Skip to content

Fix H2D staging race with host function callback on all GPUs #13275

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
anugodavar merged 3 commits into from
Oct 28, 2026
Merged

Fix H2D staging race with host function callback on all GPUs #13275

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
35b618d
Fix H2D staging race with host function callback on unified-memory APUs
anugodavar Oct 12, 2026
f133e0b
Fix H2D staging race on all GPUs, not just APUs
anugodavar Oct 22, 2026
2eaec53
Merge branch 'develop' into users/agodavar/fix-host-func-h2d-race
anugodavar Oct 26, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
17 changes: 15 additions & 2 deletions projects/clr/rocclr/device/rocm/rocblit.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1191,8 +1191,13 @@ void DmaBlitManager::releaseBuffer(BufferState& buffer) const {
bool DmaBlitManager::hsaCopyStagedOrPinned(const_address hostSrc, address hostDst, size_t size,
bool hostToDev, amd::CopyMetadata& copyMetadata,
bool enablePin) const {
// Do not skip wait here for D2H. Resolving dependent signals for SDMA engine is slow
gpu().releaseGpuMemoryFence(hostToDev || !dev().settings().blocking_blit_);
// D2H: wait when blocking_blit_ is set — resolving dependent signals for SDMA is slow.
// H2D: the staging path's CPU memcpy races with pending host callbacks (e.g.
// hipLaunchHostFunc writing to hostSrc). The pinned/SDMA path is ordered by signal
// dependencies and does not need a CPU wait, so the fence is deferred to the staging
// path below.
const bool skip_wait = hostToDev || !dev().settings().blocking_blit_;
gpu().releaseGpuMemoryFence(skip_wait);
// If Pinning is enabled, Pin host Memory for copy size > MinSizeForPinnedTransfer
// For 16KB < size <= MinSizeForPinnedTransfer Use staging buffer without pinning
bool status = true;
Expand Down Expand Up @@ -1220,6 +1225,11 @@ bool DmaBlitManager::hsaCopyStagedOrPinned(const_address hostSrc, address hostDs
}
if (hostToDev) { // H2D Path
if (outBuffer.pinnedMem_ == nullptr) { // Copy to Staging Buffer
// Wait for prior GPU work before the CPU reads hostSrc — a host-function
// callback (e.g. hipLaunchHostFunc) may still be writing to it.
if (firstTx) {
gpu().releaseGpuMemoryFence(false);
}
ClPrint(amd::LOG_DETAIL_DEBUG, amd::LOG_COPY, "memcpy stg buf=%p, host src=%p, size=%zu",
stagingBuffer, hostSrc + copyOffset, copysize);
memcpy(stagingBuffer, hostSrc + copyOffset, copysize);
Expand Down Expand Up @@ -2427,6 +2437,9 @@ bool KernelBlitManager::writeBuffer(const void* srcHost, device::Memory& dstMemo
size_t copySize = 0;
size_t stagedCopyOffset = 0;

// Wait for prior GPU work (e.g. hipLaunchHostFunc callback) before the CPU reads srcAddr.
gpu().releaseGpuMemoryFence(false);

while (totalSize > 0) {
BufferState outBuffer = {0};
// Disable pinned writes
Expand Down
Loading