This lab introduces you to CUDA computing with a few small
exercises. We have deliberately avoided larger amounts of code for this
lab, and save that for lab 5.
You should write down answers to all questions and then show us your results.
To be announced.
If you run in the lab, no installation is needed as it has been done for you! If you run on your own computer, this can be more or less complicated. Follow the instructions on NVidia's CUDA page.
However, in the lab you need to do the following:
Download this simple script:
cudaida.sh (for the IDA lab)
cudaisy.sh (for the ISY lab)
Switch to the bash shell by typing "bash" and run the script.
Now you are ready to use CUDA!
To get started, we will use the "simple.cu" example that I introduced in the lecture, arguably the simplest CUDA example. All it does is to assign every element in an array with its index.
Here is the program:
You can compile it with this simple command-line:
nvcc simple.cu -o simple
This seems to work nicely in the lab, but just in case, let me show how a slighly more elaborate line can look:
nvcc simple.cu -L /usr/local/cuda/lib -lcudart -o simple
(Feel free to build a makefile. You can base it on those from earlier labs.)
If it works properly, it should output the numbers 0 to 15 as floating-point numbers.
As it stands, the amount of computation is ridiculously small. We will soon address problems where performance is meaningful to measure. But first, let's modify the code to get used to the format.
QUESTION: How many cores will simple.cu use, max, as written?
Allocate an array of data and use as input. Calculate the square root of every element. Inspect the output so it is correct.
To upload data to the GPU, you must use the cudaMemcpy() call with cudaMemcpyHostToDevice.
Note: If you increase the data size here, please check out the comments in section 2.
QUESTION: Is the calculated square root identical to what the CPU calculates? Should we asume that this is always the case?
Now let us work on a larger dataset and make more computations. We will make fairly trivial item-by-item computations but experiment with the number of blocks and threads.
Here is a simple C program that takes two arrays (matrices) and add them component-wise.
matrix_cpu.cIt can be compiled with
gcc matrix_cpu.c -o matrix_cpu -std=c99
and run with
Write a CUDA program that performs the same thing, in parallel!
Start with a grid size of (1, 1) and a block size of (N, N). Then try a
bigger grid, with more blocks.
We must calculate an index from the thread and block numbers. That can look like this (in 1D):
int idx = blockIdx.x * blockDim.x + threadIdx.x;
QUESTION: How do you calculate the index in the array, using 2-dimensional blocks?
In order to measure execution time for your kernels you should use CUDA Events.A CUDA event variable "myEvent" is declared as
cudaEvent_t myEvent;It must beinitialized with
cudaEventCreate(&myEvent);You insert it in the CUDA stream with
cudaEventRecord(myEvent, 0);The 0 is the stream number, where 0 is the default stream.
Important! You must use cudaEventSynchronize before taking time measurements, or you don't know if the computation has ended!Finally, you get the time between two events with
cudaEventElapsedTime(&theTime, myEvent, laterEvent);where theTime is a float.
Note that in CUDA, large arrays are best allocated in C++ style: float *c = new float[N*N];
Vary N. You should be able to run at least 1024x1024 items.
Vary the block size (and consequently the grid size).
QUESTION: What happens if you use too many threads per block? (Hint: You can get misleading output if you don't clear your buffers.)
Also note that timings can be made with or without data transfers. In a serious computations, the data transfers certainly are important, but for our small examples, it is more relevant to take the time without data transfers, or both with and without.
Note: When you increase the data size, there are (at least) two problems that you will run into:
QUESTION: At what data size is the GPU faster than the CPU?
QUESTION: What block size seems like a good choice? Compared to what?
QUESTION: Write down your data size, block size and timing data for the best GPU performance you can get.
In this part, we will make a simple experiment to inspect the impact of memory coalescing, or the lack of it.
In part 2, you calculated a formula for accessing the array items based on thread and block indices. Probably, a change in X resulted in one step in the array. That will make data aligned and you get good coalescing. Swap the X and Y calculations so a change in Y results in one step in the array. Measure the computation time.
Note: The results have varied a lot in this exercise. On my 9400, I got a very significant difference, while this difference may not be as big in the lab. However, lack of difference can be caused by erroneous computation, so make sure that the output is correct.
QUESTION: How much performance did you lose by making data accesses non-coalesced?
Put in a timer in the code. Only by inspecting the code, you should find that it isn't reasonably written. Why? Fix the problem. (Hint: It is nothing that you havn't seen before, but rather a rehearsal of older problems - but a lot prettier.)
Here you will need one more CUDA event call: cudaEventDestroy(), to clean up after a CUDA event and avoid memory leaks.
QUESTION: How did you correct it?
QUESTION: What speedup did you get?
WARNING: We have discovered a problem with this lab. It can cause lockups. We are investigating the problem.
QUESTION: What kind of modifications did you implement?