[SYCL][Graph][HIP] Add HIP backend support to SYCL-Graph (#12230)

Tests HIP Support for Graph from oneapi-src/unified-runtime#1254 and updates documentation. Depends on #12753. --------- Co-authored-by: Andrey Alekseenko <andrey.alekseenko@scilifelab.se> Co-authored-by: Ewan Crawford <ewan@codeplay.com>
intel · Mar 1, 2024 · 897b270 · 897b270
1 parent be8ad8f
commit 897b270
Show file tree

Hide file tree

Showing 7 changed files with 36 additions and 14 deletions.
diff --git a/sycl/doc/design/CommandGraph.md b/sycl/doc/design/CommandGraph.md
@@ -236,7 +236,7 @@ Implementation of UR command-buffers
 for each of the supported SYCL 2020 backends.
 
 Backends which are implemented currently are: [Level Zero](#level-zero),
-[CUDA](#cuda), and partial support for [OpenCL](#opencl).
+[CUDA](#cuda), [HIP](#hip) and partial support for [OpenCL](#opencl).
 
 ### Level Zero
 
@@ -351,6 +351,27 @@ An executable CUDA Graph, which contains all commands and synchronization
 information, is saved in the UR command-buffer to allow for efficient
 graph resubmission.
 
+### HIP
+
+The HIP backend offers a Graph managemenet API very similar to CUDA Graph
+feature for batching series of operations.
+The SYCL Graph HIP backend implementation is therefore very similar to that of CUDA.
+
+UR commands (e.g. kernels) are mapped as graph nodes using the
+[HIP Management API](https://docs.amd.com/projects/HIP/en/docs-5.5.0/doxygen/html/group___graph.html).
+Synchronization between commands (UR sync-points) is implemented
+using graph dependencies.
+Executable HIP Graphs can be submitted to a HIP stream
+in the same way as regular kernels.
+The HIP backend enables enqueuing events to wait for into a stream.
+It also allows signaling the completion of a submission with an event.
+Therefore, submitting a UR command-buffer consists only of submitting to a stream
+the executable HIP Graph that represent this series of operations.
+
+An executable HIP Graph, which contains all commands and synchronization
+information, is saved in the UR command-buffer to allow for efficient
+graph resubmission.
+
 ### OpenCL
 
 SYCL-Graph is only enabled for an OpenCL backend when the

diff --git a/sycl/doc/design/images/SYCL-Graph-Architecture.svg b/sycl/doc/design/images/SYCL-Graph-Architecture.svg
diff --git a/sycl/plugins/unified_runtime/CMakeLists.txt b/sycl/plugins/unified_runtime/CMakeLists.txt
@@ -57,13 +57,14 @@ if(SYCL_PI_UR_USE_FETCH_CONTENT)
   include(FetchContent)
 
   set(UNIFIED_RUNTIME_REPO "https://github.com/oneapi-src/unified-runtime.git")
-  # commit 9b936b5748a8da86123f3f148d1cfa84e52531c4
-  # Merge: 91c6068b 767cfd1a
-  # Author: Piotr Balcer <piotr.balcer@intel.com>
-  # Date:   Mon Feb 26 10:46:55 2024 +0100
-  #     Merge pull request #1354 from AllanZyne/sanitizer-device-global
-  #     [DeviceSanitizer] Support detecting out-of-bounds error on DeviceGlobals
-  set(UNIFIED_RUNTIME_TAG 9b936b5748a8da86123f3f148d1cfa84e52531c4)
+
+  # commit a2757b2931daa2f8d7c9dd51b0fc846be1fd49a7
+  # Merge:  9b936b5 + f78d369
+  # Author: Kenneth Benzie (Benie) <k.benzie@codeplay.com>
+  # Date:   Tue Feb 27 11:34:58 2024 +0000
+  #     Merge pull request #1254 from Bensuo/cmdbuf-support-hip
+  #     [EXP][CMDBUF] HIP adapter support for command buffers
+  set(UNIFIED_RUNTIME_TAG a2757b2931daa2f8d7c9dd51b0fc846be1fd49a7 )
 
   if(SYCL_PI_UR_OVERRIDE_FETCH_CONTENT_REPO)
     set(UNIFIED_RUNTIME_REPO "${SYCL_PI_UR_OVERRIDE_FETCH_CONTENT_REPO}")

diff --git a/sycl/test-e2e/Graph/Explicit/enqueue_ordering.cpp b/sycl/test-e2e/Graph/Explicit/enqueue_ordering.cpp
@@ -1,3 +1,4 @@
+// REQUIRES: aspect-usm_shared_allocations
 // RUN: %{build} -o %t.out
 // RUN: %{run} %t.out
 // Extra run to check for leaks in Level Zero using UR_L0_LEAKS_DEBUG

diff --git a/sycl/test-e2e/Graph/Explicit/work_group_size_prop.cpp b/sycl/test-e2e/Graph/Explicit/work_group_size_prop.cpp
@@ -11,10 +11,9 @@
 // should. So the Sycl graph support cannot correctly catch the error and throw
 // the approriate exception for negative test. An issue has been reported
 // https://github.com/bashbaug/SimpleOpenCLSamples/issues/95
-// XFAIL: cuda
+// XFAIL: cuda, hip
 // UNSUPPORTED: opencl
 // Note: failing negative test with HIP in the original test
-// TODO: disable hip when HIP backend will be supported by Graph
 
 #define GRAPH_E2E_EXPLICIT
 

diff --git a/sycl/test-e2e/Graph/RecordReplay/work_group_size_prop.cpp b/sycl/test-e2e/Graph/RecordReplay/work_group_size_prop.cpp
@@ -11,10 +11,9 @@
 // should. So the Sycl graph support cannot correctly catch the error and throw
 // the approriate exception for negative test. An issue has been reported
 // https://github.com/bashbaug/SimpleOpenCLSamples/issues/95
-// XFAIL: cuda
+// XFAIL: cuda, hip
 // UNSUPPORTED: opencl
 // Note: failing negative test with HIP in the original test
-// TODO: disable hip when HIP backend will be supported by Graph
 
 #define GRAPH_E2E_RECORD_REPLAY
 

diff --git a/sycl/test-e2e/Graph/device_query.cpp b/sycl/test-e2e/Graph/device_query.cpp
@@ -20,7 +20,8 @@ int main() {
   auto Backend = Device.get_backend();
 
   if ((Backend == backend::ext_oneapi_level_zero) ||
-      (Backend == backend::ext_oneapi_cuda)) {
+      (Backend == backend::ext_oneapi_cuda) ||
+      (Backend == backend::ext_oneapi_hip)) {
     assert(SupportsGraphs == exp_ext::graph_support_level::native);
   } else if (Backend == backend::opencl) {
     // OpenCL backend support is conditional on the cl_khr_command_buffer