Added op::Relu and op::Not to GPU transformer and enabled corresponding tests (#641)

* Added op::Relu and op::Not and enabled corresponding tests.

* Removed softmax for now.

src/ngraph/runtime/gpu/gpu_cuda_kernel_ops.hpp

@@ -45,6 +45,8 @@ namespace ngraph
         class GreaterEq;
         class Less;
         class LessEq;
+        class Not;
+        class Relu;
         class Max;
         class Min;
         class Negative;
@@ -250,6 +252,20 @@ namespace ngraph
                 static constexpr const char* op = "less_equal";
                 static constexpr const char* math_kernel = "x0 <= x1";
             };
+
+            template <>
+            struct CudaOpMap<ngraph::op::Relu>
+            {
+                static constexpr const char* op = "relu";
+                static constexpr const char* math_kernel = "fmaxf(0,x0)";
+            };
+
+            template <>
+            struct CudaOpMap<ngraph::op::Not>
+            {
+                static constexpr const char* op = "not";
+                static constexpr const char* math_kernel = "!x0";
+            };
         }
     }
 }

src/ngraph/runtime/gpu/gpu_external_function.cpp

@@ -231,7 +231,7 @@ static const runtime::gpu::OpMap dispatcher{
     {TI(ngraph::op::Product), &runtime::gpu::GPU_Emitter::emit<ngraph::op::Product>},
     {TI(ngraph::op::Max), &runtime::gpu::GPU_Emitter::emit<ngraph::op::Max>},
     {TI(ngraph::op::Min), &runtime::gpu::GPU_Emitter::emit<ngraph::op::Min>},
-    {TI(ngraph::op::Relu), &runtime::gpu::GPU_Emitter::emit<ngraph::op::Relu>},
+    {TI(ngraph::op::Relu), &runtime::gpu::GPU_Emitter::EmitElementwise},
     {TI(ngraph::op::ReluBackprop), &runtime::gpu::GPU_Emitter::emit<ngraph::op::ReluBackprop>},
     {TI(ngraph::op::Softmax), &runtime::gpu::GPU_Emitter::emit<ngraph::op::Softmax>},
 };

test/backend_test.in.cpp

@@ -4868,7 +4868,6 @@ TEST(${BACKEND_NAME}, max_pool_2d_1channel_1image_strided)
 
 TEST(${BACKEND_NAME}, not)
 {
-    SKIP_TEST_FOR("GPU", "${BACKEND_NAME}");
     Shape shape{2, 2};
     auto A = make_shared<op::Parameter>(element::boolean, shape);
     auto f = make_shared<Function>(make_shared<op::Not>(A), op::ParameterVector{A});
@@ -8285,7 +8284,6 @@ TEST(${BACKEND_NAME}, min_3d_eliminate_zero_dim)
 
 TEST(${BACKEND_NAME}, relu_2Dfprop)
 {
-    SKIP_TEST_FOR("GPU", "${BACKEND_NAME}");
     auto shape_a = Shape{2, 5};
     auto A = make_shared<op::Parameter>(element::f32, shape_a);
     auto relu = make_shared<op::Relu>(A);
@@ -8308,7 +8306,6 @@ TEST(${BACKEND_NAME}, relu_2Dfprop)
 
 TEST(${BACKEND_NAME}, relu_4Dfprop)
 {
-    SKIP_TEST_FOR("GPU", "${BACKEND_NAME}");
     auto shape_a = Shape{2, 2, 2, 2};
     auto A = make_shared<op::Parameter>(element::f32, shape_a);
     auto relu = make_shared<op::Relu>(A);
@@ -8331,7 +8328,6 @@ TEST(${BACKEND_NAME}, relu_4Dfprop)
 
 TEST(${BACKEND_NAME}, fuse_max_with_constant_zero_input_as_relu)
 {
-    SKIP_TEST_FOR("GPU", "${BACKEND_NAME}");
     auto shape_a = Shape{2, 5};
     auto A = op::Constant::create(element::f32, shape_a, {0, 0, 0, 0, 0, 0, 0, 0, 0, 0});
     auto B = make_shared<op::Parameter>(element::f32, shape_a);