[NVIDIA] Fix FP16 behavior in tests on GTX1080

openvinotoolkit · Jul 4, 2023 · 1199319 · 1199319
1 parent 968c44d
commit 1199319
Show file tree

Hide file tree

Showing 3 changed files with 19 additions and 5 deletions.
diff --git a/modules/nvidia_plugin/src/transformer/cuda_graph_transformer.cpp b/modules/nvidia_plugin/src/transformer/cuda_graph_transformer.cpp
@@ -45,7 +45,7 @@ void GraphTransformer::transform(const CUDA::Device& device,
                                  const Configuration& config) const {
     auto inference_precision = config.get_inference_precision();
     if (inference_precision == ov::element::f16 && !isHalfSupported(device)) {
-        throw_ov_exception("Inference precision f16 is not supported by device!");
+        inference_precision = ov::element::f32;
     }
     auto upscale_precision = [&]() -> bool {
         return !isHalfSupported(device) || inference_precision == ov::element::f32;

diff --git a/modules/nvidia_plugin/tests/functional/skip_tests_config.cpp b/modules/nvidia_plugin/tests/functional/skip_tests_config.cpp
@@ -4,6 +4,7 @@
 
 #include "functional_test_utils/skip_tests_config.hpp"
 
+#include <cuda/runtime.hpp>
 #include <string>
 #include <vector>
 
@@ -138,5 +139,13 @@ std::vector<std::string> disabledTestPatterns() {
     retVector.emplace_back(R"(.*ReferenceTopKTest.*topk_max_sort_none)");
     retVector.emplace_back(R"(.*ReferenceTopKTest.*topk_min_sort_none)");
 #endif
+
+    if (!CUDA::isHalfSupported(CUDA::Device{})) {
+        retVector.emplace_back(
+            R"(.*OVExecGraphImportExportTest.*importExportedFunctionParameterResultOnly.*targetDevice=NVIDIA_elementType=f16.*)");
+        retVector.emplace_back(
+            R"(.*OVExecGraphImportExportTest.*importExportedIENetworkParameterResultOnly.*targetDevice=NVIDIA_elementType=f16.*)");
+    }
+
     return retVector;
-}
+}
diff --git a/modules/nvidia_plugin/tests/unit/transformations/cuda_transformations_test.cpp b/modules/nvidia_plugin/tests/unit/transformations/cuda_transformations_test.cpp
@@ -4,11 +4,12 @@
 
 #include <gtest/gtest.h>
 
+#include <cuda/runtime.hpp>
 #include <memory>
-#include "openvino/opsets/opset10.hpp"
-#include "transformer/cuda_graph_transformer.hpp"
 
 #include "common_test_utils/ngraph_test_utils.hpp"
+#include "openvino/opsets/opset10.hpp"
+#include "transformer/cuda_graph_transformer.hpp"
 
 using namespace testing;
 
@@ -23,12 +24,16 @@ TEST(TransformationTests, cuda_transformations_f16) {
         model = std::make_shared<ov::Model>(ov::NodeVector{divide}, ov::ParameterVector{data});
 
         // Run transformation
-        const CUDA::Device device{0};
+        const CUDA::Device device{};
         const auto config = ov::nvidia_gpu::Configuration(ov::AnyMap{ov::hint::inference_precision(ov::element::f16)});
         ov::nvidia_gpu::GraphTransformer().transform(device, model, config);
 
         // Check that after applying transformation all runtime info attributes was correctly propagated
         ASSERT_NO_THROW(check_rt_info(model));
+
+        if (!CUDA::isHalfSupported(device)) {
+            GTEST_SKIP() << "f16 precision isn't fully supported on the device";
+        }
     }
 
     {