1

I should write a code in CUDA which computes an array in a recursive loop. There is the possibility to precompute some intermediate steps of this recursive loop before it, i.e. to allocate some constant arrays and scalars which would avoid some computations in the loop.

The first idea was to store the constant arrays in the global memory, while scalar parameters are passed every time from the CPU to the GPU (as suggested here: CUDA and shared variables among different global functions).

I would like to try to use the GPU constant memory since it should be faster. However, the few sample codes which I found illustrate how to allocate constant memory from the host. Is it possibile to allocate some constant memory from the GPU, i.e. computing there its values (like we would do with global memory)? And could you please provide a sample code?

Edit: Since I could allocate a lot of constant arrays, maybe the texture memory could be better to use in this situation. Are there some sample codes on how to allocate memory there from the GPU?

Improve this question
2
  • 1
    Texture and constant mem space are both read only for GPU, the data in them can only be initialized from CPU Commented Jan 28, 2013 at 10:56
  • @EricShiyinKang Thank you. What do you think about the suggestion in the first answer? Commented Jan 28, 2013 at 11:10

3 Answers 3

Reset to default
2

To answer your first question: 'Is it possibile to allocate some constant memory from the GPU'. The short answer is yes, as it was answered by other people, copy data from device to device constant memory.

From here, you need to consider the access pattern and the amount of data that your problem requires.

For the constant memory, the amount of memory available is 65536 bytes and the data is broadcast if all the threads within a warp access the same element at the same time. However, 64KB of memory could not be enough.

Texture memory has special features like filtering and 2D spatial locality which is cached. So, using the texture memory for a typical filter in a 3x3 window is the typical case of use.

Finally, if you need to update the data and use them among some kernels, your options goes to the use of the global memory. In addition, you could use the surface memory (CUDA C Programming Guire, Chapter 3.2.10.2) which works as a read/write texture memory.

As you are in a phase of I should write a code in CUDA which computes an array in a recursive loop', you should try first the global memory to get a base for future improvements. When you have your kernel working, you will see which accesses can be rearranged or distributed in a different way in order to get the maximum performance of the GPU memory.

As a final note, take into account that the new Fermi and Kepler architectures have incorporated a L1 and L2 cache hierarchy for global memory accesses which can mitigate random access pattern and even outperform texture memory, as the amount of L1/L2 caches is larger.

Finally, you can find a lot of sample codes in the CUDA SDK.

Improve this answer
Sign up to request clarification or add additional context in comments.

Comments

1

As you can read here it should be possible to copy data from the gpu direct to constant memory, if you use cudaMemcpyToSymbol with the cudaMemcpyDeviceToDevice flag. But it's not possible to edit the date of the constant memeroy like you would do with global memory. You can only read from it.

Improve this answer

Comments

0

It is possible to read and write the constant memory from the host. You could only read the constant memory from the device.

Improve this answer

3 Comments

This is very bad. Hence, I should compute the constant arrays on the GPU (they are quite expensive to compute), transfer the data to the CPU and then again transfer it to the GPU constant memory?
Yes you are correct. Or you could just let the data stay in global memory and rely on the GPU caches to cache the data for you. The later is true if you have a GPU with caches.
Yes, luckily I have, but I hoped there could be the possibility to make my code faster. Now I'm trying to implement what the other answer says.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.