Sobol sampler benchmarks#275
Conversation
…sts and benchmarks, example 37 and 64 now use a single command buffer for benchmarks
# Conflicts: # 31_HLSLPathTracer/app_resources/hlsl/next_event_estimator.hlsl
| [[loop]] | ||
| for (uint16_t d = _static_cast<uint16_t>(0u); d < Depth; d++) | ||
| { |
There was a problem hiding this comment.
might want to add a mode where the depths are slightly staggered per invocation.
Explanation, SER and other things like wavefront path tracing may pack ray payloads at different depths into invocations of the same subgroup, the difference won't be wild, but it will be enough for you to have to account for having to load different matrices per invocation (which tensor cores really don't like and you need to strip with subgroup ops)
You need to add push constant dependent probability of random (depending on the triple[0].z value) break out of the depth loop to simulate paths terminating early either due to miss, absorption or russian roulette and not getting refilled. IMHO best would be a linear PDF, e.g. probability at depth=0 and probability at depth=max
…ta, staggering mask
2 participants
No description provided.