Skip to content
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.

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

Jump to bottom

[ROCm] add ROCm support for XLA RCCL thunk #36106

Merged
merged 2 commits into from Mar 30, 2020
Merged

[ROCm] add ROCm support for XLA RCCL thunk #36106

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
ab13615
add ROCm support for XLA RCCL thunk
jeffdaily Jan 21, 2020
1206aa6
use alias for nested select statements
jeffdaily Feb 5, 2020
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Jump to
Jump to file
Failed to load files.
Diff view
Diff view
28 changes: 25 additions & 3 deletions tensorflow/compiler/xla/service/gpu/BUILD
View file
Open in desktop
Expand Up @@ -11,11 +11,13 @@ load(
)
load(
"//tensorflow:tensorflow.bzl",
"if_cuda_or_rocm",
"tf_cc_test",
"tf_copts",
"tf_cuda_library",
)
load("@local_config_cuda//cuda:build_defs.bzl", "if_cuda")
load("@local_config_rocm//rocm:build_defs.bzl", "if_rocm")
load(
"//tensorflow/core/platform/default:cuda_build_defs.bzl",
"if_cuda_is_configured",
Expand Down Expand Up @@ -399,9 +401,24 @@ filegroup(
),
)

# use alias since nested select statements not possible
cc_library(
name = "empty",
)

alias(
name = "virtual_nccl",
actual = if_cuda("@local_config_nccl//:nccl", ":empty"),
)

alias(
name = "virtual_rccl",
actual = if_rocm("@local_config_rocm//rocm:rccl", ":empty"),
)

tf_cuda_library(
name = "nccl_all_reduce_thunk",
srcs = if_cuda(
srcs = if_cuda_or_rocm(
[":nccl_all_reduce_thunk_src"],
["dummy_all_reduce_thunk.cc"],
),
Expand All @@ -420,10 +437,15 @@ tf_cuda_library(
"//tensorflow/compiler/xla/service:hlo",
"//tensorflow/core:lib",
"//tensorflow/core:stream_executor_no_cuda",
] + if_cuda([
"//tensorflow/stream_executor/cuda:cuda_activation",
"//tensorflow/stream_executor/cuda:cuda_gpu_executor",
] + if_nccl([
"@local_config_nccl//:nccl",
]) + if_rocm([
"//tensorflow/stream_executor/rocm:rocm_activation",
"//tensorflow/stream_executor/rocm:rocm_gpu_executor",
]) + if_nccl([
":virtual_nccl",
":virtual_rccl",
]),
)

Expand Down
18 changes: 16 additions & 2 deletions tensorflow/compiler/xla/service/gpu/nccl_all_reduce_thunk.cc
View file
Open in desktop
Expand Up @@ -28,7 +28,11 @@ limitations under the License.
#include "absl/strings/str_join.h"
#include "absl/types/optional.h"
#include "absl/types/span.h"
#if GOOGLE_CUDA
#include "third_party/nccl/nccl.h"
#elif TENSORFLOW_USE_ROCM
#include "rocm/include/rccl/rccl.h"
#endif
#include "tensorflow/compiler/xla/layout_util.h"
#include "tensorflow/compiler/xla/refcounting_hash_map.h"
#include "tensorflow/compiler/xla/service/collective_ops_utils.h"
Expand All @@ -37,7 +41,17 @@ limitations under the License.
#include "tensorflow/compiler/xla/util.h"
#include "tensorflow/core/lib/core/blocking_counter.h"
#include "tensorflow/core/platform/mutex.h"
#include "tensorflow/stream_executor/cuda/cuda_activation.h"
#include "tensorflow/stream_executor/gpu/gpu_activation.h"

#if TENSORFLOW_USE_ROCM
// Local hipify of cuda symbols
#define cudaError_t hipError_t
#define cudaStream_t hipStream_t
#define cudaGetErrorString hipGetErrorString
#define cudaGetDevice hipGetDevice
#define cudaSetDevice hipSetDevice
#define cudaSuccess hipSuccess
#endif

namespace xla {
namespace gpu {
Expand Down Expand Up @@ -406,7 +420,7 @@ RendezvousNcclAllReduce::SubmitParticipantImpl(
CHECK(allreduce_datatype.has_value());

se::StreamExecutor* executor = participant.stream->parent();
se::cuda::ScopedActivateExecutorContext scoped_context(executor);
se::gpu::ScopedActivateExecutorContext scoped_context(executor);
cudaStream_t* cu_stream = reinterpret_cast<cudaStream_t*>(
participant.stream->implementation()->GpuStreamMemberHack());
VLOG(3) << "Using stream pointer: " << cu_stream
Expand Down