Buffer Device Address in HLSL

[devshgraphicsprogramming]

Description

Testing

TODO list:

[devshgraphicsprogramming]

its a little weird to shove non-uniform (ones that change between workgroups) parameters together with uniform

[devshgraphicsprogramming]

argh, crap I forgot you'd need to store the scan result back to smem to be able to

[devshgraphicsprogramming]

ok lets go back to the original simple histogram.atomicAdd(tid,sdata.get(tid));

I forgot the prefix sum is not already in smem, so the extra barriers will just make us slower.

[devshgraphicsprogramming]

this permamently leaves your tid changed, did you want to do this?

[devshgraphicsprogramming]

look, my snippet sometimes spawn a new temp so you leave tid unchanged after you exit the loop
#690 (comment)

[devshgraphicsprogramming]

why areyou using % operator? This is literally the most expensive operator on the GPU (unless when the rhs is a Power of Two)

[devshgraphicsprogramming]

try adding a

using key_t = decltype(declval<KeyAccessor>().get(0));

and using that throughout instead of typename Key

[devshgraphicsprogramming]

this is actually the one of the times you should use

const uint32_t RoundedUpBucketsPerThread = KeyBucketCount-1)/GroupSize+1;
for (uint32_t i=1; i<RoundedUpBucketsPerThread; i++)

formulation (you're making it clear every thread hits the loop in uniform control flow (inclusive_scan requires this)

[devshgraphicsprogramming]

I don't get why you calculate the index with %

Anyway if you go back to a uniform loop, you can compute it as i*GroupSize

[devshgraphicsprogramming]

you might do a little bit of "GroupSize not multiple of KeyBucketCount" handling here with

const uint32_t val = tid<KeyBucketCount ? sdata.get(tid):0;
sum = inclusive_scan(val,sdata);

[devshgraphicsprogramming]

ok this still makes sense to do with a toroidal shift in spite of https://github.com/Devsh-Graphics-Programming/Nabla/pull/690/files#r1621277804

but there's a different way to write this

const uint32_t shift = glsl::gl_SubgroupSize()*params.workGroupIndex;
for (uint32_t vtid=tid; vtid<KeyBucketCount; vtid+=GroupSize)
{
   // have to use modulo operator in case `KeyBucketCount<=2*GroupSize`, better hope KeyBucketCount is Power of Two
   const uint32_t shifted_tid = (vtid+shift)%KeyBucketCount;

   const uint32_t bucket_value = sdata.get(shifted_tid);
   const uint32_t firstOutputIndex = histogram.atomicSub(shifted_tid,bucket_value)-bucket_value;
   sdata.set(shifted_tid,firstOutputIndex);
}

doing a loop conditioned on shifted_tid would break unrolling during optimization, plus your %GroupSize wraps around differently than a %KeyBucketCount would (you get stuck within the [0,GroupSize) block, not really easing contetion).