Hugging Face's logo

Datasets:
Vikhrmodels
/
programming_books_eng

Modalities:
Text
Formats:
text
Size:
1M - 10M
Libraries:
Datasets
License:
Dataset card Data Studio Files
xet
 Community
programming_books_eng / CUDA /NVIDIA_OpenCL_ProgrammingGuide.pdf.txt
apsys's picture
apsys
Upload folder using huggingface_hub
08c8a6d verified
raw
history blame contribute delete
110 kB
OpenCL
Programming Guide
for the CUDA
Architecture
Version 2.3
8/27/2009
Table of Contents
Chapter 1. Introduction ..................................................................................... 5 
1.1 
1.2 
1.3 
1.4 
From Graphics Processing to General-Purpose Parallel Computing ................... 5 
CUDA™: a General-Purpose Parallel Computing Architecture ........................... 7 
CUDA’s Scalable Programming Model ............................................................. 8 
Document’s Structure ................................................................................... 9 
Chapter 2. OpenCL on the CUDA Architecture ................................................. 11 
2.1 
CUDA Architecture ...................................................................................... 11 
2.1.1 
Execution Model .................................................................................. 11 
2.1.2  Memory Model ..................................................................................... 14 
2.2 
Compilation ................................................................................................ 16 
2.2.1 
PTX..................................................................................................... 16 
2.2.2 
Volatile ................................................................................................ 17 
2.3 
Compute Capability .................................................................................... 17 
2.4  Mode Switches ........................................................................................... 18 
2.5  Matrix Multiplication Example ...................................................................... 18 
Chapter 3. Performance Guidelines ................................................................. 27 
3.1 
Instruction Performance ............................................................................. 27 
3.1.1 
Instruction Throughput ........................................................................ 27 
3.1.1.1 
Arithmetic Instructions .................................................................. 27 
3.1.1.2 
Control Flow Instructions ............................................................... 29 
3.1.1.3 
Memory Instructions ..................................................................... 30 
3.1.1.4 
Synchronization Instruction ........................................................... 30 
3.1.2  Memory Bandwidth .............................................................................. 30 
3.1.2.1 
Global Memory .............................................................................. 31 
3.1.2.2 
Local Memory ............................................................................... 38 
3.1.2.3 
Constant Memory .......................................................................... 38 
3.1.2.4 
Texture Memory ........................................................................... 38 
3.1.2.5 
Shared Memory ............................................................................ 39 
ii
NVIDIA OpenCL Programming Guide Version 2.3
3.1.2.6 
Registers ...................................................................................... 46 
3.2 
3.3 
NDRange ................................................................................................... 46 
Data Transfer between Host and Device ...................................................... 47 
3.4  Warp-Level Synchronization ........................................................................ 48 
3.5  Overall Performance Optimization Strategies ................................................ 49 
Appendix A. Technical Specifications .............................................................. 51 
A.1  General Specifications ................................................................................. 51 
A.1.1 
Specifications for Compute Capability 1.0 .............................................. 52 
A.1.2 
Specifications for Compute Capability 1.1 .............................................. 53 
A.1.3 
Specifications for Compute Capability 1.2 .............................................. 53 
A.1.4 
Specifications for Compute Capability 1.3 .............................................. 53 
A.2 
A.3 
Floating-Point Standard .............................................................................. 53 
Supported OpenCL Extensions ..................................................................... 54 
Appendix B. Mathematical Functions Accuracy ............................................... 55 
B.1 
Standard Functions ..................................................................................... 55 
B.1.1 
Single-Precision Floating-Point Functions ............................................... 55 
B.1.2 
Double-Precision Floating-Point Functions ............................................. 57 
B.2 
Native Functions ......................................................................................... 59 
NVIDIA OpenCL Programming Guide Version 2.3
iii
List of Figures
Figure 1-1.  Floating-Point Operations per Second and Memory Bandwidth for the CPU
and GPU 6 
Figure 1-2.  The GPU Devotes More Transistors to Data Processing ............................ 7 
Figure 1-3.  CUDA is Designed to Support Various Languages and Application
Programming Interfaces ...................................................................................... 8 
Figure 2-1.  Grid of Thread Blocks ........................................................................... 12 
Figure 2-2.  Automatic Scalability ............................................................................ 13 
Figure 2-3.  CUDA Architecture ............................................................................... 16 
Figure 3-1.  Examples of Coalesced Global Memory Access Patterns .......................... 34 
Figure 3-2.  Examples of Global Memory Access Patterns That Are Non-Coalesced for
Devices of Compute Capability 1.0 or 1.1 ........................................................... 35 
Figure 3-3.  Examples of Global Memory Access Patterns That Are Non-Coalesced for
Devices of Compute Capability 1.0 or 1.1 ........................................................... 36 
Figure 3-4.  Examples of Global Memory Access by Devices with Compute Capability
1.2 and Higher .................................................................................................. 37 
Figure 3-5.  Examples of Shared Memory Access Patterns without Bank Conflicts ..... 42 
Figure 3-6.  Example of a Shared Memory Access Pattern without Bank Conflicts ...... 43 
Figure 3-7.  Examples of Shared Memory Access Patterns with Bank Conflicts ........... 44 
Figure 3-8.  Example of Shared Memory Read Access Patterns with Broadcast ........... 45 
iv
NVIDIA OpenCL Programming Guide Version 2.3
Chapter 1.
Introduction
1.1
From Graphics Processing to
General-Purpose Parallel Computing
Driven by the insatiable market demand for realtime, high-definition 3D graphics,
the programmable Graphic Processor Unit or GPU has evolved into a highly
parallel, multithreaded, manycore processor with tremendous computational
horsepower and very high memory bandwidth, as illustrated by Figure 1-1.
Chapter 1. 0BIntroduction
GT200
G92
G80
Ultra
G80
NV35
NV40
NV30
G71
G70
3.0 GHz
Core2 Duo
3.2 GHz
Harpertown
Jan
Jun
Apr
2003
2004
Jun
2005
Mar
Nov
May
2006
2007
Jun
2008
GT200 = GeForce GTX 280
G71 = GeForce 7900 GTX
NV35 = GeForce FX 5950 Ultra
G92 = GeForce 9800 GTX
G70 = GeForce 7800 GTX
NV30 = GeForce FX 5800
G80 = GeForce 8800 GTX
NV40 = GeForce 6800 Ultra
G80
Ultra
G80
G71
NV40
NV30
Harpertown
Woodcrest
Prescott EE
Northwood
Figure 1-1. Floating-Point Operations per Second and Memory
Bandwidth for the CPU and GPU
The reason behind the discrepancy in floating-point capability between the CPU and
the GPU is that the GPU is specialized for compute-intensive, highly parallel
computation – exactly what graphics rendering is about – and therefore designed
such that more transistors are devoted to data processing rather than data caching
and flow control, as schematically illustrated by Figure 1-2.
6
NVIDIA OpenCL Programming Guide Version 2.3
Chapter 1. 0BIntroduction
Control
Cache
DRAM
ALU
ALU
ALU
ALU
CPU
DRAM
GPU
Figure 1-2. The GPU Devotes More Transistors to Data
Processing
More specifically, the GPU is especially well-suited to address problems that can be
expressed as data-parallel computations – the same program is executed on many
data elements in parallel – with high arithmetic intensity – the ratio of arithmetic
operations to memory operations. Because the same program is executed for each
data element, there is a lower requirement for sophisticated flow control; and
because it is executed on many data elements and has high arithmetic intensity, the
memory access latency can be hidden with calculations instead of big data caches.
Data-parallel processing maps data elements to parallel processing threads. Many
applications that process large data sets can use a data-parallel programming model
to speed up the computations. In 3D rendering, large sets of pixels and vertices are
mapped to parallel threads. Similarly, image and media processing applications such
as post-processing of rendered images, video encoding and decoding, image scaling,
stereo vision, and pattern recognition can map image blocks and pixels to parallel
processing threads. In fact, many algorithms outside the field of image rendering
and processing are accelerated by data-parallel processing, from general signal
processing or physics simulation to computational finance or computational biology.
1.2
CUDA™: a General-Purpose Parallel
Computing Architecture
In November 2006, NVIDIA introduced CUDA™, a general purpose parallel
computing architecture – with a new parallel programming model and instruction
set architecture – that leverages the parallel compute engine in NVIDIA GPUs to
solve many complex computational problems in a more efficient way than on a
CPU.
As illustrated by Figure 1-3, there are several languages and application
programming interfaces that can be used to program the CUDA architecture.
NVIDIA OpenCL Programming Guide Version 2.3
7
Chapter 1. 0BIntroduction
Figure 1-3. CUDA is Designed to Support Various Languages
and Application Programming Interfaces
1.3
CUDA’s Scalable Programming Model
The advent of multicore CPUs and manycore GPUs means that mainstream
processor chips are now parallel systems. Furthermore, their parallelism continues
to scale with Moore’s law. The challenge is to develop application software that
transparently scales its parallelism to leverage the increasing number of processor
cores, much as 3D graphics applications transparently scale their parallelism to
manycore GPUs with widely varying numbers of cores.
CUDA’s parallel programming model is designed to overcome this challenge with
three key abstractions: a hierarchy of thread groups, a hierarchy of shared memories,
and barrier synchronization.
These abstractions provide fine-grained data parallelism and thread parallelism,
nested within coarse-grained data parallelism and task parallelism. They guide the
programmer to partition the problem into coarse sub-problems that can be solved
independently in parallel, and then into finer pieces that can be solved cooperatively
in parallel. Such a decomposition preserves language expressivity by allowing
threads to cooperate when solving each sub-problem, and at the same time enables
transparent scalability since each sub-problem can be scheduled to be solved on any
of the available processor cores: A compiled program can therefore execute on any
number of processor cores, and only the runtime system needs to know the physical
processor count.
This scalable programming model allows the CUDA architecture to span a wide
market range by simply scaling the number of processors and memory partitions:
from the high-performance enthusiast GeForce GTX 280 GPU and professional
Quadro and Tesla computing products to a variety of inexpensive, mainstream
GeForce GPUs (see Appendix A for a list of all CUDA-enabled GPUs).
8
NVIDIA OpenCL Programming Guide Version 2.3
Chapter 1. 0BIntroduction
1.4
Document’s Structure
This document is organized into the following chapters:
(cid:137) Chapter 1 is a general introduction to GPU computing and the CUDA
architecture.
(cid:137) Chapter 2 describes how the OpenCL architecture maps to the CUDA
architecture and the specifics of NVIDIA’s OpenCL implementation.
(cid:137) Chapter 3 gives some guidance on how to achieve maximum performance.
(cid:137) Appendix A lists the CUDA-enabled GPUs with their technical specifications.
(cid:137) Appendix B lists the accuracy of each mathematical function on the CUDA
architecture.
NVIDIA OpenCL Programming Guide Version 2.3
9
Chapter 2.
OpenCL on the CUDA Architecture
2.1
CUDA Architecture
2.1.1
Execution Model
The CUDA architecture is a close match to the OpenCL architecture.
A CUDA device is built around a scalable array of multithreaded Streaming
Multiprocessors (SMs). A multiprocessor corresponds to an OpenCL compute unit.
A multiprocessor executes a CUDA thread for each OpenCL work-item and a thread
block for each OpenCL work-group. A kernel is executed over an OpenCL
NDRange by a grid of thread blocks. As illustrated in Figure 2-1, each of the thread
blocks that execute a kernel is therefore uniquely identified by its work-group ID,
and each thread by its global ID or by a combination of its local ID and work-group
ID.
Chapter 2. 1BOpenCL on the CUDA Architecture
Grid
Block (0, 0)
Block (1, 0)
Block (2, 0)
Block (0, 1)
Block (1, 1)
Block (2, 1)
Block (1, 1)
Thread (0, 0) Thread (1, 0) Thread (2, 0) Thread (3, 0)
Thread (0, 1) Thread (1, 1) Thread (2, 1) Thread (3, 1)
Thread (0, 2) Thread (1, 2) Thread (2, 2) Thread (3, 2)
A kernel is executed over an NDRange by a grid of thread blocks.
Figure 2-1. Grid of Thread Blocks
A thread is also given a unique thread ID within its block. The local ID of a thread
and its thread ID relate to each other in a straightforward way: For a one-
dimensional block, they are the same; for a two-dimensional block of size (Dx, Dy),
the thread ID of a thread of index (x, y) is (x + y Dx); for a three-dimensional block
of size (Dx, Dy, Dz), the thread ID of a thread of index (x, y, z) is
(x + y Dx + z Dx Dy).
When an OpenCL program on the host invokes a kernel, the work-groups are
enumerated and distributed as thread blocks to the multiprocessors with available
execution capacity. The threads of a thread block execute concurrently on one
multiprocessor. As thread blocks terminate, new blocks are launched on the vacated
multiprocessors.
Thread blocks are required to execute independently: It must be possible to execute
them in any order, in parallel or in series. This independence requirement allows
thread blocks to be scheduled in any order across any number of cores, enabling
12
NVIDIA OpenCL Programming Guide Version 2.3
programmers to write code that scales with the number of cores, as illustrated in
Figure 2-2.
Chapter 1
Kernel
Block 0
Block 1
Block 2
Block 3
Block 4
Block 5
Block 5
Block 6
Block 6
Block 7
Device with 2 SMs
Device with 4 SMs
SM 0
SM 1
SM 0
SM 1
SM 2
SM 3
Block 0
Block 1
Block 0
Block 1
Block 2
Block 3
Block 4
Block 5
Block 6
Block 7
Block 2
Block 3
Block 4
Block 5
Block 6
Block 7
A device with more multiprocessors will automatically execute a kernel in less time than a device with
fewer multiprocessors.
Figure 2-2. Automatic Scalability
A multiprocessor consists of eight Scalar Processor (SP) cores, two special function
units for transcendentals, a multithreaded instruction unit, and on-chip shared
memory, which is used to implement OpenCL local memory. The multiprocessor
creates, manages, and executes concurrent threads in hardware with zero scheduling
overhead. It implements the work-group barrier function with a single instruction.
Fast barrier synchronization together with lightweight thread creation and zero-
overhead thread scheduling efficiently support very fine-grained parallelism,
allowing, for example, a low granularity decomposition of problems by assigning
one thread to each data element (such as a pixel in an image, a voxel in a volume, a
cell in a grid-based computation).
To manage hundreds of threads running several different programs, the
multiprocessor employs a new architecture we call SIMT (single-instruction,
multiple-thread). The multiprocessor maps each thread to one SP, and each scalar
NVIDIA OpenCL Programming Guide Version 2.3
13
Chapter 2. 1BOpenCL on the CUDA Architecture
thread executes independently with its own instruction address and register state.
The multiprocessor SIMT unit creates, manages, schedules, and executes threads in
groups of 32 parallel threads called warps. (This term originates from weaving, the
first parallel thread technology. A half-warp is either the first or second half of a
warp.) Individual threads composing a SIMT warp start together at the same
program address but are otherwise free to branch and execute independently.
When a multiprocessor is given one or more thread blocks to execute, it splits them
into warps that get scheduled by the SIMT unit. The way a block is split into warps
is always the same; each warp contains threads of consecutive, increasing thread IDs
with the first warp containing the thread of thread ID zero.
Every instruction issue time, the SIMT unit selects a warp that is ready to execute
and issues the next instruction to the active threads of the warp. A warp executes
one common instruction at a time, so full efficiency is realized when all 32 threads
of a warp agree on their execution path. If threads of a warp diverge via a data-
dependent conditional branch, the warp serially executes each branch path taken,
disabling threads that are not on that path, and when all paths complete, the threads
converge back to the same execution path. Branch divergence occurs only within a
warp; different warps execute independently regardless of whether they are
executing common or disjointed code paths.
SIMT architecture is akin to SIMD (Single Instruction, Multiple Data) vector
organizations in that a single instruction controls multiple processing elements. A
key difference is that SIMD vector organizations expose the SIMD width to the
software, whereas SIMT instructions specify the execution and branching behavior
of a single thread. In contrast with SIMD vector machines, SIMT enables
programmers to write thread-level parallel code for independent, scalar threads, as
well as data-parallel code for coordinated threads. For the purposes of correctness,
the programmer can essentially ignore the SIMT behavior; however, substantial
performance improvements can be realized by taking care that the code seldom
requires threads in a warp to diverge. In practice, this is analogous to the role of
cache lines in traditional code: Cache line size can be safely ignored when designing
for correctness but must be considered in the code structure when designing for
peak performance. Vector architectures, on the other hand, require the software to
coalesce loads into vectors and manage divergence manually.
2.1.2
Memory Model
As illustrated by Figure 2-3, each multiprocessor has on-chip memory of the four
following types:
(cid:137) One set of local 32-bit registers per processor,
(cid:137) A parallel data cache or shared memory that is shared by all scalar processor cores
and is where OpenCL local memory resides,
(cid:137) A read-only constant cache that is shared by all scalar processor cores and speeds
up reads from OpenCL constant memory,
(cid:137) A read-only texture cache that is shared by all scalar processor cores and speeds up
reads from OpenCL image objects; each multiprocessor accesses the texture
cache via a texture unit that implements the various addressing modes and data
filtering specified by OpenCL sampler objects; the region of device memory
addressed by image objects is referred to a texture memory.
14
NVIDIA OpenCL Programming Guide Version 2.3
Chapter 1
There is also a global memory address space that is used for OpenCL global
memory and a local memory address space that is private to each thread (and should
not be confused with OpenCL local memory). Both memory spaces are read-write
regions of device memory and are not cached.
A variable in OpenCL private memory generally resides in a register. However in
some cases the compiler might choose to place it in CUDA local memory, which
can have adverse performance consequences because of local memory high latency
and bandwidth (see Section 3.1.2.2). Variables that are likely to be placed in CUDA
local memory are large structures or arrays that would consume too much register
space, and arrays for which the compiler cannot determine that they are indexed
with constant quantities.
The number of blocks a multiprocessor can process at once – referred to as the
number of active blocks per multiprocessor – depends on how many registers per
thread and how much shared memory per block are required for a given kernel since
the multiprocessor’s registers and shared memory are split among all the threads of
the active blocks. If there are not enough registers or shared memory available per
multiprocessor to process at least one block, the kernel will fail to launch. The
maximum number of active blocks per multiprocessor, as well as the maximum
number of active warps and maximum number of active threads are given in
Appendix A.
If a non-atomic instruction executed by a warp writes to the same location in global
or shared memory for more than one of the threads of the warp, the number of
serialized writes that occur to that location and the order in which they occur is
undefined, but one of the writes is guaranteed to succeed. If an atomic instruction
executed by a warp reads, modifies, and writes to the same location in global
memory for more than one of the threads of the warp, each read, modify, write to
that location occurs and they are all serialized, but the order in which they occur is
undefined.
NVIDIA OpenCL Programming Guide Version 2.3
15
Chapter 2. 1BOpenCL on the CUDA Architecture
Device
Multiprocessor N
Multiprocessor 2
Multiprocessor 1
Shared Memory
Registers
Registers
Registers
Processor 1
Processor 2
…
Processor M
Instruction
Unit
Constant
Cache
Texture
Cache
Device Memory
A set of SIMT multiprocessors with on-chip shared memory.
Figure 2-3. CUDA Architecture
2.2
Compilation
2.2.1
PTX
Kernels written in OpenCL C are compiled into PTX, which is CUDA’s instruction
set architecture and is described in a separate document.
Currently, the PTX intermediate representation can be obtained by calling
clGetProgramInfo() with CL_PROGRAM_BINARIES and can be passed to
16
NVIDIA OpenCL Programming Guide Version 2.3
2.2.2
Chapter 1
clCreateProgramWithBinary() to create a program object, but this will likely
not be supported in future versions.
Volatile
Only after the execution of barrier(), mem_fence(), read_mem_fence(), or
write_mem_fence() are prior writes to global or shared memory of a given
thread guaranteed to be visible by other threads. As long as this requirement is met,
the compiler is free to optimize reads and writes to global or shared memory. For
example, in the code sample below, the first reference to myArray[tid] compiles
into a global or shared memory read instruction, but the second reference does not
as the compiler simply reuses the result of the first read.
// myArray is an array of non-zero integers
// located in global or shared memory
__kernel void myKernel(__global int* result) {
int tid = get_local_id(0);
int ref1 = myArray[tid] * 1;
myArray[tid + 1] = 2;
int ref2 = myArray[tid] * 1;
result[tid] = ref1 * ref2;
}
Therefore, ref2 cannot possibly be equal to 2 in thread tid as a result of thread
tid-1 overwriting myArray[tid] by 2.
This behavior can be changed using the volatile keyword: If a variable located in
global or shared memory is declared as volatile, the compiler assumes that its value
can be changed at any time by another thread and therefore any reference to this
variable compiles to an actual memory read instruction.
Note that even if myArray is declared as volatile in the code sample above, there is
no guarantee, in general, that ref2 will be equal to 2 in thread tid since thread
tid might read myArray[tid] into ref2 before thread tid-1 overwrites its
value by 2. Synchronization is required as mentioned in Section 3.4.
2.3
Compute Capability
The compute capability of a CUDA device is defined by a major revision number and a
minor revision number.
Devices with the same major revision number are of the same core architecture. The
devices listed in Appendix A are all of compute capability 1.x (Their major revision
number is 1).
The minor revision number corresponds to an incremental improvement to the core
architecture, possibly including new features.
The technical specifications of the various compute capabilities are given in
Appendix A.
NVIDIA OpenCL Programming Guide Version 2.3
17
Chapter 2. 1BOpenCL on the CUDA Architecture
2.4
Mode Switches
GPUs dedicate some DRAM memory to the so-called primary surface, which is used
to refresh the display device whose output is viewed by the user. When users initiate
a mode switch of the display by changing the resolution or bit depth of the display
(using NVIDIA control panel or the Display control panel on Windows), the
amount of memory needed for the primary surface changes. For example, if the user
changes the display resolution from 1280x1024x32-bit to 1600x1200x32-bit, the
system must dedicate 7.68 MB to the primary surface rather than 5.24 MB. (Full-
screen graphics applications running with anti-aliasing enabled may require much
more display memory for the primary surface.) On Windows, other events that may
initiate display mode switches include launching a full-screen DirectX application,
hitting Alt+Tab to task switch away from a full-screen DirectX application, or
hitting Ctrl+Alt+Del to lock the computer.
If a mode switch increases the amount of memory needed for the primary surface,
the system may have to cannibalize memory allocations dedicated to OpenCL
applications. Therefore, a mode switch results in any call to the OpenCL runtime to
fail and return an invalid context error.
2.5
Matrix Multiplication Example
The following matrix multiplication example illustrates the typical data-parallel
approach used by OpenCL applications to achieve good performance on GPUs. It
also illustrates the use of OpenCL local memory that maps to shared memory on
the CUDA architecture. Shared memory is much faster than global memory as
mentioned in Section 2.1.2 and detailed in Section 3.1.2.5, so any opportunity to
replace global memory accesses by shared memory accesses should be exploited.
The following code sample is a straightforward implementation of matrix
multiplication that does not take advantage of shared memory. Each thread reads
one row of A and one column of B and computes the corresponding element of C
as illustrated in Figure 2-4. A is therefore read B.width times from global memory
and B is read A.height times.
// Host code
// Matrices are stored in row-major order:
// M(row, col) = *(M.elements + row * M.width + col)
typedef struct {
int width;
int height;
cl_mem elements;
} Matrix;
// Thread block size
#define BLOCK_SIZE 16
// Matrix multiplication - Host code
// Matrix dimensions are assumed to be multiples of BLOCK_SIZE
void MatMulHost(const Matrix A, const Matrix B, Matrix C,
18
NVIDIA OpenCL Programming Guide Version 2.3
Chapter 1
const cl_context context,
const cl_kernel matMulKernel,
const cl_command_queue queue)
{
// Load A and B to device memory
Matrix d_A;
d_A.width = A.width; d_A.height = A.height;
size_t size = A.width * A.height * sizeof(float);
d_A.elements = clCreateBuffer(context,
CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR,
size, A.elements, 0);
Matrix d_B;
d_B.width = B.width; d_B.height = B.height;
size = B.width * B.height * sizeof(float);
d_B.elements = clCreateBuffer(context,
CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR,
size, B.elements, 0);
// Allocate C in device memory
Matrix d_C;
d_C.width = C.width; d_C.height = C.height;
size = C.width * C.height * sizeof(float);
d_C.elements = clCreateBuffer(context,
CL_MEM_WRITE_ONLY, size, 0, 0);
// Invoke kernel
cl_uint i = 0;
clSetKernelArg(matMulKernel, i++,
sizeof(d_A.width), (void*)&d_A.width);
clSetKernelArg(matMulKernel, i++,
sizeof(d_A.height), (void*)&d_A.height);
clSetKernelArg(matMulKernel, i++,
sizeof(d_A.elements), (void*)&d_A.elements);
clSetKernelArg(matMulKernel, i++,
sizeof(d_B.width), (void*)&d_B.width);
clSetKernelArg(matMulKernel, i++,
sizeof(d_B.height), (void*)&d_B.height);
clSetKernelArg(matMulKernel, i++,
sizeof(d_B.elements), (void*)&d_B.elements);
clSetKernelArg(matMulKernel, i++,
sizeof(d_C.width), (void*)&d_C.width);
clSetKernelArg(matMulKernel, i++,
sizeof(d_C.height), (void*)&d_C.height);
clSetKernelArg(matMulKernel, i++,
sizeof(d_C.elements), (void*)&d_C.elements);
size_t localWorkSize[] = { BLOCK_SIZE, BLOCK_SIZE };
size_t globalWorkSize[] =
{ B.width / dimBlock.x, A.height / dimBlock.y };
clEnqueueNDRangeKernel(queue, matMulKernel, 2, 0,
globalWorkSize, localWorkSize,
0, 0, 0);
// Read C from device memory
clEnqueueReadBuffer(queue, d_C.elements, CL_TRUE, 0, size,
C.elements, 0, 0, 0);
// Free device memory
NVIDIA OpenCL Programming Guide Version 2.3
19
Chapter 2. 1BOpenCL on the CUDA Architecture
clReleaseMemObject(d_A.elements);
clReleaseMemObject(d_C.elements);
clReleaseMemObject(d_B.elements);
}
// Kernel code
// Matrices are stored in row-major order:
// M(row, col) = *(M.elements + row * M.width + col)
typedef struct {
int width;
int height;
__global float* elements;
} Matrix;
// Thread block size
#define BLOCK_SIZE 16
// Matrix multiplication function called by MatMulKernel()
void MatMul(Matrix A, Matrix B, Matrix C)
{
float Cvalue = 0;
int row = get_global_id(1);
int col = get_global_id(0);
for (int e = 0; e < A.width; ++e)
Cvalue += A.elements[row * A.width + e]
* B.elements[e * B.width + col];
C.elements[row * C.width + col] = Cvalue;
}
// Matrix multiplication kernel called by MatMulHost()
__kernel void MatMulKernel(
int Awidth, int Aheight, __global float* Aelements,
int Bwidth, int Bheight, __global float* Belements,
int Cwidth, int Cheight, __global float* Celements)
{
Matrix A = { Awidth, Aheight, Aelements };
Matrix B = { Bwidth, Bheight, Belements };
Matrix C = { Cwidth, Cheight, Celements };
matrixMul(A, B, C);
}
20
NVIDIA OpenCL Programming Guide Version 2.3
Chapter 1
1
-
h
t
d
w
B
.
i
col
0
B
C
t
h
g
i
e
h
B
.
t
h
g
i
e
h
A
.
A.width
B.width
0
A
row
A.height-1
Figure 2-4. Matrix Multipliation without Shared Memory
The following code sample is an implementation of matrix multiplication that does
take advantage of shared memory. In this implementation, each thread block is
responsible for computing one square sub-matrix Csub of C and each thread within
the block is responsible for computing one element of Csub. As illustrated in Figure
2-5, Csub is equal to the product of two rectangular matrices: the sub-matrix of A of
dimension (A.width, block_size) that has the same line indices as Csub, and the sub-
matrix of B of dimension (block_size, A.width) that has the same column indices as
Csub. In order to fit into the device’s resources, these two rectangular matrices are
divided into as many square matrices of dimension block_size as necessary and Csub is
computed as the sum of the products of these square matrices. Each of these
products is performed by first loading the two corresponding square matrices from
global memory to shared memory with one thread loading one element of each
matrix, and then by having each thread compute one element of the product. Each
thread accumulates the result of each of these products into a register and once
done writes the result to global memory.
NVIDIA OpenCL Programming Guide Version 2.3
21
Chapter 2. 1BOpenCL on the CUDA Architecture
By blocking the computation this way, we take advantage of fast shared memory
and save a lot of global memory bandwidth since A is only read (B.width / block_size)
times from global memory and B is read (A.height / block_size) times.
The Matrix type from the previous code sample is augmented with a stride field, so
that sub-matrices can be efficiently represented with the same type.
// Host code
// Matrices are stored in row-major order:
// M(row, col) = *(M.elements + row * M.stride + col)
typedef struct {
int width;
int height;
int stride;
cl_mem elements;
} Matrix;
// Thread block size
#define BLOCK_SIZE 16
// Matrix multiplication - Host code
// Matrix dimensions are assumed to be multiples of BLOCK_SIZE
void MatMulHost(const Matrix A, const Matrix B, Matrix C,
const cl_context context,
const cl_kernel matMulKernel,
const cl_command_queue queue)
{
// Load A and B to device memory
Matrix d_A;
d_A.width = d_A.stride = A.width; d_A.height = A.height;
size_t size = A.width * A.height * sizeof(float);
d_A.elements = clCreateBuffer(context,
CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR,
size, A.elements, 0);
Matrix d_B;
d_B.width = d_B.stride = B.width; d_B.height = B.height;
size = B.width * B.height * sizeof(float);
d_B.elements = clCreateBuffer(context,
CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR,
size, B.elements, 0);
// Allocate C in device memory
Matrix d_C;
d_C.width = d_C.stride = C.width; d_C.height = C.height;
size = C.width * C.height * sizeof(float);
d_C.elements = clCreateBuffer(context,
CL_MEM_WRITE_ONLY, size, 0, 0);
// Invoke kernel
cl_uint i = 0;
clSetKernelArg(matMulKernel, i++,
sizeof(d_A.width), (void*)&d_A.width);
clSetKernelArg(matMulKernel, i++,
sizeof(d_A.height), (void*)&d_A.height);
clSetKernelArg(matMulKernel, i++,
22
NVIDIA OpenCL Programming Guide Version 2.3
Chapter 1
sizeof(d_A.stride), (void*)&d_A.stride);
clSetKernelArg(matMulKernel, i++,
sizeof(d_A.elements), (void*)&d_A.elements);
clSetKernelArg(matMulKernel, i++,
sizeof(d_B.width), (void*)&d_B.width);
clSetKernelArg(matMulKernel, i++,
sizeof(d_B.height), (void*)&d_B.height);
clSetKernelArg(matMulKernel, i++,
sizeof(d_B. stride), (void*)&d_B.stride);
clSetKernelArg(matMulKernel, i++,
sizeof(d_B.elements), (void*)&d_B.elements);
clSetKernelArg(matMulKernel, i++,
sizeof(d_C.width), (void*)&d_C.width);
clSetKernelArg(matMulKernel, i++,
sizeof(d_C.height), (void*)&d_C.height);
clSetKernelArg(matMulKernel, i++,
sizeof(d_C.stride), (void*)&d_C.stride);
clSetKernelArg(matMulKernel, i++,
sizeof(d_C.elements), (void*)&d_C.elements);
size_t localWorkSize[] = { BLOCK_SIZE, BLOCK_SIZE };
size_t globalWorkSize[] =
{ B.width / dimBlock.x, A.height / dimBlock.y };
clEnqueueNDRangeKernel(queue, matMulKernel, 2, 0,
globalWorkSize, localWorkSize,
0, 0, 0);
// Read C from device memory
clEnqueueReadBuffer(queue, d_C.elements, CL_TRUE, 0, size,
C.elements, 0, 0, 0);
// Free device memory
clReleaseMemObject(d_A.elements);
clReleaseMemObject(d_C.elements);
clReleaseMemObject(d_B.elements);
}
// Kernel code
// Matrices are stored in row-major order:
// M(row, col) = *(M.elements + row * M.stride + col)
typedef struct {
int width;
int height;
int stride;
__global float* elements;
} Matrix;
// Thread block size
#define BLOCK_SIZE 16
// Get a matrix element
float GetElement(const Matrix A, int row, int col)
{
return A.elements[row * A.stride + col];
}
NVIDIA OpenCL Programming Guide Version 2.3
23
Chapter 2. 1BOpenCL on the CUDA Architecture
// Set a matrix element
void SetElement(Matrix A, int row, int col, float value)
{
A.elements[row * A.stride + col] = value;
}
// Get the BLOCK_SIZExBLOCK_SIZE sub-matrix Asub of A that is
// located col sub-matrices to the right and row sub-matrices down
// from the upper-left corner of A
Matrix GetSubMatrix(Matrix A, int row, int col)
{
Matrix Asub;
Asub.width = BLOCK_SIZE;
Asub.height = BLOCK_SIZE;
Asub.stride = A.stride;
Asub.elements =
&A.elements[A.stride * BLOCK_SIZE * row + BLOCK_SIZE * col];
return Asub;
}
// Matrix multiplication function called by MatMulKernel()
void MatMul(Matrix C, Matrix A, Matrix B,
__local float As[BLOCK_SIZE][BLOCK_SIZE],
__local float Bs[BLOCK_SIZE][BLOCK_SIZE])
{
// Block row and column
int blockRow = get_group_id(1);
int blockCol = get_group_id(0);
// Each thread block computes one sub-matrix Csub of C
Matrix Csub = GetSubMatrix(C, blockRow, blockCol);
// Each thread computes one element of Csub
// by accumulating results into Cvalue
float Cvalue = 0;
// Thread row and column within Csub
int row = get_local_id(1);
int col = get_local_id(0);
// Loop over all the sub-matrices of A and B that are
// required to compute Csub
// Multiply each pair of sub-matrices together
// and accumulate the results
for (int m = 0; m < (A.width / BLOCK_SIZE); ++m) {
// Get sub-matrix Asub of A
Matrix Asub = GetSubMatrix(A, blockRow, m);
// Get sub-matrix Bsub of B
Matrix Bsub = GetSubMatrix(B, m, blockCol);
// Load Asub and Bsub from device memory to shared memory
// Each thread loads one element of each sub-matrix
As[row][col] = GetElement(Asub, row, col);
Bs[row][col] = GetElement(Bsub, row, col);
24
NVIDIA OpenCL Programming Guide Version 2.3
Chapter 1
// Synchronize to make sure the sub-matrices are loaded
// before starting the computation
barrier(CLK_LOCAL_MEM_FENCE);
// Multiply Asub and Bsub together
for (int e = 0; e < BLOCK_SIZE; ++e)
Cvalue += As[row][e] * Bs[e][col];
// Synchronize to make sure that the preceding
// computation is done before loading two new
// sub-matrices of A and B in the next iteration
barrier(CLK_LOCAL_MEM_FENCE);
}
// Write Csub to device memory
// Each thread writes one element
SetElement(Csub, row, col, Cvalue);
}
// Matrix multiplication kernel called by MatMulHost()
__kernel void matrixMulKernel(
int Cwidth, int Cheight, int Cstride, __global float* Celements,
int Awidth, int Aheight, int Astride, __global float* Aelements,
int Bwidth, int Bheight, int Bstride, __global float* Belements,
__local float As[BLOCK_SIZE][BLOCK_SIZE],
__local float Bs[BLOCK_SIZE][BLOCK_SIZE])
{
Matrix C = { Cwidth, Cheight, Cstride, Celements };
Matrix A = { Awidth, Aheight, Astride, Aelements };
Matrix B = { Bwidth, Bheight, Bstride, Belements };
MatMul(A, B, C, As, Bs);
}
NVIDIA OpenCL Programming Guide Version 2.3
25
Chapter 2. 1BOpenCL on the CUDA Architecture
B
C
blockCol
1
-
E
Z
I
S
_
K
C
O
L
B
0
col
0
Csub
row
BLOCK_SIZE-1
E
Z
I
S
_
K
C
O
L
B
E
Z
I
S
_
K
C
O
L
B
E
Z
I
S
_
K
C
O
L
B
t
h
g
i
e
h
B
.
t
h
g
i
e
h
A
.
A
w
o
R
k
c
o
b
l
BLOCK_SIZE
BLOCK_SIZE
BLOCK_SIZE
A.width
B.width
Figure 2-5. Matrix Multipliation with Shared Memory
26
NVIDIA OpenCL Programming Guide Version 2.3
Chapter 3.
Performance Guidelines
3.1
Instruction Performance
To process an instruction for a warp of threads, a multiprocessor must:
(cid:137) Read the instruction operands for each thread of the warp,
(cid:137) Execute the instruction,
(cid:137) Write the result for each thread of the warp.
Therefore, the effective instruction throughput depends on the nominal instruction
throughput as well as the memory latency and bandwidth. It is maximized by:
(cid:137) Minimizing the use of instructions with low throughput (see Section 3.1.1),
(cid:137) Maximizing the use of the available memory bandwidth for each category of
memory (see Section 3.1.2),
(cid:137) Allowing the thread scheduler to overlap memory transactions with
mathematical computations as much as possible, which requires that:
(cid:190) The program executed by the threads is of high arithmetic intensity, that is,
has a high number of arithmetic operations per memory operation;
(cid:190) There are many active threads per multiprocessor, as detailed in Section 3.2.
3.1.1
3.1.1.1
Instruction Throughput
In this section, throughputs are given in number of operations per clock cycle per
multiprocessor. For a warp size of 32, an instruction is made of 32 operations.
Therefore, if T is the number of operations per clock cycle, the instruction
throughput is one instruction every 32/T clock cycles.
All throughputs are for one multiprocessor. They must be multiplied by the number
of multiprocessors in the device to get throughput for the whole device.
Arithmetic Instructions
For single-precision floating-point code, we highly recommend use of the float
type and the single-precision floating-point mathematical functions. When
compiling for devices without native double-precision floating-point support, such
as devices of compute capability 1.2 and lower, each double variable gets
converted to single-precision floating-point format (but retains its size of 64 bits)
Chapter 3. Performance Guidelines
and double-precision floating-point arithmetic gets demoted to single-precision
floating-point arithmetic.
We also recommend using native_* functions wherever possible and the
-cl-mad-enable build option, both of them can lead to large performance gains.
Single-Precision Floating-Point Basic Arithmetic
Throughput of single-precision floating-point add, multiply, and multiply-add is 8
operations per clock cycle.
Throughput of reciprocal is 2 operations per clock cycle.
Throughput of single-precision floating-point division is 0.88 operations per clock
cycle, but native_divide(x, y) provides a faster version with a throughput of
1.6 operations per clock cycle.
Single-Precision Floating-Point Square Root and Reciprocal Square
Root
Throughput of reciprocal square root is 2 operations per clock cycle.
Single-precision floating-point square root is implemented as a reciprocal square
root followed by a reciprocal instead of a reciprocal square root followed by a
multiplication, so that it gives correct results for 0 and infinity. Therefore, its
throughput is 1 operation per clock cycle.
Single-Precision Floating-Point Logarithm
Throughput of native_log(x) (see Section B.2) is 2 operations per clock cycle.
Sine and Cosine
Throughput of native_sin(x), native_cos(x), native_exp(x) is 1
operation per clock cycle.
sin(x), cos(x), tan(x), sincos(x) are much more expensive and even more
so if the absolute value of x needs to be reduced.
More precisely, the argument reduction code comprises two code paths referred to
as the fast path and the slow path, respectively.
The fast path is used for arguments sufficiently small in magnitude and essentially
consists of a few multiply-add operations. The slow path is used for arguments large
in magnitude, and consists of lengthy computations required to achieve correct
results over the entire argument range.
At present, the argument reduction code for the trigonometric functions selects the
fast path for arguments whose magnitude is less than 48039.0f for the single-
precision functions, and less than 2147483648.0 for the double-precision functions.
As the slow path requires more registers than the fast path, an attempt has been
made to reduce register pressure in the slow path by storing some intermediate
variables in CUDA local memory, which may affect performance because of local
memory high latency and bandwidth (see Section 3.1.2.2). At present, 28 bytes of
CUDA local memory are used by single-precision functions, and 44 bytes are used
by double-precision functions. However, the exact amount is subject to change.
28
NVIDIA OpenCL Programming Guide Version 2.3
Chapter 3. Performance Guidelines
Due to the lengthy computations and use of CUDA local memory in the slow path,
the trigonometric functions throughput is lower by one order of magnitude when
the slow path reduction is used as opposed to the fast path reduction.
Integer Arithmetic
Throughput of integer add is 8 operations per clock cycle.
Throughput of 32-bit integer multiplication is 2 operations per clock cycle, but
mul24 provide 24-bit integer multiplication with a troughput of 8 operations per
clock cycle. On future architectures however, mul24 will be slower than 32-bit
integer multiplication, so we recommend to provide two kernels, one using mul24
and the other using generic 32-bit integer multiplication, to be called appropriately
by the application.
Integer division and modulo operation are particularly costly and should be avoided
if possible or replaced with bitwise operations whenever possible: If n is a power of
2, (i/n) is equivalent to (i>>log2(n)) and (i%n) is equivalent to (i&(n-1));
the compiler will perform these conversions if n is literal.
Comparison
Throughput of compare, min, max is 8 operations per clock cycle.
Bitwise Operations
Throughput of any bitwise operation is 8 operations per clock cycle.
Type Conversion
Throughput of type conversion operations is 8 operations per clock cycle.
Sometimes, the compiler must insert conversion instructions, introducing additional
execution cycles. This is the case for:
(cid:137) Functions operating on char or short whose operands generally need to be
converted to int,
(cid:137) Double-precision floating-point constants (defined without any type suffix) used
as input to single-precision floating-point computations.
This last case can be avoided by using single-precision floating-point constants,
defined with an f suffix such as 3.141592653589793f, 1.0f, 0.5f.
Control Flow Instructions
Any flow control instruction (if, switch, do, for, while) can significantly
impact the effective instruction throughput by causing threads of the same warp to
diverge, that is, to follow different execution paths. If this happens, the different
executions paths have to be serialized, increasing the total number of instructions
executed for this warp. When all the different execution paths have completed, the
threads converge back to the same execution path.
To obtain best performance in cases where the control flow depends on the thread
ID, the controlling condition should be written so as to minimize the number of
divergent warps. This is possible because the distribution of the warps across the
block is deterministic as mentioned in Section 2.1.1. A trivial example is when the
controlling condition only depends on (get_local_id(0) / WSIZE) where
WSIZE is the warp size. In this case, no warp diverges since the controlling
condition is perfectly aligned with the warps.
3.1.1.2
NVIDIA OpenCL Programming Guide Version 2.3
29
Chapter 3. Performance Guidelines
Sometimes, the compiler may unroll loops or it may optimize out if or switch
statements by using branch predication instead, as detailed below. In these cases, no
warp can ever diverge.
When using branch predication none of the instructions whose execution depends
on the controlling condition gets skipped. Instead, each of them is associated with a
per-thread condition code or predicate that is set to true or false based on the
controlling condition and although each of these instructions gets scheduled for
execution, only the instructions with a true predicate are actually executed.
Instructions with a false predicate do not write results, and also do not evaluate
addresses or read operands.
The compiler replaces a branch instruction with predicated instructions only if the
number of instructions controlled by the branch condition is less or equal to a
certain threshold: If the compiler determines that the condition is likely to produce
many divergent warps, this threshold is 7, otherwise it is 4.
3.1.1.3
Memory Instructions
Memory instructions include any instruction that reads from or writes to CUDA
shared, local, or global memory.
Throughput of memory operations is 8 operations per clock cycle. When accessing
CUDA local or global memory, there are, in addition, 400 to 600 clock cycles of
memory latency.
As an example, the throughput for the assignment operator in the following sample
code:
__local float shared[32];
__global float device[32];
shared[threadIdx.x] = device[threadIdx.x];
is 8 operations per clock cycle for the read from global memory, 8 operations per
clock cycle for the write to shared memory, but above all, there is a latency of 400 to
600 clock cycles to read data from global memory.
Much of this global memory latency can be hidden by the thread scheduler if there
are sufficient independent arithmetic instructions that can be issued while waiting
for the global memory access to complete.
3.1.1.4
Synchronization Instruction
Throughput for the barrier function is 8 operations per clock cycle in the case
where no thread has to wait for any other threads.
3.1.2
Memory Bandwidth
The effective bandwidth of each memory space depends significantly on the
memory access pattern as detailed in the following sub-sections.
Since device memory is of much higher latency and lower bandwidth than on-chip
memory, device memory accesses should be minimized. A typical programming
pattern is to stage data coming from device memory into shared memory; in other
words, to have each thread of a block:
(cid:137) Load data from device memory to shared memory,
30
NVIDIA OpenCL Programming Guide Version 2.3
Chapter 3. Performance Guidelines
(cid:137) Synchronize with all the other threads of the block so that each thread can safely
read shared memory locations that were written by different threads,
(cid:137) Process the data in shared memory,
(cid:137) Synchronize again if necessary to make sure that shared memory has been
updated with the results,
(cid:137) Write the results back to device memory.
3.1.2.1
Global Memory
Global memory is not cached, so it is all the more important to follow the right
access pattern to get maximum memory bandwidth, especially given how costly
accesses to device memory are.
First, the device is capable of reading 4-byte, 8-byte, or 16-byte words from global
memory into registers in a single instruction. To have assignments such as:
__global type device[32];
type data = device[tid];
compile to a single load instruction, type must be such that sizeof(type) is
equal to 4, 8, or 16 and variables of type type must be aligned to sizeof(type)
bytes (that is, have their address be a multiple of sizeof(type)).
The alignment requirement is automatically fulfilled for built-in types.
For structures, the size and alignment requirements can be enforced by the compiler
using the alignment specifiers __attribute__ ((aligned(8))) or
__attribute__ ((aligned(16))), such as
struct {
float a;
float b;
} __attribute__ ((aligned(8)));
or
struct {
float a;
float b;
float c;
} __attribute__ ((aligned(16)));
For structures larger than 16 bytes, the compiler generates several load instructions.
To ensure that it generates the minimum number of instructions, such structures
should be defined with __attribute__ ((aligned(16))) , such as
struct {
float a;
float b;
float c;
float d;
float e;
} __attribute__ ((aligned(16)));
which is compiled into two 16-byte load instructions instead of five 4-byte load
instructions.
Any address of a variable residing in global memory or returned by one of the
memory allocation routines from the driver or runtime API is always aligned to at
least 256 bytes.
NVIDIA OpenCL Programming Guide Version 2.3
31
Chapter 3. Performance Guidelines
Second, global memory bandwidth is used most efficiently when the simultaneous
memory accesses by threads in a half-warp (during the execution of a single read or
write instruction) can be coalesced into a single memory transaction of 32, 64, or 128
bytes.
The rest of this section describes the various requirements for memory accesses to
coalesce based on the compute capability of the device. If a half-warp fulfills these
requirements, coalescing is achieved even if the warp is divergent and some threads
of the half-warp do not actually access memory.
For the purpose of the following discussion, global memory is considered to be
partitioned into segments of size equal to 32, 64, or 128 bytes and aligned to this
size.
Coalescing on Devices with Compute Capability 1.0 and 1.1
The global memory access by all threads of a half-warp is coalesced into one or two
memory transactions if it satisfies the following three conditions:
(cid:137) Threads must access
(cid:137) Either 4-byte words, resulting in one 64-byte memory transaction,
(cid:137) Or 8-byte words, resulting in one 128-byte memory transaction,
(cid:137) Or 16-byte words, resulting in two 128-byte memory transactions;
(cid:137) All 16 words must lie in the same segment of size equal to the memory
transaction size (or twice the memory transaction size when accessing 16-byte
words);
(cid:137) Threads must access the words in sequence: The kth thread in the half-warp must
access the kth word.
If a half-warp does not fulfill all the requirements above, a separate memory
transaction is issued for each thread and throughput is significantly reduced.
Figure 3-1 shows some examples of coalesced memory accesses, while Figure 3-2
and Figure 3-3 show some examples of memory accesses that are non-coalesced for
devices of compute capability 1.0 or 1.1.
Coalesced 8-byte accesses deliver a little lower bandwidth than coalesced 4-byte
accesses and coalesced 16-byte accesses deliver a noticeably lower bandwidth than
coalesced 4-byte accesses. But, while bandwidth for non-coalesced accesses is
around an order of magnitude lower than for coalesced accesses when these
accesses are 4-byte, it is only around four times lower when they are 8-byte and
around two times when they are 16-byte.
Coalescing on Devices with Compute Capability 1.2 and Higher
The global memory access by all threads of a half-warp is coalesced into a single
memory transaction as soon as the words accessed by all threads lie in the same
segment of size equal to:
(cid:137) 32 bytes if all threads access 1-byte words,
(cid:137) 64 bytes if all threads access 2-byte words,
(cid:137) 128 bytes if all threads access 4-byte or 8-byte words.
Coalescing is achieved for any pattern of addresses requested by the half-warp,
including patterns where multiple threads access the same address. This is in
32
NVIDIA OpenCL Programming Guide Version 2.3
Chapter 3. Performance Guidelines
contrast with devices of lower compute capabilities where threads need to access
words in sequence.
If a half-warp addresses words in n different segments, n memory transactions are
issued (one for each segment), whereas devices with lower compute capabilities
would issue 16 transactions as soon as n is greater than 1. In particular, if threads
access 16-byte words, at least two memory transactions are issued.
Unused words in a memory transaction are still read, so they waste bandwidth. To
reduce waste, hardware will automatically issue the smallest memory transaction that
contains the requested words. For example, if all the requested words lie in one half
of a 128-byte segment, a 64-byte transaction will be issued.
More precisely, the following protocol is used to issue a memory transaction for a
half-warp:
(cid:137) Find the memory segment that contains the address requested by the lowest
numbered active thread. Segment size is 32 bytes for 1-byte data, 64 bytes for
2-byte data, 128 bytes for 4-, 8- and 16-byte data.
(cid:137) Find all other active threads whose requested address lies in the same segment.
(cid:137) Reduce the transaction size, if possible:
(cid:137) If the transaction size is 128 bytes and only the lower or upper half is used,
reduce the transaction size to 64 bytes;
(cid:137) If the transaction size is 64 bytes and only the lower or upper half is used,
reduce the transaction sizez to 32 bytes.
(cid:137) Carry out the transaction and mark the serviced threads as inactive.
(cid:137) Repeat until all threads in the half-warp are serviced.
Figure 3-4 shows some examples of global memory accesses for devices of compute
capability 1.2 and higher.
NVIDIA OpenCL Programming Guide Version 2.3
33
Chapter 3. Performance Guidelines
Thread 0
Address 128
Thread 0
Address 128
Thread 1
Address 132
Thread 1
Address 132
Thread 2
Address 136
Thread 2
Address 136
Thread 3
Address 140
Thread 3
Address 140
Thread 4
Address 144
Thread 4
Address 144
Thread 5
Address 148
Thread 5
Address 148
Thread 6
Address 152
Thread 6
Address 152
Thread 7
Address 156
Thread 7
Address 156
Thread 8
Address 160
Thread 8
Address 160
Thread 9
Address 164
Thread 9
Address 164
Thread 10
Address 168
Thread 10
Address 168
Thread 11
Address 172
Thread 11
Address 172
Thread 12
Address 176
Thread 12
Address 176
Thread 13
Address 180
Thread 13
Address 180
Thread 14
Address 184
Thread 14
Address 184
Thread 15
Address 188
Thread 15
Address 188
Left: coalesced float memory access, resulting in a single memory transaction.
Right: coalesced float memory access (divergent warp), resulting in a single memory transaction.
Figure 3-1. Examples of Coalesced Global Memory Access
Patterns
34
NVIDIA OpenCL Programming Guide Version 2.3
Chapter 3. Performance Guidelines
Thread 0
Address 128
Thread 0
Address 128
Thread 1
Address 132
Thread 1
Address 132
Thread 2
Address 136
Thread 2
Address 136
Thread 3
Address 140
Thread 3
Address 140
Thread 4
Address 144
Thread 4
Address 144
Thread 5
Address 148
Thread 5
Address 148
Thread 6
Address 152
Thread 6
Address 152
Thread 7
Address 156
Thread 7
Address 156
Thread 8
Address 160
Thread 8
Address 160
Thread 9
Address 164
Thread 9
Address 164
Thread 10
Address 168
Thread 10
Address 168
Thread 11
Address 172
Thread 11
Address 172
Thread 12
Address 176
Thread 12
Address 176
Thread 13
Address 180
Thread 13
Address 180
Thread 14
Address 184
Thread 14
Address 184
Thread 15
Address 188
Thread 15
Address 188
Left: non-sequential float memory access, resulting in 16 memory transactions.
Right: access with a misaligned starting address, resulting in 16 memory transactions.
Figure 3-2. Examples of Global Memory Access Patterns That
Are Non-Coalesced for Devices of Compute
Capability 1.0 or 1.1
NVIDIA OpenCL Programming Guide Version 2.3
35
Chapter 3. Performance Guidelines
Thread 0
Address 128
Thread 0
Address 128
Thread 1
Address 132
Thread 2
Address 136
Thread 3
Address 140
Thread 1
Address 140
Thread 4
Address 144
Thread 5
Address 148
Thread 6
Address 152
Thread 2
Address 152
Thread 7
Address 156
Thread 8
Address 160
Thread 9
Address 164
Thread 3
Address 164
Thread 10
Address 168
Thread 11
Address 172
Thread 12
Address 176
Thread 4
Address 176
Thread 13
Address 180
Thread 14
Address 184
Thread 15
Address 188
Thread 5
Address 188
Left: non-contiguous float memory access, resulting in 16 memory transactions.
Right: non-coalesced float3 memory access, resulting in 16 memory transactions.
Figure 3-3. Examples of Global Memory Access Patterns That
Are Non-Coalesced for Devices of Compute
Capability 1.0 or 1.1
36
NVIDIA OpenCL Programming Guide Version 2.3
Chapter 3. Performance Guidelines
6
4
B
s
e
g
m
e
n
t
Thread
0
Thread
1
Thread
2
Thread
3
Thread
4
Thread
5
Thread
6
Thread
7
Thread
8
Thread
9
Thread
10
Thread
11
Thread
12
Thread
13
Thread
14
Thread
15
Thread
0
Thread
1
Thread
2
Thread
3
Thread
4
Thread
5
Thread
6
Thread
7
Thread
8
Thread
9
Thread
10
Thread
11
Thread
12
Thread
13
Thread
14
Thread
15
Address
120
Address
124
Address
128
Address
132
Address
136
Address
140
Address
144
Address
148
Address
152
Address
156
Address
160
Address
164
Address
168
Address
172
Address
176
Address
180
Address
184
Address
188
Address
192
Address
196
Address
200
Address
204
Address
208
Address
212
Address
120
Address
124
Address
128
Address
132
Address
136
Address
140
Address
144
Address
148
Address
152
Address
156
Address
160
Address
164
Address
168
Address
172
Address
176
Address
180
Address
184
Address
188
Address
192
Address
19 6
…
Address
204
Address
252
Address
256
1
2
8
B
s
e
g
m
e
n
t
Thread
0
Thread
1
Thread
2
Thread
3
Thread
4
Thread
5
Thread
6
Thread
7
Thread
8
Thread
9
Thread
10
Thread
11
Thread
12
Thread
13
Thread
14
Thread
15
3
2
B
s
e
g
m
e
n
t
6
4
B
s
e
g
m
e
n
t
Address
96
Address
100
Address
104
Address
108
Address
112
Address
116
Address
120
Address
124
Address
128
Address
132
Address
136
Address
140
Address
144
Address
148
Address
152
Address
156
Address
160
Address
164
Address
168
Address
172
Address
176
Address
180
Address
184
Address
188
Left: random float memory access within a 64B segment, resulting in one memory transaction.
Center: misaligned float memory access, resulting in one transaction.
Right: misaligned float memory access, resulting in two transactions.
Figure 3-4. Examples of Global Memory Access by Devices
with Compute Capability 1.2 and Higher
NVIDIA OpenCL Programming Guide Version 2.3
37
Chapter 3. Performance Guidelines
Common Access Patterns
Array of Structures
A common global memory access pattern is when each thread of thread ID tid
accesses one element of an array located at address BaseAddress of type type*
using the following address:
BaseAddress + tid
To get memory coalescing, type must meet the size and alignment requirements
discussed above. In particular, this means that if type is a structure larger than 16
bytes, it should be split into several structures that meet these requirements and the
data should be laid out in memory as a list of several arrays of these structures
instead of a single array of type type*.
Two-Dimensional Array
Another common global memory access pattern is when each thread of index
(tx,ty) accesses one element of a 2D array located at address BaseAddress of
type type* and of width width using the following address:
BaseAddress + width * ty + tx
In such a case, one gets memory coalescing for all half-warps of the thread block
only if:
(cid:137) The width of the thread block is a multiple of half the warp size;
(cid:137) width is a multiple of 16.
In particular, this means that an array whose width is not a multiple of 16 will be
accessed much more efficiently if it is actually allocated with a width rounded up to
the closest multiple of 16 and its rows padded accordingly.
Local Memory
Like global memory, CUDA local memory is not cached, so accesses to local
memory are as expensive as accesses to global memory. Local memory accesses are
always coalesced though since they are per-thread by definition.
Constant Memory
Constant memory is cached so a read from constant memory costs one memory
read from device memory only on a cache miss, otherwise it just costs one read
from the constant cache.
For all threads of a half-warp, reading from the constant cache is as fast as reading
from a register as long as all threads read the same address. The cost scales linearly
with the number of different addresses read by all threads. We recommend having
all threads of the entire warp read the same address as opposed to all threads within
each of its halves only, as future devices will require it for full speed read.
Texture Memory
Texture memory is cached so an image read costs one memory read from device
memory only on a cache miss, otherwise it just costs one read from the texture
cache. The texture cache is optimized for 2D spatial locality, so threads of the same
warp that read image addresses that are close together will achieve best
performance. Also, it is designed for streaming reads with a constant latency, i.e. a
cache hit reduces DRAM bandwidth demand, but not read latency.
3.1.2.2
3.1.2.3
3.1.2.4
38
NVIDIA OpenCL Programming Guide Version 2.3
Chapter 3. Performance Guidelines
Reading device memory through image objects present some benefits that can make
it an advantageous alternative to reading device memory from global or constant
memory:
(cid:137) If the memory reads do not follow the access patterns that global or constant
memory reads must respect to get good performance (see Sections 3.1.2.1 and
3.1.2.3), higher bandwidth can be achieved providing that there is locality in the
image reads;
(cid:137) The latency of addressing calculations is hidden better, possibly improving
performance for applications that perform random accesses to the data;
(cid:137) Packed data may be broadcast to separate variables in a single operation;
(cid:137) 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].
However, within the same kernel call, the texture cache is not kept coherent with
respect to image writes, so that any image read to an address that has been written
to via an image write in the same kernel call returns undefined data. In other words,
a thread can safely read via an image object some memory location only if this
memory location has been updated by a previous kernel call or memory copy, but
not if it has been previously updated by the same thread or another thread from the
same kernel call.
3.1.2.5
Shared Memory
Shared memory is where OpenCL local memory resides.
Because it is on-chip, shared memory is much faster than local and global memory.
In fact, for all threads of a warp, accessing shared memory is as fast as accessing a
register as long as there are no bank conflicts between the threads, as detailed below.
To achieve high memory bandwidth, shared memory is divided into equally-sized
memory modules, called banks, which can be accessed simultaneously. So, any
memory read or write request made of n addresses that fall in n distinct memory
banks can be serviced simultaneously, yielding an effective bandwidth that is n times
as high as the bandwidth of a single module.
However, if two addresses of a memory request fall in the same memory bank, there
is a bank conflict and the access has to be serialized. The hardware splits a memory
request with bank conflicts into as many separate conflict-free requests as necessary,
decreasing the effective bandwidth by a factor equal to the number of separate
memory requests. If the number of separate memory requests is n, the initial
memory request is said to cause n-way bank conflicts.
To get maximum performance, it is therefore important to understand how memory
addresses map to memory banks in order to schedule the memory requests so as to
minimize bank conflicts.
In the case of shared memory, the banks are organized such that successive 32-bit
words are assigned to successive banks and each bank has a bandwidth of 32 bits
per two clock cycles.
For devices of compute capability 1.x, the warp size is 32 and the number of banks
is 16 (see Section 5.1); a shared memory request for a warp is split into one request
for the first half of the warp and one request for the second half of the warp. As a
consequence, there can be no bank conflict between a thread belonging to the first
half of a warp and a thread belonging to the second half of the same warp.
NVIDIA OpenCL Programming Guide Version 2.3
39
Chapter 3. Performance Guidelines
A common case is for each thread to access a 32-bit word from an array indexed by
the thread ID tid and with some stride s:
__local float shared[32];
float data = shared[BaseIndex + s * tid];
In this case, the threads tid and tid+n access the same bank whenever s*n is a
multiple of the number of banks m or equivalently, whenever n is a multiple of m/d
where d is the greatest common divisor of m and s. As a consequence, there will be
no bank conflict only if half the warp size is less than or equal to m/d. For devices
of compute capability 1.x, this translates to no bank conflict only if d is equal to 1,
or in other words, only if s is odd since m is a power of two.
Figure 3-5 and Figure 3-6 show some examples of conflict-free memory accesses
while Figure 3-7 shows some examples of memory accesses that cause bank
conflicts.
Other cases worth mentioning are when each thread accesses an element that is
smaller or larger than 32 bits in size. For example, there are bank conflicts if an array
of char is accessed the following way:
__local char shared[32];
char data = shared[BaseIndex + tid];
because shared[0], shared[1], shared[2], and shared[3], for example,
belong to the same bank. There are no bank conflicts however, if the same array is
accessed the following way:
char data = shared[BaseIndex + 4 * tid];
There are also 2-way bank conflicts for arrays of double:
__local double shared[32];
double data = shared[BaseIndex + tid];
since the memory request is compiled into two separate 32-bit requests. One way to
avoid bank conflicts in this case is two split the double operands like in the
following sample code:
__local int shared_lo[32];
__local int shared_hi[32];
double dataIn;
shared_lo[BaseIndex + tid] = __double2loint(dataIn);
shared_hi[BaseIndex + tid] = __double2hiint(dataIn);
double dataOut =
__hiloint2double(shared_hi[BaseIndex + tid],
shared_lo[BaseIndex + tid]);
It might not always improve performance though and will perform worse on future
architectures.
A structure assignment is compiled into as many memory requests as necessary for
each member in the structure, so the following code, for example:
__local struct type shared[32];
struct type data = shared[BaseIndex + tid];
results in:
(cid:137) Three separate memory reads without bank conflicts if type is defined as
struct type {
40
NVIDIA OpenCL Programming Guide Version 2.3
Chapter 3. Performance Guidelines
float x, y, z;
};
since each member is accessed with a stride of three 32-bit words;
(cid:137) Two separate memory reads with bank conflicts if type is defined as
struct type {
float x, y;
};
since each member is accessed with a stride of two 32-bit words;
(cid:137) Two separate memory reads with bank conflicts if type is defined as
struct type {
float f;
char c;
};
since each member is accessed with a stride of five bytes.
Finally, shared memory also features a broadcast mechanism whereby a 32-bit word
can be read and broadcast to several threads simultaneously when servicing one
memory read request. This reduces the number of bank conflicts when several
threads of a half-warp read from an address within the same 32-bit word. More
precisely, a memory read request made of several addresses is serviced in several
steps over time – one step every two clock cycles – by servicing one conflict-free
subset of these addresses per step until all addresses have been serviced; at each
step, the subset is built from the remaining addresses that have yet to be serviced
using the following procedure:
(cid:137) Select one of the words pointed to by the remaining addresses as the broadcast
word,
(cid:137) Include in the subset:
(cid:137) All addresses that are within the broadcast word,
(cid:137) One address for each bank pointed to by the remaining addresses.
Which word is selected as the broadcast word and which address is picked up for
each bank at each cycle are unspecified.
A common conflict-free case is when all threads of a half-warp read from an address
within the same 32-bit word.
Figure 3-8 shows some examples of memory read accesses that involve the
broadcast mechanism.
NVIDIA OpenCL Programming Guide Version 2.3
41
Chapter 3. Performance Guidelines
Thread 0
Bank 0
Thread 0
Bank 0
Thread 1
Bank 1
Thread 1
Bank 1
Thread 2
Bank 2
Thread 2
Bank 2
Thread 3
Bank 3
Thread 3
Bank 3
Thread 4
Bank 4
Thread 4
Bank 4
Thread 5
Bank 5
Thread 5
Bank 5
Thread 6
Bank 6
Thread 6
Bank 6
Thread 7
Bank 7
Thread 7
Bank 7
Thread 8
Bank 8
Thread 8
Bank 8
Thread 9
Bank 9
Thread 9
Bank 9
Thread 10
Bank 10
Thread 10
Bank 10
Thread 11
Bank 11
Thread 11
Bank 11
Thread 12
Bank 12
Thread 12
Bank 12
Thread 13
Bank 13
Thread 13
Bank 13
Thread 14
Bank 14
Thread 14
Bank 14
Thread 15
Bank 15
Thread 15
Bank 15
Left: linear addressing with a stride of one 32-bit word.
Right: random permutation.
Figure 3-5. Examples of Shared Memory Access Patterns
without Bank Conflicts
42
NVIDIA OpenCL Programming Guide Version 2.3
Chapter 3. Performance Guidelines
Thread 0
Bank 0
Thread 1
Bank 1
Thread 2
Bank 2
Thread 3
Bank 3
Thread 4
Bank 4
Thread 5
Bank 5
Thread 6
Bank 6
Thread 7
Bank 7
Thread 8
Bank 8
Thread 9
Bank 9
Thread 10
Bank 10
Thread 11
Bank 11
Thread 12
Bank 12
Thread 13
Bank 13
Thread 14
Bank 14
Thread 15
Bank 15
Linear addressing with a stride of three 32-bit words.
Figure 3-6. Example of a Shared Memory Access Pattern
without Bank Conflicts
NVIDIA OpenCL Programming Guide Version 2.3
43
Chapter 3. Performance Guidelines
Thread 0
Bank 0
Thread 0
Bank 0
Thread 1
Bank 1
Thread 1
Bank 1
Thread 2
Bank 2
Thread 2
Bank 2
Thread 3
Bank 3
Thread 3
Bank 3
Thread 4
Bank 4
Thread 4
Bank 4
Thread 5
Bank 5
Thread 5
Bank 5
Thread 6
Bank 6
Thread 6
Bank 6
Thread 7
Bank 7
Thread 7
Bank 7
Thread 8
Bank 8
Thread 8
Bank 8
Thread 9
Bank 9
Thread 9
Bank 9
Thread 10
Bank 10
Thread 10
Bank 10
Thread 11
Bank 11
Thread 11
Bank 11
Thread 12
Bank 12
Thread 12
Bank 12
Thread 13
Bank 13
Thread 13
Bank 13
Thread 14
Bank 14
Thread 14
Bank 14
Thread 15
Bank 15
Thread 15
Bank 15
Left: Linear addressing with a stride of two 32-bit words causes 2-way bank conflicts.
Right: Linear addressing with a stride of eight 32-bit words causes 8-way bank conflicts.
Figure 3-7. Examples of Shared Memory Access Patterns with
Bank Conflicts
44
NVIDIA OpenCL Programming Guide Version 2.3
Chapter 3. Performance Guidelines
Thread 0
Bank 0
Thread 0
Bank 0
Thread 1
Bank 1
Thread 1
Bank 1
Thread 2
Bank 2
Thread 2
Bank 2
Thread 3
Bank 3
Thread 3
Bank 3
Thread 4
Bank 4
Thread 4
Bank 4
Thread 5
Bank 5
Thread 5
Bank 5
Thread 6
Bank 6
Thread 6
Bank 6
Thread 7
Bank 7
Thread 7
Bank 7
Thread 8
Bank 8
Thread 8
Bank 8
Thread 9
Bank 9
Thread 9
Bank 9
Thread 10
Bank 10
Thread 10
Bank 10
Thread 11
Bank 11
Thread 11
Bank 11
Thread 12
Bank 12
Thread 12
Bank 12
Thread 13
Bank 13
Thread 13
Bank 13
Thread 14
Bank 14
Thread 14
Bank 14
Thread 15
Bank 15
Thread 15
Bank 15
Left: This access pattern is conflict-free since all threads read from an address within the same 32-bit
word.
Right: This access pattern causes either no bank conflicts if the word from bank 5 is chosen as the
broadcast word during the first step or 2-way bank conflicts, otherwise.
Figure 3-8. Example of Shared Memory Read Access Patterns
with Broadcast
NVIDIA OpenCL Programming Guide Version 2.3
45
Chapter 3. Performance Guidelines
3.1.2.6
Registers
Generally, accessing a register is zero extra clock cycles per instruction, but delays
may occur due to register read-after-write dependencies and register memory bank
conflicts.
The delays introduced by read-after-write dependencies can be ignored as soon as
there are at least 192 active threads per multiprocessor to hide them.
The compiler and thread scheduler schedule the instructions as optimally as possible
to avoid register memory bank conflicts. They achieve best results when the number
of threads per block is a multiple of 64. Other than following this rule, an
application has no direct control over these bank conflicts. In particular, there is no
need to pack data into float4 or int4 types.
3.2
NDRange
How the NDRange affects the execution time of a kernel launch generally depends
on the kernel code. Experimentation is therefore recommended and applications
should set the work-group size explicitly as opposed to rely on the OpenCL
implementation to determine the right size (by setting local_work_size to NULL in
clEnqueueNDRangeKernel()). There are however general guidelines, described
in this section.
For a start, the kernel will simply fail to launch if the number of threads per block
either is above the maximum number of threads per block as specified in
Appendix A, or requires too many registers or shared memory than available per
multiprocessor as mentioned in Section 2.1.2. The total number of registers required
for a block is equal to
ceil
(
R
×
ceil
),32,(
T
maxR
32
)
)
R
ceil
yx
,(
maxR
is the number of registers required for the kernel,
is equal to x rounded up to the nearest multiple of
where
is the number of
registers per multiprocessor given in Appendix A, T is the number of threads per
. The total
block, and
amount of shared memory required for a block is equal to the sum of the amount of
statically allocated shared memory, the amount of dynamically allocated shared
memory, and the amount of shared memory used to pass the kernel’s arguments.
Note that each double or long long variable uses two registers. However,
devices of compute capability 1.2 and higher have twice as many registers per
multiprocessor as devices with lower compute capability.
y
Then, given a total number of threads per grid, the number of threads per block
might be dictated by the need to have enough blocks in the grid to maximize the
utilization of the available computing resources. First, there should be at least as
many blocks as there are multiprocessors in the device. Then, running only one
block per multiprocessor will force the multiprocessor to idle during thread
synchronization and also during device memory reads if there are not enough
threads per block to cover the load latency. It is therefore usually better to allow for
two or more blocks to be active on each multiprocessor to allow overlap between
blocks that wait and blocks that can run. For this to happen, not only should there
be at least twice as many blocks as there are multiprocessors in the device, but also
46
NVIDIA OpenCL Programming Guide Version 2.3
Chapter 3. Performance Guidelines
the amount of registers and shared memory required per block must be low enough
to allow for more than one active block (see Section 2.1.2). More thread blocks
stream in pipeline fashion through the device and amortize overhead even more.
The number of blocks per grid should be at least 100 if one wants it to scale to
future devices; 1000 blocks will scale across several generations.
With a high enough number of blocks, the number of threads per block should be
chosen as a multiple of the warp size to avoid swasting computing resources with
under-populated warps, or better, a multiple of 64 for the reason invoked in
Section 3.1.2.6. Allocating more threads per block is better for efficient time slicing,
but the more threads per block, the fewer registers are available per thread, which
might prevent the kernel invocation from succeeding.
Usually, 64 threads per block is minimal and makes sense only if there are multiple
active blocks per multiprocessor; 192 or 256 threads per block is better and usually
allows for enough registers to compile.
The ratio of the number of active warps per multiprocessor to the maximum
number of active warps (given in Appendix A) is called the multiprocessor occupancy.
In order to maximize occupancy, the compiler attempts to minimize register usage
while keeping the number of instructions and CUDA local memory usage to a
minimum. The CUDA Software Development Kit provides a spreadsheet to assist
programmers in choosing thread block size based on shared memory and register
requirements.
3.3
Data Transfer between Host and Device
The bandwidth between device memory and the device is much higher than the
bandwidth between device memory and host memory. Therefore, one should strive
to minimize data transfer between the host and the device, for example, by moving
more code from the host to the device, even if that means running kernels with low
parallelism computations. Intermediate data structures may be created in device
memory, operated on by the device, and destroyed without ever being mapped by
the host or copied to host memory.
Also, because of the overhead associated with each transfer, batching many small
transfers into a big one always performs much better than making each transfer
separately.
Finally, higher performance for data transfers between host and device is achieved
for memory objects allocated in page-locked (also known as pinned) host memory (as
opposed to regular pageable host memory allocated by malloc()), which has
several benefits:
(cid:137) Bandwidth between host memory and device memory is higher if host memory
is allocated as page-locked.
(cid:137) For some devices, copies between page-locked host memory and device memory
can be performed concurrently with kernel execution.
(cid:137) For some devices, page-locked host memory can be mapped into the device’s
address space. In this case, there is no need to allocate any device memory and to
explicitly copy data between device and host memory. Data transfers are
implicitly performed each time the kernel accesses the mapped memory. For
maximum performance, these memory accesses must be coalesced like if they
NVIDIA OpenCL Programming Guide Version 2.3
47
Chapter 3. Performance Guidelines
were accesses to global memory (see Section 3.1.2.1). Assuming that they are an
that the mapped memory is read or written only once, avoiding explicit copies
between device and host memory can be a win performance-wise. It is always a
win on integrated systems where device memory and host memory are physically
the same and therefore any copy between host and device memory is
superfluous.
enCL applicat
ions do not have direct control over whether memory objects are
Op
allocated in page-locked memory or not, but they can create objects using the
CL_MEM_ALLOC_HOST_PTR flag and such objects are likely to be allocated in
locked memory by the driver for best performance.
page-
d
3.4
Warp-Level Synchronization
Because a warp executes one common instruction at a time, threads within a warp
are implicitly synchronized and this can be used to omit calls to the barrier()
function for better performance.
example, both calls to barrier() are required to
In the following code sample, for
get the expected result (i.e. result[i] = 2 * myArray[i] for i > 0).
Without synchronization, any of the two references to myArray[tid] could
return either 2 or the value initially stored in myArray, depending on whether t
memory read occurs before or after the memory write from
myArray[tid + 1] = 2.
he
// myArray is an array of integers located in global or shared
// memory
__kernel void myKernel(__global int* result) {
int tid = get_local_id(0);
...
int ref1 = myArray[tid] * 1;
barrier(CLK_LOCAL_MEM_FENCE|CLK_GLOBAL_MEM_FENCE);
myArray[tid + 1] = 2;
barrier(CLK_LOCAL_MEM_FENCE|CLK_GLOBAL_MEM_FENCE);
int ref2 = myArray[tid] * 1;
result[tid] = ref1 * ref2;
...
}
However, in the following slightly modified code sample, threads are guaranteed to
belong to the same warp, so that there is no need for any barrier() call.
// myArray is an array of integers located in global or shared
// memory
__kernel void myKernel(__global int* result) {
int tid = get_local_id(0);
...
if (tid < warpSize) {
int ref1 = myArray[tid] * 1;
myArray[tid + 1] = 2;
int ref2 = myArray[tid] * 1;
result[tid] = ref1 * ref2;
}
...
}
48
NVIDIA OpenCL Programming Guide Version 2.3
Chapter 3. Performance Guidelines
Simply removing the call to barrier() is not enough however; myArray also
needs to be declared as volatile as described in Section 2.2.2.
3.5
Overall Performance Optimization Strategies
Performance optimization revolves around three basic strategies:
(cid:137) Maximizing parallel execution;
(cid:137) Optimizing memory usage to achieve maximum memory bandwidth;
(cid:137) Optimizing instruction usage to achieve maximum instruction throughput.
Maximizing parallel execution starts with structuring the algorithm in a way that
exposes as much data parallelism as possible. At points in the algorithm where
parallelism is broken because some threads need to synchronize in order to share
data between each other, there are two cases: Either these threads belong to the
same block, in which case they should use the barrier() function and share data
through shared memory within the same kernel call, or they belong to different
blocks, in which case they must share data through global memory using two
separate kernel invocations, one for writing to and one for reading from global
memory.
Once the parallelism of the algorithm has been exposed it needs to be mapped to
the hardware as efficiently as possible. This is done by carefully choosing the
NDRange of each kernel invocation as detailed in Section 3.2.
The application should also maximize parallel execution at a higher level by
explicitly exposing concurrent execution on the device through queues, as well as
maximizing concurrent execution between host and device.
Optimizing memory usage starts with minimizing data transfers with low-
bandwidth. That means minimizing data transfers between the host and the device,
as detailed in Section 3.3, since these have much lower bandwidth than data
transfers between device and global memory. That also means minimizing data
transfers between device and global memory by maximizing use of shared memory
on the device, as mentioned in Section 3.1.2. Sometimes, the best optimization
might even be to avoid any data transfer in the first place by simply recomputing the
data instead whenever it is needed.
As detailed in Sections 3.1.2.1, 3.1.2.3, 3.1.2.4, and 3.1.2.5, the effective bandwidth
can vary by an order of magnitude depending on access pattern for each type of
memory. The next step in optimizing memory usage is therefore to organize
memory accesses as optimally as possible based on the optimal memory access
patterns. This optimization is especially important for global memory accesses as
global memory bandwidth is low and its latency is hundreds of clock cycles (see
Section 3.1.1.3). Shared memory accesses, on the other hand, are usually worth
optimizing only in case they have a high degree of bank conflicts.
As for optimizing instruction usage, the use of arithmetic instructions with low
throughput (see Section 3.1.1.1) should be minimized. This includes trading
precision for speed when it does not affect the end result, such as using intrinsic
instead of regular functions (intrinsic functions are listed in Section B.2) or single-
precision instead of double-precision. Particular attention must be paid to control
flow instructions due to the SIMT nature of the device as detailed in Section 3.1.1.2.
NVIDIA OpenCL Programming Guide Version 2.3
49
Appendix A.
Technical Specifications
A.1
General Specifications
The general specifications and features of a compute device depend on its compute
capability (see Section 2.3).
The following sections describe the technical specifications and features associated
to each compute capability. The specifications for a given compute capability are the
same as for the compute capability just below unless otherwise mentioned. Similarly,
any feature supported for a given compute capability is supported for any higher
compute capability.
The compute capability and number of multiprocessors of all CUDA-enabled
devices are given in the following table:
Number of
Multiprocessors
Compute
Capability
GeForce GTX 295
GeForce GTX 285, GTX 280
GeForce GTX 260
GeForce 9800 GX2
GeForce GTS 250, GTS 150, 9800 GTX,
9800 GTX+, 8800 GTS 512
GeForce 8800 Ultra, 8800 GTX
GeForce 9800 GT, 8800 GT, 9800M GTX
GeForce GT 130, 9600 GSO, 8800 GS,
8800M GTX, 9800M GT
GeForce 8800 GTS
GeForce 9600 GT, 8800M GTS, 9800M GTS
GeForce 9700M GT
GeForce GT 120, 9500 GT, 8600 GTS, 8600 GT,
9700M GT, 9650M GS, 9600M GT, 9600M GS,
9500M GS, 8700M GT, 8600M GT, 8600M GS
GeForce G100, 8500 GT, 8400 GS, 8400M GT,
9500M G, 9300M G, 8400M GS, 9400 mGPU,
9300 mGPU, 8300 mGPU, 8200 mGPU,
(1 Multiprocessor
= 8 Processors)
2x30
30
24
2x16
16
16
14
12
12
8
6
4
2
1.3
1.3
1.3
1.1
1.1
1.0
1.1
1.1
1.0
1.1
1.1
1.1
1.1
Appendix A. Technical Specifications
8100 mGPU
GeForce 9300M GS, 9200M GS, 9100M G,
8400M G
Tesla S1070
Tesla C1060
Tesla S870
Tesla D870
Tesla C870
Quadro Plex 2200 D2
Quadro Plex 2100 D4
Quadro Plex 2100 Model S4
Quadro Plex 1000 Model IV
Quadro FX 5800
Quadro FX 4800
Quadro FX 4700 X2
Quadro FX 3700M
Quadro FX 5600
Quadro FX 3700
Quadro FX 3600M
Quadro FX 4600
Quadro FX 2700M
Quadro FX 1700, FX 570, NVS 320M, FX 1700M,
FX 1600M, FX 770M, FX 570M
Quadro FX 370, NVS 290, NVS 140M, NVS 135M,
FX 360M
Quadro FX 370M, NVS 130M
1
4x30
30
4x16
2x16
16
2x30
4x14
4x16
2x16
30
24
2x14
16
16
14
12
12
6
4
2
1
1.1
1.3
1.3
1.0
1.0
1.0
1.3
1.1
1.0
1.0
1.3
1.3
1.1
1.1
1.0
1.1
1.1
1.0
1.1
1.1
1.1
1.1
The number of multiprocessors, the clock frequency and the total amount of device
memory can be queried using the runtime.
A.1.1
Specifications for Compute Capability 1.0
(cid:137) The maximum number of threads per block is 512;
(cid:137) The maximum sizes of the x-, y-, and z-dimension of a thread block are 512, 512,
and 64, respectively;
(cid:137) The maximum size of each dimension of a grid of thread blocks is 65535;
(cid:137) The warp size is 32 threads;
(cid:137) The number of 32-bit registers per multiprocessor is 8192;
(cid:137) The amount of shared memory available per multiprocessor is 16 KB organized
into 16 banks (see Section 3.1.2.5);
(cid:137) The total amount of constant memory is 64 KB;
(cid:137) The total amount of local memory per thread is 16 KB;
(cid:137) The cache working set for constant memory is 8 KB per multiprocessor;
(cid:137) The cache working set for texture memory varies between 6 and 8 KB per
multiprocessor;
52
NVIDIA OpenCL Programming Guide Version 2.3
Appendix A. Technical Specifications
(cid:137) The maximum number of active blocks per multiprocessor is 8;
(cid:137) The maximum number of active warps per multiprocessor is 24;
(cid:137) The maximum number of active threads per multiprocessor is 768;
(cid:137) The limit on kernel size is 2 million PTX instructions;
A.1.2
A.1.3
Specifications for Compute Capability 1.1
(cid:137) Support for atomic functions operating on 32-bit words in global memory.
Specifications for Compute Capability 1.2
(cid:137) Support for atomic functions operating in shared memory and atomic functions
operating on 64-bit words in global memory;
(cid:137) Support for warp vote functions;
(cid:137) The number of registers per multiprocessor is 16384;
(cid:137) The maximum number of active warps per multiprocessor is 32;
(cid:137) The maximum number of active threads per multiprocessor is 1024.
A.1.4
Specifications for Compute Capability 1.3
(cid:137) Support for double-precision floating-point numbers.
A.2
Floating-Point Standard
All compute devices follow the IEEE-754 standard for binary floating-point
arithmetic with the following deviations:
(cid:137) There is no dynamically configurable rounding mode; however, most of the
operations support IEEE rounding modes, exposed via device functions;
(cid:137) There is no mechanism for detecting that a floating-point exception has occurred
and all operations behave as if the IEEE-754 exceptions are always masked, and
deliver the masked response as defined by IEEE-754 if there is an exceptional
event; for the same reason, while SNaN encodings are supported, they are not
signaling;
(cid:137) Absolute value and negation are not compliant with IEEE-754 with respect to
NaNs; these are passed through unchanged;
(cid:137) For single-precision floating-point numbers only:
(cid:137) Denormalized numbers are not supported; floating-point arithmetic and
comparison instructions convert denormalized operands to zero prior to the
floating-point operation;
(cid:137) Underflowed results are flushed to zero;
(cid:137) The result of an operation involving one or more input NaNs is the quiet
NaN of bit pattern 0x7fffffff; note that;
(cid:137) Some instructions are not IEEE-compliant:
NVIDIA OpenCL Programming Guide Version 2.3
53
Appendix A. Technical Specifications
(cid:137) Addition and multiplication are often combined into a single multiply-add
instruction (FMAD), which truncates the intermediate result of the
multiplication;
(cid:137) Division is implemented via the reciprocal in a non-standard-compliant
way;
(cid:137) Square root is implemented via the reciprocal square root in a non-
standard-compliant way;
(cid:137) For addition and multiplication, only round-to-nearest-even and
round-towards-zero are supported via static rounding modes; directed
rounding towards +/- infinity is not supported;
But, IEEE-compliant software (and therefore slower) implementations are
provided through the following intrinsics from Appendix B:
(cid:137) fma(float, float, float): single-precision fused multiply-add
with IEEE rounding modes,
(cid:137) native_recip(float): single-precision reciprocal with IEEE
rounding modes,
(cid:137) native_divide(float, float): single-precision division with
IEEE rounding modes,
(cid:137) native_sqrt(float): single-precision square root with IEEE
rounding modes;
(cid:137) For double-precision floating-point numbers only:
(cid:137) Round-to-nearest-even is the only supported IEEE rounding mode for
reciprocal, division, and square root.
In accordance to the IEEE-754R standard, if one of the input parameters to
fmin() or fmax() is NaN, but not the other, the result is the non-NaN
parameter.
(cid:137) The conversion of a floating-point value to an integer value in the case where the
floating-point value falls outside the range of the integer format is left undefined
by IEEE-754. For compute devices, the behavior is to clamp to the end of the
supported range. This is unlike the x86 architecture behaves.
A.3
Supported OpenCL Extensions
All compute devices supports the cl_khr_byte_addressable_store extension.
Devices of compute capability 1.1 and higher support the
cl_khr_global_int32_base_atomics, cl_khr_global_int32_extended_atomics,
cl_khr_local_int32_base_atomics, and cl_khr_local_int32_extended_atomics
extensions.
54
NVIDIA OpenCL Programming Guide Version 2.3
Appendix B.
Mathematical Functions Accuracy
B.1
Standard Functions
Error bounds in this section are generated from extensive but not exhaustive tests,
so they are not guaranteed bounds.
B.1.1
Single-Precision Floating-Point Functions
Table C-1 lists errors for the standard single-precision floating-point functions.
The recommended way to round a single-precision floating-point operand to an
integer, with the result being a single-precision floating-point number is rint(),
not round(). The reason is that round() maps to an 8-instruction sequence on
the device, whereas rint() maps to a single instruction. trunc(), ceil(), and
floor() each map to a single instruction as well.
Table C-1. Mathematical Standard Library Functions with
Maximum ULP Error
The maximum error is stated as the absolute value of the difference
in ulps between a correctly rounded single-precision result and the
result returned by the CUDA library function.
Function
x+y
x*y
x/y
1/x
1/sqrt(x)
rsqrt(x)
sqrt(x)
cbrt(x)
hypot(x,y)
exp(x)
Maximum ulp error
0 (IEEE-754 round-to-nearest-even)
(except when merged into an FMAD)
0 (IEEE-754 round-to-nearest-even)
(except when merged into an FMAD)
2 (full range)
1 (full range)
2 (full range)
3 (full range)
1 (full range)
3 (full range)
2 (full range)
Appendix B. Mathematical Functions Accuracy
Function
exp2(x)
exp10(x)
expm1(x)
log(x)
log2(x)
log10(x)
log1p(x)
sin(x)
cos(x)
tan(x)
sincos(x,cptr)
asin(x)
acos(x)
atan(x)
atan2(y,x)
sinh(x)
cosh(x)
tanh(x)
asinh(x)
acosh(x)
atanh(x)
pow(x,y)
erf(x)
erfc(x)
erfinv(x)
erfcinv(x)
lgamma(x)
tgamma(x)
fma(x,y,z)
frexp(x,exp)
ldexp(x,exp)
scalbn(x,n)
scalbln(x,l)
logb(x)
ilogb(x)
fmod(x,y)
remainder(x,y)
remquo(x,y,iptr)
modf(x,iptr)
fdim(x,y)
trunc(x)
round(x)
Maximum ulp error
2 (full range)
2 (full range)
1 (full range)
1 (full range)
3 (full range)
3 (full range)
2 (full range)
2 (full range)
2 (full range)
4 (full range)
2 (full range)
4 (full range)
3 (full range)
2 (full range)
3 (full range)
3 (full range)
2 (full range)
2 (full range)
3 (full range)
4 (full range)
3 (full range)
8 (full range)
3 (full range)
8 (full range)
5 (full range)
7 (full range)
6 (outside interval -10.001 ... -2.264; larger inside)
11 (full range)
0 (full range)
0 (full range)
0 (full range)
0 (full range)
0 (full range)
0 (full range)
0 (full range)
0 (full range)
0 (full range)
0 (full range)
0 (full range)
0 (full range)
0 (full range)
0 (full range)
56
NVIDIA OpenCL Programming Guide Version 2.3
Appendix B. Mathematical Functions Accuracy
Function
rint(x)
nearbyint(x)
ceil(x)
floor(x)
lrint(x)
lround(x)
llrint(x)
llround(x)
Maximum ulp error
0 (full range)
0 (full range)
0 (full range)
0 (full range)
0 (full range)
0 (full range)
0 (full range)
0 (full range)
B.1.2
Double-Precision Floating-Point Functions
Table C-2 lists errors for the standard double-precision floating-point functions.
These errors only apply when compiling for devices with native double-precision
support. When compiling for devices without such support, such as devices of
compute capability 1.2 and lower, the double type gets demoted to float by
default and the double-precision math functions are mapped to their single-
precision equivalents.
The recommended way to round a double-precision floating-point operand to an
integer, with the result being a double-precision floating-point number is rint(),
not round(). The reason is that round() maps to an 8-instruction sequence on
the device, whereas rint() maps to a single instruction. trunc(), ceil(), and
floor() each map to a single instruction as well.
Table C-2. Mathematical Standard Library Functions with
Maximum ULP Error
The maximum error is stated as the absolute value of the difference
in ulps between a correctly rounded double-precision result and the
result returned by the CUDA library function.
Function
x+y
x*y
x/y
1/x
sqrt(x)
rsqrt(x)
cbrt(x)
hypot(x,y)
exp(x)
exp2(x)
exp10(x)
expm1(x)
log(x)
log2(x)
Maximum ulp error
0 (IEEE-754 round-to-nearest-even)
0 (IEEE-754 round-to-nearest-even)
0 (IEEE-754 round-to-nearest-even)
0 (IEEE-754 round-to-nearest-even)
0 (IEEE-754 round-to-nearest-even)
1 (full range)
1 (full range)
2 (full range)
1 (full range)
1 (full range)
1 (full range)
1 (full range)
1 (full range)
1 (full range)
NVIDIA OpenCL Programming Guide Version 2.3
57
Appendix B. Mathematical Functions Accuracy
Function
log10(x)
log1p(x)
sin(x)
cos(x)
tan(x)
sincos(x,sptr,cptr)
asin(x)
acos(x)
atan(x)
atan2(y,x)
sinh(x)
cosh(x)
tanh(x)
asinh(x)
acosh(x)
atanh(x)
pow(x,y)
erf(x)
erfc(x)
erfinv(x)
erfcinv(x)
lgamma(x)
tgamma(x)
fma(x,y,z)
frexp(x,exp)
ldexp(x,exp)
scalbn(x,n)
scalbln(x,l)
logb(x)
ilogb(x)
fmod(x,y)
remainder(x,y)
remquo(x,y,iptr)
modf(x,iptr)
fdim(x,y)
trunc(x)
round(x)
rint(x)
nearbyint(x)
ceil(x)
floor(x)
lrint(x)
Maximum ulp error
1 (full range)
1 (full range)
2 (full range)
2 (full range)
2 (full range)
2 (full range)
2 (full range)
2 (full range)
2 (full range)
2 (full range)
1 (full range)
1 (full range)
1 (full range)
2 (full range)
2 (full range)
2 (full range)
2 (full range)
2 (full range)
7 (full range)
8 (full range)
8 (full range)
4 (outside interval -11.0001 ... -2.2637; larger inside)
8 (full range)
0 (IEEE-754 round-to-nearest-even)
0 (full range)
0 (full range)
0 (full range)
0 (full range)
0 (full range)
0 (full range)
0 (full range)
0 (full range)
0 (full range)
0 (full range)
0 (full range)
0 (full range)
0 (full range)
0 (full range)
0 (full range)
0 (full range)
0 (full range)
0 (full range)
58
NVIDIA OpenCL Programming Guide Version 2.3
Appendix B. Mathematical Functions Accuracy
Function
lround(x)
llrint(x)
llround(x)
Maximum ulp error
0 (full range)
0 (full range)
0 (full range)
B.2
Native Functions
Table C-3 lists the native single-precision floating-point functions supported on the
CUDA architecture.
Both the regular floating-point division and native_divide(x,y) have the same
accuracy, but for 2126 < y < 2128, native_divide(x,y) delivers a result of zero,
whereas the regular division delivers the correct result to within the accuracy stated
in Table C-3. Also, for 2126 < y < 2128, if x is infinity, native_divide(x,y)
delivers a NaN (as a result of multiplying infinity by zero), while the regular division
returns infinity.
Table C-3. Single-Precision Floating-Point Native Functions
with Respective Error Bounds
Function
native_recip(x)
native_sqrt(x)
Error bounds
IEEE-compliant.
IEEE-compliant.
native_divide(x,y)
For y in [2-126, 2126], the maximum ulp error is 2.
native_exp(x)
native_exp10(x)
native_log(x)
native_log2(x)
native_log10(x)
native_sin(x)
native_cos(x)
native_tan(x)
native_pow(x,y)
The maximum ulp error is
2 + floor(abs(1.16 * x)).
The maximum ulp error is
2 + floor(abs(2.95 * x)).
For x in [0.5, 2], the maximum absolute error is 2-
21.41, otherwise, the maximum ulp error is 3.
For x in [0.5, 2], the maximum absolute error is 2-22,
otherwise, the maximum ulp error is 2.
For x in [0.5, 2], the maximum absolute error is 2-24,
otherwise, the maximum ulp error is 3.
For x in [-π, π], the maximum absolute error is 2-21.41,
and larger otherwise.
For x in [-π, π], the maximum absolute error is 2-21.19,
and larger otherwise.
Derived from its implementation as
native_sin(x) * (1 / native_cos(x)).
Derived from its implementation as
exp2(y * native_log2(x)).
NVIDIA OpenCL Programming Guide Version 2.3
59
Notice
ALL NVIDIA DESIGN SPECIFICATIONS, REFERENCE BOARDS, FILES, DRAWINGS, DIAGNOSTICS, LISTS, AND
OTHER DOCUMENTS (TOGETHER AND SEPARATELY, “MATERIALS”) ARE BEING PROVIDED “AS IS.” NVIDIA
MAKES NO WARRANTIES, EXPRESSED, IMPLIED, STATUTORY, OR OTHERWISE WITH RESPECT TO THE
MATERIALS, AND EXPRESSLY DISCLAIMS ALL IMPLIED WARRANTIES OF NONINFRINGEMENT,
MERCHANTABILITY, AND FITNESS FOR A PARTICULAR PURPOSE.
Information furnished is believed to be accurate and reliable. However, NVIDIA Corporation assumes no
responsibility for the consequences of use of such information or for any infringement of patents or other
rights of third parties that may result from its use. No license is granted by implication or otherwise under any
patent or patent rights of NVIDIA Corporation. Specifications mentioned in this publication are subject to
change without notice. This publication supersedes and replaces all information previously supplied. NVIDIA
Corporation products are not authorized for use as critical components in life support devices or systems
without express written approval of NVIDIA Corporation.
Trademarks
NVIDIA, the NVIDIA logo, GeForce, Tesla, and Quadro are trademarks or registered trademarks of NVIDIA
Corporation. Other company and product names may be trademarks of the respective companies with which
they are associated.
Copyright
© 2007-2008 NVIDIA Corporation. All rights reserved.
This work incorporates portions of on an earlier work: Scalable Parallel Programming with CUDA, in ACM
Queue, VOL 6, No. 2 (March/April 2008), © ACM, 2008. http://mags.acm.org/queue/20080304/?u1=texterity"
NVIDIA Corporation
2701 San Tomas Expressway
Santa Clara, CA 95050
www.nvidia.com