NVIDIA CUDA™

CUDA C Best Practices Guide

Version 3.2
8/20/2010
CUDA C Best Practices Guide Version 3.2 ii
 Table of Contents

Preface ............................................................................................................ vii

What Is This Document? .................................................................................................... vii
Who Should Read This Guide?............................................................................................ vii
Recommendations and Best Practices ................................................................................ viii
Contents Summary ........................................................................................................... viii
Chapter 1. Parallel Computing with CUDA ....................................................... 1
1.1 Heterogeneous Computing with CUDA .......................................................................1
1.1.1 Differences Between Host and Device ................................................................1
1.1.2 What Runs on a CUDA-Enabled Device? .............................................................2
1.1.3 Maximum Performance Benefit ..........................................................................3
1.2 Understanding the Programming Environment ...........................................................4
1.2.1 CUDA Compute Capability ..................................................................................4
1.2.2 Additional Hardware Data ..................................................................................5
1.2.3 C Runtime for CUDA and Driver API Version .......................................................5
1.2.4 Which Version to Target ....................................................................................6
1.3 CUDA APIs ...............................................................................................................6
1.3.1 C Runtime for CUDA ..........................................................................................7
1.3.2 CUDA Driver API ...............................................................................................7
1.3.3 When to Use Which API ....................................................................................7
1.3.4 Comparing Code for Different APIs .....................................................................8
Chapter 2. Performance Metrics .................................................................... 11
2.1 Timing ................................................................................................................... 11
2.1.1 Using CPU Timers ........................................................................................... 11
2.1.2 Using CUDA GPU Timers .................................................................................. 12
2.2 Bandwidth ............................................................................................................. 12
2.2.1 Theoretical Bandwidth Calculation .................................................................... 13
2.2.2 Effective Bandwidth Calculation ....................................................................... 13
2.2.3 Throughput Reported by cudaprof ................................................................... 14
Chapter 3. Memory Optimizations ................................................................. 15
3.1 Data Transfer Between Host and Device .................................................................. 15

iii CUDA C Best Practices Guide Version 3.2

 3.1.1 Pinned Memory ............................................................................................... 16
3.1.2 Asynchronous Transfers and Overlapping Transfers with Computation ............... 16
3.1.3 Zero Copy....................................................................................................... 18
3.2 Device Memory Spaces ........................................................................................... 19
3.2.1 Coalesced Access to Global Memory ................................................................. 20
3.2.1.1 A Simple Access Pattern............................................................................... 21
3.2.1.2 A Sequential but Misaligned Access Pattern ................................................... 22
3.2.1.3 Effects of Misaligned Accesses...................................................................... 23
3.2.1.4 Strided Accesses ......................................................................................... 25
3.2.2 Shared Memory .............................................................................................. 26
3.2.2.1 Shared Memory and Memory Banks .............................................................. 26
3.2.2.2 Shared Memory in Matrix Multiplication (C = AB) ........................................... 27
3.2.2.3 Shared Memory in Matrix Multiplication (C = AAT) ......................................... 31
3.2.2.4 Shared Memory Use by Kernel Arguments .................................................... 33
3.2.3 Local Memory ................................................................................................. 33
3.2.4 Texture Memory ............................................................................................. 33
3.2.4.1 Textured Fetch vs. Global Memory Read ....................................................... 34
3.2.4.2 Additional Texture Capabilities ..................................................................... 35
3.2.5 Constant Memory ............................................................................................ 36
3.2.6 Registers ........................................................................................................ 36
3.2.6.1 Register Pressure ........................................................................................ 36
3.3 Allocation ............................................................................................................... 37
Chapter 4. Execution Configuration Optimizations ........................................ 39
4.1 Occupancy ............................................................................................................. 39
4.2 Calculating Occupancy ............................................................................................ 39
4.3 Hiding Register Dependencies ................................................................................. 41
4.4 Thread and Block Heuristics .................................................................................... 41
4.5 Effects of Shared Memory ....................................................................................... 43
Chapter 5. Instruction Optimizations ............................................................ 45
5.1 Arithmetic Instructions ............................................................................................ 45
5.1.1 Division and Modulo Operations ....................................................................... 45
5.1.2 Reciprocal Square Root ................................................................................... 45
5.1.3 Other Arithmetic Instructions ........................................................................... 46
5.1.4 Math Libraries ................................................................................................. 46
5.2 Memory Instructions ............................................................................................... 47

CUDA C Best Practices Guide Version 3.2 iv

Chapter 6. Control Flow ................................................................................. 49
6.1 Branching and Divergence ...................................................................................... 49
6.2 Branch Predication ................................................................................................. 49
6.3 Loop counters signed vs. unsigned .......................................................................... 50
Chapter 7. Getting the Right Answer ............................................................. 51
7.1 Debugging ............................................................................................................. 51
7.2 Numerical Accuracy and Precision............................................................................ 51
7.2.1 Single vs. Double Precision .............................................................................. 51
7.2.2 Floating-Point Math Is Not Associative .............................................................. 52
7.2.3 Promotions to Doubles and Truncations to Floats .............................................. 52
7.2.4 IEEE 754 Compliance ...................................................................................... 52
7.2.5 x86 80-bit Computations ................................................................................. 53
Chapter 8. Multi-GPU Programming .............................................................. 55
8.1 Introduction to Multi-GPU ....................................................................................... 55
8.2 Multi-GPU Programming .......................................................................................... 55
8.3 Selecting a GPU...................................................................................................... 56
8.4 Inter-GPU communication ....................................................................................... 56
8.5 Compiling Multi-GPU Applications ............................................................................ 57
8.6 Infiniband .............................................................................................................. 57
Appendix A. Recommendations and Best Practices ....................................... 59
A.1 Overall Performance Optimization Strategies ............................................................ 59
A.2 High-Priority Recommendations............................................................................... 60
A.3 Medium-Priority Recommendations .......................................................................... 60
A.4 Low-Priority Recommendations ............................................................................... 60
Appendix B. Useful NVCC Compiler Switches ................................................. 62
B.1 NVCC ..................................................................................................................... 62
Appendix C. Revision History ......................................................................... 63
C.1 Version 3.0 ............................................................................................................ 63
C.2 Version 3.1 ............................................................................................................ 63
C.3 Version 3.2 ............................................................................................................ 64

v CUDA C Best Practices Guide Version 3.2

CUDA C Best Practices Guide Version 3.2 vi
 Preface

What Is This Document?

This Best Practices Guide is a manual to help developers obtain the best performance
from the NVIDIA® CUDA™ architecture using version 3.2 of the CUDA Toolkit.
It presents established optimization techniques and explains coding metaphors and
idioms that can greatly simplify programming for the CUDA architecture.
While the contents can be used as a reference manual, you should be aware that
some topics are revisited in different contexts as various programming and
configuration topics are explored. As a result, it is recommended that first-time
readers proceed through the guide sequentially. This approach will greatly improve
your understanding of effective programming practices and enable you to better use
the guide for reference later.

Who Should Read This Guide?

This guide is intended for programmers who have a basic familiarity with the
CUDA programming environment. You have already downloaded and installed the
CUDA Toolkit and have written successful programs using it. The discussions in
this guide all use the C programming language, so you should be comfortable
reading C.
This guide refers to and relies on several other documents that you should have at
your disposal for reference, all of which are available at no cost from the CUDA
website (http://www.nvidia.com/object/cuda_develop.html). The following
documents are especially important resources:
 CUDA Quickstart Guide
 CUDA C Programming Guide
 CUDA Reference Manual
Be sure to download the correct manual for the CUDA Toolkit version and
operating system you are using.

vii CUDA C Best Practices Guide Version 3.2

Recommendations and Best Practices
Throughout this guide, specific recommendations are made regarding the design
and implementation of CUDA C code. These recommendations are categorized by
priority, which is a blend of the effect of the recommendation and its scope. Actions
that present substantial improvements for most CUDA applications have the
highest priority, while small optimizations that affect only very specific situations are
given a lower priority.
Before implementing lower priority recommendations, it is good practice to make
sure all higher priority recommendations that are relevant have already been applied.
This approach will tend to provide the best results for the time invested and will
avoid the trap of premature optimization.
The criteria of benefit and scope for establishing priority will vary depending on the
nature of the program. In this guide, they represent a typical case. Your code might
reflect different priority factors. Regardless of this possibility, it is good practice to
verify that no higher-priority recommendations have been overlooked before
undertaking lower-priority items.
Appendix A of this document lists all the recommendations and best practices,
grouping them by priority and adding some additional helpful observations.
Code samples throughout the guide omit error checking for conciseness. Production
code should, however, systematically check the error code returned by each API call
and check for failures in kernel launches (or groups of kernel launches in the case
of concurrent kernels) by calling cudaGetLastError().

Contents Summary
The remainder of this guide is divided into the following sections:
 Parallel Computing with CUDA: Important aspects of the parallel
programming architecture.
 Performance Metrics: How should performance be measured in CUDA
applications and what are the factors that most influence performance?
 Memory Optimizations: Correct memory management is one of the most
effective means of improving performance. This chapter explores the different
kinds of memory available to CUDA applications.
 Execution Configuration Optimizations: How to make sure your CUDA
application is exploiting all the available resources on the GPU.
 Instruction Optimizations: Certain operations run faster than others. Using
faster operations and avoiding slower ones often confers remarkable benefits.
 Control Flow: Carelessly designed control flow can force parallel code into
serial execution; whereas thoughtfully designed control flow can help the
hardware perform the maximum amount of work per clock cycle.
 Getting the Right Answer: How to debug code and how to handle differences
in how the CPU and GPU represent floating-point values.

CUDA C Best Practices Guide Version 3.2 viii

 Chapter 1.
Parallel Computing with CUDA

This chapter reviews heterogeneous computing with CUDA, explains the limits of
performance improvement, and helps you choose the right version of CUDA and
which application programming interface (API) to use when programming.

1.1 Heterogeneous Computing with CUDA

CUDA C programming involves running code on two different platforms
concurrently: a host system with one or more CPUs and one or more devices
(frequently graphics adapter cards) with CUDA-enabled NVIDIA GPUs.
While NVIDIA devices are frequently associated with rendering graphics, they are
also powerful arithmetic engines capable of running thousands of lightweight
threads in parallel. This capability makes them well suited to computations that can
leverage parallel execution well.
However, the device is based on a distinctly different design from the host system,
and it’s important to understand those differences and how they determine the
performance of CUDA applications to use CUDA effectively.

1.1.1 Differences Between Host and Device

The primary differences occur in threading and memory access:
 Threading resources. Execution pipelines on host systems can support a
limited number of concurrent threads. Servers that have four quad-core
processors today can run only 16 threads concurrently (32 if the CPUs support
HyperThreading.) By comparison, the smallest executable unit of parallelism on a
CUDA device comprises 32 threads (a warp). All NVIDIA GPUs can support at
least 768 concurrently active threads per multiprocessor, and some GPUs
support 1,024 or more active threads per multiprocessor (see Section G.1 of the
CUDA C Programming Guide). On devices that have 30 multiprocessors (such as
the NVIDIA® GeForce® GTX 280), this leads to more than 30,000 active
threads.
 Threads. Threads on a CPU are generally heavyweight entities. The operating
system must swap threads on and off of CPU execution channels to provide
multithreading capability. Context switches (when two threads are swapped) are
therefore slow and expensive. By comparison, threads on GPUs are extremely

1 CUDA C Best Practices Guide Version 3.2

Chapter 1.
Parallel Computing with CUDA

lightweight. In a typical system, thousands of threads are queued up for work (in
warps of 32 threads each). If the GPU must wait on one warp of threads, it
simply begins executing work on another. Because separate registers are
allocated to all active threads, no swapping of registers or state need occur
between GPU threads. Resources stay allocated to each thread until it completes
its execution.
 RAM. Both the host system and the device have RAM. On the host system,
RAM is generally equally accessible to all code (within the limitations enforced
by the operating system). On the device, RAM is divided virtually and physically
into different types, each of which has a special purpose and fulfills different
needs. The types of device RAM are explained in the CUDA C Programming
Guide and in Chapter 3 of this document.
These are the primary hardware differences between CPU hosts and GPU devices
with respect to parallel programming. Other differences are discussed as they arise
elsewhere in this document.

1.1.2 What Runs on a CUDA-Enabled Device?

Because of the considerable differences between the host and the device, it’s
important to partition applications so that each hardware system is doing the work it
does best. The following issues should be considered when determining what parts
of an application to run on the device:
 The device is ideally suited for computations that can be run on numerous data
elements simultaneously in parallel. This typically involves arithmetic on large
data sets (such as matrices) where the same operation can be performed across
thousands, if not millions, of elements at the same time. This is a requirement
for good performance on CUDA: the software must use a large number of
threads. The support for running numerous threads in parallel derives from the
CUDA architecture’s use of a lightweight threading model.
 There should be some coherence in memory access by device code. Certain
memory access patterns enable the hardware to coalesce groups of reads or
writes of multiple data items into one operation. Data that cannot be laid out so
as to enable coalescing, or that doesn’t have enough locality to use textures or
L1 efficiently, will not enjoy much of a performance benefit when used in
computations on CUDA.
 To use CUDA, data values must be transferred from the host to the device
along the PCI Express (PCIe) bus. These transfers are costly in terms of
performance and should be minimized. (See Section 3.1.) This cost has several
ramifications:
 The complexity of operations should justify the cost of moving data to and
from the device. Code that transfers data for brief use by a small number of
threads will see little or no performance benefit. The ideal scenario is one in
which many threads perform a substantial amount of work.
For example, transferring two matrices to the device to perform a matrix
addition and then transferring the results back to the host will not realize
much performance benefit. The issue here is the number of operations
performed per data element transferred. For the preceding procedure,
assuming matrices of size N×N, there are N2 operations (additions) and 3N2

CUDA C Best Practices Guide Version 3.2 2

 Chapter 1.
Parallel Computing with CUDA

elements transferred, so the ratio of operations to elements transferred is 1:3

or O(1). Performance benefits can be more readily achieved when this ratio
is higher. For example, a matrix multiplication of the same matrices requires
N3 operations (multiply-add), so the ratio of operations to elements
transferred is O(N), in which case the larger the matrix the greater the
performance benefit. The types of operations are an additional factor, as
additions have different complexity profiles than, for example, trigonometric
functions. It is important to include the overhead of transferring data to and
from the device in determining whether operations should be performed on
the host or on the device.
 Data should be kept on the device as long as possible. Because transfers
should be minimized, programs that run multiple kernels on the same data
should favor leaving the data on the device between kernel calls, rather than
transferring intermediate results to the host and then sending them back to
the device for subsequent calculations. So, in the previous example, had the
two matrices to be added already been on the device as a result of some
previous calculation, or if the results of the addition would be used in some
subsequent calculation, the matrix addition should be performed locally on
the device. This approach should be used even if one of the steps in a
sequence of calculations could be performed faster on the host. Even a
relatively slow kernel may be advantageous if it avoids one or more PCIe
transfers. Section 3.1 provides further details, including the measurements of
bandwidth between the host and the device versus within the device proper.

1.1.3 Maximum Performance Benefit

High Priority: To get the maximum benefit from CUDA, focus first on finding ways to
parallelize sequential code.

The amount of performance benefit an application will realize by running on

CUDA depends entirely on the extent to which it can be parallelized. As mentioned
previously, code that cannot be sufficiently parallelized should run on the host,
unless doing so would result in excessive transfers between the host and the device.
Amdahl’s law specifies the maximum speed-up that can be expected by parallelizing
portions of a serial program. Essentially, it states that the maximum speed-up (S) of
a program is

where P is the fraction of the total serial execution time taken by the portion of code
that can be parallelized and N is the number of processors over which the parallel
portion of the code runs.
The larger N is (that is, the greater the number of processors), the smaller the P/N
fraction. It can be simpler to view N as a very large number, which essentially
transforms the equation into S = 1 / 1 P. Now, if ¾ of a program is parallelized,
the maximum speed-up over serial code is 1 / (1 – ¾) = 4.

3 CUDA C Best Practices Guide Version 3.2

Chapter 1.
Parallel Computing with CUDA

For most purposes, the key point is that the greater P is, the greater the speed-up.
An additional caveat is implicit in this equation, which is that if P is a small number
(so not substantially parallel), increasing N does little to improve performance. To
get the largest lift, best practices suggest spending most effort on increasing P; that
is, by maximizing the amount of code that can be parallelized.

1.2 Understanding the Programming

Environment
With each generation of NVIDIA processors, new features are added to the GPU
that CUDA can leverage. Consequently, it’s important to understand the
characteristics of the architecture.
Programmers should be aware of two version numbers. The first is the compute
capability, and the second is the version number of the runtime and driver APIs.

1.2.1 CUDA Compute Capability

The compute capability describes the features of the hardware and reflects the set of
instructions supported by the device as well as other specifications, such as the
maximum number of threads per block and the number of registers per
multiprocessor. Higher compute capability versions are supersets of lower (that is,
earlier) versions, and so they are backward compatible.
The compute capability of the GPU in the device can be queried programmatically
as illustrated in the CUDA SDK in the deviceQuery sample. The output for that
program is shown in Figure 1.1. This information is obtained by calling
cudaGetDeviceProperties() and accessing the information in the structure it
returns.

Figure 1.1 Sample CUDA configuration data reported by deviceQuery

CUDA C Best Practices Guide Version 3.2 4

 Chapter 1.
Parallel Computing with CUDA

The major and minor revision numbers of the compute capability are shown on the
third and fourth lines of Figure 1.1. Device 0 of this system has compute capability
1.1.
More details about the compute capabilities of various GPUs are in Appendix A of
the CUDA C Programming Guide. In particular, developers should note the number of
multiprocessors on the device, the number of registers and the amount of memory
available, and any special capabilities of the device.

1.2.2 Additional Hardware Data

Certain hardware features are not described by the compute capability. For example,
the ability to overlap kernel execution with asynchronous data transfers between the
host and the device is available on most but not all GPUs with compute capability
1.1. In such cases, call cudaGetDeviceProperties() to determine whether the
device is capable of a certain feature. For example, the deviceOverlap field of the
device property structure indicates whether overlapping kernel execution and data
transfers is possible (displayed in the ―Concurrent copy and execution‖ line of
Figure 1.1); likewise, the canMapHostMemory field indicates whether zero-copy data
transfers can be performed.

1.2.3 C Runtime for CUDA and Driver API Version

The CUDA driver API and the C runtime for CUDA are two of the programming
interfaces to CUDA. Their version number enables developers to check the features
associated with these APIs and decide whether an application requires a newer
(later) version than the one currently installed. This is important because the CUDA
driver API is backward compatible but not forward compatible, meaning that applications,
plug-ins, and libraries (including the C runtime for CUDA) compiled against a
particular version of the driver API will continue to work on subsequent (later)
driver releases. However, applications, plug-ins, and libraries (including the C
runtime for CUDA) compiled against a particular version of the driver API may not
work on earlier versions of the driver, as illustrated in Figure 1.2.

Apps, Apps, Apps,

Libs & Libs & Libs &
Plug-ins Plug-ins Plug-ins ...

1.0 1.1 2.0

Driver Driver Driver ...
Compatible Incompatible

Figure 1.2 Compatibility of CUDA versions

5 CUDA C Best Practices Guide Version 3.2

Chapter 1.
Parallel Computing with CUDA

1.2.4 Which Version to Target

When in doubt about the compute capability of the hardware that will be present at
runtime, it is best to assume a compute capability of 1.0 as defined in the CUDA C
Programming Guide, Section G.1.
To target specific versions of NVIDIA hardware and CUDA software, use the
–arch, -code, and –gencode options of nvcc. Code that contains double-precision
arithmetic, for example, must be compiled with ―-arch=sm_13‖ (or higher compute
capability), otherwise double-precision arithmetic will get demoted to single-
precision arithmetic (see Section 7.2.1). This and other compiler switches are
discussed further in Appendix B.

1.3 CUDA APIs

The host runtime component of the CUDA software environment can be used only
by host functions. It provides functions to handle the following:
 Device management
 Context management
 Memory management
 Code module management
 Execution control
 Texture reference management
 Interoperability with OpenGL and Direct3D
It comprises two APIs:
 A low-level API called the CUDA driver API
 A higher-level API called the C runtime for CUDA that is implemented on top
of the CUDA driver API
These APIs are mutually exclusive: An application should use one or the other.
The C runtime for CUDA, which is the more commonly used API, eases device
code management by providing implicit initialization, context management, and
module management. The C host code generated by nvcc is based on the C runtime
for CUDA, so applications that link to this code must use the C runtime for CUDA.
In contrast, the CUDA driver API requires more code and is somewhat harder to
program and debug, but it offers a better level of control. 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 (<<<…>>>). Note that
the APIs relate only to host code; the kernels that are executed on the device are the
same, regardless of which API is used.
The two APIs can be easily distinguished, because the CUDA driver API is
delivered through the nvcuda dynamic library and all its entry points are prefixed

CUDA C Best Practices Guide Version 3.2 6

 Chapter 1.
Parallel Computing with CUDA

with cu; while the C runtime for CUDA is delivered through the cudart dynamic
library and all its entry points are prefixed with cuda.

1.3.1 C Runtime for CUDA

The C runtime for CUDA handles kernel loading and setting up kernel parameters
and launch configuration before the kernel is launched. The implicit code
initialization, CUDA context management, CUDA module management (cubin to
function mapping), kernel configuration, and parameter passing are all performed by
the C runtime for CUDA.
It comprises two principal parts:
 The low-level functions (cuda_runtime_api.h) have a C-style interface.
 The high-level functions (cuda_runtime.h) have a C++-style interface built on
top of the low-level functions.
The functions that make up this API are explained in the CUDA Reference Manual.

1.3.2 CUDA Driver API

The driver API is a lower-level API than the runtime API. When compared with the
runtime API, the driver API has these advantages:
 No dependency on the runtime library
 More control over devices (for example, only the driver API enables one CPU
thread to control multiple GPUs; see Chapter 8)
 No C extensions in the host code, so the host code can be compiled with
compilers other than nvcc and the host compiler it calls by default
Its primary disadvantages, as mentioned in Section 1.3, are as follows:
 Verbose code
 Greater difficulty in debugging
A key point is that for every runtime API function, there is an equivalent driver API
function. The driver API, however, includes other functions missing in the runtime
API, such as those for migrating a context from one host thread to another.
For more information on the driver API, refer to Section 3.3 of the CUDA C
Programming Guide.

1.3.3 When to Use Which API

The previous section lists some of the salient differences between the two APIs.
Additional considerations include the following:
Driver API–only features:
 Context management
 Support for 16-bit floating-point textures

7 CUDA C Best Practices Guide Version 3.2

Chapter 1.
Parallel Computing with CUDA

 Access to the MCL image processing library

In most cases, these points tend to steer developers strongly toward one API. In
cases where they do not, favor the runtime API because it is higher level and easier
to use. In addition, because runtime functions all have driver API equivalents, it is
easy to migrate runtime code to the driver API should that later become necessary.

1.3.4 Comparing Code for Different APIs

To illustrate the difference in code between the runtime and driver APIs, compare
Listings 1.1 and 1.2, which are examples of a vector addition in which two arrays are
added.
const unsigned int cnBlockSize = 512;
const unsigned int cnBlocks = 3;
const unsigned int cnDimension = cnBlocks * cnBlockSize;

// create CUDA device & context

cudaSetDevice( 0 ); // pick first device

// allocate host vectors

float * pA = new float[cnDimension];
float * pB = new float[cnDimension];
float * pC = new float[cnDimension];

// initialize host memory

randomInit(pA, cnDimension);
randomInit(pB, cnDimension);

// allocate device memory

float *pDeviceMemA, *pDeviceMemB, *pDeviceMemC;
cudaMalloc(&pDeviceMemA, cnDimension * sizeof(float));
cudaMalloc(&pDeviceMemB, cnDimension * sizeof(float));
cudaMalloc(&pDeviceMemC, cnDimension * sizeof(float));

// copy host vectors to device

cudaMemcpy(pDeviceMemA, pA, cnDimension * sizeof(float),
cudaMemcpyHostToDevice);
cudaMemcpy(pDeviceMemB, pB, cnDimension * sizeof(float),
cudaMemcpyHostToDevice);

vectorAdd<<<cnBlocks, cnBlockSize>>> (pDeviceMemA, pDeviceMemB,

pDeviceMemC);

// copy result from device to host

cudaMemcpy ((void *) pC, pDeviceMemC, cnDimension * sizeof(float),
cudaMemcpyDeviceToHost);

delete[] pA;
delete[] pB;
delete[] pC;

cudaFree(pDeviceMemA);
cudaFree(pDeviceMemB);
cudaFree(pDeviceMemC);

Listing 1.1 Host code for adding two vectors using the C runtime for
CUDA

CUDA C Best Practices Guide Version 3.2 8

 Chapter 1.
Parallel Computing with CUDA

Listing 1.1 consists of 27 lines of code. Listing 1.2 shows the same functionality
implemented using the CUDA driver API.
const unsigned int cnBlockSize = 512;
const unsigned int cnBlocks = 3;
const unsigned int cnDimension = cnBlocks * cnBlockSize;

CUdevice hDevice;
CUcontext hContext;
CUmodule hModule;
CUfunction hFunction;

// create CUDA device & context

cuInit(0);
cuDeviceGet(&hContext, 0); // pick first device
cuCtxCreate(&hContext, 0, hDevice));

cuModuleLoad(&hModule, “vectorAdd.cubin”);
cuModuleGetFunction(&hFunction, hModule, "vectorAdd");

// allocate host vectors

float * pA = new float[cnDimension];
float * pB = new float[cnDimension];
float * pC = new float[cnDimension];

// initialize host memory

randomInit(pA, cnDimension);
randomInit(pB, cnDimension);

// allocate memory on the device

CUdeviceptr pDeviceMemA, pDeviceMemB, pDeviceMemC;
cuMemAlloc(&pDeviceMemA, cnDimension * sizeof(float));
cuMemAlloc(&pDeviceMemB, cnDimension * sizeof(float));
cuMemAlloc(&pDeviceMemC, cnDimension * sizeof(float));

// copy host vectors to device

cuMemcpyHtoD(pDeviceMemA, pA, cnDimension * sizeof(float));
cuMemcpyHtoD(pDeviceMemB, pB, cnDimension * sizeof(float));

// set up parameter values

cuFuncSetBlockShape(cuFunction, cnBlockSize, 1, 1);
#define ALIGN_UP(offset, alignment) \
(offset) = ((offset) + (alignment) – 1) & ~((alignment) – 1)
int offset = 0;
ALIGN_UP(offset, __alignof(pDeviceMemA));
cuParamSetv(cuFunction, offset, &ptr, sizeof(pDeviceMemA));
offset += sizeof(pDeviceMemA);
ALIGN_UP(offset, __alignof(pDeviceMemB));
cuParamSetv(cuFunction, offset, &ptr, sizeof(pDeviceMemB));
offset += sizeof(pDeviceMemB);
ALIGN_UP(offset, __alignof(pDeviceMemC));
cuParamSetv(cuFunction, offset, &ptr, sizeof(pDeviceMemC));
offset += sizeof(pDeviceMemC);
cuParamSetSize(cuFunction, offset);

// execute kernel
cuLaunchGrid(cuFunction, cnBlocks, 1);

// copy the result from device back to host

cuMemcpyDtoH((void *) pC, pDeviceMemC,
cnDimension * sizeof(float));

delete[] pA;

9 CUDA C Best Practices Guide Version 3.2

Chapter 1.
Parallel Computing with CUDA

delete[] pB;
delete[] pC;

cuMemFree(pDeviceMemA);
cuMemFree(pDeviceMemB);
cuMemFree(pDeviceMemC);

Listing 1.2 Host code for adding two vectors using the CUDA driver API

Listing 1.2 contains 50 lines of code and performs several lower-level operations
than the runtime API. These additional calls are evident in several places, especially
the setup necessary in the driver API prior to the kernel call.

CUDA C Best Practices Guide Version 3.2 10

 Chapter 2.
Performance Metrics

When attempting to optimize CUDA code, it pays to know how to measure

performance accurately and to understand the role that bandwidth plays in
performance measurement. This chapter discusses how to correctly measure
performance using CPU timers and CUDA events. It then explores how bandwidth
affects performance metrics and how to mitigate some of the challenges it poses.

2.1 Timing
CUDA calls and kernel executions can be timed using either CPU or GPU timers.
This section examines the functionality, advantages, and pitfalls of both approaches.

2.1.1 Using CPU Timers

Any CPU timer can be used to measure the elapsed time of a CUDA call or kernel
execution. The details of various CPU timing approaches are outside the scope of
this document, but developers should always be aware of the resolution their timing
calls provide.
When using CPU timers, it is critical to remember that many CUDA API functions
are asynchronous; that is, they return control back to the calling CPU thread prior to
completing their work. All kernel launches are asynchronous, as are memory-copy
functions with the Async suffix on their names. Therefore, to accurately measure
the elapsed time for a particular call or sequence of CUDA calls, it is necessary to
synchronize the CPU thread with the GPU by calling cudaThreadSynchronize()
immediately before starting and stopping the CPU timer.
cudaThreadSynchronize()blocks the calling CPU thread until all CUDA calls
previously issued by the thread are completed.
Although it is also possible to synchronize the CPU thread with a particular stream
or event on the GPU, these synchronization functions are not suitable for timing
code in streams other than the default stream. cudaStreamSynchronize() blocks
the CPU thread until all CUDA calls previously issued into the given stream have
completed. cudaEventSynchronize() blocks until a given event in a particular
stream has been recorded by the GPU. Because the driver may interleave execution
of CUDA calls from other non-default streams, calls in other streams may be
included in the timing.

11 CUDA C Best Practices Guide Version 3.2

Chapter 2.
Performance Metrics

Because the default stream, stream 0, exhibits synchronous behavior (an operation
in the default stream can begin only after all preceding calls in any stream have
completed; and no subsequent operation in any stream can begin until it finishes),
these functions can be used reliably for timing in the default stream.
Be aware that CPU-to-GPU synchronization points such as those mentioned in this
section imply a stall in the GPU’s processing pipeline and should thus be used
sparingly to minimize their performance impact.

2.1.2 Using CUDA GPU Timers

The CUDA event API provides calls that create and destroy events, record events
(via timestamp), and convert timestamp differences into a floating-point value in
milliseconds. Listing 2.1 illustrates their use.
cudaEvent_t start, stop;
float time;

cudaEventCreate(&start);
cudaEventCreate(&stop);

cudaEventRecord( start, 0 );
kernel<<<grid,threads>>> ( d_odata, d_idata, size_x, size_y,
NUM_REPS);
cudaEventRecord( stop, 0 );
cudaEventSynchronize( stop );

cudaEventElapsedTime( &time, start, stop );

cudaEventDestroy( start );
cudaEventDestroy( stop );

Listing 2.1 How to time code using CUDA events

Here cudaEventRecord() is used to place the start and stop events into the
default stream, stream 0. The device will record a timestamp for the event when it
reaches that event in the stream. The cudaEventElapsedTime() function returns
the time elapsed between the recording of the start and stop events. This value is
expressed in milliseconds and has a resolution of approximately half a microsecond.
Like the other calls in this listing, their specific operation, parameters, and return
values are described in the CUDA Reference Manual. Note that the timings are
measured on the GPU clock, so the timing resolution is operating-system-
independent.

2.2 Bandwidth
Bandwidth—the rate at which data can be transferred—is one of the most
important gating factors for performance. Almost all changes to code should be
made in the context of how they affect bandwidth. As described in Chapter 3 of this
guide, bandwidth can be dramatically affected by the choice of memory in which
data is stored, how the data is laid out and the order in which it is accessed, as well
as other factors.

CUDA C Best Practices Guide Version 3.2 12

 Chapter 2.
Performance Metrics

To measure performance accurately, it is useful to calculate theoretical and effective

bandwidth. When the latter is much lower than the former, design or
implementation details are likely to reduce bandwidth, and it should be the primary
goal of subsequent optimization efforts to increase it.

High Priority: Use the effective bandwidth of your computation as a metric when
measuring performance and optimization benefits.

2.2.1 Theoretical Bandwidth Calculation

Theoretical bandwidth can be calculated using hardware specifications available in
the product literature. For example, the NVIDIA GeForce GTX 280 uses DDR
(double data rate) RAM with a memory clock rate of 1,107 MHz and a 512-bit wide
memory interface.
Using these data items, the peak theoretical memory bandwidth of the NVIDIA
GeForce GTX 280 is 141.6 GB/sec:
( 1107 x 106 x ( 512/8 ) x 2 ) / 109 = 141.6 GB/sec
In this calculation, the memory clock rate is converted in to Hz, multiplied by the
interface width (divided by 8, to convert bits to bytes) and multiplied by 2 due to the
double data rate. Finally, this product is divided by 109 to convert the result to
GB/sec (GBps).
Note that some calculations use 1,0243 instead of 109 for the final calculation. In
such a case, the bandwidth would be 131.9 GBps. It is important to use the same
divisor when calculating theoretical and effective bandwidth so that the comparison
is valid.

2.2.2 Effective Bandwidth Calculation

Effective bandwidth is calculated by timing specific program activities and by
knowing how data is accessed by the program. To do so, use this equation:
Effective bandwidth = (( Br + Bw ) / 109 ) / time
Here, the effective bandwidth is in units of GBps, Br is the number of bytes read
per kernel, Bw is the number of bytes written per kernel, and time is given in
seconds.
For example, to compute the effective bandwidth of a 2048 x 2048 matrix copy, the
following formula could be used:
Effective bandwidth = (( 20482 x 4 x 2 ) / 109 ) / time
The number of elements is multiplied by the size of each element (4 bytes for a
float), multiplied by 2 (because of the read and write), divided by 109 (or 1,0243) to
obtain GB of memory transferred. This number is divided by the time in seconds to
obtain GBps.

13 CUDA C Best Practices Guide Version 3.2

Chapter 2.
Performance Metrics

2.2.3 Throughput Reported by cudaprof

The memory throughput reported in the summary table of cudaprof, the CUDA
Visual Profiler, differs from the effective bandwidth obtained by the calculation in
Section 2.2.2 in several respects.
The first difference is that cudaprof measures throughput using a subset of the
GPU’s multiprocessors and then extrapolates that number to the entire GPU, thus
reporting an estimate of the data throughput.
The second and more important difference is that because the minimum memory
transaction size is larger than most word sizes, the memory throughput reported by
the profiler includes the transfer of data not used by the kernel.
The effective bandwidth calculation in Section 2.2.2, however, includes only data
transfers that are relevant to the algorithm. As such, the effective bandwidth will be
smaller than the memory throughput reported by cudaprof and is the number to use
when optimizing memory performance.
However, it’s important to note that both numbers are useful. The profiler memory
throughput shows how close the code is to the hardware limit, and the comparison
of the effective bandwidth with the profiler number presents a good estimate of
how much bandwidth is wasted by suboptimal coalescing of memory accesses.

CUDA C Best Practices Guide Version 3.2 14

 Chapter 3.
Memory Optimizations

Memory optimizations are the most important area for performance. The goal is to
maximize the use of the hardware by maximizing bandwidth. Bandwidth is best
served by using as much fast memory and as little slow-access memory as possible.
This chapter discusses the various kinds of memory on the host and device and how
best to set up data items to use the memory effectively.

3.1 Data Transfer Between Host and Device

The peak bandwidth between the device memory and the GPU is much higher
(141 GBps on the NVIDIA GeForce GTX 280, for example) than the peak
bandwidth between host memory and device memory (8 GBps on the PCIe ×16
Gen2). Hence, for best overall application performance, it is important to minimize
data transfer between the host and the device, even if that means running kernels on
the GPU that do not demonstrate any speed-up compared with running them on
the host CPU.

High Priority: Minimize data transfer between the host and the device, even if it
means running some kernels on the device that do not show performance gains when
compared with running them on the host CPU.

Intermediate data structures should 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 one larger transfer performs significantly better than making each
transfer separately.
Finally, higher bandwidth between the host and the device is achieved when using
page-locked (or pinned) memory, as discussed in the CUDA C Programming Guide and
Section 3.1.1 of this document.

15 CUDA C Best Practices Guide Version 3.2

Chapter 3.
Memory Optimizations

3.1.1 Pinned Memory

Page-locked or pinned memory transfers attain the highest bandwidth between the
host and the device. On PCIe ×16 Gen2 cards, for example, pinned memory can
attain greater than 5 GBps transfer rates.
Pinned memory is allocated using the cudaMallocHost()or cudaHostAlloc()
functions in the runtime API. The bandwidthTest.cu program in the CUDA SDK
shows how to use these functions as well as how to measure memory transfer
performance.
Pinned memory should not be overused. Excessive use can reduce overall system
performance because pinned memory is a scarce resource. How much is too much
is difficult to tell in advance, so as with all optimizations, test the applications and
the systems they run on for optimal performance parameters.

3.1.2 Asynchronous Transfers and Overlapping

Transfers with Computation
Data transfers between the host and the device using cudaMemcpy() are blocking
transfers; that is, control is returned to the host thread only after the data transfer is
complete. The cudaMemcpyAsync() function is a non-blocking variant of
cudaMemcpy() in which control is returned immediately to the host thread. In
contrast with cudaMemcpy(), the asynchronous transfer version requires pinned host
memory (see Section 3.1.1), and it contains an additional argument, a stream ID. A
stream is simply a sequence of operations that are performed in order on the device.
Operations in different streams can be interleaved and in some cases overlapped—a
property that can be used to hide data transfers between the host and the device.
Asynchronous transfers enable overlap of data transfers with computation in two
different ways. On all CUDA-enabled devices, it is possible to overlap host
computation with asynchronous data transfers and with device computations. For
example, Listing 3.1 demonstrates how host computation in the routine
cpuFunction() is performed while data is transferred to the device and a kernel
using the device is executed.
cudaMemcpyAsync(a_d, a_h, size, cudaMemcpyHostToDevice, 0);
kernel<<<grid, block>>>(a_d);
cpuFunction();

Listing 3.1 Overlapping computation and data transfers

The last argument to the cudaMemcpyAsync() function is the stream ID, which in
this case uses the default stream, stream 0. The kernel also uses the default stream,
and it will not begin execution until the memory copy completes; therefore, no
explicit synchronization is needed. Because the memory copy and the kernel both
return control to the host immediately, the host function cpuFunction() overlaps
their execution.
In Listing 3.1, the memory copy and kernel execution occur sequentially. On devices
that are capable of ―concurrent copy and execute,‖ it is possible to overlap kernel
execution on the device with data transfers between the host and the device.
Whether a device has this capability is indicated by the deviceOverlap field of a

CUDA C Best Practices Guide Version 3.2 16

 Chapter 3.
Memory Optimizations

cudaDeviceProp variable (or listed in the output of the deviceQuery SDK sample).
On devices that have this capability, the overlap once again requires pinned host
memory, and, in addition, the data transfer and kernel must use different, non-
default streams (streams with non-zero stream IDs). Non-default streams are
required for this overlap because memory copy, memory set functions, and kernel
calls that use the default stream begin only after all preceding calls on the device (in
any stream) have completed, and no operation on the device (in any stream)
commences until they are finished.
Listing 3.2 illustrates the basic technique.
cudaStreamCreate(&stream1);
cudaStreamCreate(&stream2);
cudaMemcpyAsync(a_d, a_h, size, cudaMemcpyHostToDevice, stream1);
kernel<<<grid, block, 0, stream2>>>(otherData_d);

Listing 3.2 Concurrent copy and execute

In this code, two streams are created and used in the data transfer and kernel
executions as specified in the last arguments of the cudaMemcpyAsync call and the
kernel’s execution configuration.
Listing 3.2 demonstrates how to overlap kernel execution with asynchronous data
transfer. This technique could be used when the data dependency is such that the
data can be broken into chunks and transferred in multiple stages, launching
multiple kernels to operate on each chunk as it arrives. Listings 3.3a and 3.3b
demonstrate this. They produce equivalent results. The first segment shows the
reference sequential implementation, which transfers and operates on an array of N
floats (where N is assumed to be evenly divisible by nThreads).
cudaMemcpy(a_d, a_h, N*sizeof(float), dir);
kernel<<<N/nThreads, nThreads>>>(a_d);

Listing 3.3a Sequential copy and execute

Listing 3.3b shows how the transfer and kernel execution can be broken up into
nStreams stages. This approach permits some overlapping of the data transfer and
execution.
size=N*sizeof(float)/nStreams;
for (i=0; i<nStreams; i++) {
offset = i*N/nStreams;
cudaMemcpyAsync(a_d+offset, a_h+offset, size, dir, stream[i]);
}
for (i=0; i<nStreams; i++) {
offset = i*N/nStreams;
kernel<<<N/(nThreads*nStreams), nThreads,
0, stream[i]>>>(a_d+offset);
}

Listing 3.3b Staged concurrent copy and execute

(In Listing 3.3b, it is assumed that N is evenly divisible by nThreads*nStreams.)

Because execution within a stream occurs sequentially, none of the kernels will
launch until the data transfers in their respective streams complete. Current
hardware can simultaneously process an asynchronous data transfer and execute
kernels. (It should be mentioned that it is not possible to overlap a blocking transfer
with an asynchronous transfer, because the blocking transfer occurs in the default
stream, and so it will not begin until all previous CUDA calls complete. It will not

17 CUDA C Best Practices Guide Version 3.2

Chapter 3.
Memory Optimizations

allow any other CUDA call to begin until it has completed.) A diagram depicting the
timeline of execution for the two code segments is shown in Figure 3.1, and
nStreams=4 for Listing 3.3b is shown in the bottom half.

Figure 3.1 Comparison of timelines for sequential (top) and concurrent

(bottom) copy and kernel execution

For this example, it is assumed that the data transfer and kernel execution times are
comparable. In such cases, and when the execution time (tE) exceeds the transfer
time (tT), a rough estimate for the overall time is tE + tT/nStreams for the staged
version versus tE + tT for the sequential version. If the transfer time exceeds the
execution time, a rough estimate for the overall time is tT + tE/nStreams.

3.1.3 Zero Copy

Zero copy is a feature that was added in version 2.2 of the CUDA Toolkit. It enables
GPU threads to directly access host memory. For this purpose, it requires mapped
pinned (non-pageable) memory. On integrated GPUs, mapped pinned memory is
always a performance gain because it avoids superfluous copies as integrated GPU
and CPU memory are physically the same. On discrete GPUs, mapped pinned
memory is advantageous only in certain cases. Because the data is not cached on the
GPU on devices of compute capability 1.x, mapped pinned memory should be read
or written only once, and the global loads and stores that read and write the memory
should be coalesced. Zero copy can be used in place of streams because kernel-
originated data transfers automatically overlap kernel execution without the
overhead of setting up and determining the optimal number of streams.

Low Priority: On version 2.2 of the CUDA Toolkit (and later), use zero-copy operations
on integrated GPUs.

The host code in Listing 3.4 shows how zero copy is typically set up.
float *a_h, *a_map;
…
cudaGetDeviceProperties(&prop, 0);
if (!prop.canMapHostMemory)
exit(0);

CUDA C Best Practices Guide Version 3.2 18

 Chapter 3.
Memory Optimizations

cudaSetDeviceFlags(cudaDeviceMapHost);
cudaHostAlloc(&a_h, nBytes, cudaHostAllocMapped);
cudaHostGetDevicePointer(&a_map, a_h, 0);
kernel<<<gridSize, blockSize>>>(a_map);

Listing 3.4 Zero-copy host code

In this code, the canMapHostMemory field of the structure returned by

cudaGetDeviceProperties() is used to check that the device supports mapping
host memory to the device’s address space. Page-locked memory mapping is
enabled by calling cudaSetDeviceFlags() with cudaDeviceMapHost. Note that
cudaSetDeviceFlags() must be called prior to setting a device or making a
CUDA call that requires state (that is, essentially, before a context is created). Page-
locked mapped host memory is allocated using cudaHostAlloc(), and the pointer
to the mapped device address space is obtained via the function
cudaHostGetDevicePointer(). In the code in Listing 3.4, kernel() can
reference the mapped pinned host memory using the pointer a_map in exactly the
same was as it would if a_map referred to a location in device memory.

3.2 Device Memory Spaces

CUDA devices use several memory spaces, which have different characteristics that
reflect their distinct usages in CUDA applications. These memory spaces include
global, local, shared, texture, and registers, as shown in Figure 3.2.

To Host

Figure 3.2 The various memory spaces on a CUDA device

Of these different memory spaces, global and texture memory are the most
plentiful; see Section G.1 of the CUDA C Programming Guide for the amounts of
memory available in each memory space at each compute capability level. Global,
local, and texture memory have the greatest access latency, followed by constant
memory, registers, and shared memory.
The various principal traits of the memory types are shown in Table 3.1.

19 CUDA C Best Practices Guide Version 3.2

Chapter 3.
Memory Optimizations

Table 3.1 Salient features of device memory

Memory Location Cached Access Scope Lifetime

on/off chip

Register On n/a R/W 1 thread Thread

Local Off † R/W 1 thread Thread

Shared On n/a R/W All threads in block Block

Global Off † R/W All threads + host Host allocation

Constant Off Yes R All threads + host Host allocation

Texture Off Yes R All threads + host Host allocation

† Cached only on devices of compute capability 2.x.

In the case of texture access, if a texture reference is bound to a linear (and, as of

version 2.2 of the CUDA Toolkit, pitch-linear) array in global memory, then the
device code can write to the underlying array. Reading from a texture while writing
to its underlying global memory array in the same kernel launch should be avoided
because the texture caches are read-only and are not invalidated when the associated
global memory is modified.

3.2.1 Coalesced Access to Global Memory

High Priority: Ensure global memory accesses are coalesced whenever possible.

Perhaps the single most important performance consideration in programming for

the CUDA architecture is coalescing global memory accesses. Global memory loads
and stores by threads of a half warp (for devices of compute capability 1.x) or of a
warp (for devices of compute capability 2.x) are coalesced by the device into as few
as one transaction when certain access requirements are met.
To understand these access requirements, global memory should be viewed in terms
of aligned segments of 16 and 32 words. Figure 3.3 helps explain coalescing of a half
warp of 32-bit words, such as floats. It shows global memory as rows of 64-byte
aligned segments (16 floats). Two rows of the same color represent a 128-byte
aligned segment. A half warp of threads that accesses the global memory is indicated
at the bottom of the figure. Note that this figure assumes a device of compute
capability 1.x.

CUDA C Best Practices Guide Version 3.2 20

 Chapter 3.
Memory Optimizations

Figure 3.3 Linear memory segments and threads in a half warp

The access requirements for coalescing depend on the compute capability of the
device:
 On devices of compute capability 1.0 or 1.1, the k-th thread in a half warp must
access the k-th word in a segment aligned to 16 times the size of the elements
being accessed; however, not all threads need to participate.
 On devices of compute capability 1.2 or 1.3, coalescing is achieved for any
pattern of accesses that fits into a segment size of 32 bytes for 8-bit words,
64 bytes for 16-bit words, or 128 bytes for 32- and 64-bit words. Smaller
transactions may be issued to avoid wasting bandwidth. More precisely, the
following protocol is used to issue a memory transaction for a half warp:
 Find the memory segment that contains the address requested by the lowest
numbered active thread. Segment size is 32 bytes for 8-bit data, 64 bytes for
16-bit data, and 128 bytes for 32-, 64-, and 128-bit data.
 Find all other active threads whose requested address lies in the same
segment, and reduce the transaction size if possible:
 If the transaction is 128 bytes and only the lower or upper half is used,
reduce the transaction size to 64 bytes.
 If the transaction is 64 bytes and only the lower or upper half is used,
reduce the transaction size to 32 bytes.
 Carry out the transaction and mark the serviced threads as inactive.
 Repeat until all threads in the half warp are serviced.
 On devices of compute capability 2.x, memory accesses by the threads of a
warp are coalesced into the minimum number of L1-cache-line-sized aligned
transactions necessary to satisfy all threads; see Section G.4.2 of the CUDA C
Programming Guide.
These concepts are illustrated in the following simple examples.

3.2.1.1 A Simple Access Pattern

The first and simplest case of coalescing can be achieved by any CUDA-enabled
device: the k-th thread accesses the k-th word in a segment; the exception is that not
all threads need to participate. (See Figure 3.4. Note that this figure assumes a device
of compute capability 1.x, but that the figure would be much the same except twice
as wide for devices of compute capability 2.x.)

21 CUDA C Best Practices Guide Version 3.2

Chapter 3.
Memory Optimizations

Figure 3.4 Coalesced access in which all threads but one access the
corresponding word in a segment

This access pattern results in a single 64-byte transaction, indicated by the red
rectangle. Note that even though one word is not requested, all data in the segment
are fetched. If accesses by threads were permuted within this segment, still one 64-
byte transaction would be performed by a device with compute capability 1.2 or
higher, but 16 serialized transactions would be performed by a device with compute
capability 1.1 or lower.

3.2.1.2 A Sequential but Misaligned Access Pattern

If sequential threads in a half warp access memory that is sequential but not aligned
with the segments, then a separate transaction results for each element requested on
a device with compute capability 1.1 or lower. On a device with compute capability
1.2 or higher, several different scenarios can arise depending on whether all
addresses for a half warp fall within a single 128-byte segment. If the addresses fall
within a 128-byte segment, then a single 128-byte transaction is performed, as
shown in Figure 3.5. Again, this figure assumes a device of compute capability 1.x.

Figure 3.5 Unaligned sequential addresses that fit within a single 128-
byte segment

If a half warp accesses memory that is sequential but split across two 128-byte
segments, then two transactions are performed. In the following case, illustrated in
Figure 3.6, one 64-byte transaction and one 32-byte transaction result. Again, this
figure assumes a device of compute capability 1.x.

CUDA C Best Practices Guide Version 3.2 22

 Chapter 3.
Memory Optimizations

Figure 3.6 Misaligned sequential addresses that fall within two 128-byte
segments

Memory allocated through the runtime API, such as via cudaMalloc(), is

guaranteed to be aligned to at least 256 bytes. Therefore, choosing sensible thread
block sizes, such as multiples of 16, facilitates memory accesses by half warps that
are aligned to segments. In addition, the qualifiers __align__(8) and
__align__(16) can be used when defining structures to ensure alignment to
segments.

3.2.1.3 Effects of Misaligned Accesses

It is easy and informative to explore the ramifications of misaligned accesses using a
simple copy kernel, such as the one in Listing 3.5.
__global__ void offsetCopy(float *odata, float* idata, int offset)
{
int xid = blockIdx.x * blockDim.x + threadIdx.x + offset;
odata[xid] = idata[xid];
}

Listing 3.5 A copy kernel that illustrates misaligned accesses

In Listing 3.5, data is copied from the input array idata to the output array, both of
which exist in global memory. The kernel is executed within a loop in host code that
varies the parameter offset from 1 to 32. (Figures 3.5 and 3.6 correspond to
offsets of 1 and 17, respectively.) The effective bandwidth for the copy with various
offsets on an NVIDIA GeForce GTX 280 (with compute capability 1.3) and an
NVIDIA GeForce GTX 8800 (compute capability 1.0) are shown in Figure 3.7.

23 CUDA C Best Practices Guide Version 3.2

Chapter 3.
Memory Optimizations

Figure 3.7 Performance of offsetCopy kernel

For the NVIDIA GeForce GTX 8800 device, global memory accesses with no
offset or with offsets that are multiples of 16 result in a single transaction per half
warp and an effective bandwidth of approximately 74 GBps. Otherwise, 16
transactions are issued per half warp resulting in an effective bandwidth of
approximately 7 GBps. This roughly 8x performance degradation is due to the fact
that 32 bytes, the minimum transaction size, are fetched for each thread. However,
only 4 bytes of data are used for each 32 bytes fetched—resulting in the 4/32=1/8
performance relative to the fully coalesced case. The two numbers also reflect the
different data represented by effective bandwidth (4 bytes) versus actual bandwidth
(32 bytes).
Because of this possible performance degradation, memory coalescing is the most
critical aspect of performance optimization of device memory. For devices of
compute capability 1.2 and 1.3, the situation is less dire for misaligned accesses
because, in all cases, access by a half warp of threads in this kernel results in either
one or two transactions.
On the NVIDIA GeForce GTX 280 device, this results in an effective bandwidth
of between 120 GBps for a single transaction and 70 GBps for two transactions per
half warp. The number of transactions issued for a half warp of threads depends on
the offset and whether the warp is even- or odd-numbered. For offsets of 0 or 16,
each half warp results in a single 64-byte transaction (Figure 3.4). For offsets of 1
through 7 or 9 through 15, even-numbered warps result in a single 128-byte
transaction (Figure 3.5) and odd-numbered warps result in two transactions: one 64-
byte and one 32-byte (Figure 3.6). For offsets of 8, even-numbered warps result in
one 128-byte transaction and odd-numbered warps result in two 32-byte
transactions. The two 32-byte transactions, rather than a 64- and a 32-byte
transaction, are responsible for the blip at the offset of 8 in Figure 3.7.

CUDA C Best Practices Guide Version 3.2 24

 Chapter 3.
Memory Optimizations

3.2.1.4 Strided Accesses

Although the relaxed coalescing restrictions for devices with compute capability 1.2
or higher achieve one-half full bandwidth for the offset copy case just described,
performance on such devices can degrade when successive threads access memory
locations that have non-unit strides. This pattern occurs frequently when dealing
with multidimensional data or matrices; for example, when a half warp of threads
accesses matrix elements column-wise and the matrix is stored in row-major order.
To illustrate the effect of strided access on effective bandwidth, see the following
kernel strideCopy(), which copies data with a stride of stride elements between
threads from idata to odata.
__global__ void strideCopy(float *odata, float* idata, int stride)
{
int xid = (blockIdx.x*blockDim.x + threadIdx.x)*stride;
odata[xid] = idata[xid];
}

Listing 3.6 A kernel to illustrate non-unit stride data copy

Figure 3.8 illustrates a situation that can be created using the code in Listing 3.6;
namely, threads within a half warp access memory with a stride of 2. This action is
coalesced into a single 128-byte transaction on an NVIDIA GeForce GTX 280
(compute capability 1.3).

Figure 3.8 A half warp accessing memory with a stride of 2

Although a stride of 2 results in a single transaction, note that half the elements in
the transaction are not used and represent wasted bandwidth. As the stride
increases, the effective bandwidth decreases until the point where 16 transactions
are issued for the 16 threads in a half warp, as indicated in Figure 3.9.

25 CUDA C Best Practices Guide Version 3.2

Chapter 3.
Memory Optimizations

Figure 3.9 Performance of strideCopy kernel

Note, however, that on the NVIDIA GTX 8800 device (compute capability 1.0),
any non-unit stride results in 16 separate transactions per half warp.
As illustrated in Figure 3.9, non-unit stride global memory accesses should be
avoided whenever possible. One method for doing so utilizes shared memory,
which is discussed in the next section.

3.2.2 Shared Memory

Because it is on-chip, shared memory is much faster than local and global memory.
In fact, uncached shared memory latency is roughly 100x lower than global memory
latency—provided there are no bank conflicts between the threads, as detailed in the
following section.

3.2.2.1 Shared Memory and Memory Banks

To achieve high memory bandwidth for concurrent accesses, shared memory is
divided into equally sized memory modules (banks) that can be accessed
simultaneously. Therefore, any memory load or store of n addresses that spans n
distinct memory banks can be serviced simultaneously, yielding an effective
bandwidth that is n times as high as the bandwidth of a single bank.
However, if multiple addresses of a memory request map to the same memory bank,
the accesses are serialized. The hardware splits a memory request that has 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.
The one exception here is when all threads in a half warp address the same shared
memory location, resulting in a broadcast. Devices of compute capability 2.x have
the additional ability to multicast shared memory accesses.

CUDA C Best Practices Guide Version 3.2 26

 Chapter 3.
Memory Optimizations

To minimize bank conflicts, it is important to understand how memory addresses

map to memory banks and how to optimally schedule memory requests.

Medium Priority: Accesses to shared memory should be designed to avoid serializing

requests due to bank conflicts.

Shared memory banks are organized such that successive 32-bit words are assigned
to successive banks and each bank has a bandwidth of 32 bits per clock cycle. The
bandwidth of shared memory is 32 bits per bank per clock cycle.
For devices of compute capability 1.x, the warp size is 32 threads and the number of
banks is 16. 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. Note that no
bank conflict occurs if only one memory location per bank is accessed by a half
warp of threads.
For devices of compute capability 2.x, the warp size is 32 threads and the number of
banks is also 32. A shared memory request for a warp is not split as with devices of
compute capability 1.x, meaning that bank conflicts can occur between threads in the
first half of a warp and threads in the second half of the same warp (see Section
G.4.3 of the CUDA C Programming Guide).
Refer to the CUDA C Programming Guide for more information on how accesses and
banks can be matched to avoid conflicts.

3.2.2.2 Shared Memory in Matrix Multiplication (C = AB)

Shared memory enables cooperation between threads in a block. When multiple
threads in a block use the same data from global memory, shared memory can be
used to access the data from global memory only once. Shared memory can also be
used to avoid uncoalesced memory accesses by loading and storing data in a
coalesced pattern from global memory and then reordering it in shared memory.
Aside from memory bank conflicts, there is no penalty for non-sequential or
unaligned accesses by a half warp in shared memory.
The use of shared memory is illustrated via the simple example of a matrix
multiplication C = AB for the case with A of dimension M×16, B of dimension
16×N, and C of dimension M×N. To keep the kernels simple, M and N are
multiples of 16. A natural decomposition of the problem is to use a block and tile
size of 16×16 threads. Therefore, in terms of 16×16 tiles, A is a column matrix, B is
a row matrix, and C is their outer product. (See Figure 3.10.) A grid of N/16 by
M/16 blocks is launched, where each thread block calculates the elements of a
different tile in C from a single tile of A and a single tile of B.
Note that the example discussed throughout this section assumes compute
capability 1.x; the example would be much the same for compute capability 2.x,
except that it would have a width of 32 instead of 16.

27 CUDA C Best Practices Guide Version 3.2

Chapter 3.
Memory Optimizations

Figure 3.10 A block-column matrix (A) multiplied by a block-row matrix

(B) and the resulting product matrix (C)

To do this, the simpleMultiply kernel (Listing 3.7) calculates the output elements
of a tile of matrix C.
__global__ void simpleMultiply(float *a, float* b, float *c,
int N)
{
int row = blockIdx.y * blockDim.y + threadIdx.y;
int col = blockIdx.x * blockDim.x + threadIdx.x;
float sum = 0.0f;
for (int i = 0; i < TILE_DIM; i++) {
sum += a[row*TILE_DIM+i] * b[i*N+col];
}
c[row*N+col] = sum;
}

Listing 3.7 Unoptimized matrix multiplication

In Listing 3.7, a, b, and c are pointers to global memory for the matrices A, B, and
C, respectively; blockDim.x, blockDim.y, and TILE_DIM are all 16. Each thread in
the 16×16 block calculates one element in a tile of C. row and col are the row and
column of the element in C being calculated by a particular thread. The for loop
over i multiplies a row of A by a column of B, which is then written to C.
The effective bandwidth of this kernel is only 8.7 GBps on an NVIDIA GeForce
GTX 280 and 0.7 GBps on an NVIDIA GeForce GTX 8800. To analyze
performance, it is necessary to consider how half warps of threads access global
memory in the for loop. Each half warp of threads calculates one row of a tile of C,
which depends on a single row of A and an entire tile of B as illustrated in Figure
3.11.

CUDA C Best Practices Guide Version 3.2 28

 Chapter 3.
Memory Optimizations

Figure 3.11 Computing a row (half warp) of a tile in C using one row of A
and an entire tile of B

For each iteration i of the for loop, all threads in a half warp read the same value
from global memory (the index row*TILE_DIM+i is constant within a half warp),
resulting in 16 transactions for compute capability 1.1 or lower, and 1 transaction
for compute capability 1.2 or higher. Even though the operation requires only 1
transaction for compute capability 1.2 or higher, there is wasted bandwidth in the
transaction because only 4 bytes out of a 32-byte transaction are used. For each
iteration, the 16 threads in a half warp read a row of the B tile, which is a sequential
and coalesced access for all compute capabilities.
The performance on a device of any compute capability can be improved by reading
a tile of A into shared memory as shown in Listing 3.8.
__global__ void coalescedMultiply(float *a, float* b, float *c,
int N)
{
__shared__ float aTile[TILE_DIM][TILE_DIM];

int row = blockIdx.y * blockDim.y + threadIdx.y;

int col = blockIdx.x * blockDim.x + threadIdx.x;
float sum = 0.0f;
aTile[threadIdx.y][threadIdx.x] = a[row*TILE_DIM+threadIdx.x];
for (int i = 0; i < TILE_DIM; i++) {
sum += aTile[threadIdx.y][i]* b[i*N+col];
}
c[row*N+col] = sum;
}

Listing 3.8 Using shared memory to improve the global memory load
efficiency in matrix multiplication

In Listing 3.8, each element in a tile of A is read from global memory only once, in a
fully coalesced fashion (with no wasted bandwidth), to shared memory. Within each

29 CUDA C Best Practices Guide Version 3.2

Chapter 3.
Memory Optimizations

iteration of the for loop, a value in shared memory is broadcast to all threads in a
half warp.
In Listing 3.8, a __syncthreads()synchronization barrier call is not needed after
reading the tile of A into shared memory because only threads within the half warp
that write the data into shared memory read the data. This kernel has an effective
bandwidth of 14.3 GBps on an NVIDIA GeForce GTX 280, and 8.2 GBps on an
NVIDIA GeForce GTX 8800.
A further improvement can be made to how Listing 3.8 deals with matrix B. In
calculating a tile’s row of matrix C, the entire tile of B is read. The repeated reading
of the B tile can be eliminated by reading it into shared memory once (Listing 3.9).
__global__ void sharedABMultiply(float *a, float* b, float *c,
int N)
{
__shared__ float aTile[TILE_DIM][TILE_DIM],
bTile[TILE_DIM][TILE_DIM];
int row = blockIdx.y * blockDim.y + threadIdx.y;
int col = blockIdx.x * blockDim.x + threadIdx.x;
float sum = 0.0f;
aTile[threadIdx.y][threadIdx.x] = a[row*TILE_DIM+threadIdx.x];
bTile[threadIdx.y][threadIdx.x] = b[threadIdx.y*N+col];
__syncthreads();
for (int i = 0; i < TILE_DIM; i++) {
sum += aTile[threadIdx.y][i]* bTile[i][threadIdx.x];
}
c[row*N+col] = sum;
}

Listing 3.9 Improvement by reading additional data into shared memory

Note that in Listing 3.9, a __syncthreads() call is required after reading the B tile
because a warp reads data from shared memory that were written to shared memory
by different warps. The effective bandwidth of this routine is 29.7 GBps on an
NVIDIA GeForce GTX 280 and 15.7 GBps on an NVIDIA GeForce GTX 8800.
Note that the performance improvement is not due to improved coalescing in either
case, but to avoiding redundant transfers from global memory.
The results of the various optimizations are summarized in Table 3.2.
Table 3.2 Performance improvements optimizing C = AB matrix multiply

Optimization NVIDIA GeForce NVIDIA GeForce

GTX 280 GTX 8800

No optimization 8.7 GBps 0.7 GBps

Coalesced using shared

14.3 GBps 8.2 GBps
memory to store a tile of A

Using shared memory to

eliminate redundant reads 29.7 GBps 15.7 GBps
of a tile of B

CUDA C Best Practices Guide Version 3.2 30

 Chapter 3.
Memory Optimizations

Medium Priority: Use shared memory to avoid redundant transfers from global
memory.

3.2.2.3 Shared Memory in Matrix Multiplication (C = AAT)

A variant of the previous matrix multiplication can be used to illustrate how strided
accesses to global memory, as well as shared memory bank conflicts, are handled.
This variant simply uses the transpose of A rather than B, or C = AAT.
As in the previous section, this example assumes compute capability 1.x; it would
have a width of 32 rather than 16 for compute capability 2.x.
A simple implementation for C = AAT is shown in Listing 3.10.
__global__ void simpleMultiply(float *a, float *c, int M)
{
int row = blockIdx.y * blockDim.y + threadIdx.y;
int col = blockIdx.x * blockDim.x + threadIdx.x;
float sum = 0.0f;
for (int i = 0; i < TILE_DIM; i++) {
sum += a[row*TILE_DIM+i] * a[col*TILE_DIM+i];
}
c[row*M+col] = sum;
}

Listing 3.10 Unoptimized handling of strided accesses to global memory

In Listing 3.10, the row-th, col-th element of C is obtained by taking the dot product
of the row-th and col-th rows of A. The effective bandwidth for this kernel is
1.1 GBps on an NVIDIA GeForce GTX 280 and 0.5 GBps on an NVIDIA
GeForce GTX 8800. These results are substantially lower than the corresponding
measurements for the C = AB kernel. The difference is in how threads in a half
warp access elements of A in the second term, a[col*TILE_DIM+i], for each
iteration i. For a half warp of threads, col represents sequential columns of the
transpose of A, and therefore col*TILE_DIM represents a strided access of global
memory with a stride of 16. This results in uncoalesced memory accesses on devices
with compute capability 1.1 or lower and plenty of wasted bandwidth on devices
with compute capability 1.2 or higher. The way to avoid strided access is to use
shared memory as before, except in this case a half warp reads a row of A into a
column of a shared memory tile, as shown in Listing 3.11.
__global__ void coalescedMultiply(float *a, float *c, int M)
{
__shared__ float aTile[TILE_DIM][TILE_DIM],
transposedTile[TILE_DIM][TILE_DIM];
int row = blockIdx.y * blockDim.y + threadIdx.y;
int col = blockIdx.x * blockDim.x + threadIdx.x;
float sum = 0.0f;
aTile[threadIdx.y][threadIdx.x] = a[row*TILE_DIM+threadIdx.x];
transposedTile[threadIdx.x][threadIdx.y] =
a[(blockIdx.x*blockDim.x + threadIdx.y)*TILE_DIM +
threadIdx.x];
__syncthreads();

31 CUDA C Best Practices Guide Version 3.2

Chapter 3.
Memory Optimizations

for (int i = 0; i < TILE_DIM; i++) {

sum += aTile[threadIdx.y][i]* transposedTile[i][threadIdx.x];
}
c[row*M+col] = sum;
}

Listing 3.11 An optimized version of Listing 3.10 using coalesced reads

from global memory

Listing 3.11 uses the shared transposedTile to avoid uncoalesced accesses in the
second term in the dot product, and the shared aTile technique from the previous
example to avoid uncoalesced accesses in the first term. The effective bandwidth of
this kernel is 24.9 GBps on an NVIDIA GeForce GTX 280 and 13.2 GBps on an
NVIDIA GeForce GTX 8800. These results are slightly lower than those obtained
by the final kernel for C = AB. The cause of the difference is shared memory bank
conflicts.
The reads of elements in transposedTile within the for loop are free of conflicts,
because threads of each half warp read across rows of the tile, resulting in unit stride
across the banks. However, bank conflicts occur when copying the tile from global
memory into shared memory. To enable the loads from global memory to be
coalesced, data are read from global memory sequentially. However, this requires
writing to shared memory in columns, and because of the use of 16×16 tiles in
shared memory, this results in a stride between threads of 16 banks. These 16-way
bank conflicts are very expensive. The simple remedy is to pad the shared memory
array so that it has an extra column, as in the following line of code.
__shared__ float transposedTile[TILE_DIM][TILE_DIM+1];

This padding eliminates the conflicts entirely, because now the stride between
threads is 17 banks (33 banks for compute capability 2.x), which, due to modular
arithmetic used to compute bank indices, is equivalent to a unit stride. After this
change, the effective bandwidth is 30.4 GBps on an NVIDIA GeForce GTX 280
and 15.6 GBps on an NVIDIA GeForce GTX 8800, which is comparable to the
results from the last C = AB kernel.
The results of these optimizations are summarized in Table 3.3.
Table 3.3 Performance improvements optimizing C = AAT matrix
multiplication

Optimization NVIDIA GeForce NVIDIA GeForce

GTX 280 GTX 8800

No optimization 1.1 GBps 0.5 GBps

Using shared memory to

24.9 GBps 13.2 GBps
coalesce global reads

Removing bank conflicts 30.4 GBps 15.6 GBps

These results should be compared with those in Table 3.2. As can be seen from
these tables, judicious use of shared memory can dramatically improve performance.

CUDA C Best Practices Guide Version 3.2 32

 Chapter 3.
Memory Optimizations

The examples in this section have illustrated three ways to use shared memory:
 To enable coalesced accesses to global memory, especially to avoid large strides
(for general matrices, strides are much larger than 16)
 To eliminate (or reduce) redundant loads from global memory
 To avoid wasted bandwidth

3.2.2.4 Shared Memory Use by Kernel Arguments

Shared memory holds the parameters or arguments that are passed to kernels at
launch. In kernels with long argument lists, it can be valuable to put some
arguments into constant memory (and reference them there) rather than consume
shared memory.

Low Priority: For kernels with long argument lists, place some arguments into
constant memory to save shared memory.

3.2.3 Local Memory

Local memory is so named because its scope is local to the thread, not because of its
physical location. In fact, local memory is off-chip. Hence, access to local memory is
as expensive as access to global memory. Like global memory, local memory is not
cached on devices of compute capability 1.x. In other words, the term ―local‖ in the
name does not imply faster access.
Local memory is used only to hold automatic variables. This is done by the nvcc
compiler when it determines that there is insufficient register space to hold the
variable. Automatic variables that are likely to be placed in local memory are large
structures or arrays that would consume too much register space and arrays that the
compiler determines may be indexed dynamically.
Inspection of the PTX assembly code (obtained by compiling with -ptx or -keep
command-line options to nvcc) reveals whether a variable has been placed in local
memory during the first compilation phases. If it has, 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, if they find
the variable consumes too much register space for the targeted architecture. There is
no way to check this for a specific variable, but the compiler reports total local
memory usage per kernel (lmem) when run with the --ptxas-options=-v option.

3.2.4 Texture Memory

The read-only texture memory space is cached. Therefore, a texture fetch costs one
device memory read 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. Texture memory is also designed for streaming fetches with a constant
latency; that is, a cache hit reduces DRAM bandwidth demand, but not fetch
latency.

33 CUDA C Best Practices Guide Version 3.2

Chapter 3.
Memory Optimizations

In certain addressing situations, reading device memory through texture fetching can
be an advantageous alternative to reading device memory from global or constant
memory.

3.2.4.1 Textured Fetch vs. Global Memory Read

Device memory reads through texture fetching present several potential benefits
over reads from global memory:
 They are cached, unlike global memory reads in devices of compute capability
1.x. For these devices, this results in potentially exhibit higher bandwidth if
there is 2D locality in the texture fetches and can be used to avoid uncoalesced
loads from global memory. (Note that the L1 cache in devices of compute
capability 2.x has a higher bandwidth than the texture cache, obviating this
benefit in these devices.)
 Packed data can be unpacked into separate variables in a single operation.
 8-bit and 16-bit integer input data may be optionally converted to 32-bit
floating-point values in the range [0.0, 1.0] or [-1.0, 1.0].
Listings 3.12 and 3.13 illustrate how textures can be used to avoid uncoalesced
global memory accesses in the following variation of the offsetCopy kernel. This
copy performs a shift in data, as demonstrated in the following kernel.
__global__ void shiftCopy(float *odata, float *idata, int shift)
{
int xid = blockIdx.x * blockDim.x + threadIdx.x;
odata[xid] = idata[xid+shift];
}

Listing 3.12 Unoptimized data shifts

This copy kernel applies a shift to the global memory location when reading from
idata, but writes to unshifted global memory locations in odata. The amount of
shift is specified as a function argument to the kernel. Some degradation of
performance occurs when the shift is neither zero nor a multiple of 16 because
reading from idata will be either uncoalesced (compute capability 1.1 or lower) or
result in transactions with wasted bandwidth (compute capability 1.2 or higher).
Note that regardless of compute capability, writing to odata is fully coalesced.
The version of this code that uses textures to perform the shifted read is shown in
Listing 3.13.
__global__ void textureShiftCopy(float *odata, float *idata,
int shift)
{
int xid = blockIdx.x * blockDim.x + threadIdx.x;
odata[xid] = tex1Dfetch(texRef, xid+shift);
}

Listing 3.13 Data shifts optimized by use of texture memory

Here, the texture reference texRef is bound to the idata array in the host code and
the function tex1Dfetch() reads the shifted memory locations of idata via a
texture fetch. The results of both kernels (using global memory and textures for
loads) on an NVIDIA GeForce GTX 280 and an NVIDIA GeForce GTX 8800 are
given in Figure 3.12.

CUDA C Best Practices Guide Version 3.2 34

 Chapter 3.
Memory Optimizations

Figure 3.12 Results of using texture memory to avoid uncoalesced global

memory access

The benefit of using textures for cases that are not optimally coalesced is clear.
Textured reads can maintain effective bandwidth of the unshifted, fully coalesced
cases within a few percent. Note that shifts that are neither zero nor multiples of 16
show greater effective bandwidth than the offsetCopy kernel in Figure 3.7. Because
all the stores in the shift kernels are fully coalesced with no wasted bandwidth, the
shift applies only to the loads.

3.2.4.2 Additional Texture Capabilities

If textures are fetched using tex1D(), tex2D(), or tex3D() rather than
tex1Dfetch(), the hardware provides other capabilities that might be useful for
some applications, such as image processing. (See Table 3.4.)
Table 3.4 Useful features for tex1D(), tex2D(), and tex3D() fetches

Feature Use Caveat

Filtering Fast, low-precision Valid only if the texture

interpolation between texels reference returns floating-point
data

Normalized texture Resolution-independent coding

coordinates

Addressing modes Automatic handling of Can be used only with

boundary cases¹ normalized texture coordinates

¹The automatic handling of boundary cases in the bottom row of Table 3.4 refers to how a texture coordinate is
resolved when it falls outside the valid addressing range. There are two options: clamp and wrap. If x is the
coordinate and N is the number of texels for a one-dimensional texture, then with clamp, x is replaced by 0 if x < 0
and by 1-1/N if 1 ≤x. With wrap, x is replaced by frac(x) where frac(x) = x – floor(x). Floor returns the largest
integer less than or equal to x. So, in clamp mode where N = 1, an x of 1.3 is clamped to 1.0; whereas in wrap
mode, it is converted to 0.3

35 CUDA C Best Practices Guide Version 3.2

Chapter 3.
Memory Optimizations

Within a kernel call, the texture cache is not kept coherent with respect to global
memory writes, so texture fetches from addresses that have been written via global
stores in the same kernel call return undefined data. That is, a thread can safely read
a memory location via texture if the location has been updated by a previous kernel
call or memory copy, but not if it has been previously updated by the same thread or
another thread within the same kernel call. This is relevant only when fetching from
linear or pitch-linear memory because a kernel cannot write to CUDA arrays.

3.2.5 Constant Memory

There is a total of 64 KB constant memory on a device. The constant memory space
is cached. As a result, 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. Accesses to different
addresses by threads within a half warp are serialized, so cost scales linearly with the
number of different addresses read by all threads within a half warp.
Alternatively, on devices of compute capability 2.x, programs use the LoaD
Uniform (LDU) operation; see Section G.4.4 of the CUDA C Programming Guide for
details.

3.2.6 Registers
Generally, accessing a register consumes zero extra clock cycles per instruction, but
delays may occur due to register read-after-write dependencies and register memory
bank conflicts.
The latency of read-after-write dependencies is approximately 24 cycles, but this
latency is completely hidden on multiprocessors that have at least 192 active threads
(that is, 6 warps) for devices of compute capability 1.x (8 CUDA cores per
multiprocessor * 24 cycles of latency = 192 active threads to cover that latency). For
devices of compute capability 2.0, which have 32 CUDA cores per multiprocessor,
as many as 768 threads might be required to completely hide latency.
The compiler and hardware thread scheduler will schedule instructions as optimally
as possible to avoid register memory bank conflicts. They achieve the 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 register-related reason to pack data into float4 or int4 types.

3.2.6.1 Register Pressure

Register pressure occurs when there are not enough registers available for a given
task. Even though each multiprocessor contains thousands of 32-bit registers (see
Section G.1 of the CUDA C Programming Guide), these are partitioned among
concurrent threads. To prevent the compiler from allocating too many registers, use
the –maxrregcount=N compiler command-line option (see Section B.1 below) or
the launch bounds kernel definition qualifier (see Section B.17 of the CUDA C

CUDA C Best Practices Guide Version 3.2 36

 Chapter 3.
Memory Optimizations

Programming Guide) to control the maximum number of registers to allocated per

thread.

3.3 Allocation
Device memory allocation and de-allocation via cudaMalloc() and cudaFree() (or
their Driver API equivalents) are expensive operations, so device memory should be
reused and/or sub-allocated by the application wherever possible to minimize the
impact of allocations on overall performance.

37 CUDA C Best Practices Guide Version 3.2

Chapter 3.
Memory Optimizations

CUDA C Best Practices Guide Version 3.2 38

 Chapter 4.
Execution Configuration Optimizations

One of the keys to good performance is to keep the multiprocessors on the device
as busy as possible. A device in which work is poorly balanced across the
multiprocessors will deliver suboptimal performance. Hence, it’s important to
design your application to use threads and blocks in a way that maximizes hardware
utilization and to limit practices that impede the free distribution of work. A key
concept in this effort is occupancy, which is explained in the following sections.
Another important concept is the management of system resources allocated for a
particular task. How to manage this resource utilization is discussed in the final
sections of this chapter.

4.1 Occupancy
Thread instructions are executed sequentially in CUDA, and, as a result, executing
other warps when one warp is paused or stalled is the only way to hide latencies and
keep the hardware busy. Some metric related to the number of active warps on a
multiprocessor is therefore important in determining how effectively the hardware is
kept busy. This metric is occupancy.
Occupancy is the ratio of the number of active warps per multiprocessor to the
maximum number of possible active warps. (To determine the latter number, see
the deviceQuery.cu program in the CUDA SDK or refer to Appendix A in the
CUDA C Programming Guide.) Another way to view occupancy is the percentage of
the hardware’s ability to process warps that is actively in use.
Higher occupancy does not always equate to higher performance—there is a point
above which additional occupancy does not improve performance. However, low
occupancy always interferes with the ability to hide memory latency, resulting in
performance degradation.

4.2 Calculating Occupancy

One of several factors that determine occupancy is register availability. Register
storage enables threads to keep local variables nearby for low-latency access.
However, the set of registers (known as the register file) is a limited commodity that
all threads resident on a multiprocessor must share. Registers are allocated to an

39 CUDA C Best Practices Guide Version 3.2

Chapter 4.
Execution Configuration Optimizations

entire block all at once. So, if each thread block uses many registers, the number of
thread blocks that can be resident on a multiprocessor is reduced, thereby lowering
the occupancy of the multiprocessor. The maximum number of registers per thread
can be set manually at compilation time per-file using the –maxrregcount option or
per-kernel using the __launch_bounds__ qualifier (see Section 3.2.6.1).
For purposes of calculating occupancy, the number of registers used by each thread
is one of the key factors. For example, devices with compute capability 1.0 and 1.1
have 8,192 32-bit registers per multiprocessor and can have a maximum of 768
simultaneous threads resident (24 warps x 32 threads per warp). This means that in
one of these devices, for a multiprocessor to have 100% occupancy, each thread can
use at most 10 registers. However, this approach of determining how register count
affects occupancy does not take into account the register allocation granularity. For
example, on a device of compute capability 1.0, a kernel with 128-thread blocks
using 12 registers per thread results in an occupancy of 83% with 5 active 128-
thread blocks per multiprocessor, whereas a kernel with 256-thread blocks using the
same 12 registers per thread results in an occupancy of 66% because only two 256-
thread blocks can reside on a multiprocessor. Furthermore, register allocations are
rounded up to the nearest 256 registers per block on devices with compute
capability 1.0 and 1.1.
The number of registers available, the maximum number of simultaneous threads
resident on each multiprocessor, and the register allocation granularity vary over
different compute capabilities. Because of these nuances in register allocation and
the fact that a multiprocessor’s shared memory is also partitioned between resident
thread blocks, the exact relationship between register usage and occupancy can be
difficult to determine. The --ptxas-options=-v option of nvcc details the
number of registers used per thread for each kernel. See Section 4.2 of the CUDA C
Programming Guide for the register allocation formulas for devices of various compute
capabilities and Section G.1 of the programming guide for the total number of
registers available on those devices. Alternatively, NVIDIA provides an occupancy
calculator in the form of an Excel spreadsheet that enables developers to hone in on
the optimal balance and to test different possible scenarios more easily. This
spreadsheet, shown in Figure 4.1, is called CUDA_Occupancy_calculator.xls and
is located in the tools directory of the CUDA SDK.

CUDA C Best Practices Guide Version 3.2 40

 Chapter 4.
Execution Configuration Optimizations

Figure 4.1 Use the CUDA GPU Occupancy Calculator to project occupancy

In addition to the calculator spreadsheet, occupancy can be determined using the

CUDA profiler.

4.3 Hiding Register Dependencies

Medium Priority: To hide latency arising from register dependencies, maintain

sufficient numbers of active threads per multiprocessor (i.e., sufficient occupancy).

Register dependencies arise when an instruction uses a result stored in a register

written by an instruction before it. The latency on current CUDA-enabled GPUs is
approximately 24 cycles, so threads must wait 24 cycles before using an arithmetic
result. However, this latency can be completely hidden by the execution of threads
in other warps. See Section 3.2.6 for details.

4.4 Thread and Block Heuristics

Medium Priority: The number of threads per block should be a multiple of 32 threads,
because this provides optimal computing efficiency and facilitates coalescing.

41 CUDA C Best Practices Guide Version 3.2

Chapter 4.
Execution Configuration Optimizations

The dimension and size of blocks per grid and the dimension and size of threads
per block are both important factors. The multidimensional aspect of these
parameters allows easier mapping of multidimensional problems to CUDA and does
not play a role in performance. As a result, this section discusses size but not
dimension.
Latency hiding and occupancy depend on the number of active warps per
multiprocessor, which is implicitly determined by the execution parameters along
with resource (register and shared memory) constraints. Choosing execution
parameters is a matter of striking a balance between latency hiding (occupancy) and
resource utilization.
Choosing the execution configuration parameters should be done in tandem;
however, there are certain heuristics that apply to each parameter individually. When
choosing the first execution configuration parameter—the number of blocks per
grid, or grid size—the primary concern is keeping the entire GPU busy. The number
of blocks in a grid should be larger than the number of multiprocessors so that all
multiprocessors have at least one block to execute. Furthermore, there should be
multiple active blocks per multiprocessor so that blocks that aren’t waiting for a
__syncthreads() can keep the hardware busy. This recommendation is subject to
resource availability; therefore, it should be determined in the context of the second
execution parameter—the number of threads per block, or block size—as well as
shared memory usage. To scale to future devices, the number of blocks per kernel
launch should be in the thousands.
When choosing the block size, it is important to remember that multiple concurrent
blocks can reside on a multiprocessor, so occupancy is not determined by block size
alone. In particular, a larger block size does not imply a higher occupancy. For
example, on a device of compute capability 1.1 or lower, a kernel with a maximum
block size of 512 threads results in an occupancy of 66 percent because the
maximum number of threads per multiprocessor on such a device is 768. Hence,
only a single block can be active per multiprocessor. However, a kernel with 256
threads per block on such a device can result in 100 percent occupancy with three
resident active blocks.
As mentioned in Section 4.1, higher occupancy does not always equate to better
performance. For example, improving occupancy from 66 percent to 100 percent
generally does not translate to a similar increase in performance. A lower occupancy
kernel will have more registers available per thread than a higher occupancy kernel,
which may result in less register spilling to local memory. Typically, once an
occupancy of 50 percent has been reached, additional increases in occupancy do not
translate into improved performance.
There are many such factors involved in selecting block size, and inevitably some
experimentation is required. However, a few rules of thumb should be followed:
 Threads per block should be a multiple of warp size to avoid wasting
computation on under-populated warps and to facilitate coalescing.
 A minimum of 64 threads per block should be used, but only if there are
multiple concurrent blocks per multiprocessor.
 Between 128 and 256 threads per block is a better choice and a good initial
range for experimentation with different block sizes.

CUDA C Best Practices Guide Version 3.2 42

 Chapter 4.
Execution Configuration Optimizations

 Use several (3 to 4) smaller thread blocks rather than one large thread block per
multiprocessor if latency affects performance. This is particularly beneficial to
kernels that frequently call __syncthreads().
Note that when a thread block allocates more registers than are available on a
multiprocessor, the kernel launch fails, as it will when too much shared memory or
too many threads are requested.

4.5 Effects of Shared Memory

Shared memory can be helpful in several situations, such as helping to coalesce or
eliminate redundant access to global memory. However, it also can act as a
constraint on occupancy. In many cases, the amount of shared memory required by
a kernel is related to the block size that was chosen, but the mapping of threads to
shared memory elements does not need to be one-to-one. For example, it may be
desirable to use a 32×32 element shared memory array in a kernel, but because the
maximum number of threads per block is 512, it is not possible to launch a kernel
with 32×32 threads per block. In such cases, kernels with 32×16 or 32×8 threads
can be launched with each thread processing two or four elements, respectively, of
the shared memory array. The approach of using a single thread to process multiple
elements of a shared memory array can be beneficial even if limits such as threads
per block are not an issue. This is because some operations common to each
element can be performed by the thread once, amortizing the cost over the number
of shared memory elements processed by the thread.
A useful technique to determine the sensitivity of performance to occupancy is
through experimentation with the amount of dynamically allocated shared memory,
as specified in the third parameter of the execution configuration. By simply
increasing this parameter (without modifying the kernel), it is possible to effectively
reduce the occupancy of the kernel and measure its effect on performance.
As mentioned in the previous section, once an occupancy of more than 50 percent
has been reached, it generally does not pay to optimize parameters to obtain higher
occupancy ratios. The previous technique can be used to determine whether such a
plateau has been reached.

43 CUDA C Best Practices Guide Version 3.2

Chapter 4.
Execution Configuration Optimizations

CUDA C Best Practices Guide Version 3.2 44

 Chapter 5.
Instruction Optimizations

Awareness of how instructions are executed often permits low-level optimizations

that can be useful, especially in code that is run frequently (the so-called hot spot in
a program). Best practices suggest that this optimization be performed after all
higher-level optimizations have been completed.

5.1 Arithmetic Instructions

Single-precision floats provide the best performance, and their use is highly
encouraged.
The throughput of individual arithmetic operations on devices of compute
capability 1.x is detailed in Section G.3 of the CUDA C Programming Guide, and the
throughput of these operations on devices of compute capability 2.x is detailed in
Section G.4 of the programming guide.

5.1.1 Division and Modulo Operations

Low Priority: Use shift operations to avoid expensive division and modulo calculations.

Integer division and modulo operations are particularly costly and should be avoided
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. (For further information,
refer to Chapter 5 of the CUDA C Programming Guide).

5.1.2 Reciprocal Square Root

The reciprocal square root should always be invoked explicitly as rsqrtf() for
single precision and rsqrt() for double precision. The compiler optimizes
1.0f/sqrtf(x) into rsqrtf() only when this does not violate IEEE-754
semantics.

45 CUDA C Best Practices Guide Version 3.2

Chapter 5.
Instruction Optimizations

5.1.3 Other Arithmetic Instructions

Low Priority: Avoid automatic conversion of doubles to floats.

The compiler must on occasion 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 an int
 Double-precision floating-point constants (defined without any type suffix)
used as input to single-precision floating-point computations
The latter case can be avoided by using single-precision floating-point constants,
defined with an f suffix such as 3.141592653589793f, 1.0f, 0.5f. This suffix has
accuracy implications in addition to its ramifications on performance. The effects on
accuracy are discussed in Chapter 7. Note that this distinction is particularly
important to performance on devices of compute capability 2.x.
For single-precision code, use of the float type and the single-precision math
functions are highly recommended. When compiling for devices without native
double-precision support such as devices of compute capability 1.2 and earlier, each
double-precision floating-point variable is converted to single-precision floating-
point format (but retains its size of 64 bits) and double-precision arithmetic is
demoted to single-precision arithmetic.
It should also be noted that the CUDA math library’s complementary error
function, erfcf(), is particularly fast with full single-precision accuracy.

5.1.4 Math Libraries

Medium Priority: Use the fast math library whenever speed trumps precision.

Two types of runtime math operations are supported. They can be distinguished by
their names: some have names with prepended underscores, whereas others do not
(e.g., __functionName() versus functionName()). Functions following the
__functionName() naming convention map directly to the hardware level. They
are faster but provide somewhat lower accuracy (e.g., __sinf(x) and __expf(x)).
Functions following functionName() naming convention are slower but have
higher accuracy (e.g., sinf(x) and expf(x)). The throughput of __sinf(x),
__cosf(x), and __expf(x) is much greather than that of sinf(x), cosf(x),
tanf(x). The latter become even more expensive (about an order of magnitude
slower) if the magnitude of the argument x needs to be reduced. Moreover, in such
cases, the argument-reduction code uses local memory, which can affect
performance even more because of the high latency of local memory. More details
are available in the CUDA C Programming Guide.
Note also that whenever sine and cosine of the same argument are computed, the
sincos… family of instructions should be used to optimize performance:

 __sincosf() for single-precision fast math (see next paragraph)

CUDA C Best Practices Guide Version 3.2 46

 Chapter 5.
Instruction Optimizations

 sincosf() for regular single-precision

 sincos() for double precision
The –use_fast_math compiler option of nvcc coerces every functionName() call
to the equivalent __functionName() call. This switch should be used whenever
accuracy is a lesser priority than the performance. This is frequently the case with
transcendental functions. Note this switch is effective only on single-precision
floating point.

Medium Priority: Prefer faster, more specialized math functions over slower, more
general ones when possible.

For exponentiation using base 2 or 10, use the functions exp2() or expf2() and
exp10() or expf10() rather than the functions pow() or powf(). Both pow() and
powf() are heavy-weight functions in terms of register pressure and instruction
count due to the numerous special cases arising in general exponentiation and the
difficulty of achieving good accuracy across the entire ranges of the base and the
exponent. The functions exp2(), exp2f(), exp10(), and exp10f(), on the other
hand, are similar to exp() and expf() in terms of performance, and can be as
much as ten times faster than their pow()/powf() equivalents.

5.2 Memory Instructions

High Priority: Minimize the use of global memory. Prefer shared memory access
where possible.

Memory instructions include any instruction that reads from or writes to shared,
local, or global memory. When accessing uncached local or global memory, there are
400 to 600 clock cycles of memory latency.
As an example, the assignment operator in the following sample code has a high
throughput, but, crucially, there is a latency of 400 to 600 clock cycles to read data
from global memory:
__shared__ float shared[32];
__device__ float device[32];
shared[threadIdx.x] = device[threadIdx.x];

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. However, it is best to avoid accessing
global memory whenever possible.

47 CUDA C Best Practices Guide Version 3.2

Chapter 5.
Instruction Optimizations

CUDA C Best Practices Guide Version 3.2 48

 Chapter 6.
Control Flow

6.1 Branching and Divergence

High Priority: Avoid different execution paths within the same warp.

Any flow control instruction (if, switch, do, for, while) can significantly affect
the instruction throughput by causing threads of the same warp to diverge; that is,
to follow different execution paths. If this happens, the different execution paths
must 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 4.1 of the CUDA C Programming Guide. A
trivial example is when the controlling condition depends only on (threadIdx /
WSIZE) where WSIZE is the warp size.

In this case, no warp diverges because the controlling condition is perfectly aligned
with the warps.

6.2 Branch Predication

Low Priority: Make it easy for the compiler to use branch predication in lieu of loops
or control statements.

Sometimes, the compiler may unroll loops or optimize out if or switch statements
by using branch predication instead. In these cases, no warp can ever diverge. The
programmer can also control loop unrolling using
#pragma unroll

For more information on this pragma, refer to the CUDA C Programming Guide.

49 CUDA C Best Practices Guide Version 3.2

Chapter 6.
Control Flow

When using branch predication, none of the instructions whose execution depends
on the controlling condition is skipped. Instead, each such instruction is associated
with a per-thread condition code or predicate that is set to true or false according to
the controlling condition. Although each of these instructions is scheduled for
execution, only the instructions with a true predicate are actually executed.
Instructions with a false predicate do not write results, and they 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 than 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.

6.3 Loop counters signed vs. unsigned

Medium Priority: Use signed integers rather than unsigned integers as loop counters.

In the C language standard, unsigned integer overflow semantics are well defined,
whereas signed integer overflow causes undefined results. Therefore, the compiler
can optimize more aggressively with signed arithmetic than it can with unsigned
arithmetic. This is of particular note with loop counters: since it is common for loop
counters to have values that are always positive, it may be tempting to declare the
counters as unsigned. For slightly better performance, however, they should instead
be declared as signed.
For example, consider the following code:
for (i = 0; i < n; i++) {
out[i] = in[offset + stride*i];
}

Here, the sub-expression stride*i could overflow a 32-bit integer, so if i is

declared as unsigned, the overflow semantics prevent the compiler from using some
optimizations that might otherwise have applied, such as strength reduction. If
instead i is declared as signed, where the overflow semantics are undefined, the
compiler has more leeway to use these optimizations.

CUDA C Best Practices Guide Version 3.2 50

 Chapter 7.
Getting the Right Answer

Obtaining the right answer is clearly the principal goal of all computation. On
parallel systems, it is possible to run into difficulties not typically found in traditional
serial-oriented programming. These include threading issues, unexpected values due
to the way floating-point values are computed, and challenges arising from
differences in the way CPU and GPU processors operate. This chapter examines
issues that can affect the correctness of returned data and points to appropriate
solutions.

7.1 Debugging
The CUDA debugger, CUDA-GDB, is a valuable debugging tool. It is a port of the
GNU Debugger version 6.6 and runs on 32-bit and 64-bit Linux. See the CUDA-
GDB User Manual for more details.

7.2 Numerical Accuracy and Precision

Incorrect or unexpected results arise principally from issues of floating-point
accuracy due to the way floating-point values are computed and stored. The
following sections explain the principal items of interest. Other peculiarities of
floating-point arithmetic are presented in Section G.2 of the CUDA C Programming
Guide.

7.2.1 Single vs. Double Precision

Devices of compute capability 1.3 and higher provide native support for double-
precision floating-point values (that is, values 64 bits wide). Results obtained using
double-precision arithmetic will frequently differ from the same operation
performed via single-precision arithmetic due to the greater precision of the former
and due to rounding issues. Therefore, it is important to be sure to compare like
with like and to express the results within a certain tolerance rather than expecting
them to be exact.

51 CUDA C Best Practices Guide Version 3.2

Chapter 7.
Getting the Right Answer

Whenever doubles are used, use at least the –arch=sm_13 switch on the nvcc
command line; see Sections 3.1.3 and 3.1.4 of the CUDA C Programming Guide for
more details.

7.2.2 Floating-Point Math Is Not Associative

Each floating-point arithmetic operation involves a certain amount of rounding.
Consequently, the order in which arithmetic operations are performed is important.
If A, B, and C are floating-point values, (A+B)+C is not guaranteed to equal
A+(B+C) as it is in symbolic math. When you parallelize computations, you
potentially change the order of operations and therefore the parallel results might
not match sequential results. This limitation is not specific to CUDA, but an
inherent part of parallel computation on floating-point values.

7.2.3 Promotions to Doubles and Truncations to Floats

When comparing the results of computations of float variables between the host
and device, make sure that promotions to double precision on the host do not
account for different numerical results. For example, if the code segment
float a;
…
a = a*1.02;

were performed on a device of compute capability 1.2 or less, or on a device with

compute capability 1.3 but compiled without enabling double precision (as
mentioned above), then the multiplication would be performed in single precision.
However, if the code were performed on the host, the literal 1.02 would be
interpreted as a double-precision quantity and a would be promoted to a double, the
multiplication would be performed in double precision, and the result would be
truncated to a float—thereby yielding a slightly different result. If, however, the
literal 1.02 were replaced with 1.02f, the result would be the same in all cases
because no promotion to doubles would occur. To ensure that computations use
single-precision arithmetic, always use float literals.
In addition to accuracy, the conversion between doubles and floats (and vice versa)
has a detrimental effect on performance, as discussed in Chapter 5.

7.2.4 IEEE 754 Compliance

All CUDA compute devices follow the IEEE 754 standard for binary floating-point
representation, with some small exceptions. These exceptions, which are detailed in
Section G.2 of the CUDA C Programming Guide, can lead to results that differ from
IEEE 754 values computed on the host system.
One of the key differences is the fused multiply-add (FMAD) instruction, which
combines multiply-add operations into a single instruction execution and truncates
the intermediate result of the multiplication. Its result will differ at times from
results obtained by doing the two operations separately.

CUDA C Best Practices Guide Version 3.2 52

 Chapter 7.
Getting the Right Answer

7.2.5 x86 80-bit Computations

x86 processors can use an 80-bit ―double extended precision‖ math when
performing floating-point calculations. The results of these calculations can
frequently differ from pure 64-bit operations performed on the CUDA device. To
get a closer match between values, set the x86 host processor to use regular double
or single precision (64 bits and 32 bits, respectively). This is done with the FLDCW
assembly instruction or the equivalent operating system API.

53 CUDA C Best Practices Guide Version 3.2

Chapter 7.
Getting the Right Answer

CUDA C Best Practices Guide Version 3.2 54

 Chapter 8.
Multi-GPU Programming

8.1 Introduction to Multi-GPU

Programming a multi-GPU application is no different from programming an
application to utilize multiple cores or sockets because CUDA is completely
orthogonal to CPU thread management or message passing APIs. The only new
concept is selecting the correct GPU, which in most cases is a free (without a
context) GPU. In fact, GPU acceleration can be straightforwardly added to existing
multi-threaded CPU code, whether it uses lightweight or heavyweight threads – all
one has to do is port the compute intensive portions of the code to GPU and add
calls to transfer data between CPU and GPU, leaving the inter-CPU-thread
communication code unchanged.

8.2 Multi-GPU Programming

In order to issue work to a GPU, a context is established between a CPU thread and
the GPU. Only one context can be active on GPU at a time. Similarly, a CPU
thread can have one active context at a time. A context is established during the
program’s first call to a function that changes state (such as cudaMalloc(), etc.), so
one can force the creation of a context by calling cudaFree(0). Note that a
context is created on GPU 0 by default, unless another GPU is selected explicitly
prior to context creation with a cudaSetDevice() call. Context is destroyed either
with a cudaThreadExit() call, or when the controlling CPU thread exits.
CUDA driver API allows a single CPU thread to manage multiple contexts (and
therefore multiple GPUs) by pushing/popping contexts. In the remainder of the
document we will focus on CUDA runtime API, which currently allows strictly one
context per CPU thread.
In order to issue work to p GPUs concurrently, a program needs p CPU threads,
each with its own context. Threads can be lightweight (pthreads, OpenMP, etc.) or
heavyweight (MPI). Note that any CPU multi-threading or message-passing API or
library can be used, as CPU thread management is completely orthogonal to CUDA.
For example, one can add GPU processing to an existing MPI application by
porting the compute-intensive portions of the code without changing the
communication structure.

55 CUDA C Best Practices Guide Version 3.2

Chapter 8.
Multi-GPU Programming

Even though a GPU can execute calls from one context at a time, it can belong to
multiple contexts. For example, it is possible for several CPU threads to establish
contexts with the same GPU. This allows developing multi-GPU applications on a
single GPU. GPU driver manages GPU switching between the contexts, as well as
partitioning memory among the contexts (GPU memory allocated in one context
cannot be accessed from another context).

8.3 Selecting a GPU

In many cases, the intent is for each CPU thread to control a different GPU.
Achieving this is straightforward if a program spawns as many lightweight threads as
there are GPUs – one can derive GPU index from thread ID. For example,
OpenMP thread ID can be readily used to select GPUs.
MPI rank can be used to choose a GPU reliably as long as all MPI processes are
launched on a single host node. However, when MPI processes are launched across
several cluster nodes, the assignment of processes to host nodes depends on the
MPI implementation and launch configuration (hostfile), preventing reliable
selection of a unique GPU. To address this issue, one could implement a
negotiation protocol among MPI processes running on the same host (using
MPI_Get_processor_name() call), so that each one claims a unique GPU. A less
robust option is for the GPU selection to rely on a specific allocation MPI processes
to nodes (for example, allocating consecutive processes to the same node).
Starting with CUDA 2.2, Linux drivers provide an alternative and much simpler way
to ensure that each GPU is assigned to at most one context – the administrator can
select exclusive mode via the SMI (System Management Interface) tool that comes
with the driver. In exclusive mode, if the application does not specify a GPU with a
cudaSetDevice() call, the context is no longer created on GPU 0, but on a GPU
without an active context. If there are no available GPUs, or if cudaSetDevice()
specifies a GPU already having an active context, the first CUDA call that attempts
to change the device state will fail and return an error.

8.4 Inter-GPU communication

All inter-GPU communication takes place via host nodes. GPU and the controlling
CPU thread communicate via memcopies, while CPU threads exchange data using
the same methods as applications not accelerated with GPUs. Thus, best
performance is achieved when one follows best practices for the CPU-GPU
communication as well as CPU-CPU communication. Note that the two are
independent and orthogonal.
Communication between CPU and GPU is most efficient when using pinned
memory on the CPU. Pinned memory enables asynchronous memory copies
(allowing for overlap with both CPU and GPU execution), as well as improves PCIe
throughput on FSB systems. Please refer to the CUDA C Programming Guide for
more details and examples of pinned memory usage.

CUDA C Best Practices Guide Version 3.2 56

 Chapter 8.
Multi-GPU Programming

Lightweight CPU threads exchange data most efficiently via shared memory. Note
that in order for a pinned memory region to be viewed as pinned by CPU threads
other than the one that allocated it, one must call cudaHostAlloc() with the
cudaHostAllocPortable flag. A common communication pattern will be for one
CPU thread to copy data from its GPU to a shared host memory region, after which
another CPU thread will copy the data to its GPU. Users of NUMA systems will
have to follow the same best practices as for communication between non-GPU
accelerated CPU threads.
Communication between heavy-weight processes takes place via message passing,
for example MPI. Once data has been copied from GPU to CPU it is transferred to
another process by calling one of the MPI functions. For example, one possible
pattern when exchanging data between two GPUs is for a CPU thread to call a
device-to-host cudaMemcpy(), then MPI_Sendrecv(), then a host-to-device
cudaMemcpy(). Note that performance of the MPI function is not dependent on
the fact that data originated at or is destined for a GPU. Since MPI provides several
variations for most of its communication functions, the choice of a function should
be dictated by the best practices guide for the MPI implementation as well as the
system and network.

8.5 Compiling Multi-GPU Applications

Code that includes GPU kernels or CUDA runtime kernel launch syntax must be
compiled with nvcc, the compiler driver. nvcc invokes either gcc (on Linux) or
Microsoft Visual C++ compiler (on Windows) for non-GPU code. One can pass
options to the host C/C++ compiler through nvcc by using –Xcompiler option
(for more details please refer to nvcc documentation included in the CUDA
toolkit). For example, one would specify nvcc –Xcompiler /openmp in order to
pass the -fopenmp option to the host gcc compiler (the same option for MSVC
compiler is /openmp).
MPI applications are typically compiled using mpicc, the MPI compiler driver.
While it is possible to pass all the necessary flags through nvcc, it is often simpler to
compile MPI code and CUDA kernel code into separate object files. This approach
also fits many larger projects well, since they separate the communication and
computation source code already.

8.6 Infiniband
NVIDIA GPUDirect™ technology allows the sharing of CUDA pinned host
memory with other devices. This allows accelerated transfers of GPU data to other
devices, such as supported Infiniband network adapters. If GPUDirect support is
not available for your network device, network transfer throughput can be reduced.
A possible workaround is to disable RDMA. For the OpenMPI implementation,
this can be achieved by passing the flag –mca btl_openib_flags 1 to mpirun.

57 CUDA C Best Practices Guide Version 3.2

Chapter 8.
Multi-GPU Programming

CUDA C Best Practices Guide Version 3.2 58

 Appendix A.
Recommendations and Best Practices

This appendix contains a list of all the recommendations for optimization and the
list of best practices that are explained in this document.

A.1 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. 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
launch. The application should also maximize parallel execution at a higher level by
explicitly exposing concurrent execution on the device through streams, as well as
maximizing concurrent execution between the host and the device.
Optimizing memory usage starts with minimizing data transfers between the host
and the device because those transfers have much lower bandwidth than internal
device data transfers. Kernel access to global memory also should be minimized by
maximizing the use of shared memory on the device. Sometimes, the best
optimization might even be to avoid any data transfer in the first place by simply
recomputing the data whenever it is needed.
The effective bandwidth can vary by an order of magnitude depending on the access
pattern for each type of memory. The next step in optimizing memory usage is
therefore to organize memory accesses according to the optimal memory access
patterns. This optimization is especially important for global memory accesses,
because latency of access costs hundreds of clock cycles. Shared memory accesses,
in counterpoint, are usually worth optimizing only when there exists a high degree
of bank conflicts.
As for optimizing instruction usage, the use of arithmetic instructions that have low
throughput should be avoided. This suggests trading precision for speed when it
does not affect the end result, such as using intrinsics instead of regular functions or
single precision instead of double precision. Finally, particular attention must be

59 CUDA C Best Practices Guide Version 3.2

Appendix A.
Recommendations and Best Practices

paid to control flow instructions due to the SIMT (single instruction multiple
thread) nature of the device.

A.2 High-Priority Recommendations

 To get the maximum benefit from CUDA, focus first on finding ways to
parallelize sequential code. (Section 1.1.3)
 Use the effective bandwidth of your computation as a metric when measuring
performance and optimization benefits. (Section 2.2)
 Minimize data transfer between the host and the device, even if it means
running some kernels on the device that do not show performance gains when
compared with running them on the host CPU. (Section 3.1)
 Ensure global memory accesses are coalesced whenever possible. (Section 3.2.1)
 Minimize the use of global memory. Prefer shared memory access where
possible. (Section 5.2)
 Avoid different execution paths within the same warp. (Section 6.1)

A.3 Medium-Priority Recommendations

 Accesses to shared memory should be designed to avoid serializing requests due
to bank conflicts. (Section 3.2.2.1)
 Use shared memory to avoid redundant transfers from global memory. (Section
3.2.2.2)
 To hide latency arising from register dependencies, maintain sufficient numbers
of active threads per multiprocessor (i.e., sufficient occupancy). (Sections 3.2.6
and 4.3)
 The number of threads per block should be a multiple of 32 threads, because
this provides optimal computing efficiency and facilitates coalescing. (Section
4.4)
 Use the fast math library whenever speed trumps precision. (Section 5.1.4)
 Prefer faster, more specialized math functions over slower, more general ones
when possible. (Section 5.1.4)
 Use signed integers rather than unsigned integers as loop counters. (Section 6.3)

A.4 Low-Priority Recommendations

 On version 2.2 of the CUDA Toolkit (and later), use zero-copy operations on
integrated GPUs. (Section 3.1.3)

CUDA C Best Practices Guide Version 3.2 60

 Appendix A.
Recommendations and Best Practices

 For kernels with long argument lists, place some arguments into constant
memory to save shared memory. (Section 3.2.2.4)
 Use shift operations to avoid expensive division and modulo calculations.
(Section 5.1.1)
 Avoid automatic conversion of doubles to floats. (Section 5.1.3)
 Make it easy for the compiler to use branch predication in lieu of loops or
control statements. (Section 6.2)

61 CUDA C Best Practices Guide Version 3.2

 Appendix B.
Useful NVCC Compiler Switches

B.1 NVCC
nvcc is the compiler that converts .cu files into C for the host system and CUDA
assembly or binary instructions for the device. It supports a spate of switches, of
which the following are especially useful for optimization and related best practices:
 -arch=sm_13 or higher is required for double precision. See Section 7.2.1.
 –maxrregcount=N specifies the maximum number of registers kernels can use
at a per-file level. See Section 3.2.6.1. (See also the __launch_bounds__
qualifier discussed in Section B.17 of the CUDA C Programming Guide to control
the number of registers used on a per-kernel basis.)
 --ptxas-options=-v or -Xptxas=-v lists per-kernel register, shared, and
constant memory usage.
 –use_fast_math compiler option of nvcc coerces every functionName() call
to the equivalent __functionName() call. This makes the code run faster at the
cost of slightly diminished precision and accuracy. See Section 5.1.4.

62 CUDA C Best Practices Guide Version 3.2

 Appendix C.
Revision History

C.1 Version 3.0

 Added compute capability 2.0.
 Devices of compute capability 2.0 have an L1 cache of configurable size for
global and local memory (see Section G.4.1 of the CUDA C Programming Guide).
Note that the L1 cache has higher bandwidth than the texture cache and
therefore obviates some of the benefits of texture fetch seen in devices of
compute capability 1.x. See Section 3.2.4.1.
 Removed references to device emulation mode, which is deprecated in CUDA
Toolkit 3.0 and will be removed in CUDA Toolkit 3.1.
 The CUFFT and CUBLAS libraries can now interoperate with applications that
use the driver API.
 Added a recommendation for the use of exp2() or exp10() in preference to
pow() to Section 5.1.4.

 The __launch_bounds__ kernel definition qualifier can now be used to control

register usage on a per-kernel basis (in contrast to the -maxrregcount
command-line option of nvcc, which controls this on a per-file basis). See
Section 3.2.6.1 of this guide and Section B.14 of the CUDA C Programming
Guide.
 Note importance of distinguishing single-precision floating-point constants and
arithmetic from double-precision constants and arithmetic for performance on
devices of compute capability 2.0 in Section 5.1.3.
 Devices of compute capability 2.0 can multicast shared memory accesses in
addition to the broadcast mechanism in devices of compute capability 1.x.
 Added Chapter 8 on multi-GPU programming.

C.2 Version 3.1

 Added Section 3.3 on memory allocation.
 Added fourth bullet to rules of thumb for block size selection in Section 4.4.

63 CUDA C Best Practices Guide Version 3.2

Appendix C.
Revision History

 Added Section 6.3 on signed versus unsigned loop counters.

C.3 Version 3.2

 CUdeviceptr is now a 64-bit value in 64-bit applications, so it is no longer
necessary to cast CUdeviceptr values with (void*)(size_t) when passing
them as arguments to kernels using the CUDA Driver API; this casting is now
effectively a no-op. As such, the type casts have been removed from the
example code in Section 1.3.4.
 Corrected discussion in Section 3.2.6 of the necessary numbers of active threads
per multiprocessor to hide latency on devices of compute capability 2.0.

CUDA C Best Practices Guide Version 3.2 64

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, CUDA, GeForce, NVIDIA Quadro, and Tesla 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
© 2010 NVIDIA Corporation. All rights reserved.