Merge pull request #1594 from bratpiorka/rrudnick_ze_relaxed_alloc

Always use relaxed allocation limits for L0 host mem

Author: bratpiorka

.github/workflows/reusable_compatibility.yml

@@ -385,16 +385,22 @@ jobs:
         UMF_LOG: level:warning;flush:debug;output:stderr;pid:no
         LD_LIBRARY_PATH: ${{github.workspace}}/latest_version/build/lib/
       run: |
-        ctest --verbose -E "test_memoryProvider|test_disjoint_pool"
+        ctest --verbose -E "test_memoryProvider|test_disjoint_pool|test_provider_level_zero|test_provider_level_zero_dlopen_global|test_provider_level_zero_dlopen_local"
 
     - name: Run disabled tests individually with latest UMF libs (warnings enabled)
       working-directory: ${{github.workspace}}/tag_version/build
       env:
         UMF_LOG: level:warning;flush:debug;output:stderr;pid:no
         LD_LIBRARY_PATH: ${{github.workspace}}/latest_version/build/lib/
+        MATRIX_PROVIDER: ${{matrix.provider}}
       run: |
         test/test_memoryProvider --gtest_filter="-*Trace"
         test/test_disjoint_pool --gtest_filter="-test.internals"
+        if [[ "$MATRIX_PROVIDER" == "LEVEL_ZERO" ]]; then
+          test/test_provider_level_zero --gtest_filter="-*allocInvalidSize/2"
+          test/test_provider_level_zero_dlopen_global --gtest_filter="-*allocInvalidSize/2"
+          test/test_provider_level_zero_dlopen_local --gtest_filter="-*allocInvalidSize/2"
+        fi
 
     # Browse all folders in the examples directory, build them using the 
     # latest UMF version, and run them, excluding those in the exclude list.

src/provider/provider_level_zero.c

@@ -158,10 +158,10 @@ static void store_last_native_error(int32_t native_error) {
 struct ctl ze_memory_ctl_root;
 static UTIL_ONCE_FLAG ctl_initialized = UTIL_ONCE_FLAG_INIT;
 
-static ze_relaxed_allocation_limits_exp_desc_t relaxed_device_allocation_desc =
-    {.stype = ZE_STRUCTURE_TYPE_RELAXED_ALLOCATION_LIMITS_EXP_DESC,
-     .pNext = NULL,
-     .flags = ZE_RELAXED_ALLOCATION_LIMITS_EXP_FLAG_MAX_SIZE};
+static ze_relaxed_allocation_limits_exp_desc_t relaxed_allocation_desc = {
+    .stype = ZE_STRUCTURE_TYPE_RELAXED_ALLOCATION_LIMITS_EXP_DESC,
+    .pNext = NULL,
+    .flags = ZE_RELAXED_ALLOCATION_LIMITS_EXP_FLAG_MAX_SIZE};
 
 static ze_external_memory_export_desc_t memory_export_desc = {
     .stype = ZE_STRUCTURE_TYPE_EXTERNAL_MEMORY_EXPORT_DESC,
@@ -691,6 +691,13 @@ static umf_result_t ze_memory_provider_alloc_helper(void *provider, size_t size,
             .stype = ZE_STRUCTURE_TYPE_HOST_MEM_ALLOC_DESC,
             .pNext = NULL,
             .flags = 0};
+        void *lastNext = &host_desc.pNext;
+
+        // add relaxed allocation desc to the pNext chain
+        ze_relaxed_allocation_limits_exp_desc_t relaxed_allocation_desc_copy =
+            relaxed_allocation_desc;
+        *(void **)lastNext = &relaxed_allocation_desc_copy;
+
         ze_result = g_ze_ops.zeMemAllocHost(ze_provider->context, &host_desc,
                                             size, alignment, resultPtr);
         break;
@@ -703,13 +710,12 @@ static umf_result_t ze_memory_provider_alloc_helper(void *provider, size_t size,
             .ordinal = ze_provider->device_ordinal};
         void *lastNext = &dev_desc.pNext;
 
-        ze_relaxed_allocation_limits_exp_desc_t
-            relaxed_device_allocation_desc_copy =
-                relaxed_device_allocation_desc;
+        ze_relaxed_allocation_limits_exp_desc_t relaxed_allocation_desc_copy =
+            relaxed_allocation_desc;
         if (use_relaxed_allocation(ze_provider, size)) {
             // add relaxed allocation desc to the pNext chain
-            *(void **)lastNext = &relaxed_device_allocation_desc_copy;
-            lastNext = &relaxed_device_allocation_desc_copy.pNext;
+            *(void **)lastNext = &relaxed_allocation_desc_copy;
+            lastNext = &relaxed_allocation_desc_copy.pNext;
         }
 
         // check if the allocation should use import / export mechanism
@@ -734,7 +740,7 @@ static umf_result_t ze_memory_provider_alloc_helper(void *provider, size_t size,
         ze_device_mem_alloc_desc_t dev_desc = {
             .stype = ZE_STRUCTURE_TYPE_DEVICE_MEM_ALLOC_DESC,
             .pNext = use_relaxed_allocation(ze_provider, size)
-                         ? &relaxed_device_allocation_desc
+                         ? &relaxed_allocation_desc
                          : NULL,
             .flags = 0,
             .ordinal = ze_provider->device_ordinal};

test/providers/provider_level_zero.cpp

@@ -568,6 +568,13 @@ TEST(umfLevelZeroProviderOps, default_name_null_handle) {
 }
 
 TEST_P(umfLevelZeroProviderTest, allocInvalidSize) {
+    if (GetParam() == UMF_MEMORY_TYPE_HOST) {
+        // For host memory, even with relaxed limits, the allocation may
+        // succeed or fail depending on the system's available memory,
+        // so we skip this test for host memory type.
+        GTEST_SKIP() << "Skipping test for host memory type.";
+    }
+
     umf_memory_provider_handle_t provider = nullptr;
     umf_result_t umf_result = umfMemoryProviderCreate(
         umfLevelZeroMemoryProviderOps(), params, &provider);
@@ -577,6 +584,7 @@ TEST_P(umfLevelZeroProviderTest, allocInvalidSize) {
     void *ptr = nullptr;
     umf_result = umfMemoryProviderAlloc(
         provider, std::numeric_limits<size_t>::max(), 0, &ptr);
+
     ASSERT_EQ(umf_result, UMF_RESULT_ERROR_MEMORY_PROVIDER_SPECIFIC);
     const char *message;
     int32_t error;