Timothy Lochner

Im not sure I’ve ever seen InterlockedAdd (hlsl) work correctly on Nvidia. Works on AMD tho. Can’t help but wonder if I’m doing something wrong, and AMD is more permissive. Seen a few reports of other people noticing the same thing.

InterlockedAdd is bread and butter in a Compute Shader, used all the time by thousands of titles. If NVIDIA have a problem with it it'll be very scenario or driver specific. Anything to illustrate what you're doing?

See, that's my line of thought. Always question your own code first, right? Just something to the gist of: RWBuffer<uint> someBuffer; [numthreads(64,1,1)] void main( uint3 dtId : ... ) { //... InterlockedAdd(someBuffer[dtId.x], someVal); }

64 threads is one wavefront on AMD but two warps on NVidia. What do you do with someBuffer after? You're probably missing memory barriers.

I'm willing to accept a memory barrier is missing. Every time I feel like I understand barriers, they always seem to have another hand to reveal. SomeBuffer is used in another compute shader immediately following this one.

It'd be unusual to need a memory barrier if you're using atomics on UAVs. But if you're trying to read straight back out of someBuffer that could be iffy. Are you sure you didn't want to use the 3 operand version of InterlockedAdd with the 'old' return value?

I use the 3 operand version (and store the result in another buffer). Twitter just didn't give me enough space to write it all out.