softmax & convolution memory primitive cacheing (#1290)

* Updated softmax.

* Formatting.

* Updated convolution.

* Use build_primitive overloading. Add helper to emit type_string given a node.

* Formatting.

* Update ConvolutionBackpropData.

* convolution backprop & max pool memory primitive cacheing (#1303)

* Updated ConvolutionBackpropFilters.
* Update MaxPool.

* Update Max and Min. (#1307)

src/ngraph/runtime/gpu/cuda_emitter.hpp

@@ -22,6 +22,10 @@
 #include "ngraph/runtime/gpu/gpu_shape.hpp"
 #include "ngraph/strides.hpp"
 
+#include "ngraph/op/convolution.hpp"
+#include "ngraph/op/max_pool.hpp"
+#include "ngraph/op/softmax.hpp"
+
 namespace ngraph
 {
     class GPUShape;
@@ -37,6 +41,11 @@ namespace ngraph
             {
                 friend class GPUPrimitiveEmitter;
 
+            public:
+                size_t build_primitive(const op::Softmax* node);
+                size_t build_primitive(const op::Convolution* node);
+                size_t build_primitive(const op::MaxPool* node);
+
             public:
                 size_t build_pad(const std::array<std::string, 2>& dtypes,
                                  GPUShape input_shape,

src/ngraph/runtime/gpu/cudnn_emitter.hpp

@@ -30,6 +30,11 @@
 #include "ngraph/runtime/gpu/gpu_runtime_context.hpp"
 #include "ngraph/shape.hpp"
 
+#include "ngraph/op/convolution.hpp"
+#include "ngraph/op/max.hpp"
+#include "ngraph/op/max_pool.hpp"
+#include "ngraph/op/min.hpp"
+
 namespace ngraph
 {
     namespace runtime
@@ -48,6 +53,14 @@ namespace ngraph
             {
                 friend class GPUPrimitiveEmitter;
 
+            public:
+                size_t build_primitive(const op::Convolution* node);
+                size_t build_primitive(const op::ConvolutionBackpropData* node);
+                size_t build_primitive(const op::ConvolutionBackpropFilters* node);
+                size_t build_primitive(const op::MaxPool* node);
+                size_t build_primitive(const op::Max* node);
+                size_t build_primitive(const op::Min* node);
+
             public:
                 enum class Prop
                 {

src/ngraph/runtime/gpu/gpu_kernel_emitters.cpp

@@ -232,3 +232,19 @@ void runtime::gpu::kernel::emit_cudnnReduceTensor(codegen::CodeWriter& writer,
     writer << "                  " << out.get_name() << "));\n";
     writer << "ngraph::runtime::gpu::free_gpu_buffer(workspace_ptr);\n";
 }
+
+std::string runtime::gpu::kernel::emit_type_string(const Node* node)
+{
+    std::stringstream ss;
+    for (auto const& input : node->get_inputs())
+    {
+        ss << input.get_element_type().c_type_string() << "_";
+    }
+    for (auto const& output : node->get_outputs())
+    {
+        ss << output.get_element_type().c_type_string() << "_";
+    }
+    std::string types = ss.str();
+    std::replace(types.begin(), types.end(), ' ', '_');
+    return types;
+}

src/ngraph/runtime/gpu/gpu_kernel_emitters.hpp

@@ -82,6 +82,8 @@ namespace ngraph
                                             const std::string& output_desc,
                                             const float& alpha,
                                             const float& beta);
+
+                std::string emit_type_string(const Node* node);
             }
         }
     }