Isolate the device counter to fix the device_reset issue

.gitlab-ci.yml

@@ -315,17 +315,17 @@ build/cuda91/clang/all/release/shared:
 
 # cuda 9.2 and friends
 build/cuda92/gcc/all/release/shared:
-  <<: *default_build
+  <<: *default_build_with_test
   extends:
-    - .full_test_condition
+    - .quick_test_condition
     - .use_gko-cuda92-gnu7-llvm50-intel2017
   variables:
     <<: *default_variables
     BUILD_OMP: "ON"
     BUILD_CUDA: "ON"
     BUILD_HIP: "ON"
     BUILD_TYPE: "Release"
-    CUDA_ARCH: 35
+    CUDA_ARCH: 61
 
 # cuda 10.0 and friends
 # Make sure that our jobs run when using self-installed

.gitlab/image.yml

@@ -30,8 +30,7 @@
   image: ginkgohub/cuda:92-gnu7-llvm50-intel2017
   tags:
     - private_ci
-    - controller
-    - cpu
+    - nvidia-gpu
 
 .use_gko-cuda100-gnu7-llvm60-intel2018:
   image: ginkgohub/cuda:100-gnu7-llvm60-intel2018

cuda/CMakeLists.txt

@@ -65,7 +65,7 @@ find_library(CURAND curand
     HINT ${CMAKE_CUDA_IMPLICIT_LINK_DIRECTORIES})
 
 add_library(ginkgo_cuda $<TARGET_OBJECTS:ginkgo_cuda_device> "")
-set(GKO_CUDA_COMMON_SOURCES    
+set(GKO_CUDA_COMMON_SOURCES
     ../common/components/precision_conversion.cpp
     ../common/matrix/dense_kernels.cpp
     ../common/solver/bicg_kernels.cpp

cuda/base/executor.cpp

@@ -40,6 +40,7 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 
 #include <ginkgo/config.hpp>
+#include <ginkgo/core/base/device.hpp>
 #include <ginkgo/core/base/exception_helpers.hpp>
 
 
@@ -64,8 +65,12 @@ std::shared_ptr<CudaExecutor> CudaExecutor::create(
                          alloc_mode),
         [device_id](CudaExecutor *exec) {
             auto device_reset = exec->get_device_reset();
+            std::lock_guard<std::mutex> guard(
+                device_class::get_mutex(device_id));
             delete exec;
-            if (!CudaExecutor::get_num_execs(device_id) && device_reset) {
+            auto &num_execs = device_class::get_num_execs(device_id);
+            num_execs--;
+            if (!num_execs && device_reset) {
                 cuda::device_guard g(device_id);
                 cudaDeviceReset();
             }

cuda/test/base/CMakeLists.txt

@@ -1,5 +1,6 @@
 ginkgo_create_cuda_test(array)
 ginkgo_create_cuda_test(cuda_executor)
+ginkgo_create_thread_test(cuda_executor_reset)
 if(GINKGO_HAVE_HWLOC)
     find_package(NUMA REQUIRED)
     ginkgo_create_cuda_test(cuda_executor_topology NUMA::NUMA)

cuda/test/base/cuda_executor_reset.cpp

@@ -0,0 +1,87 @@
+/*******************************<GINKGO LICENSE>******************************
+Copyright (c) 2017-2021, the Ginkgo authors
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions
+are met:
+
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in the
+documentation and/or other materials provided with the distribution.
+
+3. Neither the name of the copyright holder nor the names of its
+contributors may be used to endorse or promote products derived from
+this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS
+IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
+TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
+PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+******************************<GINKGO LICENSE>*******************************/
+
+#include <ginkgo/core/base/executor.hpp>
+
+
+#include <thread>
+
+
+#include <gtest/gtest.h>
+
+
+namespace {
+
+
+#define GTEST_ASSERT_NO_EXIT(statement) \
+    ASSERT_EXIT({ {statement} exit(0); }, ::testing::ExitedWithCode(0), "")
+
+
+TEST(DeviceReset, HipCuda)
+{
+    GTEST_ASSERT_NO_EXIT({
+        auto ref = gko::ReferenceExecutor::create();
+        auto hip = gko::HipExecutor::create(0, ref, true);
+        auto cuda = gko::CudaExecutor::create(0, ref, true);
+    });
+}
+
+
+TEST(DeviceReset, CudaHip)
+{
+    GTEST_ASSERT_NO_EXIT({
+        auto ref = gko::ReferenceExecutor::create();
+        auto cuda = gko::CudaExecutor::create(0, ref, true);
+        auto hip = gko::HipExecutor::create(0, ref, true);
+    });
+}
+
+
+void func()
+{
+    auto ref = gko::ReferenceExecutor::create();
+    auto exec = gko::CudaExecutor::create(0, ref, true);
+}
+
+
+TEST(DeviceReset, CudaCuda)
+{
+    GTEST_ASSERT_NO_EXIT({
+        std::thread t1(func);
+        std::thread t2(func);
+        t1.join();
+        t2.join();
+    });
+}
+
+
+}  // namespace

devices/CMakeLists.txt

@@ -16,7 +16,7 @@ function(ginkgo_add_library name)
     set_target_properties(${name} PROPERTIES POSITION_INDEPENDENT_CODE ON)
 endfunction()
 
-ginkgo_add_library(ginkgo_device machine_topology.cpp)
+ginkgo_add_library(ginkgo_device machine_topology.cpp device.cpp)
 ginkgo_install_library(ginkgo_device)
 
 add_subdirectory(cuda)

devices/cuda/executor.cpp

@@ -64,10 +64,4 @@ bool CudaExecutor::verify_memory_to(const HipExecutor *dest_exec) const
 }
 
 
-unsigned CudaExecutor::num_execs[max_devices];
-
-
-std::mutex CudaExecutor::mutex[max_devices];
-
-
 }  // namespace gko

devices/device.cpp

@@ -0,0 +1,71 @@
+/*******************************<GINKGO LICENSE>******************************
+Copyright (c) 2017-2021, the Ginkgo authors
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions
+are met:
+
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in the
+documentation and/or other materials provided with the distribution.
+
+3. Neither the name of the copyright holder nor the names of its
+contributors may be used to endorse or promote products derived from
+this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS
+IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
+TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
+PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+******************************<GINKGO LICENSE>*******************************/
+
+#include <memory>
+#include <mutex>
+
+
+#include <ginkgo/core/base/device.hpp>
+
+
+namespace gko {
+
+
+std::mutex &NvidiaDevice::get_mutex(int i)
+{
+    static std::mutex mutex[max_devices];
+    return mutex[i];
+}
+
+
+int &NvidiaDevice::get_num_execs(int i)
+{
+    static int num_execs[max_devices];
+    return num_execs[i];
+}
+
+
+std::mutex &AmdDevice::get_mutex(int i)
+{
+    static std::mutex mutex[max_devices];
+    return mutex[i];
+}
+
+
+int &AmdDevice::get_num_execs(int i)
+{
+    static int num_execs[max_devices];
+    return num_execs[i];
+}
+
+
+}  // namespace gko

devices/hip/executor.cpp

@@ -61,10 +61,4 @@ bool HipExecutor::verify_memory_to(const CudaExecutor *dest_exec) const
 }
 
 
-int HipExecutor::num_execs[max_devices];
-
-
-std::mutex HipExecutor::mutex[max_devices];
-
-
 }  // namespace gko

hip/base/executor.hip.cpp

@@ -40,6 +40,7 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 
 #include <ginkgo/config.hpp>
+#include <ginkgo/core/base/device.hpp>
 #include <ginkgo/core/base/exception_helpers.hpp>
 
 
@@ -63,8 +64,12 @@ std::shared_ptr<HipExecutor> HipExecutor::create(
         new HipExecutor(device_id, std::move(master), device_reset, alloc_mode),
         [device_id](HipExecutor *exec) {
             auto device_reset = exec->get_device_reset();
+            std::lock_guard<std::mutex> guard(
+                device_class::get_mutex(device_id));
             delete exec;
-            if (!HipExecutor::get_num_execs(device_id) && device_reset) {
+            auto &num_execs = device_class::get_num_execs(device_id);
+            num_execs--;
+            if (!num_execs && device_reset) {
                 hip::device_guard g(device_id);
                 hipDeviceReset();
             }

hip/test/base/CMakeLists.txt

@@ -1,4 +1,5 @@
 ginkgo_create_hip_test(hip_executor)
+ginkgo_create_thread_test(hip_executor_reset)
 if(GINKGO_HAVE_HWLOC)
     find_package(NUMA REQUIRED)
     ginkgo_create_hip_test(hip_executor_topology NUMA::NUMA)

hip/test/base/hip_executor_reset.cpp

@@ -0,0 +1,87 @@
+/*******************************<GINKGO LICENSE>******************************
+Copyright (c) 2017-2021, the Ginkgo authors
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions
+are met:
+
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in the
+documentation and/or other materials provided with the distribution.
+
+3. Neither the name of the copyright holder nor the names of its
+contributors may be used to endorse or promote products derived from
+this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS
+IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
+TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
+PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+******************************<GINKGO LICENSE>*******************************/
+
+#include <ginkgo/core/base/executor.hpp>
+
+
+#include <thread>
+
+
+#include <gtest/gtest.h>
+
+
+namespace {
+
+
+#define GTEST_ASSERT_NO_EXIT(statement) \
+    ASSERT_EXIT({ {statement} exit(0); }, ::testing::ExitedWithCode(0), "")
+
+
+TEST(DeviceReset, HipCuda)
+{
+    GTEST_ASSERT_NO_EXIT({
+        auto ref = gko::ReferenceExecutor::create();
+        auto hip = gko::HipExecutor::create(0, ref, true);
+        auto cuda = gko::CudaExecutor::create(0, ref, true);
+    });
+}
+
+
+TEST(DeviceReset, CudaHip)
+{
+    GTEST_ASSERT_NO_EXIT({
+        auto ref = gko::ReferenceExecutor::create();
+        auto cuda = gko::CudaExecutor::create(0, ref, true);
+        auto hip = gko::HipExecutor::create(0, ref, true);
+    });
+}
+
+
+void func()
+{
+    auto ref = gko::ReferenceExecutor::create();
+    auto exec = gko::HipExecutor::create(0, ref, true);
+}
+
+
+TEST(DeviceReset, HipHip)
+{
+    GTEST_ASSERT_NO_EXIT({
+        std::thread t1(func);
+        std::thread t2(func);
+        t1.join();
+        t2.join();
+    });
+}
+
+
+}  // namespace

include/ginkgo/config.hpp.in

@@ -67,6 +67,14 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #cmakedefine GINKGO_MIXED_PRECISION
 
 
+/* Should we compile cuda kenrels for Ginkgo? */
+#cmakedefine GINKGO_BUILD_CUDA
+
+
+/* Should we compile hip kenrels for Ginkgo? */
+#cmakedefine GINKGO_BUILD_HIP
+
+
 /* What is HIP compiled for, hcc or nvcc? */
 // clang-format off
 #define GINKGO_HIP_PLATFORM_HCC @GINKGO_HIP_PLATFORM_HCC@