7 - Communication & Synchronisation

Question 1

How do work items/threads communicate?

Answer 1

Through memory

Question 2

What is the idea case for memory?

Answer 2

One type that is large, cheap and fast

Question 3

What are the attributes of large, cheap and fast memory?

Answer 3

Large = slow/expensive

Cheap = small/slow

Fast = small/expensive

Question 4

What are 4 types of GPU memory types?

Answer 4

Private memory, local memory, global memory, constant memory

Question 5

What are the attributes of private memory?

Answer 5

Very fast, only accessible by a single work item, registers, 10/100 bytes

Question 6

What are the attributes of local memory?

Answer 6

Fast, accessible by all work items within a single work group, user accessible cache, K/MB

Question 7

What are the attributes of global memory?

Answer 7

Slow, accessible by threads from all work groups, DRAM, GB

Question 8

What are the attributes of constant memory?

Answer 8

Fast, also accessible, by all threads, part of global memory but cached, not writable, relatively small, KB

Question 9

What should you minimse time spent on?

Answer 9

Memory operations

Question 10

How do you minimise time spent on memory operations?

Answer 10

Move frequently accessed data to a faster memory

Question 11

What is the order of fast memory?

Answer 11

host&raquo_space; global&raquo_space; local&raquo_space; private

Question 12

What doesn’t benefit from moving frequently accessed data to a faster memory?

Answer 12

Single or sporadic accesses

Question 13

When does data become global memory?

Answer 13

When it is transferred from host to device

Question 14

What is local memory?

Answer 14

Making a local copy of the input to make accesses faster

Question 15

Why do you need synchronisation?

Answer 15

Accesses to shared locations need to be correctly synchronised/coordinated to avoid race conditions

Question 16

What are 3 types of synchronisation mechanisms?

Answer 16

Barriers/memory fences
Atomic operations
Separate kernel launches

Question 17

What do barriers do?

Answer 17

Ensure that all work items within the same work group reach the same point

Question 18

Which has lower overhead, global or local memory barriers?

Answer 18

Local

Question 19

Where should you avoid putting barriers?

Answer 19

In conditional statements, should always apply to all work items from the group otherwise deadlock

Question 20

What is impossible in modern GPU/CPU hardware?

Answer 20

Synchronise different work groups

Question 21

How do you synchronise different workgroups?

Answer 21

By writing and launching separate kernels

Question 22

What do Atomic functions do?

Answer 22

Provide a mechanism for atomic (without interruption) memory operations

Question 23

What do Atomic functions guarantee?

Answer 23

Race free execution

Question 24

How are Atomic updates performed?

Answer 24

Serially, so performance penalty

Question 25

What is the order in Atomic functions?

Answer 25

The order is unspecific, so can only be used with associative and commutative operators

Question 26

What are the limitations of Atomic functions?

Answer 26

Atomics are slower than normal accesses Performance degrades with many simultaneous attempts to perform atomic operations on the same data

Question 27

What is the usage for Atomic functions?

Answer 27

For infrequent, sparse and/or unpredictable global communication Attempt to use shared memory and structure algorithms to avoid synchronisation whenever possible

Question 28

What is does global memory reads by GPU involve?

Answer 28

Reading entire blocks of data

Question 29

What is memory coalescing?

Answer 29

Sequential data access for better performance. When another value is requested and is from the same block then no additional memory access is required

Question 30

What are the effects of the Stride in Strided memory access?

Answer 30

A stride affects the access the pattern, if the stride is larger than the block size then the benefits of blocking are gone