More than once I have noticed that when I have a fixed size read-only array which is small enough to fit in constant memory, it instead seems to be read faster when accessed via const T* restrict Arr (via parameter in kernel call). This is for the Tesla K20c… Would there be any reason why this is …

Are your array accesses uniform across the warp, in other words, do all threads in a warp access the same address? The constant cache has a single port that broadcasts data to each thread in a warp. If more than one address is accessed across the warp, the cache access needs to be repeated as many t…

Dear aduC, Could I support njuffa. In my experience in practice constant memory is surprisingly slow and difficult to use. See previous discussion [image] Constant memory 64KB only 8KB usable? Appendix G of the CUDA 3.1 C programming guide says My largest kernel uses 55878 bytes of …

In my current situation there is a small long long array which caches the powers of 7 up to 7^17. Each thread reads multiple times from this array, though each will read a subset of that array which will not be the same across the warp. This specific code takes hours to run, so I will test both met…

You might want to consider yet another variant where the required powers are computed on the fly. With the maximum power being 17 it should take at most five steps. It is hard to predict whether that would be faster (64-bit integers are emulated), but a general observation is that with modern GPUs, …

My first run actually only took 47 minutes, which was less than I expected given the huge number of calculations. The problem I solved was one where you have a 4x4 grid with numbers from 0 to 6 inclusive. The goal is to determine which board arrangement results in the most rows+columns+diagonals wh…

Are all load/store indices into “int Arr[16];” constants? If so then is NVCC correctly representing the entire array in 16 registers or is it spilling to local memory (bad!)?

I assume what was meant is “are all indexes into this array compile-time constants”, which is a necessary but not sufficient condition for a local array to be mapped to registers. Another consideration is the size of such an array. One could reasonably assume that sixteen 64-bit long longs (equivale…

I would claim perspective is important: The default storage for a local array is local memory, where “local” means “thread-local”. The compiler may, as an optimization, promote the local array to register storage. Whether such promotion takes place is subject to certain requirements (e.g. indexes m…

Just trying to figure out the best/fastest way to compute the powers of 7 (inside the kernel) for a long long type. Should I use the double function pow(double, double) from the cuda math library and cast to long long, or may that cause some rounding issues?

const restrict read faster than constant ?

Accelerated Computing CUDA CUDA Programming and Performance

allanmac January 15, 2014, 9:04pm 9

Yes, I understand it’s not the same as spilling but I didn’t have a better word for, “… variables that you really think should be held in registers but are being unexpectedly demoted to local memory by the compiler.” :)

Topic		Replies	Views
Constant Arrays CUDA Programming and Performance	13	30852	November 24, 2007
Really slow constant memory Random access to constant memory CUDA Programming and Performance	13	4686	December 4, 2009
How fast is a const array defined in a kernel? CUDA Programming and Performance	9	1423	April 21, 2022
__constant__ memory in function scope CUDA Programming and Performance	13	4963	June 1, 2015
Using lookup table in constant memory CUDA Programming and Performance	5	3543	January 10, 2019
constant cache no faster than global mem? constant memory access CUDA Programming and Performance	5	5804	March 6, 2008
Constant memory accesses by a threads of a half warp CUDA Programming and Performance	10	2085	October 27, 2014
Fast constant memory access from all threads in CUDA CUDA Programming and Performance	0	409	September 11, 2019
Constant memory 64KB only 8KB usable? CUDA Programming and Performance	13	7230	February 15, 2011
Warp Serialisation and Constant Memory Performance Surprise CUDA Programming and Performance	7	4015	March 3, 2009

const __restrict__ read faster than __constant__ ?

Related topics

const restrict read faster than constant ?