GPUs

Question 1

What is a GPU?

Answer 1

Graphics Processing Unit
Specialised hardware for graphics computation, highly parallel multiprocessors

Question 2

List the 6 stages in the graphics pipeline

Answer 2

1. Input assembler
2. Vertex shader
3. Geometry shader
4. Setup & rasterise
5. Pixel shader
6. Output merger

Question 3

Which 3 stages in the graphics pipeline are programmable?

Answer 3

1. Vertex shader
2. Geometry shader
3. Pixel shader
   (all the shaders!!)

Question 4

Draw the graphics pipeline, showing the unified processor array

Answer 4

.

Question 5

Tasks pass through the unified processor array ? ?

Answer 5

several times

Question 6

GPUs take advantage of ? level parallelism

Answer 6

data

Question 7

Which can support more threads: CPU or GPU?

Answer 7

GPU - supports multithreading whereas CPU is designed for single-threading

Question 8

What is the design goal of a CPU?

Answer 8

Designed for single-threaded performance

Question 9

What is the design goal of a GPU?

Answer 9

Designed for throughput
- individual thread performance is not critical
- each core is relatively simple

Question 10

Write out the code for a DAXPY (Double A X Plus Y) loop

Answer 10

for (i=0; i<1024; ++i) {
Z[i] = a*X[i] + Y[i]
}
Multiple vector of doubles X by constant a then add to vector of doubles Y

Question 11

Why can we parallelise a DAXPY loop?

Answer 11

All iterations are independent because there are no data dependencies between them
eg. Z[0] = aX[0] + Y[0] is independent of Z[1] = aX[1] + Y[1]

Question 12

How do GPUs parallelise code?

Answer 12

Using SIMT

Question 13

Describe SIMT

Answer 13

Single Instruction, Multiple Threads
Each processor runs many threads concurrently, each with its own registers and memory
All threads execute the same instructions but on different data

Question 14

What is 1 similarity between SIMD and SIMT?

Answer 14

Both take advantage of data-level parallelism

Question 15

What are the differences between SIMD and SIMT?

Answer 15

SIMD is a form of vector processing - vector registers each hold several items, vector operations work on vector registers

In SIMT, each instruction operates on threads - same instruction issued to each thread, each with separate state (registers and memory)

ie. one register storing lots of data items vs lots of registers storing 1 data item

Question 16

Draw a diagram showing SIMT execution of the DAXPY loop with 4 threads in lock-step

Answer 16

.

Question 17

In the case of 4 threads, how many times is the DAXPY kernel called?

Answer 17

Called 256 times with 4 threads each time

Question 18

Each thread executes the DAXPY ?

Answer 18

kernel

Question 19

What is a warp?

Answer 19

Group of threads that are executed simultaneously
All threads in a warp execute the same instruction

Question 20

What support is there in a GPU for conditional execution?

Answer 20

Use a predicate register to hold conditions, then only execute something if the condition holds - predication

Question 21

What is a downside of predication?

Answer 21

The thread sits idle if the predicate is false, leading to inefficiency

Question 22

Draw a diagram to show how conditional branches in GPUs are treated

Answer 22

.

Question 23

When might we need more complex branching?

Answer 23

Procedure call and return, nested if-then-else etc.

Question 24

How is more complex branching handled?

Answer 24

Handled by hardware with special instructions
Each thread has its own stack that predicates are pushed to and popped from

Question 25

What is branch divergence?

Answer 25

When some but not all threads branch

Question 26

When is complex branching inefficient?

Answer 26

When branch divergence occurs

Question 27

What is the relationship between branch nesting and efficiency?

Answer 27

Efficiency decreases as branch nesting increases

Question 28

Write a simple block of code and work out its efficiency

Answer 28

.

Question 29

What are the disadvantages of data parallelism?

Answer 29

- High memory bandwidth requirements - Memory accesses can be high latency

Question 30

When do we get stalls?

Answer 30

From long latency operations such as accessing memory (when a thread's data is not in cache)

Question 31

How can synchronisation between threads cause delays?

Answer 31

May need to wait for all threads to execute the instructions before their memory barrier before continuing

Question 32

What happens if there is a stall?

Answer 32

Have to wait, even if it is just 1 thread that has stalled, since all threads need to proceed at once

Question 33

How can we overcome a stall / hide latency by doing useful work?

Answer 33

Organising execution better - execute other warps of threads whilst waiting

Question 34

Draw a diagram showing one way of executing a DAXPY loop that will cause a stall

Answer 34

.

Question 35

Draw a diagram showing a better way of organising DAXPY execution that will avoid a stall

Answer 35

.

Question 36

What is the multithreading support in a GPU

Answer 36

Much more thread state than processing resource (ie. more registers than ALUs) so can swap threads in to hide latency

Question 37

What are the trade-offs with multithreading?

Answer 37

+ Support many threads, enables a lot of latency hiding - Overhead in area from supporting lots of threads because need places to keep their thread state

Question 38

What is a warp scheduler?

Answer 38

Responsible for choosing which warps to execute Selects based on when operands are ready, for latency hiding

Question 39

How are warps scheduled? Draw a diagram to clarify

Answer 39

One instruction from each warp is executed Then the next instruction from each warp is executed - this can be in a different order This keeps all threads busy because we are executing instructions on all groups of threads (warps)

Question 40

Multithreading hides ?

Answer 40

latency