Atomic Increment with Threshold Check

Accelerated Computing CUDA CUDA Programming and Performance
1

Hello,

i want to increment my acculumator cell and i am using atomicInc(…).
My accumulator provides 2 Byte for each cell, so i have to check if
the value i want to increment is still not higher than 32767.
What’s the best way to do this?

So i need an atomicInc function with a condition which checks the threshold or something like this.

And of course i know that an atomicInc is not the best opportunity for
incrementing accumulator cells, but in the first step it’s the simplest way and
i want to performance measurement with this approach.

2

increment and then decrement on overflow

3

if i increment with atomic add and the value is 32767 than overflow occurs because i can’t stop the incrementation in between.

4

yes, i was incorrect. on overflow, you just need to set result to 32767 unconditionally

alternatively, you can use atomicCAS, which is developed exactly for any complex computations

5

but i can’t stop the atomicAdd in between…

*val = 32676

atomicInc(val, 1);

so the program terminate because it’s an overflow.
If i read the value on the adress val before, it can be possible that more than one thread
reads the value and test the condition if val < 32676 and try to atomicInc the value.
It can be possible that the overflow occurs.

I need a good possibility, that each thread only increment if the actual value is <32676 and
each thread need consistent data

6

program doesn’t terminate on overflow, it just calculates 32767+1 as -32768

i also proposed the solution based on atomicCAS

7

so i need to check if the value is 32767 or <0?

how does the solution with atomicCAS works?

8

  1. the value is 32767 before increment and -32768 after increment. i don’t know which one atomicInc returns. you need to react only in this case, extra increments after overflow will be handled automatically by resetting value back to 32767

  2. read the docs and find some examples yourself

9

As already indicated, you can build “custom atomic” functions based on atomicCAS. The programming guide demonstrates a sample “custom atomic” that implements double atomicAdd on devices that don’t natively support that operation:

[url]Programming Guide :: CUDA Toolkit Documentation

that example could be modified to perform your operation. Here’s another “custom atomic” example that updates two separate 32-bit (packed) variables, based on a min-test of one of them:

[url]cuda - How can I implement a custom atomic function involving several variables? - Stack Overflow

If your 16-bit quantities are packed (e.g. an array of short, for example), then you can still do atomics on them by working with an ordinary int quantity, but masking/updating only the necessary bits.