CUDA Programming Model

Xing Zeng, Dongyue Mou

Outline

• Introduction
• Motivation
• Programming Model
• Memory Model
• CUDA API
• Example
• Pro & Contra
• Trend
Outline

• Introduction
• Motivation
• Programming Model
• Memory Model
• CUDA API
• Example
• Pro & Contra
• Trend
Introduction

What is CUDA?
- Compute Unified Device Architecture.
- A powerful parallel programming model for issuing and
managing computations on the GPU without mapping them to a
graphics API.

• Heterogenous - mixed serial-parallel programming

• Scalable - hierarchical thread execution model
• Accessible - minimal but expressive changes to C
Introduction
Software Stack:
• Libraries:
CUFFT & CUBLAS

• Runtime:
Common component
Device component
Host component

• Driver:
Driver API
Introduction
CUDA SDK
Outline

• Introduction
• Motivation
• Programming Model
• Memory Model
• CUDA API
• Example
• Pro & Contra
• Trend
Motivation

GPU Programming Model

GPGPU Programming Model
CUDA Programming Model
Motivation

GPU Programming Model

GPGPU Programming Model
CUDA Programming Model
Motivation
GPU Programming Model for Graphics
Motivation

GPU Programming Model

GPGPU Programming Model
CUDA Programming Model
Motivation
GPGPU Programming Model
Trick the GPU into general-purpose
computing by casting problem as graphics

• Turn data into images ("texture maps")

• Turn algorithms into image synthesis ("rending passes")
Drawback:
• Tough learning curve
• potentially high overhead of graphics API
• highly constrained memory layout & access model
• Need for many passes drives up bandwidth consumption
Motivation
GPGPU Programming to do A + B
Motivation
What's wrong with GPGPU 1
APIs are specific to Graphics

Limited texture size and

dimension

Limited Instruction set

No thread Communication Limited local storage

Limited shader outputs

No scatter
Motivation
What's wrong with GPGPU 2
Motivation

GPU Programming Model

GPGPU Programming Model
CUDA Programming Model
Outline

• Introduction
• Motivation
• Programming Model
• Memory Model
• CUDA API
• Example
• Pro & Contra
• Trend
Programming Model
CUDA: Unified Design
Advantage:

HW: fully generally data-parallel arch-

tecture.
• General thread launch
• Global load-store
• Parallel data cache
• Scalar architecture
• Integers, bit operation

SW: program the GPU in C

• Scalable data parallel execuation/
memory model
• C with minimal yet powerful
extensions
Motivation
From GPGPU to CUDA Programming Model
Programming Model
Feature 1:
• Thread not pixel
• Full Integer and Bit Instructions
• No limits on branching, looping
• 1D, 2D, 3D threadID allocation

Feature 2:
• Fully general load/store to
GPU memory
• Untyped, not fixed texture types
• Pointer support

Feature 3:
• Dedicated on-chip memory
• Shared between threads for
inter-threads communication
• Explicitly managed
• As fast as registers
Programming Model
Important Concepts:
• Device: GPU, viewed as a
co-processor.
• Host: CPU
• Kernel: data-parallel,
computed-intensive positions
of application running on the
device.
Programming Model
Important Concepts:
• Thread: basic execution unit
• Thread block:
A batch of thread. Threads
in a block cooperate together,
efficiently share data.
Thread/block have unique id
• Grid:
A batch of thread block.
that excuate same kernel.
Threads in different block in
the same grid cannot directly
communicate with each other
Programming Model
Simple example ( Matrx addition ):
cpu c program: cuda program:
Programming Model
Hardware implementation:

A set of SIMD
Multiprocessors with On-
Chip shared memory
Programming Model
G80 Example:
• 16 Multiprocessors, 128 Thread Processors
• Up to 12,288 parallel threads active
• Per-block shared memory accelerates processing.
Programming Model
Streaming Multiprocessor (SM)
• Processing elements
o 8 scalar thread processors
o 32 GFLOPS peak at 1.35GHz
o 8192 32-bit registers (32KB)
o usual ops: float, int, branch...

• Hardware multithreading
o up to 8 blocks (3 active) residents at once
o up to 768 active threads in total

• 16KB on-chip memory

• supports thread communication
• shared amongst threads of a block
Programming Model
Execution Model:
Programming Model
Single Instruction Multiple Thread (SIMT) Execution:
• Groups of 32 threads formed
into warps
o always executing same instruction
o share instruction fetch/dispatch
o some become inactive
when code path diverges
o hardware automatically handles divergence

• Warps are primitive unit of scheduling

• pick 1 of 24 warps for each instruction slot.
• all warps from all active blocks are time-sliced
Outline

• Introduction
• Motivation
• Programming Model
• Memory Model
• CUDA API
• Example
• Pro & Contra
• Trend
Memory Model
There are 6 Memory Types :
Memory Model
There are 6 Memory Types :

• Registers
o on chip
o fast access
o per thread
o limited amount
o 32 bit
Memory Model
There are 6 Memory Types :

• Registers
• Local Memory
o in DRAM
o slow
o non-cached
o per thread
o relative large
Memory Model
There are 6 Memory Types :

• Registers
• Local Memory
• Shared Memory
o on chip
o fast access
o per block
o 16 KByte
o synchronize between
threads
Memory Model
There are 6 Memory Types :

• Registers
• Local Memory
• Shared Memory
• Global Memory
o in DRAM
o slow
o non-cached
o per grid
o communicate between
grids
Memory Model
There are 6 Memory Types :

• Registers
• Local Memory
• Shared Memory
• Global Memory
• Constant Memory
o in DRAM
o cached
o per grid
o read-only
Memory Model
There are 6 Memory Types :

• Registers
• Local Memory
• Shared Memory
• Global Memory
• Constant Memory
• Texture Memory
o in DRAM
o cached
o per grid
o read-only
Memory Model
• Registers
• Shared Memory
o on chip

• Local Memory
• Global Memory
• Constant Memory
• Texture Memory
o in Device Memory
Memory Model
• Global Memory
• Constant Memory
• Texture Memory
o managed by host code
o persistent across kernels
Outline

• Introduction
• Motivation
• Programming Model
• Memory Model
• CUDA API
• Example
• Pro & Contra
• Trend
CUDA API

CUDA API provides a easily path for users to write programs for
GPU device .

It consists of:

• A minimal set of extensions to C/C++

o type qualifiers
o call-syntax
o build-in variables

• A runtime library to support the execution

o host component
o device component
o common component
CUDA API

CUDA C/C++ Extensions:

• New function type qualifiers
__host__ void HostFunc(...); //executable on host
__global__ void KernelFunc(...); //callable from host
__device__ void DeviceFunc(...); //callable from device only
o Restrictions for device code (__global__ / __device__)
 no recursive call
 no static variable
 no function pointer
 __global__ function is asynchronous invoked
 __global__ function must have void return type
CUDA API

CUDA C/C++ Extensions:

• New variable type qualifiers
__device__ int GlobalVar; //in global memory, lifetime of app
__const__ int ConstVar; //in constant memory, lifetime of app
__shared__ int SharedVar; //in shared memory, lifetime of blocks
o Restrictions
 no external usage
 only file scope
 no combination with struct or union
 no initialization for __shared__
CUDA API

CUDA C/C++ Extensions:

• New syntax to invoke the device code
KernelFunc<<< Dg, Db, Ns, S >>>(...);
o Dg: dimension of grid
o Db: dimension of block
o Ns: optional, shared memory for external variables
o S : optional, associated stream

• New build-in variables for indexing the threads

o gridDim: dimension of the whole grid
o blockIdx: index of the current block
o blockDim: dimension of each block in the grid
o threadIdx: index of the current thread
CUDA API

CUDA Runtime Library:

• Common component
o Vector/Texture Types
o Mathematical/Time Functions

• Device component
o Mathematical/Time/Texture Functions
o Synchronization Function
o __syncthreads()
o Type Conversion/Casting Functions
CUDA API

CUDA Runtime Library:

• Host component

o Structure
 Driver API
 Runtime API

o Functions
 Device, Context, Memory, Module, Texture management
 Execution control
 Interoperability with OpenGL and Direct3D
CUDA API

The CUDA source file uses .cu as

extension. It contains host and
device source codes.

The CUDA Compiler Driver nvcc can

compile it and generate CPU/PTX
binary code.
(PTX: Parallel Thread Execution, a
device independent VM code)

PTX code may be further translated

for special GPU-Arch.
Outline

• Introduction
• Motivation
• Programming Model
• Memory Model
• CUDA API
• Example
• Pro & Contra
• Trend
Programming Model
Simple example ( Matrx addition ):
cpu c program: cuda program:
Outline

• Introduction
• Motivation
• Programming Model
• Memory Model
• CUDA API
• Example
• Pro & Contra
• Trend
Pro & Contra

CUDA allows
• massive parallel computing
• with a relative low price
• high integrated solution
• personal supercomputing
• ecofriendly production
• easy to learn
Pro & Contra

Problem ......
• slightly low precision
• limited support for IEEE-754
• no recursive function call
• hard to use for irregular join/fork logic
• no concurrency between jobs
Outline

• Introduction
• Motivation
• Programming Model
• Memory Model
• CUDA API
• Example
• Pro & Contra
• Trend
Trend

• More cores on-chip

• Better support for float point
• Flexiber configuration & control/data flow
• Lower price
• Support higher level programming language
References

[1] CUDA Programming Guide, nVidia Corp.

[2] The CUDA Compiler Driver, nVidia Corp.
[3] Parallel Thread Execution, nVidia Corp.
[4] CUDA: A Heterogeneous Parallel Programming Model for
Manycore Computing, ASPLOS 2008, gpgpu.org
Question?