Pinned host memory from a process-wide pool; the compressed chunk uploaded from pinned staging

The GPU engines are built per worker per step, and each took its page-locked host buffers with
cudaHostAlloc / cudaHostRegister and gave them back with cudaFreeHost / cudaHostUnregister. Those
calls map and lock pages under the process's memory-map lock, so all workers queue behind each
other: on the 5080 workstation a myoglobin run made ~3400 of them (6-22 thread-s, depending on the
load), 9gdj ~10400 (124 thread-s); on a four-socket host they cost 20-30 ms each.

- CudaHostPtr takes its memory from a pool: a freed buffer is kept in a power-of-two size class for
  the next engine, up to 512 MiB in all; a buffer above 64 MiB is never kept. Rugnux::Run hands the
  pool back at its end.
- PinnedVector (std::vector with the pool's allocator) replaces the registered vectors of
  AdaptiveSpotFinderGPU and ImagePreprocessorGPU; the Bragg predictors download into a pooled
  pinned buffer and copy into their output, instead of registering it.
- BSLZ4DecoderGPU copies the compressed chunk into a pooled pinned buffer before the upload, so it
  is one DMA instead of the driver's pageable staging (9gdj: 163 GB pageable H2D, 52 s of copy
  time, -> 147 GB pinned, 13 s).

Measured with nsys on the workstation: myob 3398 -> 1396 calls (674 of them the release at the
end), cytc 2986 -> 1270, 9gdj 10428 -> 1748 (124 -> 3.7 thread-s). Output identical (p.hkl md5 on
myob, cytc, 9gdj; CI tier battery).

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EizBKhTqrAYago9KJG3dA3
This commit is contained in:
leonarski_fandClaude Opus 5.5 committed 2026-10-11 10:37:23 +02:00
1 parent 54f8e07591
commit 09c41d067d
13 files changed
+175 -45

No files matched your search

+24
View File
@@ -103,6 +103,30 @@ TEST_CASE("CudaDevicePtr_HandledPoolFailureLeavesNoError", "[CUDAMemHelpers]") {
cudaMemPoolDestroy(capped);
}
// Engines built one after another take the pinned buffers the previous one gave back - also a little
// larger ones, within the size class - while a frame-sized buffer is not kept at all.
TEST_CASE("CudaPinnedPool_ReusedByTheNextEngine", "[CUDAMemHelpers]") {
if (get_gpu_count() == 0)
SKIP("No CUDA GPU present");
cuda_release_pinned_pool();
void *first = nullptr;
{
CudaHostPtr<float> a(1000);
first = a.get();
}
{
PinnedVector<float> b(1020);
CHECK(static_cast<void *>(b.data()) == first);
}
{
CudaHostPtr<uint8_t> large(CUDA_PINNED_POOL_MAX_BUFFER + 1);
large.get()[CUDA_PINNED_POOL_MAX_BUFFER] = 1;
}
CHECK(cuda_release_pinned_pool() == 4096);
CHECK(cuda_release_pinned_pool() == 0);
}
#endif
// A merge that finds the card full waits for the GPU work running beside it: until the last one