Skip to content

added optimized merged pool embedding script #122

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
wants to merge 1 commit into
base: mi350_develop_0811
Choose a base branch
from
Open

added optimized merged pool embedding script #122

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
196 changes: 94 additions & 102 deletions fbgemm_gpu/src/merge_pooled_embedding_ops/merge_pooled_embedding_ops_gpu.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -159,19 +159,13 @@ void all_to_one(
// the target device. Since target device can be different across each call,
// we store all the events in a 2-dimentionsal vector.
using PerDeviceEventList = std::vector<at::cuda::CUDAEvent>;
static thread_local std::vector<PerDeviceEventList> two_hop_copy_begin_events;
static thread_local std::vector<PerDeviceEventList> copy_begin_events;
static thread_local std::vector<PerDeviceEventList> copy_completion_events;
static thread_local std::once_flag flag1;
std::call_once(flag1, [num_gpus]() {
two_hop_copy_begin_events.reserve(num_gpus);
copy_begin_events.reserve(num_gpus);
copy_completion_events.reserve(num_gpus);

for (auto i = 0; i < num_gpus; i++) {
two_hop_copy_begin_events.emplace_back(num_gpus);
copy_begin_events.emplace_back(num_gpus);
copy_completion_events.emplace_back(num_gpus);
copy_begin_events.push_back(PerDeviceEventList(num_gpus));
copy_completion_events.push_back(PerDeviceEventList(num_gpus));
}
});

Expand All @@ -196,37 +190,27 @@ void all_to_one(
auto intermediate_node =
intermediate_nodes(src_device_id, target_device_index);
if (intermediate_node != -1) {
Tensor dst = at::empty_like(
src, src.options().device(at::Device(at::kCUDA, intermediate_node)));

// This is a cross-device copy from the src device to the dst device.
// We need to synchronize src with dst to ensure that the dst memory
// is ready; Write-after-Write or Write-after-Read dependencies are
// handled. E.g. CudaCachingAllocator might reuse existing tensor memory
// space that has ongoing kernel ops
//
// To address this, we use cudaEvent to make src stream wait on all
// kernels on the dst stream (after dst memory initialization) to be
// completed prior to the copy
const auto source_device_id = src.get_device();
at::cuda::CUDAStream copy_stream =
at::cuda::getCurrentCUDAStream(source_device_id);

auto& dst_ready =
two_hop_copy_begin_events[intermediate_node][source_device_id];
dst_ready.record(at::cuda::getCurrentCUDAStream(intermediate_node));
dst_ready.block(copy_stream);

two_hop_transfers.push_back(
{.intermediate_tensor = dst,
{.intermediate_tensor = at::empty(
src.sizes(),
src.options().device(at::Device(at::kCUDA, intermediate_node))),
.output_idx = i,
.transfer_cuda_event =
std::make_unique<at::cuda::CUDAEvent>(cudaEventDisableTiming)});
auto& dst = two_hop_transfers.back().intermediate_tensor;
at::cuda::CUDAStream copy_stream =
at::cuda::getCurrentCUDAStream(src_device_id);
void* tmp1 = nullptr;
void* tmp2 = nullptr;
size_t srcPitchBytes;
size_t dstPitchBytes;
AT_CUDA_CHECK(cudaMallocPitch(&tmp1, &srcPitchBytes, src.size(1) * src.element_size(), src.size(0)));
AT_CUDA_CHECK(cudaMallocPitch(&tmp2, &dstPitchBytes, dst.size(1) * dst.element_size(), dst.size(0)));
AT_CUDA_CHECK(cudaMemcpy2DAsync(
dst.data_ptr(),
dst.stride(0) * dst.element_size(),
dstPitchBytes,//dst.stride(0) * dst.element_size(),
src.data_ptr(),
src.stride(0) * src.element_size(),
srcPitchBytes,//src.stride(0) * src.element_size(),
src.size(1) * src.element_size(),
src.size(0),
cudaMemcpyDeviceToDevice,
Expand All @@ -248,11 +232,13 @@ void all_to_one(
}

// synchronize source streams and launch copies on source stream.
at::cuda::CUDAGuard device_guard(src_device);
// We always perform the copy on the source device, using the current
// stream on the source device, and we fully synchronize on both src and
// dst's current streams for completion of the copy. We have to explicitly
// do this for non-contig copies. This mimics the behavior of cross-device
// cudaMemcpyAsync on the default stream.

at::cuda::CUDAStream copy_stream =
at::cuda::getCurrentCUDAStream(device_id);
// This is a cross-device copy on the src current stream and dst current
Expand All @@ -262,7 +248,9 @@ void all_to_one(
// that no one is operating on the dst memory when we perform the copy.
// src waits on dst barrier (src already waits on src)
auto& dst_ready = copy_begin_events[target_device_index][device_id];
device_guard.set_device(target_device);
dst_ready.record(at::cuda::getCurrentCUDAStream(target_device_index));
device_guard.set_device(src_device);
dst_ready.block(copy_stream);
for (const auto i : c10::irange(input_tensors.size())) {
const auto metadata = is_two_hop_transfer.at(i);
Expand All @@ -279,11 +267,17 @@ void all_to_one(

auto& dst = output_tensors[i];
// on source device, launch memcpy.
void* tmp1 = nullptr;
void* tmp2 = nullptr;
size_t srcPitchBytes;
size_t dstPitchBytes;
AT_CUDA_CHECK(cudaMallocPitch(&tmp1, &srcPitchBytes, src.size(1) * src.element_size(), src.size(0)));
AT_CUDA_CHECK(cudaMallocPitch(&tmp2, &dstPitchBytes, dst.size(1) * dst.element_size(), dst.size(0)));
AT_CUDA_CHECK(cudaMemcpy2DAsync(
dst.data_ptr(),
dst.stride(0) * dst.element_size(),
dstPitchBytes,//dst.stride(0) * dst.element_size(),
src.data_ptr(),
src.stride(0) * src.element_size(),
srcPitchBytes,//src.stride(0) * src.element_size(),
src.size(1) * src.element_size(),
src.size(0),
cudaMemcpyDeviceToDevice,
Expand All @@ -300,24 +294,34 @@ void all_to_one(
continue;
}

// source rank stream
// intermediate rank
at::cuda::CUDAGuard device_guard(src_device);
// intermediate rank stream
at::cuda::CUDAStream copy_stream =
at::cuda::getCurrentCUDAStream(src_device_id);
// wait on first hop transfer
two_hop_transfer.transfer_cuda_event->block(copy_stream);
// synchronize with target rank
auto& dst_ready = copy_begin_events[target_device_index][src_device_id];
device_guard.set_device(target_device);
dst_ready.record(at::cuda::getCurrentCUDAStream(target_device_index));
device_guard.set_device(src_device);
dst_ready.block(copy_stream);
// originating tensor output position
const auto output_index = two_hop_transfer.output_idx;
auto& dst = output_tensors.at(output_index);
// on source device, launch memcpy.
size_t srcPitchBytes;
void* tmp1 = nullptr;
AT_CUDA_CHECK(cudaMallocPitch(&tmp1, &srcPitchBytes, src.size(1) * src.element_size(), src.size(0)));
size_t dstPitchBytes;
void* tmp2 = nullptr;
AT_CUDA_CHECK(cudaMallocPitch(&tmp2, &dstPitchBytes, dst.size(1) * dst.element_size(), dst.size(0)));
AT_CUDA_CHECK(cudaMemcpy2DAsync(
dst.data_ptr(),
dst.stride(0) * dst.element_size(),
dstPitchBytes,//dst.stride(0) * dst.element_size(),
src.data_ptr(),
src.stride(0) * src.element_size(),
srcPitchBytes,//src.stride(0) * src.element_size(),
src.size(1) * src.element_size(),
src.size(0),
cudaMemcpyDeviceToDevice,
Expand All @@ -333,11 +337,17 @@ void all_to_one(
// single device memcpy, not that src_device == dst_device.
at::cuda::CUDAStream copy_stream =
at::cuda::getCurrentCUDAStream(target_device_index);
void* tmp1 = nullptr;
void* tmp2 = nullptr;
size_t srcPitchBytes;
size_t dstPitchBytes;
AT_CUDA_CHECK(cudaMallocPitch(&tmp1, &srcPitchBytes, src.size(1) * src.element_size(), src.size(0)));
AT_CUDA_CHECK(cudaMallocPitch(&tmp2, &dstPitchBytes, dst.size(1) * dst.element_size(), dst.size(0)));
AT_CUDA_CHECK(cudaMemcpy2DAsync(
dst.data_ptr(),
dst.stride(0) * dst.element_size(),
dstPitchBytes,//dst.stride(0) * dst.element_size(),
src.data_ptr(),
src.stride(0) * src.element_size(),
srcPitchBytes,//src.stride(0) * src.element_size(),
src.size(1) * src.element_size(),
src.size(0),
cudaMemcpyDeviceToDevice,
Expand All @@ -350,12 +360,15 @@ void all_to_one(
for (const auto device_id : c10::irange(num_gpus)) {
if (device_id != target_device_index) {
auto src_device = at::Device(at::kCUDA, device_id);
// record stream event
// Still on src_device, record stream event
at::cuda::CUDAGuard device_guard(src_device);
at::cuda::CUDAStream copy_stream =
at::cuda::getCurrentCUDAStream(device_id);

auto& src_ready = copy_completion_events[target_device_index][device_id];
src_ready.record(copy_stream);

device_guard.set_device(target_device);
src_ready.block(at::cuda::getCurrentCUDAStream(target_device_index));
}
}
Expand All @@ -371,19 +384,11 @@ Tensor sum_reduce_to_one(
// the target device. Since target device can be different across each call,
// we store all the events in a 2-dimentionsal vector.
using PerDeviceEventList = std::vector<at::cuda::CUDAEvent>;
static thread_local std::vector<PerDeviceEventList> two_hop_copy_begin_events;
static thread_local std::vector<PerDeviceEventList> copy_begin_events;
static thread_local std::vector<PerDeviceEventList> copy_completion_events;
static thread_local std::once_flag flag1;
std::call_once(flag1, [num_gpus]() {
two_hop_copy_begin_events.reserve(num_gpus);
copy_begin_events.reserve(num_gpus);
copy_completion_events.reserve(num_gpus);

for (auto i = 0; i < num_gpus; i++) {
two_hop_copy_begin_events.emplace_back(num_gpus);
copy_begin_events.emplace_back(num_gpus);
copy_completion_events.emplace_back(num_gpus);
copy_completion_events.push_back(PerDeviceEventList(num_gpus));
}
});

Expand Down Expand Up @@ -412,8 +417,6 @@ Tensor sum_reduce_to_one(
ten.device() == input_tensors[j].device()) {
ten.add_(input_tensors[j]);
// Replace with a dummy tensor without storage to mark reduced away
// With CudaCachingAllocator, ops follow stream semantics. Thus data
// will not be corrupted.
input_tensors[j] = Tensor();
}
}
Expand All @@ -434,37 +437,33 @@ Tensor sum_reduce_to_one(
if (intermediate_node == -1) {
continue;
}
auto intermediate_device = at::Device(at::kCUDA, intermediate_node);
Tensor dst = at::empty_like(src, intermediate_device);

// This is a cross-device copy from the src device to the dst device.
// We need to synchronize src with dst to ensure that the dst memory
// is ready; Write-after-Write or Write-after-Read dependencies are handled.
// E.g. CudaCachingAllocator might reuse existing tensor memory space that
// has ongoing kernel ops
//
// To address this, we use cudaEvent to make src stream wait on all kernels
// on the dst stream (after dst memory initialization) to be completed prior
// to the copy
const auto intermediate_device = at::Device(at::kCUDA, intermediate_node);
copied_tensors[i] = at::empty_like(src, intermediate_device);

const auto src_device_idx = src.get_device();
const auto& copy_stream = at::cuda::getCurrentCUDAStream(src_device_idx);
auto& dst_ready =
two_hop_copy_begin_events[intermediate_node][src_device_idx];
dst_ready.record(at::cuda::getCurrentCUDAStream(intermediate_node));
dst_ready.block(copy_stream);
// This is a cross-device copy on the src current stream and dst current
// stream.
// Unlike all_to_one case, we don't need to wait for dst ready to worry
// about write-after-write and write-after-read dependencies because we're
// creating a temp tensor, dst.

at::cuda::CUDAGuard device_guard(src.device());
// on source device, launch memcpy.
auto& dst = copied_tensors[i];
void* tmp1 = nullptr;
void* tmp2 = nullptr;
size_t srcPitchBytes;
size_t dstPitchBytes;
AT_CUDA_CHECK(cudaMallocPitch(&tmp1, &srcPitchBytes, src.size(1) * src.element_size(), src.size(0)));
AT_CUDA_CHECK(cudaMallocPitch(&tmp2, &dstPitchBytes, dst.size(1) * dst.element_size(), dst.size(0)));
AT_CUDA_CHECK(cudaMemcpy2DAsync(
dst.data_ptr(),
dst.stride(0) * dst.element_size(),
dstPitchBytes,//dst.stride(0) * dst.element_size(),
src.data_ptr(),
src.stride(0) * src.element_size(),
srcPitchBytes,//src.stride(0) * src.element_size(),
src.size(1) * src.element_size(),
src.size(0),
cudaMemcpyDeviceToDevice,
copy_stream));
at::cuda::getCurrentCUDAStream(src.get_device())));
copied_tensors[i] = dst;
}

// Wait for cross-device copies to complete, then reduce
Expand All @@ -476,11 +475,15 @@ Tensor sum_reduce_to_one(
auto intermediate_device = at::Device(at::kCUDA, intermediate_node);

auto src_device = at::Device(at::kCUDA, device_id);
// Still on src_device, record stream event
at::cuda::CUDAGuard device_guard(src_device);
at::cuda::CUDAStream copy_stream =
at::cuda::getCurrentCUDAStream(device_id);

auto& src_ready = copy_completion_events[target_device_index][device_id];
src_ready.record(copy_stream);

device_guard.set_device(intermediate_device);
src_ready.block(at::cuda::getCurrentCUDAStream(intermediate_node));

// Find any tensor in the intermediate GPU to reduce to.
Expand All @@ -501,8 +504,6 @@ Tensor sum_reduce_to_one(
continue;
}
if (ten_at_intermediate_node.has_storage()) {
// With CudaCachingAllocator, ops follow stream semantics. Thus data
// will not be corrupted.
ten_at_intermediate_node.add_(ten);
input_tensors[i] = Tensor();
} else {
Expand All @@ -520,45 +521,40 @@ Tensor sum_reduce_to_one(
continue;
}

// This is a cross-device copy from the src device to the dst device.
// We need to synchronize src with dst to ensure that the dst memory
// is ready; Write-after-Write or Write-after-Read dependencies are handled.
// E.g. CudaCachingAllocator might reuse existing tensor memory space that
// has ongoing kernel ops
//
// To address this, we use cudaEvent to make src stream wait on all kernels
// on the dst stream (after dst memory initialization) to be completed prior
// to the copy
copied_tensors[i] = at::empty_like(src, target_device);

const auto src_idx = src.get_device();
const auto& copy_stream = at::cuda::getCurrentCUDAStream(src_idx);
auto& dst_ready = copy_begin_events[target_device_index][src_idx];
dst_ready.record(at::cuda::getCurrentCUDAStream(target_device_index));
dst_ready.block(copy_stream);
Tensor dst = at::empty_like(src, target_device);

auto& dst = copied_tensors[i];
at::cuda::CUDAGuard device_guard(src.device());
void* tmp1 = nullptr;
void* tmp2 = nullptr;
size_t srcPitchBytes;
size_t dstPitchBytes;
AT_CUDA_CHECK(cudaMallocPitch(&tmp1, &srcPitchBytes, src.size(1) * src.element_size(), src.size(0)));
AT_CUDA_CHECK(cudaMallocPitch(&tmp2, &dstPitchBytes, dst.size(1) * dst.element_size(), dst.size(0)));
AT_CUDA_CHECK(cudaMemcpy2DAsync(
dst.data_ptr(),
dst.stride(0) * dst.element_size(),
dstPitchBytes,//.stride(0) * dst.element_size(),
src.data_ptr(),
src.stride(0) * src.element_size(),
srcPitchBytes,//.stride(0) * src.element_size(),
src.size(1) * src.element_size(),
src.size(0),
cudaMemcpyDeviceToDevice,
copy_stream));
at::cuda::getCurrentCUDAStream(src.get_device())));
copied_tensors[i] = dst;
}

// Wait for cross-device copies to complete, then reduce
for (const auto device_id : c10::irange(num_gpus)) {
if (device_id != target_device_index) {
auto src_device = at::Device(at::kCUDA, device_id);
// Still on src_device, record stream event
at::cuda::CUDAGuard device_guard(src_device);
at::cuda::CUDAStream copy_stream =
at::cuda::getCurrentCUDAStream(device_id);

auto& src_ready = copy_completion_events[target_device_index][device_id];
src_ready.record(copy_stream);

device_guard.set_device(target_device);
src_ready.block(at::cuda::getCurrentCUDAStream(target_device_index));

for (const auto i : c10::irange(input_tensors.size())) {
Expand Down Expand Up @@ -636,12 +632,8 @@ Tensor cat_dim_2d(
cat_dim_2d_output_shape(tensors, uncat_dim_size, cat_dim);

auto* prop = at::cuda::getCurrentDeviceProperties();
auto output = at::empty(
output_shape,
tensors.front()
.options()
.device(output_device)
.dtype(tensors[0].dtype()));
auto output =
at::empty(output_shape, tensors.front().options().device(output_device));
TORCH_CHECK(
output.stride(0) * output.element_size() <=
static_cast<int64_t>(prop->memPitch));
Expand Down Expand Up @@ -741,4 +733,4 @@ TORCH_LIBRARY_FRAGMENT(fbgemm, m) {
"merge_pooled_embeddings", fbgemm_gpu::merge_pooled_embeddings);
DISPATCH_TO_CUDA("all_to_one_device", fbgemm_gpu::all_to_one_device);
DISPATCH_TO_CUDA("sum_reduce_to_one", fbgemm_gpu::sum_reduce_to_one_device);
}
}