GPU Fundamentals

Jeff Larkin <[email protected]>, November 14, 2016

 Who Am I?
2002 – B.S. Computer Science – Furman University
2005 – M.S. Computer Science – UT Knoxville
2002 – Graduate Teaching Assistant
2005 – Graduate Research Assistant (ICL)
2005 – 2013 – Cray, Inc
Worked on porting & optimizing HPC apps @ ORNL, User Training
2013 – Present – NVIDIA Corp.
Porting & optimizing HPC apps @ ORNL , User Training,
Representative to OpenACC & OpenMP
2
 GPU Architecture
Speed v. Throughput

AGENDA Latency Hiding

Memory Coalescing
SIMD v. SIMT

3
 GPU Architecture
Two Main Components

Global memory
Analogous to RAM in a CPU server

Accessible by both GPU and CPU

Currently up to 16 GB in Tesla products

Streaming Multiprocessors (SM)

Perform the actual computation

Each SM has its own: Control units, registers, execution pipelines, caches

7
GPU Architecture
Streaming Multiprocessor (SM)

Many CUDA Cores per SM

Architecture dependent

Special-function units
cos/sin/tan, etc.

Shared memory + L1 cache

Thousands of 32-bit registers

8
 GPU Architecture
CUDA Core
Floating point & Integer unit
IEEE 754-2008 floating-point CUDA Core
standard Dispatch Port
Operand Collector
Fused multiply-add (FMA)
instruction for both single and
FP Unit INT Unit
double precision

Logic unit Result Queue

Move, compare unit

Branch unit
9
 Execution Model
Software Hardware
Threads are executed by scalar processors
Scalar
Thread Processor
Thread blocks are executed on multiprocessors

Thread blocks do not migrate

Several concurrent thread blocks can reside on one

Thread multiprocessor - limited by multiprocessor
Block Multiprocessor
resources (shared memory and register file)

... A kernel is launched as a grid of thread blocks

Grid Device 10
 Warps

A thread block consists of

32 Threads 32-thread warps
... = 32 Threads
32 Threads A warp is executed
Thread
Block Warps Multiprocessor physically in parallel
(SIMT) on a multiprocessor

11
 GPU Memory Hierarchy Review

SM-0 SM-1 SM-N

Registers Registers Registers

L1 SMEM L1 SMEM L1 SMEM

L2

Global Memory
12
 GPU Architecture
Memory System on each SM

Extremely fast, but small, i.e., 10s of Kb

Programmer chooses whether to use cache as L1 or Shared Mem
L1
Hardware-managed—used for things like register spilling

Should NOT attempt to utilize like CPU caches

Shared Memory—programmer MUST synchronize data accesses!!!

User-managed scratch pad

Repeated access to same data or multiple threads with same data

13
 GPU Architecture
Memory system on each GPU board

Unified L2 cache (100s of Kb)

Fast, coherent data sharing across all cores in the GPU
Unified/Managed Memory
Since CUDA6 it’s possible to allocate 1 pointer (virtual address) whose physical
location will be managed by the runtime.
Pre-Pascal GPUS – managed by software, limited to GPU memory size
Pascal & Beyond – Hardware can page fault to manage location, can oversubscribe
GPU memory.

14
 Speed v. Throughput
Speed Throughput

Which is better depends on your needs…

*Images from Wikimedia Commons via Creative Commons 15

 Low Latency or High Throughput?

CPU GPU
 Optimized for low-latency access to  Optimized for data-parallel, throughput
cached data sets computation
 Control logic for out-of-order and  Tolerant of memory latency
speculative execution  More transistors dedicated to computation
 10’s of threads  10,000’s of threads
16
 Low Latency or High Throughput?
CPU architecture must minimize latency within each thread
GPU architecture hides latency with computation from other thread warps

CPU core – Low Latency Processor Computation Thread/Warp

T1 T2 T3 T4 Tn Processing

GPU Stream Multiprocessor – High Throughput Processor

Waiting for data
W4
W3 Ready to be processed
W2
W1 Context switch

17
 Memory Coalescing
Global memory access happens in
transactions of 32 or 128 bytes
The hardware will try to reduce to 0 1 31
as few transactions as possible
Coalesced access:
A group of 32 contiguous threads
(“warp”) accessing adjacent words
0 1 31
Few transactions and high utilization
Uncoalesced access:
A warp of 32 threads accessing
scattered words
Many transactions and low utilization

18
 SIMD and SIMT

LD.128b Single Instruction Multiple

Thread (SIMT)
LD. LD. LD. LD. • Scalar instructions execute
LD.128b 1 1 1 1 simultaneously by multiple
LD. LD. LD. LD. hardware threads
AD.128b +
1 1 1 1 • Contiguous data not required.
ST.128b + + + +

Single Instruction Multiple Data (SIMD) AD. AD. AD. AD.

• Vector instructions perform the same operation on 1 1 1 1
multiple data elements.
• Data must be loaded and stored in contiguous LD. LD. LD. LD.
buffers 1 1 1 1

19
 SIMD and SIMT

LD.128b Single Instruction Multiple

Thread (SIMT)
LD. LD. LD. LD. • Scalar instructions execute
LD.128b 1 1 1 1 simultaneously by multiple
LD. LD. LD. LD. hardware threads
AD.128b +
1 1 1 1 • Contiguous data not required.
+ • So if something can run in
ST.128b + + +
SIMD, it can run in SIMT, but
not necessarily the reverse.
Single Instruction Multiple Data (SIMD) AD. AD. AD. AD. • SIMT can better handle
• Vector instructions perform the same operation on 1 1 1 1 indirection
multiple data elements.
• Data must be loaded and stored in contiguous LD. LD. LD. LD.
buffers 1 1 1 1

20
 SIMD and SIMT

LD.128b Single Instruction Multiple

Thread (SIMT)
LD. LD. LD. LD. • Scalar instructions execute
LD.128b 1 1 1 1 simultaneously by multiple
LD. LD. LD. LD. hardware threads
AD.128b +
1 1 1 1 • Contiguous data not required.
+ • So if something can run in
ST.128b + + +
SIMD, it can run in SIMT, but
not necessarily the reverse.
Single Instruction Multiple Data (SIMD) AD. AD. AD. AD. • SIMT can better handle
• Vector instructions perform the same operation on 1 1 1 1 indirection
multiple data elements. • The hardware enables
• Data must be loaded and stored in contiguous LD. LD. LD. LD.
1 1 1 1 parallel execution of scalar
buffers instructions
• Either the programmer or the compiler must
generate vector instructions
21
SIMD and SIMT Branching
SIMT
SIMD

1. Execute converged 1. Execute converged

instructions instructions
2. Generate vector 2. Executed true
mask for true branch
3. Execute masked 3. Execute false
vector instruction branch
4. Generate vector 4. Continue to
mask for false execute converged
5. Execute masked instructions
vector instruction
6. Continue to
execute converged
instructions

Divergence (hopefully) handled by compiler Divergence handle by hardware through

through masks and/or gather/scatter predicated instructions.
operations. 22
 Next 2 Lectures

Wednesday – OpenACC Basics

Friday – More OpenACC?

23