15-GPU-Threads-and-Kernels

Question 1

What are OpenCL devices?

Answer 1

Belong to platforms, represent GPUs or other accelerators

Question 2

What are OpenCL contexts?

Answer 2

Coordinate interaction between host and device, one per device

Question 3

What are command queues?

Answer 3

Request work execution on device, usually one per device

Question 4

What is simpleOpenContext_GPU()?

Answer 4

Helper function that finds first GPU and creates context/queue

Question 5

What is device memory allocation?

Answer 5

Use clCreateBuffer() to allocate GPU memory

Question 6

What are memory flags?

Answer 6

CL_MEM_READ_ONLY, CL_MEM_WRITE_ONLY, CL_MEM_READ_WRITE, CL_MEM_COPY_HOST_PTR

Question 7

What is a kernel?

Answer 7

Function executed on GPU, preceded by __kernel

Question 8

What is get_global_id(0)?

Answer 8

Returns global thread index for current work item

Question 9

What is kernel compilation?

Answer 9

OpenCL kernels compiled at runtime using clCreateProgramWithSource()

Question 10

What is compileKernelFromFile()?

Answer 10

Helper function that reads, compiles, and creates kernel from file

Question 11

What are kernel arguments?

Answer 11

Set with clSetKernelArg() before enqueueing

Question 12

What is clEnqueueNDRangeKernel()?

Answer 12

Launches kernel with specified work dimensions and sizes

Question 13

What are work items?

Answer 13

Basic unit of GPU execution, maps to hardware thread

Question 14

What is work group size?

Answer 14

Number of work items per work group, affects performance

Question 15

What is NDRange?

Answer 15

N-dimensional range of all work items in kernel launch

Question 16

What is the hierarchy: work items and work groups?

Answer 16

Work items in work groups, communication only within groups

Question 17

What are local indices?

Answer 17

get_local_id() gives position within work group

Question 18

What are global indices?

Answer 18

get_global_id() gives position in entire NDRange

Question 19

What is the advantage of local memory?

Answer 19

Faster than global, shared within work group

Question 20

What is memory coalescing?

Answer 20

GPU optimization where adjacent threads access adjacent memory

Question 21

What is the copy pattern: host to device?

Answer 21

clEnqueueWriteBuffer() or CL_MEM_COPY_HOST_PTR

Question 22

What is the copy pattern: device to host?

Answer 22

clEnqueueReadBuffer() with CL_TRUE for blocking

Question 23

What is the typical GPU program flow?

Answer 23

Allocate memory → copy to GPU → launch kernel → copy results back