Merge branch 'master' into louisfeng/NGMX-296-conv_bias

src/ngraph/runtime/cpu/mkldnn_emitter.cpp

@@ -25,6 +25,12 @@
 
 using namespace ngraph::runtime::cpu;
 
+MKLDNNEmitter::~MKLDNNEmitter()
+{
+    for (auto p : m_mkldnn_primitives)
+        delete p;
+}
+
 const std::vector<mkldnn::primitive*>& MKLDNNEmitter::get_mkldnn_primitives() const
 {
     return m_mkldnn_primitives;

src/ngraph/runtime/cpu/mkldnn_emitter.hpp

@@ -39,6 +39,8 @@ namespace ngraph
             {
             public:
                 MKLDNNEmitter() {}
+                ~MKLDNNEmitter();
+
                 const std::vector<mkldnn::primitive*>& get_mkldnn_primitives() const;
 
                 size_t insert_primitive(mkldnn::primitive* primitive);

src/ngraph/runtime/gpu/gpu_emitter.hpp

@@ -24,12 +24,12 @@
 #include "ngraph/runtime/gpu/gpu_external_function.hpp"
 #include "ngraph/runtime/gpu/gpu_tensor_view_wrapper.hpp"
 
-#define EMITTER_DECL(E)                                                                            \
-    E(codegen::CodeWriter& writer,                                                                 \
-      const ngraph::Node* n,                                                                       \
-      const std::vector<ngraph::runtime::gpu::GPU_TensorViewWrapper>& args,                        \
-      const std::vector<ngraph::runtime::gpu::GPU_TensorViewWrapper>& out)
-
+#define EMITTER_DECL(op_name)                                                                      \
+    emit<op_name>(GPU_ExternalFunction * external_function,                                        \
+                  codegen::CodeWriter & writer,                                                    \
+                  const ngraph::Node* node,                                                        \
+                  const std::vector<GPU_TensorViewWrapper>& args,                                  \
+                  const std::vector<GPU_TensorViewWrapper>& out)
 namespace ngraph
 {
     namespace runtime
@@ -39,45 +39,30 @@ namespace ngraph
             class GPU_Emitter
             {
             public:
-                static void EMITTER_DECL(EmitNop);
-                static void EMITTER_DECL(EmitAdd);
-                static void EMITTER_DECL(EmitDot);
-                static void EMITTER_DECL(EmitMultiply);
-                static void EMITTER_DECL(EmitGetOutputElement);
-                static void EMITTER_DECL(EmitXLAGetTupleElement);
-                static void EMITTER_DECL(EmitUnaryElementwise);
-                static void EMITTER_DECL(EmitTuple);
-                static void EMITTER_DECL(EmitConcat);
-                static void EMITTER_DECL(EmitDivide);
-                static void EMITTER_DECL(EmitEqual);
-                static void EMITTER_DECL(EmitGreater);
-                static void EMITTER_DECL(EmitGreaterEq);
-                static void EMITTER_DECL(EmitLess);
-                static void EMITTER_DECL(EmitLessEq);
-                static void EMITTER_DECL(EmitMaximum);
-                static void EMITTER_DECL(EmitMinimum);
-                static void EMITTER_DECL(EmitNegative);
-                static void EMITTER_DECL(EmitNotEqual);
-                static void EMITTER_DECL(EmitSelect);
-                static void EMITTER_DECL(EmitSubtract);
-                static void EMITTER_DECL(EmitBroadcast);
-                static void EMITTER_DECL(EmitConvert);
-                static void EMITTER_DECL(EmitConstant);
-                static void EMITTER_DECL(EmitReshape);
-                static void EMITTER_DECL(EmitFunctionCall);
-                static void EMITTER_DECL(EmitReduce);
-                static void EMITTER_DECL(EmitSlice);
-                static void EMITTER_DECL(EmitSum);
-                static void EMITTER_DECL(EmitPower);
-                static void EMITTER_DECL(EmitReplaceSlice);
-                static void EMITTER_DECL(EmitOneHot);
-                static void EMITTER_DECL(EmitSqrt);
-                static void EMITTER_DECL(EmitConvolution);
-                static void EMITTER_DECL(EmitMaxPool);
-                static void EMITTER_DECL(EmitReverse);
-                static void EMITTER_DECL(EmitReduceWindow);
-                static void EMITTER_DECL(EmitSelectAndScatter);
-                static void EMITTER_DECL(EmitResult);
+                template <typename OP>
+                static void emit(GPU_ExternalFunction* external_function,
+                                 codegen::CodeWriter& writer,
+                                 const ngraph::Node* node,
+                                 const std::vector<GPU_TensorViewWrapper>& args,
+                                 const std::vector<GPU_TensorViewWrapper>& out)
+                {
+                    throw std::runtime_error("Unimplemented op in GPU emitter for " +
+                                             node->get_name());
+                }
+
+                static void nop(GPU_ExternalFunction* external_function,
+                                codegen::CodeWriter& writer,
+                                const ngraph::Node* node,
+                                const std::vector<GPU_TensorViewWrapper>& args,
+                                const std::vector<GPU_TensorViewWrapper>& out)
+                {
+                }
+
+                static void EmitUnaryElementwise(GPU_ExternalFunction* external_function,
+                                                 codegen::CodeWriter& writer,
+                                                 const ngraph::Node* node,
+                                                 const std::vector<GPU_TensorViewWrapper>& args,
+                                                 const std::vector<GPU_TensorViewWrapper>& out);
             };
         }
     }

src/ngraph/runtime/gpu/gpu_external_function.hpp

@@ -41,7 +41,8 @@ namespace ngraph
             class GPU_CallFrame;
 
             using OpFunction =
-                std::function<void(codegen::CodeWriter&,
+                std::function<void(GPU_ExternalFunction* external_function,
+                                   codegen::CodeWriter&,
                                    const ngraph::Node*,
                                    const std::vector<GPU_TensorViewWrapper>& inputs,
                                    const std::vector<GPU_TensorViewWrapper>& outputs)>;