ayo
6684888c0b
- Reprepare enc3 weights each launch to avoid stale caches after feature recreation - Require per-slot input markers before evaluating recorded frames - Advance split-frame timelines only through completed frames - Add cancellation and readiness tests plus recreation consistency checks - Update architecture and native kernel documentation ref #8
35 lines
1021 B
C++
35 lines
1021 B
C++
#pragma once
|
|
#include <cstdint>
|
|
#include <deque>
|
|
|
|
// A timeline signal releases every lower value too. Prep can cancel a later
|
|
// frame while the CUDA worker is still producing an earlier one, so only the
|
|
// completed prefix of admitted frames may be signalled. The caller serializes
|
|
// admission and retirement with the same mutex used for the timeline signal.
|
|
class D4rFrameCompletion
|
|
{
|
|
public:
|
|
void admit(uint32_t frame) { pending_.push_back({frame, false}); }
|
|
|
|
uint64_t retire(uint32_t frame)
|
|
{
|
|
for (auto& pending : pending_)
|
|
if (pending.frame == frame)
|
|
{
|
|
pending.done = true;
|
|
break;
|
|
}
|
|
while (!pending_.empty() && pending_.front().done)
|
|
{
|
|
completedThrough_ = pending_.front().frame;
|
|
pending_.pop_front();
|
|
}
|
|
return completedThrough_;
|
|
}
|
|
|
|
private:
|
|
struct Pending { uint32_t frame; bool done; };
|
|
std::deque<Pending> pending_;
|
|
uint64_t completedThrough_ = 0;
|
|
};
|