[rocm-libraries] ROCm/rocm-libraries#1726 (commit a589ab3)

[rocSPARSE] sync before hipFree. (#1705)

Propagate #1705 to 7.0.

Author: kliegeois

.jenkins/common.groovy

@@ -123,14 +123,6 @@ def runCoverageCommand (platform, project, gfilter, String dirmode = "release")
                 """
 
         platform.runCommand(this, command)
-
-        publishHTML([allowMissing: false,
-                    alwaysLinkToLastBuild: false,
-                    keepAll: false,
-                    reportDir: "${project.paths.project_build_prefix}/build/${dirmode}/coverage-report",
-                    reportFiles: "index.html",
-                    reportName: "Code coverage report",
-                    reportTitles: "Code coverage report"])
     }
 
 }

CHANGELOG.md

@@ -3,6 +3,12 @@
 Documentation for rocSPARSE is available at
 [https://rocm.docs.amd.com/projects/rocSPARSE/en/latest/](https://rocm.docs.amd.com/projects/rocSPARSE/en/latest/).
 
+## rocSPARSE 4.0.3 for ROCm 7.0.2
+
+### Resolved issues
+
+* Resolved an issue causing premature deallocation of internal buffers still in use.
+
 ## rocSPARSE 4.0.2 for ROCm 7.0.0
 
 ### Added

CMakeLists.txt

@@ -223,7 +223,7 @@ if( CMAKE_CXX_COMPILER_ID MATCHES "Clang" )
 endif( )
 
 # Setup version
-set(VERSION_STRING "4.0.2")
+set(VERSION_STRING "4.0.3")
 set(SOVERSION_STRING "1.0")
 
 rocm_setup_version(VERSION ${VERSION_STRING})

library/src/auxiliary/rocsparse_auxiliary.cpp

@@ -1298,6 +1298,14 @@ try
     ROCSPARSE_ROUTINE_TRACE;
 
     ROCSPARSE_CHECKARG_POINTER(0, hyb);
+
+    // Due to the changes in the hipFree introduced in HIP 7.0
+    // https://rocm.docs.amd.com/projects/HIP/en/latest/hip-7-changes.html#update-hipfree
+    // we need to introduce a device synchronize here as the below hipFree calls are now asynchronous.
+    // hipFree() previously had an implicit wait for synchronization purpose which is applicable for all memory allocations.
+    // This wait has been disabled in the HIP 7.0 runtime for allocations made with hipMallocAsync and hipMallocFromPoolAsync.
+    RETURN_IF_HIP_ERROR(hipDeviceSynchronize());
+
     // Clean up ELL part
     if(hyb->ell_col_ind != nullptr)
     {
@@ -4977,6 +4985,13 @@ try
 
     ROCSPARSE_CHECKARG_POINTER(0, descr);
 
+    // Due to the changes in the hipFree introduced in HIP 7.0
+    // https://rocm.docs.amd.com/projects/HIP/en/latest/hip-7-changes.html#update-hipfree
+    // we need to introduce a device synchronize here as the below hipFree calls are now asynchronous.
+    // hipFree() previously had an implicit wait for synchronization purpose which is applicable for all memory allocations.
+    // This wait has been disabled in the HIP 7.0 runtime for allocations made with hipMallocAsync and hipMallocFromPoolAsync.
+    RETURN_IF_HIP_ERROR(hipDeviceSynchronize());
+
     // Clean up row pointer array
     if(descr->csr_row_ptr_C != nullptr)
     {

library/src/common/rocsparse_handle.cpp

@@ -152,6 +152,13 @@ _rocsparse_handle::~_rocsparse_handle()
 {
     ROCSPARSE_ROUTINE_TRACE;
 
+    // Due to the changes in the hipFree introduced in HIP 7.0
+    // https://rocm.docs.amd.com/projects/HIP/en/latest/hip-7-changes.html#update-hipfree
+    // we need to introduce a device synchronize here as the below hipFree calls are now asynchronous.
+    // hipFree() previously had an implicit wait for synchronization purpose which is applicable for all memory allocations.
+    // This wait has been disabled in the HIP 7.0 runtime for allocations made with hipMallocAsync and hipMallocFromPoolAsync.
+    PRINT_IF_HIP_ERROR(hipDeviceSynchronize());
+
     PRINT_IF_HIP_ERROR(rocsparse_hipFree(buffer));
     PRINT_IF_HIP_ERROR(rocsparse_hipFree(sone));
     PRINT_IF_HIP_ERROR(rocsparse_hipFree(done));

library/src/common/rocsparse_mat_info.cpp

@@ -187,6 +187,13 @@ _rocsparse_mat_info::~_rocsparse_mat_info()
     // Clear csritsv info struct
     WARNING_IF_ROCSPARSE_ERROR(rocsparse::destroy_csritsv_info(this->csritsv_info));
 
+    // Due to the changes in the hipFree introduced in HIP 7.0
+    // https://rocm.docs.amd.com/projects/HIP/en/latest/hip-7-changes.html#update-hipfree
+    // we need to introduce a device synchronize here as the below hipFree calls are now asynchronous.
+    // hipFree() previously had an implicit wait for synchronization purpose which is applicable for all memory allocations.
+    // This wait has been disabled in the HIP 7.0 runtime for allocations made with hipMallocAsync and hipMallocFromPoolAsync.
+    WARNING_IF_HIP_ERROR(hipDeviceSynchronize());
+
     // Clear zero pivot
     WARNING_IF_HIP_ERROR(rocsparse_hipFree(this->zero_pivot));
 

library/src/common/rocsparse_trm_info.cpp

@@ -27,6 +27,13 @@
 
 rocsparse::trm_info_t::~trm_info_t()
 {
+    // Due to the changes in the hipFree introduced in HIP 7.0
+    // https://rocm.docs.amd.com/projects/HIP/en/latest/hip-7-changes.html#update-hipfree
+    // we need to introduce a device synchronize here as the below hipFree calls are now asynchronous.
+    // hipFree() previously had an implicit wait for synchronization purpose which is applicable for all memory allocations.
+    // This wait has been disabled in the HIP 7.0 runtime for allocations made with hipMallocAsync and hipMallocFromPoolAsync.
+    WARNING_IF_HIP_ERROR(hipDeviceSynchronize());
+
     WARNING_IF_HIP_ERROR(rocsparse_hipFree(this->row_map));
     this->row_map = nullptr;
 

library/src/conversion/rocsparse_csxsldu_preprocess.cpp

@@ -57,7 +57,7 @@ namespace rocsparse
 
         if(temp_alloc)
         {
-            RETURN_IF_HIP_ERROR(rocsparse_hipFree(temp_storage_ptr));
+            RETURN_IF_HIP_ERROR(rocsparse_hipFreeAsync(temp_storage_ptr, handle->stream));
         }
 
         return rocsparse_status_success;

library/src/conversion/rocsparse_extract.hpp

@@ -38,6 +38,13 @@ struct _rocsparse_extract_descr
     int64_t* m_device_nnz{};
     virtual ~_rocsparse_extract_descr()
     {
+        // Due to the changes in the hipFree introduced in HIP 7.0
+        // https://rocm.docs.amd.com/projects/HIP/en/latest/hip-7-changes.html#update-hipfree
+        // we need to introduce a device synchronize here as the below hipFree calls are now asynchronous.
+        // hipFree() previously had an implicit wait for synchronization purpose which is applicable for all memory allocations.
+        // This wait has been disabled in the HIP 7.0 runtime for allocations made with hipMallocAsync and hipMallocFromPoolAsync.
+        WARNING_IF_HIP_ERROR(hipDeviceSynchronize());
+
         std::ignore = rocsparse_hipFree(this->m_device_nnz);
     }
 

library/src/conversion/rocsparse_sparse_to_sparse.cpp

@@ -175,6 +175,14 @@ try
                                                         &ind_type,
                                                         &base,
                                                         &val_type));
+
+            // Due to the changes in the hipFree introduced in HIP 7.0
+            // https://rocm.docs.amd.com/projects/HIP/en/latest/hip-7-changes.html#update-hipfree
+            // we need to introduce a device synchronize here as the below hipFree calls are now asynchronous.
+            // hipFree() previously had an implicit wait for synchronization purpose which is applicable for all memory allocations.
+            // This wait has been disabled in the HIP 7.0 runtime for allocations made with hipMallocAsync and hipMallocFromPoolAsync.
+            RETURN_IF_HIP_ERROR(hipDeviceSynchronize());
+
             RETURN_IF_HIP_ERROR(rocsparse_hipFree(row));
             RETURN_IF_HIP_ERROR(rocsparse_hipFree(col));
             RETURN_IF_HIP_ERROR(rocsparse_hipFree(val));