Skip to content

Ensure 32-byte PCIe DMA alignment for host staging buffers #481

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

Open
copybara-service wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Ensure 32-byte PCIe DMA alignment for host staging buffers #481

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
24 changes: 18 additions & 6 deletions tpu_raiden/weight_sync/weight_synchronizer_base.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -270,16 +270,22 @@ absl::StatusOr<raiden::PjRtCopyFuture> WeightSynchronizerBase::H2d() {

std::vector<xla::Future<>> shard_futures;
if (is_tiled) {
auto temp_buffer = std::make_shared<std::vector<uint8_t>>(layer_size);
void* raw_aligned_ptr = nullptr;
if (posix_memalign(&raw_aligned_ptr, 32, layer_size) != 0) {
return absl::ResourceExhaustedError(
"Failed to allocate 32-byte aligned host staging buffer");
}
auto temp_buffer = std::shared_ptr<uint8_t>(
static_cast<uint8_t*>(raw_aligned_ptr), [](uint8_t* p) { free(p); });
auto status = tpu_raiden::weight_sync::TileBuffer(
shard_info.host_ptr, temp_buffer->data(),
shard_info.host_ptr, temp_buffer.get(),
shard_hold.buffer->on_device_shape(), *xla_layout);
if (!status.ok()) {
return status;
}

xla::Future<> future =
shard_hold.CopyRawHostToDevice(temp_buffer->data(), 0, layer_size);
shard_hold.CopyRawHostToDevice(temp_buffer.get(), 0, layer_size);
xla::Future<> mapped_future = future.Map([temp_buffer]() {});
shard_futures.push_back(std::move(mapped_future));
} else {
Expand Down Expand Up @@ -318,8 +324,14 @@ absl::StatusOr<raiden::PjRtCopyFuture> WeightSynchronizerBase::D2h() {

std::vector<xla::Future<>> shard_futures;
if (is_tiled) {
auto temp_buffer = std::make_shared<std::vector<uint8_t>>(layer_size);
uint8_t* temp_buffer_ptr = temp_buffer->data();
void* raw_aligned_ptr = nullptr;
if (posix_memalign(&raw_aligned_ptr, 32, layer_size) != 0) {
return absl::ResourceExhaustedError(
"Failed to allocate 32-byte aligned host staging buffer");
}
auto temp_buffer = std::shared_ptr<uint8_t>(
static_cast<uint8_t*>(raw_aligned_ptr), [](uint8_t* p) { free(p); });
uint8_t* temp_buffer_ptr = temp_buffer.get();

xla::Future<> copy_future =
shard_hold.CopyRawDeviceToHost(temp_buffer_ptr, 0, layer_size);
Expand All @@ -329,7 +341,7 @@ absl::StatusOr<raiden::PjRtCopyFuture> WeightSynchronizerBase::D2h() {
shape = shard_hold.buffer->on_device_shape(),
layout = *xla_layout]() -> absl::Status {
return tpu_raiden::weight_sync::DetileBuffer(
temp_buffer->data(), dst_host_ptr, shape, layout);
temp_buffer.get(), dst_host_ptr, shape, layout);
});

shard_futures.push_back(std::move(detile_future));
Expand Down
Loading