device associated to the host thread:
cudaSetDevice(device);
A device must be selected before any __global__ function or any function from
Appendix D is called. If this is not done by an explicit call to cudaSetDevice(),
device 0 is automatically selected and any subsequent explicit call to
cudaSetDevice() will have no effect.
4.5.2.3
Memory Management
The functions from Section D.5 are used to allocate and free device memory, access
the memory allocated for any variable declared in global memory space, and transfer
data between host and device memory.
Linear memory is allocated using cudaMalloc() or cudaMallocPitch() and
freed using cudaFree().
The following code sample allocates an array of 256 floating-point elements in linear
memory:
float* devPtr;
cudaMalloc((void**)&devPtr, 256 * sizeof(float));
cudaMallocPitch() is recommended for allocations of 2D arrays as it makes
sure that the allocation is appropriately padded to meet the alignment requirements
32
CUDA Programming Guide Version 1.1
Chapter 4. Application Programming Interface
described in Section 5.1.2.1, therefore ensuring best performance when accessing
the row addresses or performing copies between 2D arrays and other regions of
device memory (using the cudaMemcpy2D() functions). The returned pitch (or
stride) must be used to access array elements. The following code sample allocates a
width×height 2D array of floating-point values and shows how to loop over the
array elements in device code:
// host code
float* devPtr;
int pitch;
cudaMallocPitch((void**)&devPtr, &pitch,
width * sizeof(float), height);
myKernel<<<100, 512>>>(devPtr, pitch);
// device code
__global__ void myKernel(float* devPtr, int pitch)
{
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];
}
}
}
CUDA arrays are allocated using cudaMallocArray() and freed using
cudaFreeArray(). cudaMallocArray() requires a format description created
using cudaCreateChannelDesc().
The following code sample allocates a width×height CUDA array of one 32-bit
floating-point component:
cudaChannelFormatDesc channelDesc =
cudaCreateChannelDesc();
cudaArray* cuArray;
cudaMallocArray(&cuArray, &channelDesc, width, height);
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().
Section D.5 lists all the various functions used to copy memory between linear
memory allocated with cudaMalloc(), linear memory allocated with
cudaMallocPitch(), CUDA arrays, and memory allocated for variables declared
in global or constant memory space.
The following code sample copies the 2D array to the CUDA array allocated in the
previous code samples:
cudaMemcpy2DToArray(cuArray, 0, 0, devPtr, pitch,
width * sizeof(float), height,
cudaMemcpyDeviceToDevice);
The following code sample copies some host memory array to device memory:
float data[256];
int size = sizeof(data);
float* devPtr;
cudaMalloc((void**)&devPtr, size);
cudaMemcpy(devPtr, data, size, cudaMemcpyHostToDevice);
CUDA Programming Guide Version 1.1
33
Chapter 4. Application Programming Interface
The following code sample copies some host memory array to constant memory:
__constant__ float constData[256];
float data[256];
cudaMemcpyToSymbol(constData, data, sizeof(data));
4.5.2.4
Stream Management
The functions from Section D.3 are used to create and destroy streams and
determine whether all operations of a stream have completed or not.
The following code sample creates two streams:
cudaStream_t stream[2];
for (int i = 0; i < 2; ++i)
cudaStreamCreate(&stream[i]);
Each of these streams is defined by the following code sample as a sequence of one
memory copy from host to device, one kernel launch, and one memory copy from
device to host:
for (int i = 0; i < 2; ++i)
cudaMemcpyAsync(inputDevPtr + i * size, hostPtr + i * size,
size, cudaMemcpyHostToDevice, stream[i]);
for (int i = 0; i < 2; ++i)
myKernel<<<100, 512, 0, stream[i]>>>
(outputDevPtr + i * size, inputDevPtr + i * size, size);
for (int i = 0; i < 2; ++i)
cudaMemcpyAsync(hostPtr + i * size, outputDevPtr + i * size,
size, cudaMemcpyDeviceToHost, stream[i]);
cudaThreadSynchronize();
Each stream copies its portion of input array hostPtr to array inputDevPtr in
device memory, processes inputDevPtr on the device by calling myKernel(), and
copies the result outputDevPtr back to the same portion of hostPtr. Processing
hostPtr using two streams allows for the memory copies of one stream to overlap
with the kernel execution of the other stream. hostPtr must point to page-locked
host memory for any overlap to occur:
float* hostPtr;
cudaMallocHost((void**)&hostPtr, 2 * size);
is called in the end to make sure all streams are
finished before proceeding further.
cudaThreadSynchronize()
4.5.2.5
Event Management
The functions from Section D.4 are used to create, record, and destroy events and
query the elapsed time between two events.
The following code sample creates two events:
cudaEvent_t start, stop;
cudaEventCreate(&start);
cudaEventCreate(&stop);
These events can be used to time the code sample of the previous section the
following way:
cudaEventRecord(start, 0);
for (int i = 0; i < 2; ++i)
cudaMemcpyAsync(inputDev + i * size, inputHost + i * size,
size, cudaMemcpyHostToDevice, stream[i]);
34
CUDA Programming Guide Version 1.1
Chapter 4. Application Programming Interface
for (int i = 0; i < 2; ++i)
myKernel<<<100, 512, 0, stream[i]>>>
(outputDev + i * size, inputDev + i * size, size);
for (int i = 0; i < 2; ++i)
cudaMemcpyAsync(outputHost + i * size, outputDev + i * size,
size, cudaMemcpyDeviceToHost, stream[i]);
cudaEventRecord(stop, 0);
cudaEventSynchronize(stop);
float elapsedTime;
cudaEventElapsedTime(&elapsedTime, start, stop);
4.5.2.6
Texture Reference Management
The functions from Section D.6 are used to manage texture references.
The texture type defined by the high-level API is a structure publicly derived
from the textureReference type defined by the low-level API as such:
struct textureReference
{
int
enum cudaTextureFilterMode
enum cudaTextureAddressMode
struct cudaChannelFormatDesc
}
normalized;
filterMode;
addressMode[2];
channelDesc;
specifies whether texture coordinates are normalized or not; if it is
non-zero, all elements in the texture are addressed with texture coordinates in
the range [0,1] rather than in the range [0,width-1] or [0,height-1],
where width and height are the texture sizes;
filterMode specifies the filtering mode, that is how the value returned when
fetching the texture is computed based on the input texture coordinates;
filterMode is equal to cudaFilterModePoint or
cudaFilterModeLinear; if it is cudaFilterModePoint, the returned
value is the texel whose texture coordinates are the closest to the input texture
coordinates; if it is cudaFilterModeLinear, the returned value is the linear
interpolation of the two (for a one-dimensional texture) or four (for a
two-dimensional texture) texels whose texture coordinates are the closest to the
input texture coordinates; cudaFilterModeLinear is only valid for returned
values of floating-point type;
addressMode specifies the addressing mode, that is how out-of-range texture
coordinates are handled; addressMode is an array of size two whose first and
second elements specify the addressing mode for the first and second texture
coordinates, respectively; the addressing mode is equal to either
cudaAddressModeClamp, in which case out-of-range texture coordinates are
clamped to the valid range, or cudaAddressModeWrap, in which case out-ofrange texture coordinates are wrapped to the valid range;
cudaAddressModeWrap is only supported for normalized texture coordinates;
channelDesc describes the format of the value that is returned when fetching
the texture; channelDesc is of the following type:
normalized
struct cudaChannelFormatDesc {
int x, y, z, w;
enum cudaChannelFormatKind f;
};
CUDA Programming Guide Version 1.1
35
Chapter 4. Application Programming Interface
where x, y, z, and w are equal to the number of bits of each component of the
returned value and f is:
¾ cudaChannelFormatKindSigned if these components are of signed
integer type,
¾ cudaChannelFormatKindUnsigned if they are of unsigned integer
type,
¾ cudaChannelFormatKindFloat if they are of floating point type.
normalized, addressMode, and filterMode may be directly modified in host
code. They only apply to texture references bound to CUDA arrays.
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
cudaBindTextureToArray().
The following code samples bind a texture reference to linear memory pointed to by
devPtr:
Using the low-level API:
texture texRef;
textureReference* texRefPtr;
cudaGetTextureReference(&texRefPtr, “texRef”);
cudaChannelFormatDesc channelDesc =
cudaCreateChannelDesc();
cudaBindTexture(0, texRefPtr, devPtr, &channelDesc, size);
Using the high-level API:
texture texRef;
cudaBindTexture(0, texRef, devPtr, size);
The following code samples bind a texture reference to a CUDA array cuArray:
Using the low-level API:
texture texRef;
textureReference* texRefPtr;
cudaGetTextureReference(&texRefPtr, “texRef”);
cudaChannelFormatDesc channelDesc;
cudaGetChannelDesc(&channelDesc, cuArray);
cudaBindTextureToArray(texRef, cuArray, &channelDesc);
Using the high-level API:
texture 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.
cudaUnbindTexture() is used to unbind a texture reference.
4.5.2.7
OpenGL Interoperability
The functions from Section D.8 are used to control interoperability with OpenGL.
A buffer object must be registered to CUDA before it can be mapped. This is done
with cudaGLRegisterBufferObject():
GLuint bufferObj;
cudaGLRegisterBufferObject(bufferObj);
36
CUDA Programming Guide Version 1.1
Chapter 4. Application Programming Interface
Once it is registered, a buffer object can be read from or written to by kernels using
the device memory address returned by cudaGLMapBufferObject():
GLuint bufferObj;
float* devPtr;
cudaGLMapBufferObject((void**)&devPtr, bufferObj);
Unmapping is done with cudaGLUnmapBufferObject() and unregistering with
cudaGLUnregisterBufferObject().
4.5.2.8
Direct3D Interoperability
The functions from Section D.9 are used to control interoperability with Direct3D.
Interoperability with Direct3D must be initialized using cudaD3D9Begin() and
terminated using cudaD3D9End().
In between these calls, a vertex object must be registered to CUDA before it can be
mapped. This is done with cudaD3D9RegisterVertexBuffer():
LPDIRECT3DVERTEXBUFFER9 vertexBuffer;
cudaD3D9RegisterVertexBuffer(vertexBuffer);
Once it is registered, a vertex buffer can be read from or written to by kernels using
the device memory address returned by cudaD3D9MapVertexBuffer():
LPDIRECT3DVERTEXBUFFER9 vertexBuffer;
float* devPtr;
cudaD3D9MapVertexBuffer((void**)&devPtr, vertexBuffer);
Unmapping is done with cudaD3D9UnmapVertexBuffer() and unregistering
with cudaD3D9UnregisterVertexBuffer().
4.5.2.9
Debugging using the Device Emulation Mode
The programming environment does not include any native debug support for code
that runs on the device, but comes with a device emulation mode for the purpose of
debugging. When compiling an application is this mode (using the -deviceemu
option), the device code is compiled for and runs on the host, allowing the
programmer to use the host’s native debugging support to debug the application as
if it were a host application. The preprocessor macro __DEVICE_EMULATION__ is
defined in this mode. All code for an application, including any libraries used, must
be compiled consistently either for device emulation or for device execution.
Linking code compiled for device emulation with code compiled for device
execution causes the following runtime error to be returned upon initialization:
cudaErrorMixedDeviceExecution.
When running an application in device emulation mode, the programming model is
emulated by the runtime. For each thread in a thread block, the runtime creates a
thread on the host. The programmer needs to make sure that:
The host is able to run up to the maximum number of threads per block, plus
one for the master thread.
Enough memory is available to run all threads, knowing that each thread gets
256 KB of stack.
Many features provided through the device emulation mode make it a very effective
debugging tool:
By using the host’s native debugging support programmers can use all features
that the debugger supports, like setting breakpoints and inspecting data.
CUDA Programming Guide Version 1.1
37
Chapter 4. Application Programming Interface
Since device code is compiled to run on the host, the code can be augmented
with code that cannot run on the device, like input and output operations to files
or to the screen (printf(), etc.).
Since all data resides on the host, any device- or host-specific data can be read
from either device or host code; similarly, any device or host function can be
called from either device or host code.
In case of incorrect usage of the synchronization intrinsic, the runtime detects
dead lock situations.
Programmers must keep in mind that device emulation mode is emulating the
device, not simulating it. Therefore, device emulation mode is very useful in finding
algorithmic errors, but certain errors are hard to find:
When a memory location is accessed in multiple threads within the grid at
potentially the same time, the results when running in device emulation mode
potentially differ from the results when running on the device, since in emulation
mode threads execute sequentially.
When dereferencing a pointer to global memory on the host or a pointer to host
memory on the device, device execution almost certainly fails in some undefined
way, whereas device emulation can produce correct results.
Most of the time the same floating-point computation will not produce exactly
the same result when performed on the device as when performed on the host in
device emulation mode. This is expected since in general, all you need to get
different results for the same floating-point computation are slightly different
compiler options, let alone different compilers, different instruction sets, or
different architectures.
In particular, some host platforms store intermediate results of single-precision
floating-point calculations in extended precision registers, potentially resulting in
significant differences in accuracy when running in device emulation mode.
When this occurs, programmers can try any of the following methods, none of
which is guaranteed to work:
¾ Declare some floating-point variables as volatile to force single-precision
storage;
¾ Use the –ffloat-store compiler option of gcc,
¾ Use the /Op or /fp compiler options of the Visual C++ compiler,
¾ Use _FPU_GETCW() and _FPU_SETCW() on Linux or _controlfp()
on Windows to force single-precision floating-point computation for a
portion of the code by surrounding it with
unsigned int originalCW;
_FPU_GETCW(originalCW);
unsigned int cw = (originalCW & ~0x300) | 0x000;
_FPU_SETCW(cw);
or
unsigned int originalCW = _controlfp(0, 0);
_controlfp(_PC_24, _MCW_PC);
at the beginning, to store the current value of the control word and change
it to force the mantissa to be stored in 24 bits using, and with
_FPU_SETCW(originalCW);
38
CUDA Programming Guide Version 1.1
Chapter 4. Application Programming Interface
or
_controlfp(originalCW, 0xfffff);
at the end, to restore the original control word.
Unlike compute devices (see Appendix A), host platforms also usually support
denormalized numbers. This can lead to dramatically different results between
device emulation and device execution modes since some computation might
produce a finite result in one case and an infinite result in the other.
4.5.3
Driver API
The driver API 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 CUDA are summarized in Table 4-1.
Table 4-1. Objects Available in the CUDA Driver API
4.5.3.1
Object
Handle
Description
Device
CUdevice
CUDA-capable device
Context
CUcontext
Roughly equivalent to a CPU process
Module
CUmodule
Roughly equivalent to a dynamic library
Function
CUfunction
Kernel
Heap memory
CUdeviceptr
Pointer to device memory
CUDA array
CUarray
Opaque container for one-dimensional or two-dimensional
data on the device, readable via texture references
Texture reference
CUtexref
Object that describes how to interpret texture memory data
Initialization
Initialization with cuInit() is required before any function from Appendix E is
called (see Section E.1).
4.5.3.2
Device Management
The functions from Section E.2 are used to manage the devices present in the
system.
cuDeviceGetCount() and cuDeviceGet() provide a way to enumerate these
devices and other functions from Section E.2 to retrieve their properties:
int deviceCount;
cuDeviceGetCount(&deviceCount);
int device;
for (int device = 0; device < deviceCount; ++device) {
CUdevice cuDevice;
cuDeviceGet(&cuDevice, device);
int major, minor;
cuDeviceComputeCapability(&major, &minor, cuDevice);
}
4.5.3.3
Context Management
The functions from Section E.3 are used to create, attach, and detach CUDA
contexts.
CUDA Programming Guide Version 1.1
39
Chapter 4. Application Programming Interface
A CUDA context is analogous to a CPU process. All resources and actions
performed within the compute 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 references, each context has its own
distinct 32-bit address space. As a result, CUdeviceptr values from different
CUDA contexts reference different memory locations.
Contexts have a one-to-one correspondence with host threads. 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.
To facilitate interoperability between third party authored code operating in the
same context, the driver API maintains a usage count that is incremented by each
distinct client of a given context. For example, if three libraries are loaded to use the
same CUDA context, each library must call cuCtxAttach() to increment the
usage count and cuCtxDetach() to decrement the usage count when the library is
done using the context. The context is destroyed when the usage count goes to 0.
For most libraries, it is expected that the application will have created a CUDA
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.
4.5.3.4
Module Management
The functions from Section E.4 are used to load and unload modules and to retrieve
handles or pointers to variables or functions defined in the module.
Modules are dynamically loadable packages of device code and data, akin to DLLs in
Windows, that are output by nvcc (see Section 4.2.5). The names for all symbols,
including functions, global variables, and texture references, are maintained at
module scope so that modules written by independent third parties may interoperate
in the same CUDA context.
This code sample loads a module and retrieves a handle to some kernel:
CUmodule cuModule;
cuModuleLoad(&cuModule, “myModule.cubin”);
CUfunction cuFunction;
cuModuleGetFunction(&cuFunction, cuModule, “myKernel”);
4.5.3.5
Execution Control
The functions described in Section E.7 manage the execution of a kernel on the
device. cuFuncSetBlockShape() sets the number of threads per block for a
given function, and how their threadIDs are assigned. cuFuncSetSharedSize()
sets the size of shared memory for the function. The cuParam*() family of
functions is used specify the parameters that will be provided to the kernel the next
time cuLaunchGrid() or cuLaunch() is invoked to launch the kernel:
cuFuncSetBlockShape(cuFunction, blockWidth, blockHeight, 1);
int offset = 0;
int i;
cuParamSeti(cuFunction, offset, i);
offset += sizeof(i);
40
CUDA Programming Guide Version 1.1
Chapter 4. Application Programming Interface
float f;
cuParamSetf(cuFunction, offset, f);
offset += sizeof(f);
char data[32];
cuParamSetv(cuFunction, offset, (void*)data, sizeof(data));
offset += sizeof(data);
cuParamSetSize(cuFunction, offset);
cuFuncSetSharedSize(cuFunction, numElements * sizeof(float));
cuLaunchGrid(cuFunction, gridWidth, gridHeight);
4.5.3.6
Memory Management
The functions from Section E.8 are used to allocate and free device memory and
transfer data between host and device memory.
Linear memory is allocated using cuMemAlloc() or cuMemAllocPitch() and
freed using cuMemFree().
The following code sample allocates an array of 256 floating-point elements in linear
memory:
CUdeviceptr devPtr;
cuMemAlloc(&devPtr, 256 * sizeof(float));
cuMemAllocPitch() is recommended for allocations of 2D arrays as it makes
sure that the allocation is appropriately padded to meet the alignment requirements
described in Section 5.1.2.1, therefore ensuring best performance when accessing
the row addresses or performing copies between 2D arrays and other regions of
device memory (using the cuMemcpy2D()). The returned pitch (or stride) must be
used to access array elements. The following code sample allocates a
width×height 2D array of floating-point values and shows how to loop over the
array elements in device code:
// host code
CUdeviceptr devPtr;
int pitch;
cuMemAllocPitch(&devPtr, &pitch,
width * sizeof(float), height, 4);
cuModuleGetFunction(&cuFunction, cuModule, “myKernel”);
cuFuncSetBlockShape(cuFunction, 512, 1, 1);
cuParamSeti(cuFunction, 0, devPtr);
cuParamSetSize(cuFunction, sizeof(devPtr));
cuLaunchGrid(cuFunction, 100, 1);
// device code
__global__ void myKernel(float* devPtr)
{
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];
}
}
}
CUDA arrays are created using cuArrayCreate() and destroyed using
cuArrayDestroy().
The following code sample allocates a width×height CUDA array of one 32-bit
floating-point component:
CUDA Programming Guide Version 1.1
41
Chapter 4. Application Programming Interface
CUDA_ARRAY_DESCRIPTOR desc;
desc.Format = CU_AD_FORMAT_FLOAT;
desc.NumChannels = 1;
desc.Width = width;
desc.Height = height;
CUarray cuArray;
cuArrayCreate(&cuArray, &desc);
Section E.8 lists all the various functions used to copy memory between linear
memory allocated with cuMemAlloc(), linear memory allocated with
cuMemAllocPitch(), and CUDA arrays. The following code sample copies the
2D array to the CUDA array allocated in the previous code samples:
CUDA_MEMCPY2D copyParam;
memset(©Param, 0, sizeof(copyParam));
copyParam.dstMemoryType = CU_MEMORYTYPE_ARRAY;
copyParam.dstArray = cuArray;
copyParam.srcMemoryType = CU_MEMORYTYPE_DEVICE;
copyParam.srcDevice = devPtr;
copyParam.srcPitch = pitch;
copyParam.WidthInBytes = width * sizeof(float);
copyParam.Height = height;
cuMemcpy2D(©Param);
The following code sample copies some host memory array to device memory:
float data[256];
int size = sizeof(data);
CUdeviceptr devPtr;
cuMemAlloc(&devPtr, size);
cuMemcpyHtoD(devPtr, data, size);
4.5.3.7
Stream Management
The functions from Section E.5 are used to create and destroy streams and
determine whether all operations of a stream have completed or not.
The following code sample creates two streams:
CUstream stream[2];
for (int i = 0; i < 2; ++i)
cuStreamCreate(&stream[i], 0);
Each of these streams is defined by the following code sample as a sequence of one
memory copy from host to device, one kernel launch, and one memory copy from
device to host:
for (int i = 0; i < 2; ++i)
cuMemcpyHtoDAsync(inputDevPtr + i * size, hostPtr + i * size,
size, stream[i]);
for (int i = 0; i < 2; ++i) {
cuFuncSetBlockShape(cuFunction, 512, 1, 1);
int offset = 0;
cuParamSeti(cuFunction, offset, outputDevPtr);
offset += sizeof(int);
cuParamSeti(cuFunction, offset, inputDevPtr);
offset += sizeof(int);
cuParamSeti(cuFunction, offset, size);
offset += sizeof(int);
cuParamSetSize(cuFunction, offset);
cuLaunchGridAsync(cuFunction, 100, 1, stream[i]);
42
CUDA Programming Guide Version 1.1
Chapter 4. Application Programming Interface
}
for (int i = 0; i < 2; ++i)
cuMemcpyDtoHAsync(hostPtr + i * size, outputDevPtr + i * size,
size, stream[i]);
cudaCtxSynchronize();
Each stream copies its portion of input array hostPtr to array inputDevPtr in
device memory, processes inputDevPtr on the device by calling cuFunction, and
copies the result outputDevPtr back to the same portion of hostPtr. Processing
hostPtr using two streams allows for the memory copies of one stream to
potentially overlap with the kernel execution of the other stream. hostPtr must
point to page-locked host memory for any overlap to occur:
float* hostPtr;
cuMemAllocHost((void**)&hostPtr, 2 * size);
is called in the end to make sure all streams are finished
before proceeding further.
cuCtxSynchronize()
4.5.3.8
Event Management
The functions from Section E.6 are used to create, record, and destroy events and
query the elapsed time between two events.
The following code sample creates two events:
CUevent start, stop;
cuEventCreate(&start);
cuEventCreate(&stop);
These events can be used to time the code sample of the previous section the
following way:
cuEventRecord(start, 0);
for (int i = 0; i < 2; ++i)
cuMemcpyHtoDAsync(inputDevPtr + i * size, hostPtr + i * size,
size, stream[i]);
for (int i = 0; i < 2; ++i) {
cuFuncSetBlockShape(cuFunction, 512, 1, 1);
int offset = 0;
cuParamSeti(cuFunction, offset, outputDevPtr);
offset += sizeof(int);
cuParamSeti(cuFunction, offset, inputDevPtr);
offset += sizeof(int);
cuParamSeti(cuFunction, offset, size);
offset += sizeof(int);
cuParamSetSize(cuFunction, offset);
cuLaunchGridAsync(cuFunction, 100, 1, stream[i]);
}
for (int i = 0; i < 2; ++i)
cuMemcpyDtoHAsync(hostPtr + i * size, outputDevPtr + i * size,
size, stream[i]);
cuEventRecord(stop, 0);
cuEventSynchronize(stop);
float elapsedTime;
cuEventElapsedTime(&elapsedTime, start, stop);
4.5.3.9
Texture Reference Management
The functions from Section E.9 are used to manage texture references.
CUDA Programming Guide Version 1.1
43
Chapter 4. Application Programming Interface
Before a kernel can use a texture reference to read from texture memory, the texture
reference must be bound to a texture using cuTexRefSetAddress() or
cuTexRefSetArray().
If a module cuModule contains some texture reference texRef defined as
texture texRef;
the following code sample retrieves texRef‘s handle:
CUtexref cuTexRef;
cuModuleGetTexRef(&cuTexRef, cuModule, “texRef”);
The following code sample binds texRef to some linear memory pointed to by
devPtr:
cuTexRefSetAddress(NULL, cuTexRef, devPtr, size);
The following code samples bind texRef to a CUDA array cuArray:
cuTexRefSetArray(cuTexRef, cuArray, CU_TRSA_OVERRIDE_FORMAT);
Section E.9 lists various functions used to set address mode, filter mode, format,
and other flags for some texture reference. 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.
4.5.3.10
OpenGL Interoperability
The functions from Section E.10 are used to control interoperability with OpenGL.
Interoperability with OpenGL must be initialized using cuGLInit().
A buffer object must be registered to CUDA before it can be mapped. This is done
with cuGLRegisterBufferObject():
GLuint bufferObj;
cuGLRegisterBufferObject(bufferObj);
Once it is registered, a buffer object can be read from or written to by kernels using
the device memory address returned by cuGLMapBufferObject():
GLuint bufferObj;
CUdeviceptr devPtr;
int size;
cuGLMapBufferObject(&devPtr, &size, bufferObj);
Unmapping is done with cuGLUnmapBufferObject() and unregistering with
cuGLUnregisterBufferObject().
4.5.3.11
Direct3D Interoperability
The functions from Section D.9 are used to control interoperability with Direct3D.
Interoperability with Direct3D must be initialized using cuD3D9Begin() and
terminated using cuD3D9End().
In between these calls, a vertex object must be registered to CUDA before it can be
mapped. This is done with cuD3D9RegisterVertexBuffer():
LPDIRECT3DVERTEXBUFFER9 vertexBuffer;
cuD3D9RegisterVertexBuffer(vertexBuffer);
Once it is registered, a vertex buffer can be read from or written to by kernels using
the device memory address returned by cuD3D9MapVertexBuffer():
LPDIRECT3DVERTEXBUFFER9 vertexBuffer;
44
CUDA Programming Guide Version 1.1
Chapter 4. Application Programming Interface
CUdeviceptr devPtr;
int size;
cuD3D9MapVertexBuffer(&devPtr, &size, vertexBuffer);
Unmapping is done with cuD3D9UnmapVertexBuffer() and unregistering with
cuD3D9UnregisterVertexBuffer().
CUDA Programming Guide Version 1.1
45
Chapter 5.
Performance Guidelines
5.1
Instruction Performance
To process an instruction for a warp of threads, a multiprocessor must:
Read the instruction operands for each thread of the warp,
Execute the instruction,
Write the result for each thread of the warp.
Therefore, the effective instruction throughput depends on the nominal instruction
throughput as well as the memory latency and bandwidth. It is maximized by:
Minimizing the use of instructions with low throughput (see Section 5.1.1),
Maximizing the use of the available memory bandwidth for each category of
memory (see Section 5.1.2),
Allowing the thread scheduler to overlap memory transactions with
mathematical computations as much as possible, which requires that:
¾ The program executed by the threads is of high arithmetic intensity, that is,
has a high number of arithmetic operations per memory operation;
¾ There are many active threads per multiprocessor, as detailed in Section 5.2.
5.1.1
Instruction Throughput
5.1.1.1
Arithmetic Instructions
To issue one instruction for a warp, a multiprocessor takes:
4 clock cycles for floating-point add, floating-point multiply, floating-point
multiply-add, integer add, bitwise operations, compare, min, max, type
conversion instruction;
16 clock cycles for reciprocal, reciprocal square root, __log(x) (see Table B-2).
32-bit integer multiplication takes 16 clock cycles, but __mul24 and __umul24
(see Appendix B) provide signed and unsigned 24-bit integer multiplication in 4
clock cycles. On future architectures however, __[u]mul24 will be slower than 32bit integer multiplication, so we recommend to provide two kernels, one using
__[u]mul24 and the other using generic 32-bit integer multiplication, to be called
appropriately by the application.
CUDA Programming Guide Version 1.1
47
Chapter 5. Performance Guidelines
Integer division and modulo operation are particularly costly and should be avoided
if possible or replaced with bitwise operations whenever possible: 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.
Other functions take more clock cycles as they are implemented as combinations of
several instructions.
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. Therefore, it takes 32 clock cycles for a warp.
Floating-point division takes 36 clock cycles, but __fdividef(x, y) provides a
faster version at 20 clock cycles (see Appendix B).
__sin(x), __cos(x), __exp(x) take 32 clock cycles.
Sometimes, the compiler must insert conversion instructions, introducing additional
execution cycles. This is the case for:
Functions operating on char or short whose operands generally need to be
converted to int,
Double-precision floating-point constants (defined without any type suffix) used
as input to single-precision floating-point computations,
Single-precision floating-point variables used as input parameters to the doubleprecision version of the mathematical functions defined in Table B-1.
The two last cases can be avoided by using:
Single-precision floating-point constants, defined with an f suffix such as
3.141592653589793f, 1.0f, 0.5f,
The single-precision version of the mathematical functions, defined with an f
suffix as well, such as sinf(), logf(), expf().
For single precision code, we highly recommend use of the float type and the single
precision math functions. When compiling for devices without native double
precision support, such as devices of compute capability 1.x, the double type gets
demoted to float by default and the double precision math functions are mapped to
their single precision equivalents. However, on those future devices that will support
double precision, these functions will map to double precision implementations.
5.1.1.2
Control 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, that is, 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 Section 3.2. A trivial example is when the
controlling condition only depends on (threadIdx / WSIZE) where WSIZE is
the warp size. In this case, no warp diverges since the controlling condition is
perfectly aligned with the warps.
48
CUDA Programming Guide Version 1.1
Chapter 5. Performance Guidelines
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 Section 4.2.5.2).
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
per-thread 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.
5.1.1.3
Memory Instructions
Memory instructions include any instruction that reads from or writes to shared or
global memory. A multiprocessor takes 4 clock cycles to issue one memory
instruction for a warp. When accessing global memory, there are, in addition, 400 to
600 clock cycles of memory latency.
As an example, the assignment operator in the following sample code:
__shared__ float shared[32];
__device__ float device[32];
shared[threadIdx.x] = device[threadIdx.x];
takes 4 clock cycles to issue a read from global memory, 4 clock cycles to issue a
write to shared memory, but above all 400 to 600 clock cycles to read a float from
global memory.
Much of this global memory latency can be hidden by the thread scheduler if there
are sufficient independent arithmetic instructions that can be issued while waiting
for the global memory access to complete.
5.1.1.4
Synchronization Instruction
__syncthreads takes 4 clock cycles to issue for a warp if no thread has to wait
for any other threads.
5.1.2
Memory Bandwidth
The effective bandwidth of each memory space depends significantly on the
memory access pattern as detailed in the following sub-sections.
Since device memory is of much higher latency and lower bandwidth than on-chip
memory, device memory accesses should be minimized. 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 written by different threads,
CUDA Programming Guide Version 1.1
49
Chapter 5. Performance Guidelines
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.
5.1.2.1
Global Memory
The global memory space is not cached, so it is all the more important to follow the
right access pattern to get maximum memory bandwidth, especially given how
costly accesses to device memory are.
First, the device is capable of reading 32-bit, 64-bit, or 128-bit words from global
memory into registers in a single instruction. To have assignments such as:
__device__ type device[32];
type data = device[tid];
compile to a single load instruction, type must be such that sizeof(type) is
equal to 4, 8, or 16 and variables of type type must be aligned to sizeof(type)
bytes (that is, have their address be a multiple of sizeof(type)).
The alignment requirement is automatically fulfilled for built-in types of
Section 4.3.1.1 like float2 or float4.
For structures, the size and alignment requirements can be enforced by the compiler
using the alignment specifiers __align__(8) or __align__(16), such as
struct __align__(8) {
float a;
float b;
};
or
struct __align__(16) {
float a;
float b;
float c;
};
For structures larger than 16 bytes, the compiler generates several load instructions.
To ensure that it generates the minimum number of instructions, such structures
should be defined with __align__(16) , such as
struct __align__(16) {
float a;
float b;
float c;
float d;
float e;
};
which is compiled into two 128-bit load instructions instead of five 32-bit load
instructions.
Second, the global memory addresses simultaneously accessed by each thread of a
half-warp during the execution of a single read or write instruction should be
arranged so that the memory accesses can be coalesced into a single contiguous,
aligned memory access.
50
CUDA Programming Guide Version 1.1
Chapter 5. Performance Guidelines
More precisely, in each half-warp, thread number N within the half-warp should
access address
HalfWarpBaseAddress + N
where HalfWarpBaseAddress is of type type* and type is such that it meets
the size and alignment requirements discussed above. Moreover,
HalfWarpBaseAddress should be aligned to 16*sizeof(type) bytes (i.e. be a
multiple of 16*sizeof(type)). Any address BaseAddress of a variable residing
in global memory or returned by one of the memory allocation routines from
Sections D.5 or E.8 is always aligned to at least 256 bytes, so to satisfy the memory
alignment constraint, HalfWarpBaseAddress-BaseAddress should be a
multiple of 16*sizeof(type).
Note that if a half-warp fulfills all the requirements above, the per-thread memory
accesses are coalesced even if some threads of the half-warp do not actually access
memory.
We recommend fulfilling the coalescing requirements for the entire warp as
opposed to only each of its halves separately because future devices will necessitate
it for proper coalescing.
Figure 5-1 shows some examples of coalesced memory accesses while Figure 5-2
and Figure 5-3 show some examples of non-coalesced memory accesses.
Coalesced 64-bit accesses deliver a little lower bandwidth than coalesced 32-bit
accesses and coalesced 128-bit accesses deliver a noticeably lower bandwidth than
coalesced 32-bit accesses. But, while bandwidth for non-coalesced accesses is
around an order of magnitude lower than for coalesced accesses when these
accesses are 32-bit, it is only around four times lower when they are 64-bit and
around two times when they are 128-bit.
CUDA Programming Guide Version 1.1
51
Chapter 5. Performance Guidelines
Thread 0
Address 128
Thread 0
Address 128
Thread 1
Address 132
Thread 1
Address 132
Thread 2
Address 136
Thread 2
Address 136
Thread 3
Address 140
Thread 3
Address 140
Thread 4
Address 144
Thread 4
Address 144
Thread 5
Address 148
Thread 5
Address 148
Thread 6
Address 152
Thread 6
Address 152
Thread 7
Address 156
Thread 7
Address 156
Thread 8
Address 160
Thread 8
Address 160
Thread 9
Address 164
Thread 9
Address 164
Thread 10
Address 168
Thread 10
Address 168
Thread 11
Address 172
Thread 11
Address 172
Thread 12
Address 176
Thread 12
Address 176
Thread 13
Address 180
Thread 13
Address 180
Thread 14
Address 184
Thread 14
Address 184
Thread 15
Address 188
Thread 15
Address 188
Left: coalesced float memory access.
Right: coalesced float memory access (divergent warp).
Figure 5-1.
52
Examples of Coalesced Global Memory Access
Patterns
CUDA Programming Guide Version 1.1
Chapter 5. Performance Guidelines
Thread 0
Address 128
Thread 0
Address 128
Thread 1
Address 132
Thread 1
Address 132
Thread 2
Address 136
Thread 2
Address 136
Thread 3
Address 140
Thread 3
Address 140
Thread 4
Address 144
Thread 4
Address 144
Thread 5
Address 148
Thread 5
Address 148
Thread 6
Address 152
Thread 6
Address 152
Thread 7
Address 156
Thread 7
Address 156
Thread 8
Address 160
Thread 8
Address 160
Thread 9
Address 164
Thread 9
Address 164
Thread 10
Address 168
Thread 10
Address 168
Thread 11
Address 172
Thread 11
Address 172
Thread 12
Address 176
Thread 12
Address 176
Thread 13
Address 180
Thread 13
Address 180
Thread 14
Address 184
Thread 14
Address 184
Thread 15
Address 188
Thread 15
Address 188
Left: non-sequential float memory access.
Right: misaligned starting address.
Figure 5-2.
Examples of Non-Coalesced Global Memory
Access Patterns
CUDA Programming Guide Version 1.1
53
Chapter 5. Performance Guidelines
Thread 0
Address 128
Thread 1
Address 132
Thread 2
Address 136
Thread 3
Address 140
Thread 4
Address 144
Thread 5
Address 148
Thread 6
Address 152
Thread 7
Address 156
Thread 8
Address 160
Thread 9
Address 164
Thread 10
Address 168
Thread 11
Address 172
Thread 12
Address 176
Thread 13
Address 180
Thread 14
Address 184
Thread 15
Address 188
Thread 0
Address 128
Thread 1
Address 140
Thread 2
Address 152
Thread 3
Address 164
Thread 4
Address 176
Thread 5
Address 188
Left: non-contiguous float memory access.
Right: non-coalesced float3 memory access.
Figure 5-3.
54
Examples of Non-Coalesced Global Memory
Access Patterns
CUDA Programming Guide Version 1.1
Chapter 5. Performance Guidelines
A common global memory access pattern is when each thread of thread ID tid
accesses one element of an array located at address BaseAddress of type type*
using the following address:
BaseAddress + tid
To get memory coalescing, type must meet the size and alignment requirements
discussed above. In particular, this means that if type is a structure larger than 16
bytes, it should be split into several structures that meet these requirements and the
data should be laid out in memory as a list of several arrays of these structures
instead of a single array of type type*.
Another common global memory access pattern is when each thread of index
(tx,ty) accesses one element of a 2D array located at address BaseAddress of
type type* and of width width using the following address:
BaseAddress + width * ty + tx
In such a case, one gets memory coalescing for all half-warps of the thread block
only if:
The width of the thread block is a multiple of half the warp size;
width is a multiple of 16.
In particular, this means that an array whose width is not a multiple of 16 will be
accessed much more efficiently if it is actually allocated with a width rounded up to
the closest multiple of 16 and its rows padded accordingly. The
cudaMallocPitch() and cuMemAllocPitch() functions and associated
memory copy functions described in Sections D.5 and E.8 enable programmers to
write non-hardware-dependent code to allocate arrays that conform to these
constraints.
5.1.2.2
Constant Memory
The constant memory space is cached so a read from constant memory costs one
memory read from device memory only on a cache miss, otherwise it just costs one
read from the constant cache.
For all threads of a half-warp, reading from the constant cache is as fast as reading
from a register as long as all threads read the same address. The cost scales linearly
with the number of different addresses read by all threads. We recommend having
all threads of the entire warp read the same address as opposed to all threads within
each of its halves only, as future devices will require it for full speed read.
5.1.2.3
Texture Memory
The texture memory space is cached so a texture fetch costs one memory read from
device memory only on a cache miss, otherwise it just costs one read from the
texture cache. The texture cache is optimized for 2D spatial locality, so threads of
the same warp that read texture addresses that are close together will achieve best
performance. Also, it is designed for streaming fetches with a constant latency, i.e. a
cache hit reduces DRAM bandwidth demand, but not fetch latency.
Reading device memory through texture fetching can be an advantageous alternative
to reading device memory from global or constant memory as detailed in
Section 5.4.
CUDA Programming Guide Version 1.1
55
Chapter 5. Performance Guidelines
5.1.2.4
Shared Memory
Because it is on-chip, the shared memory space is much faster than the local and
global memory spaces. In fact, for all threads of a warp, accessing the shared
memory is as fast as accessing a register as long as there are no bank conflicts
between the threads, as detailed below.
To achieve high memory bandwidth, shared memory is divided into equally-sized
memory modules, called banks, which can be accessed simultaneously. So, any
memory read or write request made of n addresses that fall in n distinct memory
banks can be serviced simultaneously, yielding an effective bandwidth that is n times
as high as the bandwidth of a single module.
However, if two addresses of a memory request fall in the same memory bank, there
is a bank conflict and the access has to be serialized. The hardware splits a memory
request with bank conflicts into as many separate conflict-free requests as necessary,
decreasing the effective bandwidth by a factor equal to the number of separate
memory requests. If the number of separate memory requests is n, the initial
memory request is said to cause n-way bank conflicts.
To get maximum performance, it is therefore important to understand how memory
addresses map to memory banks in order to schedule the memory requests so as to
minimize bank conflicts.
In the case of the shared memory space, the banks are organized such that
successive 32-bit words are assigned to successive banks and each bank has a
bandwidth of 32 bits per two clock cycles.
For devices of compute capability 1.x, the warp size is 32 and the number of banks
is 16 (see Section 5.1); a shared memory request for a warp is split into one request
for the first half of the warp and one request for the second half of the warp. 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.
A common case is for each thread to access a 32-bit word from an array indexed by
the thread ID tid and with some stride s:
__shared__ float shared[32];
float data = shared[BaseIndex + s * tid];
In this case, the threads tid and tid+n access the same bank whenever s*n is a
multiple of the number of banks m or equivalently, whenever n is a multiple of m/d
where d is the greatest common divisor of m and s. As a consequence, there will be
no bank conflict only if half the warp size is less than or equal to m/d. For devices
of compute capability 1.x, this translates to no bank conflict only if d is equal to 1,
or in other words, only if s is odd since m is a power of two.
Figure 5-4 and Figure 5-5 show some examples of conflict-free memory accesses
while Figure 5-6 shows some examples of memory accesses that cause bank
conflicts.
Other cases worth mentioning are when each thread accesses an element that is
smaller or larger than 32 bits in size. For example, there will be bank conflicts if an
array of char is accessed the following way:
__shared__ char shared[32];
char data = shared[BaseIndex + tid];
56
CUDA Programming Guide Version 1.1
Chapter 5. Performance Guidelines
because shared[0], shared[1], shared[2], and shared[3], for example,
belong to the same bank. There will not be any bank conflict however, if the same
array is accessed the following way:
char data = shared[BaseIndex + 4 * tid];
A structure assignment is compiled into as many memory requests as there are
members in the structure, so the following code, for example:
__shared__ struct type shared[32];
struct type data = shared[BaseIndex + tid];
results in:
Three separate memory reads without bank conflicts if type is defined as
struct type {
float x, y, z;
};
since each member is accessed with a stride of three 32-bit words;
Two separate memory reads with bank conflicts if type is defined as
struct type {
float x, y;
};
since each member is accessed with a stride of two 32-bit words;
Two separate memory reads with bank conflicts if type is defined as
struct type {
float f;
char c;
};
since each member is accessed with a stride of five bytes.
Finally, shared memory also 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 of a half-warp 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 – one step every two clock cycles – 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 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 5-7 shows some examples of memory read accesses that involve the
broadcast mechanism.
CUDA Programming Guide Version 1.1
57
Chapter 5. Performance Guidelines
Thread 0
Bank 0
Thread 0
Bank 0
Thread 1
Bank 1
Thread 1
Bank 1
Thread 2
Bank 2
Thread 2
Bank 2
Thread 3
Bank 3
Thread 3
Bank 3
Thread 4
Bank 4
Thread 4
Bank 4
Thread 5
Bank 5
Thread 5
Bank 5
Thread 6
Bank 6
Thread 6
Bank 6
Thread 7
Bank 7
Thread 7
Bank 7
Thread 8
Bank 8
Thread 8
Bank 8
Thread 9
Bank 9
Thread 9
Bank 9
Thread 10
Bank 10
Thread 10
Bank 10
Thread 11
Bank 11
Thread 11
Bank 11
Thread 12
Bank 12
Thread 12
Bank 12
Thread 13
Bank 13
Thread 13
Bank 13
Thread 14
Bank 14
Thread 14
Bank 14
Thread 15
Bank 15
Thread 15
Bank 15
Left: linear addressing with a stride of one 32-bit word.
Right: random permutation.
Figure 5-4.
58
Examples of Shared Memory Access Patterns
without Bank Conflicts
CUDA Programming Guide Version 1.1
Chapter 5. Performance Guidelines
Thread 0
Bank 0
Thread 1
Bank 1
Thread 2
Bank 2
Thread 3
Bank 3
Thread 4
Bank 4
Thread 5
Bank 5
Thread 6
Bank 6
Thread 7
Bank 7
Thread 8
Bank 8
Thread 9
Bank 9
Thread 10
Bank 10
Thread 11
Bank 11
Thread 12
Bank 12
Thread 13
Bank 13
Thread 14
Bank 14
Thread 15
Bank 15
Linear addressing with a stride of three 32-bit words.
Figure 5-5.
Example of a Shared Memory Access Pattern
without Bank Conflicts
CUDA Programming Guide Version 1.1
59
Chapter 5. Performance Guidelines
Thread 0
Bank 0
Thread 0
Bank 0
Thread 1
Bank 1
Thread 1
Bank 1
Thread 2
Bank 2
Thread 2
Bank 2
Thread 3
Bank 3
Thread 3
Bank 3
Thread 4
Bank 4
Thread 4
Bank 4
Thread 5
Bank 5
Thread 5
Bank 5
Thread 6
Bank 6
Thread 6
Bank 6
Thread 7
Bank 7
Thread 7
Bank 7
Thread 8
Bank 8
Thread 8
Bank 8
Thread 9
Bank 9
Thread 9
Bank 9
Thread 10
Bank 10
Thread 10
Bank 10
Thread 11
Bank 11
Thread 11
Bank 11
Thread 12
Bank 12
Thread 12
Bank 12
Thread 13
Bank 13
Thread 13
Bank 13
Thread 14
Bank 14
Thread 14
Bank 14
Thread 15
Bank 15
Thread 15
Bank 15
Left: Linear addressing with a stride of two 32-bit words causes 2-way bank conflicts.
Right: Linear addressing with a stride of eight 32-bit words causes 8-way bank conflicts.
Figure 5-6.
60
Examples of Shared Memory Access Patterns
with Bank Conflicts
CUDA Programming Guide Version 1.1
Chapter 5. Performance Guidelines
Thread 0
Bank 0
Thread 0
Bank 0
Thread 1
Bank 1
Thread 1
Bank 1
Thread 2
Bank 2
Thread 2
Bank 2
Thread 3
Bank 3
Thread 3
Bank 3
Thread 4
Bank 4
Thread 4
Bank 4
Thread 5
Bank 5
Thread 5
Bank 5
Thread 6
Bank 6
Thread 6
Bank 6
Thread 7
Bank 7
Thread 7
Bank 7
Thread 8
Bank 8
Thread 8
Bank 8
Thread 9
Bank 9
Thread 9
Bank 9
Thread 10
Bank 10
Thread 10
Bank 10
Thread 11
Bank 11
Thread 11
Bank 11
Thread 12
Bank 12
Thread 12
Bank 12
Thread 13
Bank 13
Thread 13
Bank 13
Thread 14
Bank 14
Thread 14
Bank 14
Thread 15
Bank 15
Thread 15
Bank 15
Left: This access pattern is conflict-free since all threads read from an address within the same 32-bit
word.
Right: This access pattern causes either no bank conflicts if the word from bank 5 is chosen as the
broadcast word during the first step or 2-way bank conflicts, otherwise.
Figure 5-7.
Example of Shared Memory Read Access
Patterns with Broadcast
CUDA Programming Guide Version 1.1
61
Chapter 5. Performance Guidelines
5.1.2.5
Registers
Generally, accessing a register is zero extra clock cycles per instruction, but delays
may occur due to register read-after-write dependencies and register memory bank
conflicts.
The delays introduced by read-after-write dependencies can be ignored as soon as
there are at least 192 active threads per multiprocessor to hide them.
The compiler and thread scheduler schedule the instructions as optimally as possible
to avoid register memory bank conflicts. They achieve best results when the number
of threads per block is a multiple of 64. Other than following this rule, an
application has no direct control over these bank conflicts. In particular, there is no
need to pack data into float4 or int4 types.
5.2
Number of Threads per Block
Given a total number of threads per grid, the number of threads per block, or
equivalently the number of blocks, should be chosen to maximize the utilization of
the available computing resources. This means that there should be at least as many
blocks as there are multiprocessors in the device.
Furthermore, running only one block per multiprocessor will force the
multiprocessor to idle during thread synchronization and also during device memory
reads if there are not enough threads per block to cover the load latency. It is
therefore better to allow for two or more blocks to be active on each multiprocessor
to allow overlap between blocks that wait and blocks that can run. For this to
happen, not only should there be at least twice as many blocks as there are
multiprocessors in the device, but also the amount of allocated shared memory per
block should be at most half the total amount of shared memory available per
multiprocessor (see Section 3.2). More thread blocks stream in pipeline fashion
through the device and amortize overhead even more.
With a high enough number of blocks, 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, or better, a multiple of 64 for the reason invoked in
Section 5.1.2.5. Allocating more threads per block is better for efficient time slicing,
but the more threads per block, the fewer registers are available per thread. This
might prevent a kernel invocation from succeeding if the kernel compiles to more
registers than are allowed by the execution configuration. The number of registers a
kernel compiles to (as well as local, shared, and constant memory usage) is reported
by the compiler when compiling with the --ptxas-options=-v option.
For devices of compute capability 1.x, the number of registers available per thread is
equal to:
R
B × ceil (T ,32)
where R is the total number of registers per multiprocessor given in Appendix A, B
is the number of active blocks per multiprocessor, T is the number of threads per
block, and ceil(T, 32) is T rounded up to the nearest multiple of 32.
62
CUDA Programming Guide Version 1.1
Chapter 5. Performance Guidelines
64 threads per block is minimal and makes sense only if there are multiple active
blocks per multiprocessor. 192 or 256 threads per block is better and usually allows
for enough registers to compile.
The number of blocks per grid should be at least 100 if one wants it to scale to
future devices; 1000 blocks will scale across several generations.
The ratio of the number of active warps per multiprocessor to the maximum
number of active warps (given in Appendix A) is called the multiprocessor occupancy.
In order to maximize occupancy, the compiler attempts to minimize register usage
and programmers need to choose execution configurations with care. The CUDA
Software Development Kit provides a spreadsheet to assist programmers in
choosing thread block size based on shared memory and register requirements.
5.3
Data Transfer between Host and Device
The bandwidth between the device and the device memory is much higher than the
bandwidth between the device memory and the host memory. Therefore, one
should strive to minimize data transfer between the host and the device, for
example, by moving 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.
Also, because of the overhead associated with each transfer, batching many small
transfers into a big one always performs much better than making each transfer
separately.
Finally, higher bandwidth between host and device is achieved when using pagelocked memory, as mentioned in Section 4.5.1.2.
5.4
Texture Fetch versus Global or Constant
Memory Read
Device memory reads through texture fetching present several benefits over reads
from global or constant memory:
They are cached, potentially exhibiting higher bandwidth if there is locality in the
texture fetches;
They are not subject to the constraints on memory access patterns that global or
constant memory reads must respect to get good performance (see
Sections 5.1.2.1 and 5.1.2.2);
The latency of addressing calculations is hidden better, possibly improving
performance for applications that perform random accesses to the data;
Packed data may be broadcast to separate variables in a single operation;
8-bit and 16-bit integer input data may be optionally converted to 32-bit floatingpoint values in the range [0.0, 1.0] or [-1.0, 1.0] (see Section 4.3.4.1).
If the texture is a CUDA array (see Section 4.3.4.2), the hardware provides other
capabilities that may be useful for different applications, especially image processing:
CUDA Programming Guide Version 1.1
63
Chapter 5. Performance Guidelines
Feature
Useful for…
Caveat
Filtering
Fast, low-precision interpolation
between texels
Only valid if the texture reference
returns floating-point data
Normalized texture
coordinates
Resolution-independent coding
Addressing modes
Automatic handling of boundary cases
Can only be used with normalized
texture coordinates
However, within the same kernel call, the texture cache is not kept coherent with
respect to global memory writes, so that any texture fetch to an address that has
been written to via a global write in the same kernel call returns undefined data. In
other words, a thread can safely read via texture some 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 or another thread from the same
kernel call. This is only relevant when fetching from linear memory as a kernel
cannot write to CUDA arrays anyway.
5.5
Overall Performance Optimization Strategies
Performance optimization revolves around three basic strategies:
Maximizing parallel execution;
Optimizing memory usage to achieve maximum memory bandwidth;
Optimizing instruction usage to achieve maximum instruction throughput.
Maximizing parallel execution starts with structuring the algorithm in a way that
exposes as much data parallelism as possible. At points in the algorithm where
parallelism is broken because some threads need to synchronize in order to share
data between 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 call, 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.
Once the parallelism of the algorithm has been exposed it needs to be mapped to
the hardware as efficiently as possible. This is done by carefully choosing the
execution configuration of each kernel invocation as detailed in Section 5.2.
The application should also maximize parallel execution at a higher level by
explicitly exposing concurrent execution on the device through streams, as
described in Section 4.5.1.5, as well as maximizing concurrent execution between
host and device.
Optimizing memory usage starts with minimizing data transfers with lowbandwidth. That means minimizing data transfers between the host and the device,
as detailed in Section 5.3, since these have much lower bandwidth than data
transfers between the device and global memory. That also means minimizing data
transfers between the device and global memory by maximizing use of shared
memory on the device, as mentioned in Section 5.1.2. Sometimes, the best
64
CUDA Programming Guide Version 1.1
Chapter 5. Performance Guidelines
optimization might even be to avoid any data transfer in the first place by simply
recomputing the data instead whenever it is needed.
As detailed in Sections 5.1.2.1, 5.1.2.2, 5.1.2.3, and 5.1.2.4, the effective bandwidth
can vary by an order of magnitude depending on access pattern for each type of
memory. The next step in optimizing memory usage is therefore to organize
memory accesses as optimally as possible based on the optimal memory access
patterns. This optimization is especially important for global memory accesses as
global memory’s bandwidth is low and its latency is hundreds of clock cycles (see
Section 5.1.1.3). Shared memory accesses, on the other hand, are usually worth
optimizing only in case they have a high degree of bank conflicts.
As for optimizing instruction usage, the use of arithmetic instructions with low
throughput (see Section 5.1.1.1) should be minimized. This includes trading
precision for speed when it does not affect the end result, such as using intrinsic
instead of regular functions (intrinsic functions are listed in Table B-2) or single
precision instead of double precision. Particular attention must be paid to control
flow instructions due to the SIMD nature of the device as detailed in Section 5.1.1.2.
CUDA Programming Guide Version 1.1
65
Chapter 6.
Example of Matrix Multiplication
6.1
Overview
The task of computing the product C of two matrices A and B of dimensions
(wA, hA) and (wB, wA) respectively, is split among several threads in the following
way:
Each thread block is responsible for computing one square sub-matrix Csub of C;
Each thread within the block is responsible for computing one element of Csub.
The dimension block_size of Csub is chosen equal to 16, so that the number of threads
per block is a multiple of the warp size (Section 5.2) and remains below the
maximum number of threads per block (Appendix A).
As illustrated in Figure 6-1, Csub is equal to the product of two rectangular matrices:
the sub-matrix of A of dimension (wA, block_size) that has the same line indices as
Csub, and the sub-matrix of B of dimension (block_size, wA) that has the same column
indices as Csub. In order to fit into the device’s 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
matrices from global memory to shared memory with one thread loading one
element of each matrix, and then by having each thread compute one element of the
product. Each thread accumulates the result of each of these products into a register
and once done writes the result to global memory.
By blocking the computation this way, we take advantage of fast shared memory
and save a lot of global memory bandwidth since A and B are read from global
memory only (wA / block_size) times.
Nonetheless, this example has been written for clarity of exposition to illustrate
various CUDA programming principles, not with the goal of providing a
high-performance kernel for generic matrix multiplication and should not be
construed as such.
CUDA Programming Guide Version 1.1
67
Example of Matrix Multiplication
BLOCK_SIZE
Chapter 6.
wA
BLOCK_SIZE
B
BLOCK_SIZE
C
A
Csub
BLOCK_SIZE
BLOCK_SIZE
wA
hA
BLOCK_SIZE
wB
Each thread block computes one sub-matrix Csub of C. Each thread within the block
computes one element of Csub.
Figure 6-1.
68
Matrix Multiplication
CUDA Programming Guide Version 1.1
Chapter 6.
6.2
Example of Matrix Multiplication
Source Code Listing
// Thread block size
#define BLOCK_SIZE 16
// Forward declaration of the device multiplication function
__global__ void Muld(float*, float*, int, int, float*);
// Host multiplication function
// Compute C = A * B
//
hA is the height of A
//
wA is the width of A
//
wB is the width of B
void Mul(const float* A, const float* B, int hA, int wA, int wB,
float* C)
{
int size;
// Load A and B to the device
float* Ad;
size = hA * wA * sizeof(float);
cudaMalloc((void**)&Ad, size);
cudaMemcpy(Ad, A, size, cudaMemcpyHostToDevice);
float* Bd;
size = wA * wB * sizeof(float);
cudaMalloc((void**)&Bd, size);
cudaMemcpy(Bd, B, size, cudaMemcpyHostToDevice);
// Allocate C on the device
float* Cd;
size = hA * wB * sizeof(float);
cudaMalloc((void**)&Cd, size);
// Compute the execution configuration assuming
// the matrix dimensions are multiples of BLOCK_SIZE
dim3 dimBlock(BLOCK_SIZE, BLOCK_SIZE);
dim3 dimGrid(wB / dimBlock.x, hA / dimBlock.y);
// Launch the device computation
Muld<<>>(Ad, Bd, wA, wB, Cd);
// Read C from the device
cudaMemcpy(C, Cd, size, cudaMemcpyDeviceToHost);
// Free device memory
cudaFree(Ad);
cudaFree(Bd);
cudaFree(Cd);
}
CUDA Programming Guide Version 1.1
69
Chapter 6.
Example of Matrix Multiplication
// Device multiplication function called by Mul()
// Compute C = A * B
//
wA is the width of A
//
wB is the width of B
__global__ void Muld(float* A, float* B, int wA, int wB, float* C)
{
// Block index
int bx = blockIdx.x;
int by = blockIdx.y;
// Thread index
int tx = threadIdx.x;
int ty = threadIdx.y;
// Index of the first sub-matrix of A processed by the block
int aBegin = wA * BLOCK_SIZE * by;
// Index of the last sub-matrix of A processed by the block
int aEnd
= aBegin + wA - 1;
// Step size used to iterate through the sub-matrices of A
int aStep = BLOCK_SIZE;
// Index of the first sub-matrix of B processed by the block
int bBegin = BLOCK_SIZE * bx;
// Step size used to iterate through the sub-matrices of B
int bStep = BLOCK_SIZE * wB;
// The element of the block sub-matrix that is computed
// by the thread
float Csub = 0;
// Loop over all the
// compute the block
for (int a = aBegin,
a <= aEnd;
a += aStep,
sub-matrices of A and B required to
sub-matrix
b = bBegin;
b += bStep) {
// Shared memory for the sub-matrix of A
__shared__ float As[BLOCK_SIZE][BLOCK_SIZE];
// Shared memory for the sub-matrix of B
__shared__ float Bs[BLOCK_SIZE][BLOCK_SIZE];
// Load the matrices from global memory to shared memory;
// each thread loads one element of each matrix
As[ty][tx] = A[a + wA * ty + tx];
Bs[ty][tx] = B[b + wB * ty + tx];
// Synchronize to make sure the matrices are loaded
__syncthreads();
// Multiply the two matrices together;
// each thread computes one element
// of the block sub-matrix
for (int k = 0; k < BLOCK_SIZE; ++k)
70
CUDA Programming Guide Version 1.1
Chapter 6.
Example of Matrix Multiplication
Csub += As[ty][k] * Bs[k][tx];
// Synchronize to make sure that the preceding
// computation is done before loading two new
// sub-matrices of A and B in the next iteration
__syncthreads();
}
// Write the block sub-matrix to global memory;
// each thread writes one element
int c = wB * BLOCK_SIZE * by + BLOCK_SIZE * bx;
C[c + wB * ty + tx] = Csub;
}
6.3
Source Code Walkthrough
The source code contains two functions:
Mul(),
a host function serving as a wrapper to Muld();
Muld(), a kernel that executes the matrix multiplication on the device.
6.3.1
Mul()
Mul() takes as input:
Two pointers to host memory that point to the elements of A and B,
The height and width of A and the width of B,
A pointer to host memory that points where C should be written.
Mul() performs the following operations:
6.3.2
It allocates enough global memory to store A, B, and C using cudaMalloc();
It copies A and B from host memory to global memory using cudaMemcpy();
It calls Muld() to compute C on the device;
It copies C from global memory to host memory using cudaMemcpy();
It frees the global memory allocated for A, B, and C using cudaFree().
Muld()
Muld() has the same input as Mul(), except that pointers point to device memory
instead of host memory.
For each block, Muld()iterates through all the sub-matrices of A and B required to
compute Csub. At each iteration:
It loads one sub-matrix of A and one sub-matrix of B from global memory to
shared memory;
It synchronizes to make sure that both sub-matrices are fully loaded by all the
threads within the block;
It computes the product of the two sub-matrices and adds it to the product
obtained during the previous iteration;
CUDA Programming Guide Version 1.1
71
Chapter 6.
Example of Matrix Multiplication
It synchronizes again to make sure that the product of the two sub-matrices is
done before starting the next iteration.
Once all sub-matrices have been handled, Csub is fully computed and Muld() writes
it to global memory.
Muld() is written to maximize memory performance according to Section 5.1.2.1
and 5.1.2.4.
Indeed, assuming that wA and wB are multiples of 16 as suggested in Section 5.1.2.1,
global memory coalescing is ensured because a, b, and c are all multiples of
BLOCK_SIZE, which is equal to 16.
There is also no shared memory bank conflict since for each half-warp, ty and k are
the same for all threads and tx varies from 0 to 15, so each thread accesses a
different bank for the memory accesses As[ty][tx], Bs[ty][tx], and
Bs[k][tx] and the same bank for the memory access As[ty][k].
72
CUDA Programming Guide Version 1.1
Appendix A.
Technical Specifications
All devices with a 1.x compute capability (see Section 3.3) follow the technical
specifications detailed in this appendix.
Atomic functions are only available for devices of compute capability 1.1 (see
Section 4.4.6).
The number of multiprocessors and compute capability of all devices supporting
CUDA are given in the following table:
Number of
Multiprocessors
Compute
Capability
GeForce 8800 Ultra, 8800 GTX
16
1.0
GeForce 8800 GT
14
1.1
GeForce 8800M GTX
12
1.1
GeForce 8800 GTS
12
1.0
GeForce 8800M GTS
8
1.1
GeForce 8600 GTS, 8600 GT, 8700M GT,
8600M GT, 8600M GS
4
1.1
GeForce 8500 GT, 8400 GS, 8400M GT,
8400M GS
2
1.1
GeForce 8400M G
1
1.1
Tesla S870
4x16
1.0
Tesla D870
2x16
1.0
Tesla C870
16
1.0
Quadro Plex 1000 Model S4
4x16
1.0
Quadro Plex 1000 Model IV
2x16
1.0
Quadro FX 5600
16
1.0
Quadro FX 4600
12
1.0
Quadro FX 1700, FX 570, NVS 320M, FX 1600M,
FX 570M
4
1.1
Quadro FX 370, NVS 290, NVS 140M, NVS 135M,
FX 360M
2
1.1
Quadro NVS 130M
1
1.1
CUDA Programming Guide Version 1.1
73
Appendix A.
Technical Specifications
The clock frequency and total amount of device memory can be queried using the
runtime (Sections 4.5.2.2 and 4.5.3.2).
A.1
General Specifications
A.2
The maximum number of threads per block is 512;
The maximum sizes of the x-, y-, and z-dimension of a thread block are 512, 512,
and 64, respectively;
The maximum size of each dimension of a grid of thread blocks is 65535;
The warp size is 32 threads;
The number of registers per multiprocessor is 8192;
The amount of shared memory available per multiprocessor is 16 KB organized
into 16 banks (see Section 5.1.2.4);
The total amount of constant memory is 64 KB;
The cache working set for constant memory is 8 KB per multiprocessor;
The cache working set for texture memory is 8 KB per multiprocessor;
The maximum number of active blocks per multiprocessor is 8;
The maximum number of active warps per multiprocessor is 24;
The maximum number of active threads per multiprocessor is 768;
For a texture reference bound to a one-dimensional CUDA array, the maximum
width is 213;
For a texture reference bound to a two-dimensional CUDA array, the maximum
width is 216 and the maximum height is 215;
For a texture reference bound to linear memory, the maximum width is 227;
The limit on kernel size is 2 million of native instructions;
Each multiprocessor is composed of eight processors, so that a multiprocessor is
able to process the 32 threads of a warp in four clock cycles.
Floating-Point Standard
Compute devices follow the IEEE-754 standard for single-precision binary floatingpoint arithmetic with the following deviations:
74
Addition and multiplication are often combined into a single multiply-add
instruction (FMAD);
Division is implemented via the reciprocal in a non-standard-compliant way;
Square root is implemented via the reciprocal square root in a non-standardcompliant way;
For addition and multiplication, only round-to-nearest-even and
round-towards-zero are supported via static rounding modes; directed rounding
towards +/- infinity is not supported;
There is no dynamically configurable rounding mode;
CUDA Programming Guide Version 1.1
Appendix A.
Technical Specifications
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;
There is no mechanism for detecting that a floating-point exception has occurred
and floating-point exceptions are always masked, but when an exception occurs
the masked response is standard compliant;
Signaling NaNs are not supported.
The result of an operation involving one or more input NaNs is the quiet NaN
of bit pattern 0x7fffffff. Note that 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 behaves.
CUDA Programming Guide Version 1.1
75
Appendix B.
Mathematical Functions
Functions from Section B.1 can be used by both host and device functions whereas
functions from Section B.2 can only be used in device functions.
B.1
Common Runtime Component
Table B-1 below lists all the mathematical standard library functions supported by
the CUDA runtime library. It also specifies the error bounds of each function when
executed on the device. These error bounds also apply when the function is
executed on the host in the case where the host does no supply the function. They
are generated from extensive but not exhaustive tests, so they are not guaranteed
bounds.
Addition and multiplication are IEEE-compliant, so have a maximum error of
0.5 ulp. They are however often combined into a single multiply-add instruction
(FMAD), which truncates the intermediate result of the multiplication.
The recommended way to round a floating-point operand to an integer, with the
result being a floating-point number is rintf(), not roundf(). The reason is that
roundf() maps to an 8-instruction sequence, whereas rintf() maps to a single
instruction. truncf(), ceilf(), and floorf() each map to a single instruction
as well.
The CUDA runtime library also supports integer min() and max() which also
map to a single instruction.
Table B-1.
Mathematical Standard Library Functions with
Maximum ULP Error
Function
x/y
Maximum ulp error
1/x
1 (full range)
1/sqrtf(x)
rsqrtf(x)
2 (full range)
sqrtf(x)
3 (full range)
cbrtf(x)
1 (full range)
CUDA Programming Guide Version 1.1
2 (full range)
77
Appendix B.
78
Mathematical Functions
Function
hypotf(x)
Maximum ulp error
expf(x)
2 (full range)
exp2f(x)
2 (full range)
exp10f(x)
2 (full range)
expm1f(x)
1 (full range)
logf(x)
1 (full range)
log2f(x)
3 (full range)
log10f(x)
3 (full range)
log1pf(x)
2 (full range)
sinf(x)
2 (full range)
cosf(x)
2 (full range)
tanf(x)
4 (full range)
sincosf(x,sptr,cptr)
2 (full range)
asinf(x)
4 (full range)
acosf(x)
3 (full range)
atanf(x)
2 (full range)
atan2f(y,x)
3 (full range)
sinhf(x)
3 (full range)
coshf(x)
2 (full range)
tanhf(x)
2 (full range)
asinhf(x)
3 (full range)
acoshf(x)
4 (full range)
atanhf(x)
3 (full range)
powf(x,y)
16 (full range)
erff(x)
4 (full range)
erfcf(x)
8 (full range)
lgammaf(x)
6 (outside interval -10.001 ... -2.264; larger inside)
tgammaf(x)
11 (full range)
fmaf(x,y,z)
0 (full range)
frexpf(x,exp)
0 (full range)
ldexpf(x,exp)
0 (full range)
scalbnf(x,n)
0 (full range)
scalblnf(x,l)
0 (full range)
logbf(x)
0 (full range)
ilogbf(x)
0 (full range)
fmodf(x,y)
0 (full range)
remainderf(x,y)
0 (full range)
remquof(x,y,iptr)
0 (full range)
modff(x,iptr)
0 (full range)
fdimf(x,y)
0 (full range)
truncf(x)
0 (full range)
roundf(x)
0 (full range)
3 (full range)
CUDA Programming Guide Version 1.1
Appendix B.
Function
rintf(x)
Maximum ulp error
nearbyintf(x)
0 (full range)
ceilf(x)
0 (full range)
floorf(x)
0 (full range)
lrintf(x)
0 (full range)
lroundf(x)
0 (full range)
llrintf(x)
0 (full range)
llroundf(x)
0 (full range)
signbit(x)
N/A
isinf(x)
N/A
isnan(x)
N/A
isfinite(x)
N/A
copysignf(x,y)
N/A
fminf(x,y)
N/A
fmaxf(x,y)
N/A
fabsf(x)
N/A
nanf(cptr)
N/A
nextafterf(x,y)
N/A
CUDA Programming Guide Version 1.1
Mathematical Functions
0 (full range)
79
Appendix B.
B.2
Mathematical Functions
Device Runtime Component
Table B-2 lists the intrinsic functions that are only supported in device code. Their
error bounds are GPU-specific. Among these functions are the less accurate, but
faster versions of some of the functions of Table B-1; they have the same name
prefixed with __ (such as __sinf(x)).
__fadd_rz(x,y) computes the sum of floating-point parameters x and y using
the round-towards-zero rounding mode.
__fmul_rz(x,y) computes the product of floating-point parameters x and y
using the round-towards-zero rounding mode.
Both the regular floating-point division and __fdividef(x,y) have the same
accuracy, but for 2126 < y < 2128, __fdividef(x,y) delivers a result of zero,
whereas the regular division delivers the correct result to within the accuracy stated
in Table B-1. 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 regular division returns
infinity.
__[u]mul24(x,y) computes the product of the 24 least significant bits of the
integer parameters x and y and delivers the 32 least significant bits of the result. The
8 most significant bits of x or y are ignored.
__[u]mulhi(x,y) computes the product of the integer parameters x and y and
delivers the 32 most significant bits of the 64-bit result.
__[u]mul64hi(x,y) computes the product of the 64-bit integer parameters x
and y and delivers the 64 most significant bits of the 128-bit result.
__saturate(x) returns 0 if x is less than 0, 1 if x is more than 1, and x
otherwise.
__[u]sad(x,y,z) (Sum of Absolute Difference) returns the sum of integer
parameter z and the absolute value of the difference between integer parameters x
and y.
__clz(x) returns the number, between 0 and 32 inclusive, of consecutive zero bits
starting at the most significant bit (i.e. bit 31) of integer parameter x.
__clzll(x) returns the number, between 0 and 64 inclusive, of consecutive zero
bits starting at the most significant bit (i.e. bit 63) of 64-bit integer parameter x.
__ffs(x) returns the position of the first (least significant) bit set in integer
parameter x. The least significant bit is position 1. If x is 0, __ffs() returns 0.
Note that this is identical to the Linux function ffs.
__ffsll(x) returns the position of the first (least significant) bit set in 64-bit
integer parameter x. The least significant bit is position 1. If x is 0, __ffsll()
returns 0. Note that this is identical to the Linux function ffsll.
80
CUDA Programming Guide Version 1.1
Appendix B.
Table B-2.
Mathematical Functions
Intrinsic Functions Supported by the CUDA
Runtime Library with Respective Error Bounds
for Devices of Compute Capability 1.x
Function
__fadd_rz(x,y)
Error bounds
__fmul_rz(x,y)
IEEE-compliant.
__fdividef(x,y)
For y in [2-126, 2126], the maximum ulp error is 2.
__expf(x)
The maximum ulp error is
IEEE-compliant.
2 + floor(abs(1.16 * x)).
__exp10f(x)
The maximum ulp error is
2 + floor(abs(2.95 * x)).
__logf(x)
For x in [0.5, 2], the maximum absolute error is 2-21.41,
otherwise, the maximum ulp error is 3.
__log2f(x)
For x in [0.5, 2], the maximum absolute error is 2-22,
otherwise, the maximum ulp error is 2.
__log10f(x)
For x in [0.5, 2], the maximum absolute error is 2-24,
otherwise, the maximum ulp error is 3.
__sinf(x)
For x in [-π, π], the maximum absolute error is 2-21.41, and
larger otherwise.
__cosf(x)
For x in [-π, π], the maximum absolute error is 2-21.19, and
larger otherwise.
__sincosf(x,sptr,cptr)
Same as sinf(x) and cosf(x).
__tanf(x)
Derived from its implementation as
__sinf(x) * (1 / __cosf(x)).
__powf(x, y)
Derived from its implementation as
exp2f(y * __log2f(x)).
__mul24(x,y)
__umul24(x,y)
N/A
__mulhi(x,y)
__umulhi(x,y)
N/A
__int_as_float(x)
N/A
__float_as_int(x)
N/A
__saturate(x)
N/A
__sad(x,y,z)
__usad(x,y,z)
N/A
__clz(x)
N/A
__ffs(x)
N/A
CUDA Programming Guide Version 1.1
81
Appendix C.
Atomic Functions
Atomic functions can only be used in device functions.
C.1
Arithmetic Functions
C.1.1
atomicAdd()
int atomicAdd(int* address, int val);
unsigned int atomicAdd(unsigned int* address,
unsigned int val);
reads the 32-bit word old located at the address address in global memory,
computes (old + val), and stores the result back to global memory at the same
address. These three operations are performed in one atomic transaction. The
function returns old.
C.1.2
atomicSub()
int atomicSub(int* address, int val);
unsigned int atomicSub(unsigned int* address,
unsigned int val);
reads the 32-bit word old located at the address address in global memory,
computes (old - val), and stores the result back to global memory at the same
address. These three operations are performed in one atomic transaction. The
function returns old.
C.1.3
atomicExch()
int atomicExch(int* address, int
unsigned int atomicExch(unsigned
unsigned
float atomicExch(float* address,
val);
int* address,
int val);
float val);
reads the 32-bit word old located at the address address in global memory and
stores val back to global memory at the same address. These two operations are
performed in one atomic transaction. The function returns old.
CUDA Programming Guide Version 1.1
83
Appendix C.
C.1.4
Atomic Functions
atomicMin()
int atomicMin(int* address, int val);
unsigned int atomicMin(unsigned int* address,
unsigned int val);
reads the 32-bit word old located at the address address in global memory,
computes the minimum of old and val, and stores the result back to global
memory at the same address. These three operations are performed in one atomic
transaction. The function returns old.
C.1.5
atomicMax()
int atomicMax(int* address, int val);
unsigned int atomicMax(unsigned int* address,
unsigned int val);
reads the 32-bit word old located at the address address in global memory,
computes the maximum of old and val, and stores the result back to global
memory at the same address. These three operations are performed in one atomic
transaction. The function returns old.
C.1.6
atomicInc()
unsigned int atomicInc(unsigned int* address,
unsigned int val);
reads the 32-bit word old located at the address address in global memory,
computes ((old >= val) ? 0 : (old+1)), and stores the result back to
global memory at the same address. These three operations are performed in one
atomic transaction. The function returns old.
C.1.7
atomicDec()
unsigned int atomicDec(unsigned int* address,
unsigned int val);
reads the 32-bit word old located at the address address in global memory,
computes (((old == 0) | (old > val)) ? val : (old-1)), and stores
the result back to global memory at the same address. These three operations are
performed in one atomic transaction. The function returns old.
C.1.8
atomicCAS()
int atomicCAS(int* address, int
unsigned int atomicCAS(unsigned
unsigned
unsigned
compare, int val);
int* address,
int compare,
int val);
reads the 32-bit word old located at the address address in global memory,
computes (old == compare ? val : old), and stores the result back to
global memory at the same address. These three operations are performed in one
atomic transaction. The function returns old (Compare And Swap).
84
CUDA Programming Guide Version 1.1
Appendix C.
C.2
Bitwise Functions
C.2.1
atomicAnd()
Atomic Functions
int atomicAnd(int* address, int val);
unsigned int atomicAnd(unsigned int* address,
unsigned int val);
reads the 32-bit word old located at the address address in global memory,
computes (old & val), and stores the result back to global memory at the same
address. These three operations are performed in one atomic transaction. The
function returns old.
C.2.2
atomicOr()
int atomicOr(int* address, int val);
unsigned int atomicOr(unsigned int* address,
unsigned int val);
reads the 32-bit word old located at the address address in global memory,
computes (old | val), and stores the result back to global memory at the same
address. These three operations are performed in one atomic transaction. The
function returns old.
C.2.3
atomicXor()
int atomicXor(int* address, int val);
unsigned int atomicXor(unsigned int* address,
unsigned int val);
reads the 32-bit word old located at the address address in global memory,
computes (old ^ val), and stores the result back to global memory at the same
address. These three operations are performed in one atomic transaction. The
function returns old.
CUDA Programming Guide Version 1.1
85
Appendix D.
Runtime API Reference
There are two levels for the runtime API.
The low-level API (cuda_runtime_api.h) is a C-style interface that does not
require compiling with nvcc.
The high-level API (cuda_runtime.h) is a C++-style interface built on top of the
low-level API. It wraps some of the low level API routines, using overloading,
references and default arguments. These wrappers can be used from C++ code and
can be compiled with any C++ compiler. The high-level API also has some CUDAspecific wrappers that wrap low-level routines that deal with symbols, textures, and
device functions. These wrappers require the use of nvcc because they depend on
code being generated by the compiler (see Section 4.2.5). For example, the
execution configuration syntax described in Section 4.2.3 to invoke kernels is only
available in source code compiled with nvcc.
D.1
Device Management
D.1.1
cudaGetDeviceCount()
cudaError_t cudaGetDeviceCount(int* count);
returns in *count the number of devices with compute capability greater or equal
to 1.0 that are available for execution. If there is no such device,
cudaGetDeviceCount() returns 1 and device 0 only supports device emulation
mode and is of compute capability less than 1.0.
D.1.2
cudaSetDevice()
cudaError_t cudaSetDevice(int dev);
records dev as the device on which the active host thread executes the device code.
D.1.3
cudaGetDevice()
cudaError_t cudaGetDevice(int* dev);
CUDA Programming Guide Version 1.1
87
Appendix D.
Runtime API Reference
returns in *dev the device on which the active host thread executes the device
code.
D.1.4
cudaGetDeviceProperties()
cudaError_t cudaGetDeviceProperties(struct cudaDeviceProp* prop,
int dev);
returns in *prop the properties of device dev. The cudaDeviceProp structure is
defined as:
struct cudaDeviceProp {
char
name[256];
size_t totalGlobalMem;
size_t sharedMemPerBlock;
int
regsPerBlock;
int
warpSize;
size_t memPitch;
int
maxThreadsPerBlock;
int
maxThreadsDim[3];
int
maxGridSize[3];
size_t totalConstMem;
int
major;
int
minor;
int
clockRate;
size_t textureAlignment;
};
where:
is an ASCII string identifying the device;
totalGlobalMem is the total amount of global memory available on the device
in bytes;
sharedMemPerBlock is the total amount of shared memory available per
block in bytes;
regsPerBlock is the total number of registers available per block;
warpSize is the warp size;
memPitch is the maximum pitch allowed by the memory copy functions of
Section D.5 that involve memory regions allocated through
cudaMallocPitch() (Section D.5.2);
maxThreadsPerBlock is the maximum number of threads per block;
maxThreadsDim[3] is the maximum sizes of each dimension of a block;
maxGridSize[3] is the maximum sizes of each dimension of a grid;
totalConstMem is the total amount of constant memory available on the
device in bytes;
major and minor are the major and minor revision numbers defining the
device’s compute capability;
clockRate is the clock frequency in kilohertz;
textureAlignment is the alignment requirement mentioned in
Section 4.3.4.3; texture base addresses that are aligned to textureAlignment
bytes do not need an offset applied to texture fetches.
name
88
CUDA Programming Guide Version 1.1
Appendix D.
D.1.5
Runtime API Reference
cudaChooseDevice()
cudaError_t cudaChooseDevice(int* dev,
const struct cudaDeviceProp* prop);
returns in *dev the device which properties best match *prop.
D.2
Thread Management
D.2.1
cudaThreadSynchronize()
cudaError_t cudaThreadSynchronize(void);
blocks until the device has completed all preceding requested tasks.
cudaThreadSynchronize() returns an error if one of the preceding tasks failed.
D.2.2
cudaThreadExit()
cudaError_t cudaThreadExit(void);
explicitly cleans up all runtime-related resources associated with the calling host
thread. Any subsequent API call reinitializes the runtime. cudaThreadExit() is
implicitly called on host thread exit.
D.3
Stream Management
D.3.1
cudaStreamCreate()
cudaError_t cudaStreamCreate(cudaStream_t* stream);
creates a stream.
D.3.2
cudaStreamQuery()
cudaError_t cudaStreamQuery(cudaStream_t stream);
returns cudaSuccess if all operations in the stream have completed, or
cudaErrorNotReady if not.
D.3.3
cudaStreamSynchronize()
cudaError_t cudaStreamSynchronize(cudaStream_t stream);
blocks until the device has completed all operations in the stream.
D.3.4
cudaStreamDestroy()
cudaError_t cudaStreamDestroy(cudaStream_t stream);
destroys the stream.
CUDA Programming Guide Version 1.1
89
Appendix D.
Runtime API Reference
D.4
Event Management
D.4.1
cudaEventCreate()
cudaError_t cudaEventCreate(cudaEvent_t* event);
creates an event.
D.4.2
cudaEventRecord()
cudaError_t cudaEventRecord(cudaEvent_t event, CUstream stream);
records an event. If stream is non-zero, the event is recorded after all preceding
operations in the stream have been completed; otherwise, it is recorded after all
preceding operations in the CUDA context have been completed. Since this
operation is asynchronous, cudaEventQuery() and/or
cudaEventSynchronize() must be used to determine when the event has
actually been recorded.
If cudaEventRecord() has previously been called and the event has not been
recorded yet, this function returns cudaErrorInvalidValue.
D.4.3
cudaEventQuery()
cudaError_t cudaEventQuery(cudaEvent_t event);
returns cudaSuccess if the event has actually been recorded, or
cudaErrorNotReady if not. If cudaEventRecord() has not been called on
this event, the function returns cudaErrorInvalidValue.
D.4.4
cudaEventSynchronize()
cudaError_t cudaEventSynchronize(cudaEvent_t event);
blocks until the event has actually been recorded. If cudaEventRecord() has not
been called on this event, the function returns cudaErrorInvalidValue.
D.4.5
cudaEventDestroy()
cudaError_t cudaEventDestroy(cudaEvent_t event);
destroys the event.
D.4.6
cudaEventElapsedTime()
cudaError_t cudaEventElapsedTime(float* time,
cudaEvent_t start,
cudaEvent_t end);
computes the elapsed time between two events (in milliseconds with a resolution of
around 0.5 microseconds). If either event has not been recorded yet, this function
90
CUDA Programming Guide Version 1.1
Appendix D.
Runtime API Reference
returns cudaErrorInvalidValue. If either event has been recorded with a nonzero stream, the result is undefined.
D.5
Memory Management
D.5.1
cudaMalloc()
cudaError_t cudaMalloc(void** devPtr, size_t count);
allocates count bytes of linear memory on the device and returns in *devPtr a
pointer to the allocated memory. The allocated memory is suitably aligned for any
kind of variable. The memory is not cleared. cudaMalloc() returns
cudaErrorMemoryAllocation in case of failure.
D.5.2
cudaMallocPitch()
cudaError_t cudaMallocPitch(void** devPtr,
size_t* pitch,
size_t widthInBytes,
size_t height);
allocates at least widthInBytes*height bytes of linear memory on the device
and returns in *devPtr a pointer to the allocated memory. The function may pad
the allocation to ensure that corresponding pointers in any given row will continue
to meet the alignment requirements for coalescing as the address is updated from
row to row (see Section 5.1.2.1). The pitch returned in *pitch by
cudaMallocPitch() is the width in bytes of the allocation. The intended usage
of pitch is as a separate parameter of the allocation, used to compute addresses
within the 2D array. Given the row and column of an array element of type T, the
address is computed as
T* pElement = (T*)((char*)BaseAddress + Row * pitch) + Column;
For allocations of 2D arrays, it is recommended that programmers consider
performing pitch allocations using cudaMallocPitch(). Due to pitch alignment
restrictions in the hardware, this is especially true if the application will be
performing 2D memory copies between different regions of device memory
(whether linear memory or CUDA arrays).
D.5.3
cudaFree()
cudaError_t cudaFree(void* devPtr);
frees the memory space pointed to by devPtr, which must have been returned by a
previous call to cudaMalloc() or cudaMallocPitch(). Otherwise, or if
cudaFree(devPtr) has already been called before, an error is returned. If
devPtr is 0, no operation is performed. cudaFree() returns
cudaErrorInvalidDevicePointer in case of failure.
CUDA Programming Guide Version 1.1
91
Appendix D.
D.5.4
Runtime API Reference
cudaMallocArray()
cudaError_t cudaMallocArray(struct cudaArray** array,
const struct cudaChannelFormatDesc* desc,
size_t width, size_t height);
allocates a CUDA array according to the cudaChannelFormatDesc structure
desc and returns a handle to the new CUDA array in *array.
cudaChannelFormatDesc is described in Section 4.3.4.
D.5.5
cudaFreeArray()
cudaError_t cudaFreeArray(struct cudaArray* array);
frees the CUDA array array. If array is 0, no operation is performed.
D.5.6
cudaMallocHost()
cudaError_t cudaMallocHost(void** hostPtr, size_t size);
allocates size bytes of host memory that is page-locked and accessible to the
device. The driver tracks the virtual memory ranges allocated with this function and
automatically accelerates calls to functions such as cudaMemcpy*(). Since the
memory can be accessed directly by the device, it can be read or written with much
higher bandwidth than pageable memory obtained with functions such as
malloc(). Allocating excessive amounts of memory with cudaMallocHost()
may degrade system performance, since it reduces the amount of memory available
to the system for paging. As a result, this function is best used sparingly to allocate
staging areas for data exchange between host and device.
D.5.7
cudaFreeHost()
cudaError_t cudaFreeHost(void* hostPtr);
frees the memory space pointed to by hostPtr, which must have been returned by
a previous call to cudaMallocHost().
D.5.8
cudaMemset()
cudaError_t cudaMemset(void* devPtr, int value, size_t count);
fills the first count bytes of the memory area pointed to by devPtr with the
constant byte value value.
D.5.9
cudaMemset2D()
cudaError_t cudaMemset2D(void* dstPtr, size_t pitch,
int value, size_t width, size_t height);
sets to the specified value value a matrix (height rows of width bytes each)
pointed to by dstPtr. pitch is the width in memory in bytes of the 2D array
pointed to by dstPtr, including any padding added to the end of each row (see
92
CUDA Programming Guide Version 1.1
Appendix D.
Runtime API Reference
Section D.5.2). This function performs fastest when the pitch is one that has been
passed back by cudaMallocPitch().
D.5.10
cudaMemcpy()
cudaError_t cudaMemcpy(void* dst, const void* src,
size_t count,
enum cudaMemcpyKind kind);
cudaError_t cudaMemcpyAsync(void* dst, const void* src,
size_t count,
enum cudaMemcpyKind kind,
cudaStream_t stream);
copies count bytes from the memory area pointed to by src to the memory area
pointed to by dst, where kind is one of cudaMemcpyHostToHost,
cudaMemcpyHostToDevice, cudaMemcpyDeviceToHost, or
cudaMemcpyDeviceToDevice, and specifies the direction of the copy. The
memory areas may not overlap. Calling cudaMemcpy() with dst and src pointers
that do not match the direction of the copy results in an undefined behavior.
cudaMemcpyAsync() is asynchronous and can optionally be associated to a
stream by passing a non-zero stream argument. It only works on page-locked host
memory and returns an error if a pointer to pageable memory is passed as input.
D.5.11
cudaMemcpy2D()
cudaError_t cudaMemcpy2D(void* dst, size_t dpitch,
const void* src, size_t spitch,
size_t width, size_t height,
enum cudaMemcpyKind kind);
cudaError_t cudaMemcpy2DAsync(void* dst, size_t dpitch,
const void* src, size_t spitch,
size_t width, size_t height,
enum cudaMemcpyKind kind,
cudaStream_t stream);
copies a matrix (height rows of width bytes each) from the memory area pointed
to by src to the memory area pointed to by dst, where kind is one of
cudaMemcpyHostToHost, cudaMemcpyHostToDevice,
cudaMemcpyDeviceToHost, or cudaMemcpyDeviceToDevice, and specifies
the direction of the copy. dpitch and spitch are the widths in memory in bytes
of the 2D arrays pointed to by dst and src, including any padding added to the
end of each row (see Section D.5.2). The memory areas may not overlap. Calling
cudaMemcpy2D() with dst and src pointers that do not match the direction of
the copy results in an undefined behavior. cudaMemcpy2D() returns an error if
dpitch or spitch is greater than the maximum allowed (see memPitch in
Section D.1.4).
cudaMemcpy2DAsync() is asynchronous and can optionally be associated to a
stream by passing a non-zero stream argument. It only works on page-locked host
memory and returns an error if a pointer to pageable memory is passed as input.
CUDA Programming Guide Version 1.1
93
Appendix D.
Runtime API Reference
D.5.12
cudaMemcpyToArray()
cudaError_t cudaMemcpyToArray(struct cudaArray* dstArray,
size_t dstX, size_t dstY,
const void* src, size_t count,
enum cudaMemcpyKind kind);
cudaError_t cudaMemcpyToArrayAsync(struct cudaArray* dstArray,
size_t dstX, size_t dstY,
const void* src, size_t count,
enum cudaMemcpyKind kind,
cudaStream_t stream);
copies count bytes from the memory area pointed to by src to the CUDA array
dstArray starting at the upper left corner (dstX, dstY), where kind is one of
cudaMemcpyHostToHost, cudaMemcpyHostToDevice,
cudaMemcpyDeviceToHost, or cudaMemcpyDeviceToDevice, and specifies
the direction of the copy.
cudaMemcpyToArrayAsync() is asynchronous and can optionally be associated
to a stream by passing a non-zero stream argument. It only works on page-locked
host memory and returns an error if a pointer to pageable memory is passed as
input.
D.5.13
cudaMemcpy2DToArray()
cudaError_t cudaMemcpy2DToArray(struct cudaArray* dstArray,
size_t dstX, size_t dstY,
const void* src, size_t spitch,
size_t width, size_t height,
enum cudaMemcpyKind kind);
cudaError_t cudaMemcpy2DToArrayAsync(struct cudaArray* dstArray,
size_t dstX, size_t dstY,
const void* src, size_t spitch,
size_t width, size_t height,
enum cudaMemcpyKind kind,
cudaStream_t stream);
copies a matrix (height rows of width bytes each) from the memory area pointed
to by src to the CUDA array dstArray starting at the upper left corner (dstX,
dstY), where kind is one of cudaMemcpyHostToHost,
cudaMemcpyHostToDevice, cudaMemcpyDeviceToHost, or
cudaMemcpyDeviceToDevice, and specifies the direction of the copy. spitch
is the width in memory in bytes of the 2D array pointed to by src, including any
padding added to the end of each row (see Section D.5.2). cudaMemcpy2D()
returns an error if spitch is greater than the maximum allowed (see memPitch in
Section D.1.4).
cudaMemcpy2DToArrayAsync() is asynchronous and can optionally be
associated to a stream by passing a non-zero stream argument. It only works on
page-locked host memory and returns an error if a pointer to pageable memory is
passed as input.
94
CUDA Programming Guide Version 1.1
Appendix D.
D.5.14
Runtime API Reference
cudaMemcpyFromArray()
cudaError_t cudaMemcpyFromArray(void* dst,
const struct cudaArray* srcArray,
size_t srcX, size_t srcY,
size_t count,
enum cudaMemcpyKind kind);
cudaError_t cudaMemcpyFromArrayAsync(void* dst,
const struct cudaArray* srcArray,
size_t srcX, size_t srcY,
size_t count,
enum cudaMemcpyKind kind,
cudaStream_t stream);
copies count bytes from the CUDA array srcArray starting at the upper left
corner (srcX, srcY) to the memory area pointed to by dst, where kind is one of
cudaMemcpyHostToHost, cudaMemcpyHostToDevice,
cudaMemcpyDeviceToHost, or cudaMemcpyDeviceToDevice, and specifies
the direction of the copy.
cudaMemcpyFromArrayAsync() is asynchronous and can optionally be
associated to a stream by passing a non-zero stream argument. It only works on
page-locked host memory and returns an error if a pointer to pageable memory is
passed as input.
D.5.15
cudaMemcpy2DFromArray()
cudaError_t cudaMemcpy2DFromArray(void* dst, size_t dpitch,
const struct cudaArray* srcArray,
size_t srcX, size_t srcY,
size_t width, size_t height,
enum cudaMemcpyKind kind);
cudaError_t cudaMemcpy2DFromArrayAsync(void* dst, size_t dpitch,
const struct cudaArray* srcArray,
size_t srcX, size_t srcY,
size_t width, size_t height,
enum cudaMemcpyKind kind,
cudaStream_t stream);
copies a matrix (height rows of width bytes each) from the CUDA array
srcArray starting at the upper left corner (srcX, srcY) to the memory area
pointed to by dst, where kind is one of cudaMemcpyHostToHost,
cudaMemcpyHostToDevice, cudaMemcpyDeviceToHost, or
cudaMemcpyDeviceToDevice, and specifies the direction of the copy. dpitch
is the width in memory in bytes of the 2D array pointed to by dst, including any
padding added to the end of each row (see Section D.5.2). cudaMemcpy2D()
returns an error if dpitch is greater than the maximum allowed (see memPitch in
Section D.1.4).
cudaMemcpy2DFromArrayAsync() is asynchronous and can optionally be
associated to a stream by passing a non-zero stream argument. It only works on
page-locked host memory and returns an error if a pointer to pageable memory is
passed as input.
CUDA Programming Guide Version 1.1
95
Appendix D.
Runtime API Reference
D.5.16
cudaMemcpyArrayToArray()
cudaError_t cudaMemcpyArrayToArray(struct cudaArray* dstArray,
size_t dstX, size_t dstY,
const struct cudaArray* srcArray,
size_t srcX, size_t srcY,
size_t count,
enum cudaMemcpyKind kind);
copies count bytes from the CUDA array srcArray starting at the upper left
corner (srcX, srcY) to the CUDA array dstArray starting at the upper left
corner (dstX, dstY), where kind is one of cudaMemcpyHostToHost,
cudaMemcpyHostToDevice, cudaMemcpyDeviceToHost, or
cudaMemcpyDeviceToDevice, and specifies the direction of the copy.
D.5.17
cudaMemcpy2DArrayToArray()
cudaError_t cudaMemcpy2DArrayToArray(struct cudaArray* dstArray,
size_t dstX, size_t dstY,
const struct cudaArray* srcArray,
size_t srcX, size_t srcY,
size_t width, size_t height,
enum cudaMemcpyKind kind);
copies a matrix (height rows of width bytes each) from the CUDA array
srcArray starting at the upper left corner (srcX, srcY) to the CUDA array
dstArray starting at the upper left corner (dstX, dstY), where kind is one of
cudaMemcpyHostToHost, cudaMemcpyHostToDevice,
cudaMemcpyDeviceToHost, or cudaMemcpyDeviceToDevice, and specifies
the direction of the copy.
D.5.18
cudaMemcpyToSymbol()
template
cudaError_t cudaMemcpyToSymbol(const T& symbol, const void* src,
size_t count, size_t offset = 0,
enum cudaMemcpyKind kind = cudaMemcpyHostToDevice);
copies count bytes from the memory area pointed to by src to the memory area
pointed to by offset bytes from the start of symbol symbol. The memory areas
may not overlap. symbol can either be a variable that resides in global or constant
memory space, or it can be a character string, naming a variable that resides in
global or constant memory space. kind can be either
cudaMemcpyHostToDevice or cudaMemcpyDeviceToDevice.
D.5.19
cudaMemcpyFromSymbol()
template
cudaError_t cudaMemcpyFromSymbol(void *dst, const T& symbol,
size_t count, size_t offset = 0,
enum cudaMemcpyKind kind = cudaMemcpyDeviceToHost);
96
CUDA Programming Guide Version 1.1
Appendix D.
Runtime API Reference
copies count bytes from the memory area pointed to by offset bytes from the
start of symbol symbol to the memory area pointed to by dst. The memory areas
may not overlap. symbol can either be a variable that resides in global or constant
memory space, or it can be a character string, naming a variable that resides in
global or constant memory space. kind can be either
cudaMemcpyDeviceToHost or cudaMemcpyDeviceToDevice.
D.5.20
cudaGetSymbolAddress()
template
cudaError_t cudaGetSymbolAddress(void** devPtr, const T& symbol);
returns in *devPtr the address of symbol symbol on the device. symbol can
either be a variable that resides in global memory space, or it can be a character
string, naming a variable that resides in global memory space. If symbol cannot be
found, or if symbol is not declared in global memory space, *devPtr is
unchanged and an error is returned. cudaGetSymbolAddress() returns
cudaErrorInvalidSymbol in case of failure.
D.5.21
cudaGetSymbolSize()
template
cudaError_t cudaGetSymbolSize(size_t* size, const T& symbol);
returns in *size the size of symbol symbol. symbol can either be a variable that
resides in global or constant memory space, or it can be a character string, naming a
variable that resides in global or constant memory space. If symbol cannot be
found, or if symbol is not declared in global or constant memory space, *size is
unchanged and an error is returned. cudaGetSymbolSize() returns
cudaErrorInvalidSymbol in case of failure.
D.6
Texture Reference Management
D.6.1
Low-Level API
D.6.1.1
cudaCreateChannelDesc()
struct cudaChannelFormatDesc
cudaCreateChannelDesc(int x, int y, int z, int w,
enum cudaChannelFormatKind f);
returns a channel descriptor with format f and number of bits of each component
x, y, z, and w. cudaChannelFormatDesc is described in Section 4.3.4.
D.6.1.2
cudaGetChannelDesc()
cudaError_t cudaGetChannelDesc(struct cudaChannelFormatDesc* desc,
const struct cudaArray* array);
returns in *desc the channel descriptor of the CUDA array array.
D.6.1.3
cudaGetTextureReference()
cudaError_t cudaGetTextureReference(
CUDA Programming Guide Version 1.1
97
Appendix D.
Runtime API Reference
struct textureReference** texRef,
const char* symbol);
returns in *texRef the structure associated to the texture reference defined by
symbol symbol.
D.6.1.4
cudaBindTexture()
cudaError_t cudaBindTexture(size_t* offset,
const struct textureReference* texRef,
const void* devPtr,
const struct cudaChannelFormatDesc* desc,
size_t size = UINT_MAX);
binds size bytes of the memory area pointed to by devPtr to the texture
reference texRef. desc describes how the memory is interpreted when fetching
values from the texture. Any memory previously bound to texRef is unbound.
Since the hardware enforces an alignment requirement on texture base addresses,
cudaBindTexture() returns in *offset a byte offset that must be applied to
texture fetches in order to read from the desired memory. This offset must be
divided by the texel size and passed to kernels that read from the texture so they can
be applied to the tex1Dfetch() function. If the device memory pointer was
returned from cudaMalloc(), the offset is guaranteed to be 0 and NULL may be
passed as the offset parameter.
D.6.1.5
cudaBindTextureToArray()
cudaError_t cudaBindTextureToArray(
const struct textureReference* texRef,
const struct cudaArray* array,
const struct cudaChannelFormatDesc* desc);
binds the CUDA array array to the texture reference texRef. desc describes
how the memory is interpreted when fetching values from the texture. Any CUDA
array previously bound to texRef is unbound.
D.6.1.6
cudaUnbindTexture()
cudaError_t cudaUnbindTexture(
const struct textureReference* texRef);
unbinds the texture bound to texture reference texRef.
D.6.1.7
cudaGetTextureAlignmentOffset()
cudaError_t cudaGetTextureAlignmentOffset(size_t* offset,
const struct textureReference* texRef);
returns in *offset the offset that was returned when texture reference texRef
was bound.
D.6.2
High-Level API
D.6.2.1
cudaCreateChannelDesc()
template
struct cudaChannelFormatDesc cudaCreateChannelDesc();
returns a channel descriptor with format matching type T. T can be any types of
Section 4.3.1.1. 3-component types default to a 4-component format.
98
CUDA Programming Guide Version 1.1
Appendix D.
D.6.2.2
Runtime API Reference
cudaBindTexture()
template
static __inline__ __host__ cudaError_t
cudaBindTexture(size_t* offset,
const struct texture& texRef,
const void* devPtr,
const struct cudaChannelFormatDesc& desc,
size_t size = UINT_MAX);
binds size bytes of the memory area pointed to by devPtr to texture reference
texRef. desc describes how the memory is interpreted when fetching values from
the texture. The offset parameter is an optional byte offset as with the low-level
cudaBindTexture() function described in Section D.6.1.4. Any memory
previously bound to texRef is unbound.
template
static __inline__ __host__ cudaError_t
cudaBindTexture(size_t* offset,
const struct texture& texRef,
const void* devPtr,
size_t size = UINT_MAX);
binds size bytes of the memory area pointed to by devPtr to texture reference
texRef. The channel descriptor is inherited from the texture reference type. The
offset parameter is an optional byte offset as with the low-level
cudaBindTexture() function described in Section D.6.1.4. Any memory
previously bound to texRef is unbound.
D.6.2.3
cudaBindTextureToArray()
template
static __inline__ __host__ cudaError_t
cudaBindTextureToArray(
const struct texture& texRef,
const struct cudaArray* cuArray,
const struct cudaChannelFormatDesc& desc);
binds the CUDA array array to texture reference texRef. desc describes how
the memory is interpreted when fetching values from the texture. Any CUDA array
previously bound to texRef is unbound.
template
static __inline__ __host__ cudaError_t
cudaBindTextureToArray(
const struct texture& texRef,
const struct cudaArray* cuArray);
binds the CUDA array array to texture reference texRef. The channel descriptor
is inherited from the CUDA array. Any CUDA array previously bound to texRef
is unbound.
D.6.2.4
cudaUnbindTexture()
template
static __inline__ __host__ cudaError_t
cudaUnbindTexture(const struct texture& texRef);
unbinds the texture bound to texture reference texRef.
CUDA Programming Guide Version 1.1
99
Appendix D.
Runtime API Reference
D.7
Execution Control
D.7.1
cudaConfigureCall()
cudaError_t cudaConfigureCall(dim3 gridDim, dim3 blockDim,
size_t sharedMem = 0,
int tokens = 0);
specifies the grid and block dimensions for the device call to be executed similar to
the execution configuration syntax described in Section 4.2.3.
cudaConfigureCall() is stack based. Each call pushes data on top of an
execution stack. This data contains the dimension for the grid and thread blocks,
together with any arguments for the call.
D.7.2
cudaLaunch()
template cudaError_t cudaLaunch(T entry);
launches the function entry on the device. entry can either be a function that
executes on the device, or it can be a character string, naming a function that
executes on the device. entry must be declared as a __global__ function.
cudaLaunch() must be preceded by a call to cudaConfigureCall() since it
pops the data that was pushed by cudaConfigureCall() from the execution
stack.
D.7.3
cudaSetupArgument()
cudaError_t cudaSetupArgument(void* arg,
size_t count, size_t offset);
template cudaError_t cudaSetupArgument(T arg,
size_t offset);
pushes count bytes of the argument pointed to by arg at offset bytes from the
start of the parameter passing area, which starts at offset 0. The arguments are
stored in the top of the execution stack. cudaSetupArgument() must be
preceded by a call to cudaConfigureCall().
D.8
OpenGL Interoperability
D.8.1
cudaGLRegisterBufferObject()
cudaError_t cudaGLRegisterBufferObject(GLuint bufferObj);
registers the buffer object of ID bufferObj for access by CUDA. This function
must be called before CUDA can map the buffer object. While it is registered, the
buffer object cannot be used by any OpenGL commands except as a data source for
OpenGL drawing commands.
100
CUDA Programming Guide Version 1.1
Appendix D.
D.8.2
Runtime API Reference
cudaGLMapBufferObject()
cudaError_t cudaGLMapBufferObject(void** devPtr,
GLuint bufferObj);
maps the buffer object of ID bufferObj into the address space of CUDA and
returns in *devPtr the base pointer of the resulting mapping.
D.8.3
cudaGLUnmapBufferObject()
cudaError_t cudaGLUnmapBufferObject(GLuint bufferObj);
unmaps the buffer object of ID bufferObj for access by CUDA.
D.8.4
cudaGLUnregisterBufferObject()
cudaError_t cudaGLUnregisterBufferObject(GLuint bufferObj);
unregisters the buffer object of ID bufferObj for access by CUDA.
D.9
Direct3D Interoperability
D.9.1
cudaD3D9Begin()
cudaError_t cudaD3D9Begin(IDirect3DDevice9* device);
initializes interoperability with the Direct3D device device. This function must be
called before CUDA can map any objects from device. The application can then
map vertex buffers owned by the Direct3D device until cuD3D9End() is called.
D.9.2
cudaD3D9End()
cudaError_t cudaD3D9End(void);
concludes interoperability with the Direct3D device previously specified to
cuD3D9Begin().
D.9.3
cudaD3D9RegisterVertexBuffer()
cudaError_t
cudaD3D9RegisterVertexBuffer(IDirect3DVertexBuffer9* VB);
registers the vertex buffer VB for access by CUDA.
D.9.4
cudaD3D9MapVertexBuffer()
cudaError_t cudaD3D9MapVertexBuffer(void** devPtr,
IDirect3DVertexBuffer9* VB);
maps the vertex buffer VB into the address space of the current CUDA context and
returns in *devPtr the base pointer of the resulting mapping.
CUDA Programming Guide Version 1.1
101
Appendix D.
Runtime API Reference
D.9.5
cudaD3D9UnmapVertexBuffer()
cudaError_t cudaD3D9UnmapVertexBuffer(IDirect3DVertexBuffer9* VB);
unmaps the vertex buffer VB for access by CUDA.
D.9.6
cudaD3D9UnregisterVertexBuffer()
cudaError_t
cudaD3D9UnregisterVertexBuffer(IDirect3DVertexBuffer9* VB);
unregisters the vertex buffer VB for access by CUDA.
D.9.7
cudaD3D9GetDevice()
cudaError_t
cudaD3D9GetDevice(int* dev, const char* adapterName);
returns in *dev the device corresponding to the adapter name adapterName
obtained from EnumDisplayDevices or
IDirect3D9::GetAdapterIdentifier().
D.10
Error Handling
D.10.1
cudaGetLastError()
cudaError_t cudaGetLastError(void);
returns the last error that was returned from any of the runtime calls in the same
host thread and resets it to cudaSuccess.
D.10.2
cudaGetErrorString()
const char* cudaGetErrorString(cudaError_t error);
returns a message string from an error code.
102
CUDA Programming Guide Version 1.1
Appendix E.
Driver API Reference
E.1
Initialization
E.1.1
cuInit()
CUresult cuInit(unsigned int Flags);
initializes the driver API and must be called before any other function from the
driver API. Currently, the Flags parameters must be 0. If cuInit() has not been
called, any function from the driver API will return
CUDA_ERROR_NOT_INITIALIZED.
E.2
Device Management
E.2.1
cuDeviceGetCount()
CUresult cuDeviceGetCount(int* count);
returns in *count the number of devices with compute capability greater or equal
to 1.0 that are available for execution. If there is no such device,
cuDeviceGetCount() returns 1 and device 0 only supports device emulation
mode and is of compute capability less than 1.0.
E.2.2
cuDeviceGet()
CUresult cuDeviceGet(CUdevice* dev, int ordinal);
returns in *dev a device handle given an ordinal in the range
[0, cuDeviceGetCount()-1].
E.2.3
cuDeviceGetName()
CUresult cuDeviceGetName(char* name, int len, CUdevice dev);
CUDA Programming Guide Version 1.1
103
Appendix E.
Driver API Reference
returns an ASCII string identifying the device dev in the NULL-terminated string
pointed to by name. len specifies the maximum length of the string that may be
returned.
E.2.4
cuDeviceTotalMem()
CUresult cuDeviceTotalMem(unsigned int* bytes, CUdevice dev);
returns in *bytes the total amount of memory available on the device dev in
bytes.
E.2.5
cuDeviceComputeCapability()
CUresult cuDeviceComputeCapability(int* major, int* minor,
CUdevice dev);
returns in *major and *minor the major and minor revision numbers that define
the compute capability of device dev.
E.2.6
cuDeviceGetAttribute()
CUresult cuDeviceGetAttribute(int* value,
CUdevice_attribute attrib,
CUdevice dev);
returns in *value the integer value of the attribute attrib on device dev. The
supported attributes are:
CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_BLOCK:
maximum number of
threads per block;
CU_DEVICE_ ATTRIBUTE_MAX_BLOCK_DIM_X:
maximum x-dimension of a
block;
CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_Y:
maximum y-dimension of a
block;
CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_Z:
maximum z-dimension of a
block;
CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_X:
maximum x-dimension of a
grid;
CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_Y:
maximum y-dimension of a
grid;
CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_Z:
maximum z-dimension of a
grid;
CU_DEVICE_ATTRIBUTE_SHARED_MEMORY_PER_BLOCK:
total amount of
shared memory available per block in bytes;
CU_DEVICE_ATTRIBUTE_TOTAL_CONSTANT_MEMORY:
total amount of
constant memory available on the device in bytes;
CU_DEVICE_ATTRIBUTE_WARP_SIZE: warp size;
CU_DEVICE_ATTRIBUTE_MAX_PITCH: maximum pitch allowed by the
memory copy functions of Section E.8 that involve memory regions allocated
through cuMemAllocPitch() (Section E.8.3);
104
CUDA Programming Guide Version 1.1
Appendix E.
CU_DEVICE_ATTRIBUTE_REGISTERS_PER_BLOCK:
Driver API Reference
total number of registers
available per block;
CU_DEVICE_ATTRIBUTE_CLOCK_RATE:
clock frequency in kilohertz;
CU_DEVICE_ATTRIBUTE_TEXTURE_ALIGNMENT: alignment requirement
mentioned in Section 4.3.4.3; texture base addresses that are aligned to
textureAlign bytes do not need an offset applied to texture fetches;
CU_DEVICE_ATTRIBUTE_GPU_OVERLAP: 1 if the device can concurrently
copy memory between host and device while executing a kernel, or 0 if not.
E.2.7
cuDeviceGetProperties()
cudaError_t cuGetDeviceProperties(CUdevprop* prop,
CUdevice dev);
returns in *prop the properties of device dev. The CUdevprop structure is
defined as:
typedef struct CUdevprop_st {
int maxThreadsPerBlock;
int maxThreadsDim[3];
int maxGridSize[3];
int sharedMemPerBlock;
int totalConstantMemory;
int SIMDWidth;
int memPitch;
int regsPerBlock;
int clockRate;
int textureAlign;
} CUdevprop;
where:
is the maximum number of threads per block;
maxThreadsDim[3] is the maximum sizes of each dimension of a block;
maxGridSize[3] is the maximum sizes of each dimension of a grid;
sharedMemPerBlock is the total amount of shared memory available per
block in bytes;
totalConstantMemory is the total amount of constant memory available on
the device in bytes;
SIMDWidth is the warp size;
memPitch is the maximum pitch allowed by the memory copy functions of
Section E.8 that involve memory regions allocated through
cuMemAllocPitch() (Section E.8.3);
regsPerBlock is the total number of registers available per block;
clockRate is the clock frequency in kilohertz;
textureAlign is the alignment requirement mentioned in Section 4.3.4.3;
texture base addresses that are aligned to textureAlign bytes do not need an
offset applied to texture fetches.
maxThreadsPerBlock
CUDA Programming Guide Version 1.1
105
Appendix E.
Driver API Reference
E.3
Context Management
E.3.1
cuCtxCreate()
CUresult cuCtxCreate(CUcontext* pCtx, unsigned int Flags, CUdevice
dev);
creates a new context for a device and associates it with the calling thread. Currently
the Flags parameter must be 0. The context is created with a usage count of 1 and
the caller of cuCtxCreate() must call cuCtxDetach() when done using the
context. This function fails if a context is already current to the thread.
E.3.2
cuCtxAttach()
CUresult cuCtxAttach(CUcontext* pCtx, unsigned int Flags);
increments the usage count of the context and passes back a context handle in
*pCtx that must be passed to cuCtxDetach() when the application is done with
the context. cuCtxAttach() fails if there is no context current to the thread.
Currently, the Flags parameter must be 0.
E.3.3
cuCtxDetach()
CUresult cuCtxDetach(CUcontext ctx);
decrements the usage count of the context, and destroys the context if the usage
count goes to 0. The context must be a handle that was passed back by
cuCtxCreate() or cuCtxAttach(), and must be current to the calling thread.
E.3.4
cuCtxGetDevice()
CUresult cuCtxGetDevice(CUdevice* device);
returns in *device the ordinal of the current context's device.
E.3.5
cuCtxSynchronize()
CUresult cuCtxSynchronize(void);
blocks until the device has completed all preceding requested tasks.
cuCtxSynchronize() returns an error if one of the preceding tasks failed.
E.4
Module Management
E.4.1
cuModuleLoad()
CUresult cuModuleLoad(CUmodule* mod, const char* filename);
106
CUDA Programming Guide Version 1.1
Appendix E.
Driver API Reference
takes a file name filename and loads the corresponding module mod into the
current context. The CUDA driver API does not attempt to lazily allocate the
resources needed by a module; if the memory for functions and data (constant and
global) needed by the module cannot be allocated, cuModuleLoad() fails. The file
should be a cubin file as output by nvcc (see Section 4.2.5).
E.4.2
cuModuleLoadData()
CUresult cuModuleLoadData(CUmodule* mod, const void* image);
takes a pointer image and loads the corresponding module mod into the current
context. The pointer may be obtained by mapping a cubin file, passing a cubin file as a
text string, or incorporating a cubin object into the executable resources and using
operation system calls such as Windows’ FindResource() to obtain the pointer.
E.4.3
cuModuleLoadFatBinary()
CUresult cuModuleLoadFatBinary(CUmodule* mod, const void* fatBin);
takes a pointer fatBin and loads the corresponding module mod into the current
context. The pointer represents a fat binary object, which is a collection of different
cubin files, all representing the same device code but compiled and optimized for
different architectures. There is currently no documented API for constructing and
using fat binary objects by programmers, and therefore this function is an internal
function in this version of CUDA. More information can be found in the nvcc
document.
E.4.4
cuModuleUnload()
CUresult cuModuleUnload(CUmodule mod);
unloads a module mod from the current context.
E.4.5
cuModuleGetFunction()
CUresult cuModuleGetFunction(CUfunction* func,
CUmodule mod, const char* funcname);
returns in *func the handle of the function of name funcname located in module
mod. If no function of that name exists, cuModuleGetFunction() returns
CUDA_ERROR_NOT_FOUND.
E.4.6
cuModuleGetGlobal()
CUresult cuModuleGetGlobal(CUdeviceptr* devPtr,
unsigned int* bytes,
CUmodule mod, const char* globalname);
returns in *devPtr and *bytes the base pointer and size of the global of name
globalname located in module mod. If no variable of that name exists,
cuModuleGetGlobal() returns CUDA_ERROR_NOT_FOUND. Both parameters
devPtr and bytes are optional. If one of them is null, it is ignored.
CUDA Programming Guide Version 1.1
107
Appendix E.
E.4.7
Driver API Reference
cuModuleGetTexRef()
CUresult cuModuleGetTexRef(CUtexref* texRef,
CUmodule hmod, const char* texrefname);
returns in *texref the handle of the texture reference of name texrefname in
the module mod. If no texture reference of that name exists,
cuModuleGetTexRef() returns CUDA_ERROR_NOT_FOUND. This texture
reference handle should not be destroyed, since it will be destroyed when the
module is unloaded.
E.5
Stream Management
E.5.1
cuStreamCreate()
CUresult cuStreamCreate(CUstream* stream, unsigned int flags);
creates a stream. At present, flags is required to be 0.
E.5.2
cuStreamQuery()
CUresult cuStreamQuery(CUstream stream);
returns CUDA_SUCCESS if all operations in the stream have completed, or
CUDA_ERROR_NOT_READY if not.
E.5.3
cuStreamSynchronize()
CUresult cuStreamSynchronize(CUstream stream);
blocks until the device has completed all operations in the stream.
E.5.4
cuStreamDestroy()
CUresult cuStreamDestroy(CUstream stream);
destroys the stream.
E.6
Event Management
E.6.1
cuEventCreate()
CUresult cuEventCreate(CUevent* event, unsigned int flags);
creates an event. At present, flags is required to be 0.
E.6.2
cuEventRecord()
CUresult cuEventRecord(CUevent event, CUstream stream);
108
CUDA Programming Guide Version 1.1
Appendix E.
Driver API Reference
records an event. If stream is non-zero, the event is recorded after all preceding
operations in the stream have been completed; otherwise, it is recorded after all
preceding operations in the CUDA context have been completed. Since this
operation is asynchronous, cuEventQuery() and/or cuEventSynchronize()
must be used to determine when the event has actually been recorded.
If cuEventRecord() has previously been called and the event has not been
recorded yet, this function returns CUDA_ERROR_INVALID_VALUE.
E.6.3
cuEventQuery()
CUresult cuEventQuery(CUevent event);
returns CUDA_SUCCESS if the event has actually been recorded, or
CUDA_ERROR_NOT_READY if not. If cuEventRecord() has not been called on
this event, the function returns CUDA_ERROR_INVALID_VALUE.
E.6.4
cuEventSynchronize()
CUresult cuEventSynchronize(CUevent event);
blocks until the event has actually been recorded. If cuEventRecord() has not
been called on this event, the function returns CUDA_ERROR_INVALID_VALUE.
E.6.5
cuEventDestroy()
CUresult cuEventDestroy(CUevent event);
destroys the event.
E.6.6
cuEventElapsedTime()
CUresult cuEventDestroy(float* time,
CUevent start, CUevent end);
computes the elapsed time between two events (in milliseconds with a resolution of
around 0.5 microseconds). If either event has not been recorded yet, this function
returns CUDA_ERROR_INVALID_VALUE. If either event has been recorded with a
non-zero stream, the result is undefined.
E.7
Execution Control
E.7.1
cuFuncSetBlockShape()
CUresult cuFuncSetBlockShape(CUfunction func,
int x, int y, int z);
specifies the X, Y and Z dimensions of the thread blocks that are created when the
kernel given by func is launched.
CUDA Programming Guide Version 1.1
109
Appendix E.
E.7.2
Driver API Reference
cuFuncSetSharedSize()
CUresult cuFuncSetSharedSize(CUfunction func, unsigned int bytes);
sets through bytes the amount of shared memory that will be available to each
thread block when the kernel given by func is launched.
E.7.3
cuParamSetSize()
CUresult cuParamSetSize(CUfunction func, unsigned int numbytes);
sets through numbytes the total size in bytes needed by the function parameters of
function func.
E.7.4
cuParamSeti()
CUresult cuParamSeti(CUfunction func,
int offset, unsigned int value);
sets an integer parameter that will be specified the next time the kernel
corresponding to func will be invoked. offset is a byte offset.
E.7.5
cuParamSetf()
CUresult cuParamSetf(CUfunction func,
int offset, float value);
sets a floating point parameter that will be specified the next time the kernel
corresponding to func will be invoked. offset is a byte offset.
E.7.6
cuParamSetv()
CUresult cuParamSetv(CUfunction func,
int offset, void* ptr,
unsigned int numbytes);
copies an arbitrary amount of data into the parameter space of the kernel
corresponding to func. offset is a byte offset.
E.7.7
cuParamSetTexRef()
CUresult cuParamSetTexRef(CUfunction func,
int texunit, CUtexref texRef);
makes the CUDA array or linear memory bound to the texture reference texRef
available to a device program as a texture. In this version of CUDA, the texture
reference must be obtained via cuModuleGetTexRef() and the texunit
parameter must be set to CU_PARAM_TR_DEFAULT.
E.7.8
cuLaunch()
CUresult cuLaunch(CUfunction func);
110
CUDA Programming Guide Version 1.1
Appendix E.
Driver API Reference
invokes the kernel func on a 1×1 grid of blocks. The block contains the number of
threads specified by a previous call to cuFuncSetBlockShape().
E.7.9
cuLaunchGrid()
CUresult cuLaunchGrid(CUfunction func,
int grid_width, int grid_height);
CUresult cuLaunchGridAsync(CUfunction func,
int grid_width, int grid_height,
CUstream stream);
invokes the kernel on a grid_width × grid_height grid of blocks. Each
block contains the number of threads specified by a previous call to
cuFuncSetBlockShape().
cuLaunchGridAsync() can optionally be associated to a stream by passing a
non-zero stream argument. It only works on page-locked host memory and
returns an error if a pointer to pageable memory is passed as input.
E.8
Memory Management
E.8.1
cuMemGetInfo()
CUresult cuMemGetInfo(unsigned int* free, unsigned int* total);
returns in *free and *total respectively, the free and total amount of memory
available for allocation by the CUDA context, in bytes.
E.8.2
cuMemAlloc()
CUresult cuMemAlloc(CUdeviceptr* devPtr, unsigned int count);
allocates count bytes of linear memory on the device and returns in *devPtr a
pointer to the allocated memory. The allocated memory is suitably aligned for any
kind of variable. The memory is not cleared. If count is 0, cuMemAlloc() returns
CUDA_ERROR_INVALID_VALUE.
E.8.3
cuMemAllocPitch()
CUresult cuMemAllocPitch(CUdeviceptr* devPtr,
unsigned int* pitch,
unsigned int widthInBytes,
unsigned int height,
unsigned int elementSizeBytes);
allocates at least widthInBytes*height bytes of linear memory on the device
and returns in *devPtr a pointer to the allocated memory. The function may pad
the allocation to ensure that corresponding pointers in any given row will continue
to meet the alignment requirements for coalescing as the address is updated from
row to row (see Section 5.1.2.1). elementSizeBytes specifies the size of the
largest reads and writes that will be performed on the memory range.
CUDA Programming Guide Version 1.1
111
Appendix E.
Driver API Reference
elementSizeBytes may be 4, 8 or 16 (since coalesced memory transactions are
not possible on other data sizes). If elementSizeBytes is smaller than the actual
read/write size of a kernel, the kernel will run correctly, but possibly at reduced
speed. The pitch returned in *pitch by cuMemAllocPitch() is the width in
bytes of the allocation. The intended usage of pitch is as a separate parameter of the
allocation, used to compute addresses within the 2D array. Given the row and
column of an array element of type T, the address is computed as
T* pElement = (T*)((char*)BaseAddress + Row * Pitch) + Column;
The pitch returned by cuMemAllocPitch() is guaranteed to work with
cuMemcpy2D() under all circumstances. For allocations of 2D arrays, it is
recommended that programmers consider performing pitch allocations using
cuMemAllocPitch(). Due to alignment restrictions in the hardware, this is
especially true if the application will be performing 2D memory copies between
different regions of device memory (whether linear memory or CUDA arrays).
E.8.4
cuMemFree()
CUresult cuMemFree(CUdeviceptr devPtr);
frees the memory space pointed to by devPtr, which must have been returned by a
previous call to cuMemMalloc() or cuMemMallocPitch().
E.8.5
cuMemAllocHost()
CUresult cuMemAllocHost(void** hostPtr, unsigned int count);
allocates count bytes of host memory that is page-locked and accessible to the
device. The driver tracks the virtual memory ranges allocated with this function and
automatically accelerates calls to functions such as cuMemcpy(). Since the memory
can be accessed directly by the device, it can be read or written with much higher
bandwidth than pageable memory obtained with functions such as malloc().
Allocating excessive amounts of memory with cuMemAllocHost() may degrade
system performance, since it reduces the amount of memory available to the system
for paging. As a result, this function is best used sparingly to allocate staging areas
for data exchange between host and device.
E.8.6
cuMemFreeHost()
CUresult cuMemFreeHost(void* hostPtr);
frees the memory space pointed to by hostPtr, which must have been returned by
a previous call to cuMemAllocHost().
E.8.7
cuMemGetAddressRange()
CUresult cuMemGetAddressRange(CUdeviceptr* basePtr,
unsigned int* size,
CUdeviceptr devPtr);
returns the base address in *basePtr and size and *size of the allocation by
cuMemAlloc() or cuMemAllocPitch() that contains the input pointer
112
CUDA Programming Guide Version 1.1
Appendix E.
Driver API Reference
devPtr. Both parameters basePtr and size are optional. If one of them is null,
it is ignored.
E.8.8
cuArrayCreate()
CUresult cuArrayCreate(CUarray* array,
const CUDA_ARRAY_DESCRIPTOR* desc);
creates a CUDA array according to the CUDA_ARRAY_DESCRIPTOR structure
desc and returns a handle to the new CUDA array in *array. The
CUDA_ARRAY_DESCRIPTOR structure is defined as such:
typedef struct {
unsigned int Width;
unsigned int Height;
CUarray_format Format;
unsigned int NumChannels;
} CUDA_ARRAY_DESCRIPTOR;
where:
and Height are the width and height of the CUDA array (in elements);
the CUDA array is one-dimensional if height is 0, two-dimensional, otherwise;
NumChannels specifies the number of packed components per CUDA array
element.; it may be 1, 2 or 4;
Format specifies the format of the elements; CUarray_format is defined as
such:
Width
typedef enum CUarray_format_enum {
CU_AD_FORMAT_UNSIGNED_INT8 = 0x01,
CU_AD_FORMAT_UNSIGNED_INT16 = 0x02,
CU_AD_FORMAT_UNSIGNED_INT32 = 0x03,
CU_AD_FORMAT_SIGNED_INT8
= 0x08,
CU_AD_FORMAT_SIGNED_INT16
= 0x09,
CU_AD_FORMAT_SIGNED_INT32
= 0x0a,
CU_AD_FORMAT_HALF
= 0x10,
CU_AD_FORMAT_FLOAT
= 0x20
} CUarray_format;
Here are examples of CUDA array descriptions:
Description for a CUDA array of 2048 floats:
CUDA_ARRAY_DESCRIPTOR desc;
desc.Format = CU_AD_FORMAT_FLOAT;
desc.NumChannels = 1;
desc.Width = 2048;
desc.Height = 1;
Description for a 64×64 CUDA array of floats:
CUDA_ARRAY_DESCRIPTOR desc;
desc.Format = CU_AD_FORMAT_FLOAT;
desc.NumChannels = 1;
desc.Width = 64;
desc.Height = 64;
Description for a width×height CUDA array of 64-bit, 4x16-bit float16's:
CUDA_ARRAY_DESCRIPTOR desc;
desc.FormatFlags = CU_AD_FORMAT_HALF;
desc.NumChannels = 4;
CUDA Programming Guide Version 1.1
113
Appendix E.
Driver API Reference
desc.Width = width;
desc.Height = height;
Description for a width×height CUDA array of 16-bit elements, each of
which is two 8-bit unsigned chars:
CUDA_ARRAY_DESCRIPTOR arrayDesc;
desc.FormatFlags = CU_AD_FORMAT_UNSIGNED_INT8;
desc.NumChannels = 2;
desc.Width = width;
desc.Height = height;
E.8.9
cuArrayGetDescriptor()
CUresult cuArrayGetDescriptor(CUDA_ARRAY_DESCRIPTOR* arrayDesc,
CUarray array);
returns in *arrayDesc the descriptor that was used to create the CUDA array
array. It is useful for subroutines that have been passed a CUDA array, but need
to know the CUDA array parameters for validation or other purposes.
E.8.10
cuArrayDestroy()
CUresult cuArrayDestroy(CUarray array);
destroys the CUDA array array.
E.8.11
cuMemset()
CUresult cuMemsetD8(CUdeviceptr dstDevPtr,
unsigned char value, unsigned int count);
CUresult cuMemsetD16(CUdeviceptr dstDevPtr,
unsigned short value, unsigned int count);
CUresult cuMemsetD32(CUdeviceptr dstDevPtr,
unsigned int value, unsigned int count);
sets the memory range of count 8-, 16-, or 32-bit values to the specified value
value.
E.8.12
cuMemset2D()
CUresult cuMemsetD2D8(CUdeviceptr dstDevPtr,
unsigned int dstPitch,
unsigned char value,
unsigned int width, unsigned int height);
CUresult cuMemsetD2D16(CUdeviceptr dstDevPtr,
unsigned int dstPitch,
unsigned short value,
unsigned int width, unsigned int height);
CUresult cuMemsetD2D32(CUdeviceptr dstDevPtr,
unsigned int dstPitch,
unsigned int value,
unsigned int width, unsigned int height);
114
CUDA Programming Guide Version 1.1
Appendix E.
Driver API Reference
sets the 2D memory range of width 8-, 16-, or 32-bit values to the specified value
value. height specifies the number of rows to set, and dstPitch specifies the
number of bytes between each row (see Section E.8.3). These functions perform
fastest when the pitch is one that has been passed back by cuMemAllocPitch().
E.8.13
cuMemcpyHtoD()
CUresult cuMemcpyHtoD(CUdeviceptr dstDevPtr,
const void *srcHostPtr,
unsigned int count);
CUresult cuMemcpyHtoDAsync(CUdeviceptr dstDevPtr,
const void *srcHostPtr,
unsigned int count,
CUstream stream);
copies from host memory to device memory. dstDevPtr and srcHostPtr
specify the base addresses of the destination and source, respectively. count
specifies the number of bytes to copy.
cuMemcpyHtoDAsync() is asynchronous and can optionally be associated to a
stream by passing a non-zero stream argument. It only works on page-locked host
memory and returns an error if a pointer to pageable memory is passed as input.
E.8.14
cuMemcpyDtoH()
CUresult cuMemcpyDtoH(void* dstHostPtr,
CUdeviceptr srcDevPtr,
unsigned int count);
CUresult cuMemcpyDtoHAsync(void* dstHostPtr,
CUdeviceptr srcDevPtr,
unsigned int count,
CUstream stream);
copies from device to host memory. dstHostPtr and srcDevPtr specify the
base addresses of the source and destination, respectively. count specifies the
number of bytes to copy.
cuMemcpyDtoHAsync() is asynchronous and can optionally be associated to a
stream by passing a non-zero stream argument. It only works on page-locked host
memory and returns an error if a pointer to pageable memory is passed as input.
E.8.15
cuMemcpyDtoD()
CUresult cuMemcpyDtoD(CUdeviceptr dstDevPtr,
CUdeviceptr srcDevPtr,
unsigned int count);
copies from device memory to device memory. dstDevice and srcDevPtr are
the base pointers of the destination and source, respectively. count specifies the
number of bytes to copy.
CUDA Programming Guide Version 1.1
115
Appendix E.
Driver API Reference
E.8.16
cuMemcpyDtoA()
CUresult cuMemcpyDtoA(CUarray dstArray,
unsigned int dstIndex,
CUdeviceptr srcDevPtr,
unsigned int count);
copies from device memory to a 1D CUDA array. dstArray and dstIndex
specify the CUDA array handle and starting index of the destination data.
srcDevPtr specifies the base pointer of the source. count specifies the number
of bytes to copy.
E.8.17
cuMemcpyAtoD()
CUresult cuMemcpyAtoD(CUdeviceptr dstDevPtr,
CUarray srcArray,
unsigned int srcIndex,
unsigned int count);
copies from a 1D CUDA array to device memory. dstDevPtr specifies the base
pointer of the destination and must be naturally aligned with the CUDA array
elements. srcArray and srcIndex specify the CUDA array handle and the index
(in array elements) of the array element where the copy is to begin. count specifies
the number of bytes to copy and must be evenly divisible by the array element size.
E.8.18
cuMemcpyAtoH()
CUresult cuMemcpyAtoH(void* dstHostPtr,
CUarray srcArray,
unsigned int srcIndex,
unsigned int count);
CUresult cuMemcpyAtoHAsync(void* dstHostPtr,
CUarray srcArray,
unsigned int srcIndex,
unsigned int count,
CUstream stream);
copies from a 1D CUDA array to host memory. dstHostPtr specifies the base
pointer of the destination. srcArray and srcIndex specify the CUDA array
handle and starting index of the source data. count specifies the number of bytes
to copy.
cuMemcpyAtoHAsync() is asynchronous and can optionally be associated to a
stream by passing a non-zero stream argument. It only works on page-locked host
memory and returns an error if a pointer to pageable memory is passed as input.
E.8.19
cuMemcpyHtoA()
CUresult cuMemcpyHtoA(CUarray dstArray,
unsigned int dstIndex,
const void *srcHostPtr,
unsigned int count);
CUresult cuMemcpyHtoAAsync(CUarray dstArray,
116
CUDA Programming Guide Version 1.1
Appendix E.
Driver API Reference
unsigned int dstIndex,
const void *srcHostPtr,
unsigned int count,
CUstream stream);
copies from host memory to a 1D CUDA array. dstArray and dstIndex specify
the CUDA array handle and starting index of the destination data. srcHostPtr
specify the base addresse of the source. count specifies the number of bytes to
copy.
cuMemcpyHtoAAsync() is asynchronous and can optionally be associated to a
stream by passing a non-zero stream argument. It only works on page-locked host
memory and returns an error if a pointer to pageable memory is passed as input.
E.8.20
cuMemcpyAtoA()
CUresult cuMemcpyAtoA(CUarray dstArray,
unsigned int dstIndex,
CUarray srcArray,
unsigned int srcIndex,
unsigned int count);
copies from one 1D CUDA array to another. dstArray and srcArray specify
the handles of the destination and source CUDA arrays for the copy, respectively.
dstIndex and srcIndex specify the destination and source indices into the
CUDA array. These values are in the range [0, Width-1] for the CUDA array;
they are not byte offsets. count is the number of bytes to be copied. The size of
the elements in the CUDA arrays need not be the same format, but the elements
must be the same size; and count must be evenly divisible by that size.
E.8.21
cuMemcpy2D()
CUresult cuMemcpy2D(const CUDA_MEMCPY2D* copyParam);
CUresult cuMemcpy2DUnaligned(const CUDA_MEMCPY2D* copyParam);
CUresult cuMemcpy2DAsync(const CUDA_MEMCPY2D* copyParam,
CUstream stream);
perform a 2D memory copy according to the parameters specified in copyParam.
The CUDA_MEMCPY2D structure is defined as such:
typedef struct CUDA_MEMCPY2D_st {
unsigned int srcXInBytes, srcY;
CUmemorytype srcMemoryType;
const void *srcHost;
CUdeviceptr srcDevice;
CUarray srcArray;
unsigned int srcPitch;
unsigned int dstXInBytes, dstY;
CUmemorytype dstMemoryType;
void *dstHost;
CUdeviceptr dstDevice;
CUarray dstArray;
unsigned int dstPitch;
CUDA Programming Guide Version 1.1
117
Appendix E.
Driver API Reference
unsigned int WidthInBytes;
unsigned int Height;
} CUDA_MEMCPY2D;
where:
and dstMemoryType specify the type of memory of the
source and destination, respectively; Cumemorytype_enum is defined as such:
srcMemoryType
typedef enum CUmemorytype_enum {
CU_MEMORYTYPE_HOST = 0x01,
CU_MEMORYTYPE_DEVICE = 0x02,
CU_MEMORYTYPE_ARRAY = 0x03
} CUmemorytype;
If srcMemoryType is CU_MEMORYTYPE_HOST, srcHost and srcPitch
specify the (host) base address of the source data and the bytes per row to apply.
srcArray is ignored.
If srcMemoryType is CU_MEMORYTYPE_DEVICE, srcDevice and
srcPitch specify the (device) base address of the source data and the bytes per
row to apply. srcArray is ignored.
If srcMemoryType is CU_MEMORYTYPE_ARRAY, srcArray specifies the
handle of the source data. srcHost, srcDevice and srcPitch are ignored.
If dstMemoryType is CU_MEMORYTYPE_HOST, dstHost and dstPitch
specify the (host) base address of the destination data and the bytes per row to
apply. dstArray is ignored.
If dstMemoryType is CU_MEMORYTYPE_DEVICE, dstDevice and
dstPitch specify the (device) base address of the destination data and the
bytes per row to apply. dstArray is ignored.
If dstMemoryType is CU_MEMORYTYPE_ARRAY, dstArray specifies the
handle of the destination data. dstHost, dstDevice and dstPitch are
ignored.
srcXInBytes
and srcY specify the base address of the source data for the
copy.
For host pointers, the starting address is
void* Start =
(void*)((char*)srcHost+srcY*srcPitch + srcXInBytes);
For device pointers, the starting address is
CUdeviceptr Start = srcDevice+srcY*srcPitch+srcXInBytes;
For CUDA arrays, srcXInBytes must be evenly divisible by the array element
size.
and dstY specify the base address of the destination data for the
copy.
For host pointers, the base address is
dstXInBytes
void* dstStart =
(void*)((char*)dstHost+dstY*dstPitch + dstXInBytes);
For device pointers, the starting address is
CUdeviceptr dstStart = dstDevice+dstY*dstPitch+dstXInBytes;
118
CUDA Programming Guide Version 1.1
Appendix E.
Driver API Reference
For CUDA arrays, dstXInBytes must be evenly divisible by the array element
size.
and Height specify the width (in bytes) and height of the 2D
copy being performed. Any pitches must be greater than or equal to
WidthInBytes.
cuMemcpy2D() returns an error if any pitch is greater than the maximum allowed
(see CU_DEVICE_ATTRIBUTE_MAX_PITCH in Section E.2.6).
cuMemAllocPitch() passes back pitches that always work with cuMemcpy2D().
On intra-device memory copies (device ↔ device, CUDA array ↔ device, CUDA
array ↔ CUDA array), cuMemcpy2D() may fail for pitches not computed by
cuMemAllocPitch(). cuMemcpy2DUnaligned() does not have this
restriction, but may run significantly slower in the cases where cuMemcpy2D()
would have returned an error code.
WidthInBytes
cuMemcpy2DAsync() is asynchronous and can optionally be associated to a
stream by passing a non-zero stream argument. It only works on page-locked host
memory and returns an error if a pointer to pageable memory is passed as input.
E.9
Texture Reference Management
E.9.1
cuTexRefCreate()
CUresult cuTexRefCreate(CUtexref* texRef);
creates a texture reference and returns its handle in *texRef. Once created, the
application must call cuTexRefSetArray() or cuTexRefSetAddress() to
associate the reference with allocated memory. Other texture reference functions
are used to specify the format and interpretation (addressing, filtering, etc.) to be
used when the memory is read through this texture reference. To associate the
texture reference with a texture ordinal for a given function, the application should
call cuParamSetTexRef().
E.9.2
cuTexRefDestroy()
CUresult cuTexRefDestroy(CUtexref texRef);
destroys the texture reference.
E.9.3
cuTexRefSetArray()
CUresult cuTexRefSetArray(CUtexref texRef,
CUarray array,
unsigned int flags);
binds the CUDA array array to the texture reference texRef. Any previous
address or CUDA array state associated with the texture reference is superseded by
this function. flags must be set to CU_TRSA_OVERRIDE_FORMAT. Any CUDA
array previously bound to texRef is unbound.
CUDA Programming Guide Version 1.1
119
Appendix E.
E.9.4
Driver API Reference
cuTexRefSetAddress()
CUresult cuTexRefSetAddress(unsigned int* byteOffset,
CUtexref texRef,
CUdeviceptr devPtr,
int bytes);
binds a linear address range to the texture reference texRef. Any previous address
or CUDA array state associated with the texture reference is superseded by this
function. Any memory previously bound to texRef is unbound.
Since the hardware enforces an alignment requirement on texture base addresses,
cuTexRefSetAddress() passes back a byte offset in *byteOffset that must
be applied to texture fetches in order to read from the desired memory. This offset
must be divided by the texel size and passed to kernels that read from the texture so
they can be applied to the tex1Dfetch() function.
If the device memory pointer was returned from cuMemAlloc(), the offset is
guaranteed to be 0 and NULL may be passed as the ByteOffset parameter.
E.9.5
cuTexRefSetFormat()
CUresult cuTexRefSetFormat(CUtexref texRef,
CUarray_format format,
int numPackedComponents);
specifies the format of the data to be read by the texture reference texRef.
format and numPackedComponents are exactly analogous to the Format and
NumChannels members of the CUDA_ARRAY_DESCRIPTOR structure: They
specify the format of each component and the number of components per array
element.
E.9.6
cuTexRefSetAddressMode()
CUresult cuTexRefSetAddressMode(CUtexref texRef,
int dim, CUaddress_mode mode);
specifies the addressing mode mode for the given dimension of the texture
reference texRef. If dim is zero, the addressing mode is applied to the first
parameter of the functions used to fetch from the texture (see Section 4.4.5); if dim
is 1, the second, and so on. CUaddress_mode is defined as such:
typedef enum CUaddress_mode_enum {
CU_TR_ADDRESS_MODE_WRAP = 0,
CU_TR_ADDRESS_MODE_CLAMP = 1,
CU_TR_ADDRESS_MODE_MIRROR = 2,
} CUaddress_mode;
Note that this call has no effect if texRef is bound to linear memory.
E.9.7
cuTexRefSetFilterMode()
CUresult cuTexRefSetFilterMode(CUtexref texRef,
CUfilter_mode mode);
120
CUDA Programming Guide Version 1.1
Appendix E.
Driver API Reference
specifies the filtering mode mode to be used when reading memory through the
texture reference texRef. CUfilter_mode_enum is defined as such:
typedef enum CUfilter_mode_enum {
CU_TR_FILTER_MODE_POINT = 0,
CU_TR_FILTER_MODE_LINEAR = 1
} CUfilter_mode;
Note that this call has no effect if texRef is bound to linear memory.
E.9.8
cuTexRefSetFlags()
CUresult cuTexRefSetFlags(CUtexref texRef, unsigned int Flags);
specifies optional flags to control the behavior of data returned through the texture
reference. The valid flags are:
CU_TRSF_READ_AS_INTEGER,
which suppresses the default behavior of
having the texture promote integer data to floating point data in the range [0, 1];
CU_TRSF_NORMALIZED_COORDINATES, which suppresses the default
behavior of having the texture coordinates range from [0, Dim) where Dim is
the width or height of the CUDA array. Instead, the texture coordinates [0, 1.0)
reference the entire breadth of the array dimension.
E.9.9
cuTexRefGetAddress()
CUresult cuTexRefGetAddress(CUdeviceptr* devPtr, CUtexref texRef);
returns in *devPtr the base address bound to the texture reference texRef, or
returns CUDA_ERROR_INVALID_VALUE if the texture reference is not bound to
any device memory range.
E.9.10
cuTexRefGetArray()
CUresult cuTexRefGetArray(CUarray* array, CUtexref texRef);
returns in *array the CUDA array bound by the texture reference texRef, or
returns CUDA_ERROR_INVALID_VALUE if the texture reference is not bound to
any CUDA array.
E.9.11
cuTexRefGetAddressMode()
CUresult cuTexRefGetAddressMode(CUaddress_mode* mode,
CUtexref texRef,
int dim);
returns in *mode the addressing mode corresponding to the dimension dim of the
texture reference texRef. Currently the only valid values for dim are 0 and 1.
E.9.12
cuTexRefGetFilterMode()
CUresult cuTexRefGetFilterMode(CUfilter_mode* mode,
CUtexref texRef);
CUDA Programming Guide Version 1.1
121
Appendix E.
Driver API Reference
returns in *mode the filtering mode of the texture reference texRef.
E.9.13
cuTexRefGetFormat()
CUresult cuTexRefGetFormat(CUarray_format* format,
int* numPackedComponents,
CUtexref texRef);
returns in *format and *numPackedComponents the format and number of
components of the CUDA array bound to the texture reference texRef. If
format or numPackedComponents is null, it will be ignored.
E.9.14
cuTexRefGetFlags()
CUresult cuTexRefGetFlags(unsigned int* flags, CUtexref texRef);
returns in *flags the flags of the texture reference texRef.
E.10
OpenGL Interoperability
E.10.1
cuGLInit()
CUresult cuGLInit(void);
initializes OpenGL interoperability. It must be called before performing any other
OpenGL interoperability operations. It may fail if the needed OpenGL driver
facilities are not available.
E.10.2
cuGLRegisterBufferObject()
CUresult cuGLRegisterBufferObject(GLuint bufferObj);
registers the buffer object of ID bufferObj for access by CUDA. This function
must be called before CUDA can map the buffer object. While it is registered, the
buffer object cannot be used by any OpenGL commands except as a data source for
OpenGL drawing commands.
E.10.3
cuGLMapBufferObject()
CUresult cuGLMapBufferObject(CUdeviceptr* devPtr,
unsigned int* size,
GLuint bufferObj);
maps the buffer object of ID bufferObj into the address space of the current
CUDA context and returns in *devPtr and *size the base pointer and size of the
resulting mapping.
E.10.4
cuGLUnmapBufferObject()
CUresult cuGLUnmapBufferObject(GLuint bufferObj);
122
CUDA Programming Guide Version 1.1
Appendix E.
Driver API Reference
unmaps the buffer object of ID bufferObj for access by CUDA.
E.10.5
cuGLUnregisterBufferObject()
CUresult cuGLUnregisterBufferObject(GLuint bufferObj);
unregisters the buffer object of ID bufferObj for access by CUDA.
E.11
Direct3D Interoperability
E.11.1
cuD3D9Begin()
CUresult cuD3D9Begin(IDirect3DDevice9* device);
initializes interoperability with the Direct3D device device. This function must be
called before CUDA can map any objects from device. The application can then
map vertex buffers owned by the Direct3D device until cuD3D9End() is called.
E.11.2
cuD3D9End()
CUresult cuD3D9End(void);
concludes interoperability with the Direct3D device previously specified to
cuD3D9Begin().
E.11.3
cuD3D9RegisterVertexBuffer()
CUresult cuD3D9RegisterVertexBuffer(IDirect3DVertexBuffer9* VB);
registers the Direct3D vertex buffer VB for access by CUDA.
E.11.4
cuD3D9MapVertexBuffer()
CUresult cuD3D9MapVertexBuffer(CUdeviceptr* devPtr,
unsigned int* size,
IDirect3DVertexBuffer9* VB);
maps the Direct3D vertex buffer VB into the address space of the current CUDA
context and returns in *devPtr and *size the base pointer and size of the
resulting mapping.
E.11.5
cuD3D9UnmapVertexBuffer()
CUresult cuD3D9UnmapVertexBuffer(IDirect3DVertexBuffer9* VB);
unmaps the vertex buffer VB for access by CUDA.
E.11.6
cuD3D9UnregisterVertexBuffer()
CUresult cuD3D9UnregisterVertexBuffer(IDirect3DVertexBuffer9* VB);
CUDA Programming Guide Version 1.1
123
Appendix E.
Driver API Reference
unregisters the vertex buffer VB for access by CUDA.
E.11.7
cuD3D9GetDevice()
cudaError_t
cuD3D9GetDevice(CUdevice* dev, const char* adapterName);
returns in *dev the device corresponding to the adapter name adapterName
obtained from EnumDisplayDevices or
IDirect3D9::GetAdapterIdentifier().
124
CUDA Programming Guide Version 1.1
Appendix F.
Texture Fetching
This appendix gives the formula used to compute the value returned by the texture
functions of Section 4.4.5 depending on the various attributes of the texture
reference (see Section 4.3.4).
The texture bound to the texture reference is represented as an array T of N texels
for a one-dimensional texture or N × M texels for a two-dimensional texture. It is
fetched using texture coordinates x and y .
A texture coordinate must fall within T ’s valid addressing range before it can be
used to address T . The addressing mode specifies how an out-of-range texture
coordinate x is remapped to the valid range. If x is non-normalized, only the
clamp addressing mode is supported and x is replaced by 0 if x < 0 and N − 1 if
N ≤ x . If x is normalized:
In clamp addressing mode, x is replaced by 0 if x < 0 and 1 − 1 N if 1 ≤ x ,
In wrap addressing mode, x is replaced by frac(x) , where
frac( x) = x − floor ( x) and floor (x ) is the largest integer not greater than x .
In the remaining of the appendix, x and y are the non-normalized texture
coordinates remapped to T ’s valid addressing range. x and y are derived from the
normalized texture coordinates x̂ and ŷ as such: x = Nxˆ and y = Myˆ .
CUDA Programming Guide Version 1.1
125
Appendix F.
F.1
Texture Fetching
Nearest-Point Sampling
In this filtering mode, the value returned by the texture fetch is
tex ( x) = T [i ]
for a one-dimensional texture,
tex( x, y ) = T [i, j ]
for a two-dimensional texture,
where i = floor (x) and j = floor ( y ) .
Figure F-1 illustrates nearest-point sampling for a one-dimensional texture with
N =4.
For integer textures, the value returned by the texture fetch can be optionally
remapped to [0.0, 1.0] (see Section 4.3.4.1).
tex(x)
T[3]
T[0]
T[2]
T[1]
x
0
1
2
3
4
Non-Normalized
0
0.25
0.5
0.75
1
Normalized
Figure F-1.
126
Nearest-Point Sampling of a One-Dimensional
Texture of Four Texels
CUDA Programming Guide Version 1.1
Appendix F.
F.2
Texture Fetching
Linear Filtering
In this filtering mode, which is only available for floating-point textures, the value
returned by the texture fetch is
tex( x) = (1 − α )T [i ] + αT [i + 1]
for a one-dimensional texture,
tex( x, y ) = (1 − α )(1 − β )T [i, j ] + α (1 − β )T [i + 1, j ] + (1 − α ) β T [i, j + 1] + αβ T [i + 1, j + 1]
for a two-dimensional texture,
where:
i = floor ( x B ) , α = frac( x B ) , x B = x − 0.5 ,
j = floor ( y B ) , β = frac ( y B ) , y B = y − 0.5 .
α and β are stored in 9-bit fixed point format with 8 bits of fractional value.
Figure F-2 illustrates nearest-point sampling for a one-dimensional texture with
N =4.
tex(x)
T[3]
T[0]
T[2]
T[1]
x
0
1
2
3
4
Non-Normalized
0
0.25
0.5
0.75
1
Normalized
Figure F-2.
Linear Filtering of a One-Dimensional Texture of
Four Texels in Clamp Addressing Mode
CUDA Programming Guide Version 1.1
127
Appendix F.
F.3
Texture Fetching
Table Lookup
A table lookup TL(x) where x spans the interval [0, R] can be implemented as
TL( x) = tex(
N −1
x + 0.5) in order to ensure that TL(0) = T [0] and TL ( R ) = T [ N − 1] .
R
Figure F-3 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 .
TL(x)
T[3]
T[0]
T[2]
T[1]
x
0
4/3
8/3
4
0
1/3
2/3
1
Figure F-3.
128
One-Dimensional Table Lookup Using Linear
Filtering
CUDA Programming Guide Version 1.1
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 or otherwise under any
patent or patent rights of NVIDIA Corporation. Specifications mentioned in this publication are subject to
change without notice. This publication supersedes and replaces all information previously supplied. NVIDIA
Corporation products are not authorized for use as critical components in life support devices or systems
without express written approval of NVIDIA Corporation.
Trademarks
NVIDIA, the NVIDIA logo, GeForce, Tesla, and Quadro are trademarks or registered trademarks of NVIDIA
Corporation. Other company and product names may be trademarks of the respective companies with which
they are associated.
Copyright
© 2007 NVIDIA Corporation. All rights reserved.
NVIDIA Corporation
2701 San Tomas Expressway
Santa Clara, CA 95050
www.nvidia.com
Source Exif Data:
File Type : PDF File Type Extension : pdf MIME Type : application/pdf PDF Version : 1.4 Linearized : No Tagged PDF : Yes XMP Toolkit : 3.1-701 Producer : Acrobat Distiller 7.0 (Windows) Company : NVIDIA Corporation Source Modified : D:20071129104143 Creator Tool : Acrobat PDFMaker 7.0 for Word Modify Date : 2007:11:29 13:36:58+01:00 Create Date : 2007:11:29 12:33:37+01:00 Metadata Date : 2007:11:29 13:36:58+01:00 Document ID : uuid:67cfdf41-d2b2-499b-adc0-201690f57be6 Instance ID : uuid:253dd27b-f938-458d-b59e-cce98d64995e Version ID : 611 Format : application/pdf Title : NVIDIA CUDA Programming Guide Creator : czeller Subject : Headline : Page Count : 143 Page Layout : OneColumn Author : czellerEXIF Metadata provided by EXIF.tools