[GPU] Fixes after review openvinotoolkit#2

Lyamin-Roman · Jun 15, 2023 · cbeced0 · cbeced0
1 parent f44ca5c
commit cbeced0
Show file tree

Hide file tree

Showing 7 changed files with 32 additions and 57 deletions.
diff --git a/src/plugins/intel_gpu/include/intel_gpu/primitives/reorder.hpp b/src/plugins/intel_gpu/include/intel_gpu/primitives/reorder.hpp
@@ -25,13 +25,13 @@ struct WeightsReorderParams {
           _out_layout(out_layout),
           _transposed(transposed) {}
 
-    virtual size_t hash() const {
+    size_t hash() const {
         size_t seed = hash_combine(_in_layout.hash(), _out_layout.hash());
         seed = hash_combine(seed, _transposed);
         return seed;
     }
 
-    virtual bool operator==(const WeightsReorderParams& rhs) const {
+    bool operator==(const WeightsReorderParams& rhs) const {
         if (typeid(*this) != typeid(rhs))
             return false;
 
@@ -46,8 +46,6 @@ struct WeightsReorderParams {
 
     void set_input_layout(const layout& layout) { _in_layout = layout; }
 
-    virtual ~WeightsReorderParams() = default;
-
 protected:
     layout _in_layout;
     layout _out_layout;

diff --git a/src/plugins/intel_gpu/src/graph/impls/ocl/reorder.cpp b/src/plugins/intel_gpu/src/graph/impls/ocl/reorder.cpp
@@ -115,13 +115,13 @@ struct reorder_impl : typed_primitive_impl_ocl<reorder> {
         bool is_reorder_weights = format::is_weights_format(impl_param.get_input_layout().format) ||
                                   format::is_weights_format(impl_param.get_output_layout().format);
         if (is_reorder_weights) {
-            return create_reorder_weigths(impl_param);
+            return create_reorder_weights(impl_param);
         } else {
             return typed_primitive_impl_ocl<reorder>::create<reorder_impl>(arg, impl_param);
         }
     }
 
-    static std::unique_ptr<primitive_impl> create_reorder_weigths(const kernel_impl_params& impl_param) {
+    static std::unique_ptr<primitive_impl> create_reorder_weights(const kernel_impl_params& impl_param) {
         const auto& prim = impl_param.typed_desc<reorder>();
         const auto& weights_params = prim->weights_reorder_params;
         auto& kernel_selector = kernel_selector::ReorderWeightsKernelSelector::Instance();
@@ -166,7 +166,7 @@ attach_reorder_impl::attach_reorder_impl() {
     };
     implementation_map<reorder>::add(impl_types::ocl, shape_types::dynamic_shape, reorder_impl::create, types, formats);
 
-    WeightsReordersFactory::add(cldnn::impl_types::ocl, shape_types::static_shape, reorder_impl::create_reorder_weigths);
+    WeightsReordersFactory::add(cldnn::impl_types::ocl, shape_types::static_shape, reorder_impl::create_reorder_weights);
 }
 
 }  // namespace detail

diff --git a/src/plugins/intel_gpu/src/graph/impls/onednn/convolution_onednn.cpp b/src/plugins/intel_gpu/src/graph/impls/onednn/convolution_onednn.cpp
@@ -130,24 +130,17 @@ struct convolution_onednn : typed_primitive_onednn_impl<convolution> {
         return attrs;
     }
 
-    static kernel_selector::WeightsReorderParams get_weights_reorder(const kernel_impl_params& impl_params, const dnnl::primitive_desc& pd, bool rotate) {
-        kernel_selector::WeightsReorderParams weights_reorder_params;
-
+    static std::shared_ptr<WeightsReorderParams> get_weights_reorder(const kernel_impl_params& impl_params, const dnnl::primitive_desc& pd, bool rotate) {
         auto cldnn_prim = impl_params.typed_desc<convolution>();
-        auto weights_layout = impl_params.get_input_layout(1);
-        auto grouped_weights = format::is_grouped(weights_layout.format) || cldnn_prim->grouped_weights_shape;
-        cldnn::format out_fmt = onednn::find_format(pd.weights_desc(0), grouped_weights);
-        kernel_selector::WeightsLayout reqLayout = to_weights_layout(out_fmt, cldnn_prim->grouped_weights_shape);
-
-        weights_reorder_params.src = convert_weights_tensor(weights_layout, cldnn_prim->grouped_weights_shape);
-        weights_reorder_params.dest = weights_reorder_params.src.TransformIgnorePadding(reqLayout,
-                                                                                        weights_reorder_params.src.GetDType(),
-                                                                                        cldnn_prim->groups,
-                                                                                        false);
-        weights_reorder_params.rotate = rotate;
-        weights_reorder_params.is_initialized = true;
-
-        return weights_reorder_params;
+
+        auto input_weights_layout = impl_params.get_input_layout(1);
+        auto grouped_weights = format::is_grouped(input_weights_layout.format) || cldnn_prim->grouped_weights_shape;
+        format out_fmt = onednn::find_format(pd.weights_desc(0), grouped_weights);
+
+        auto output_weights_layout = input_weights_layout;
+        output_weights_layout.format = out_fmt;
+
+        return std::make_shared<WeightsReorderParams>(input_weights_layout, output_weights_layout, rotate);
     }
 
 public:

diff --git a/src/plugins/intel_gpu/src/graph/impls/onednn/deconvolution_onednn.cpp b/src/plugins/intel_gpu/src/graph/impls/onednn/deconvolution_onednn.cpp
@@ -51,24 +51,17 @@ struct deconvolution_onednn : typed_primitive_onednn_impl<deconvolution> {
         return arg.get_onednn_primitive_attributes();
     }
 
-    static kernel_selector::WeightsReorderParams get_weights_reorder(const kernel_impl_params& impl_params, const dnnl::primitive_desc& pd) {
-        kernel_selector::WeightsReorderParams weights_reorder_params;
-
+    static std::shared_ptr<WeightsReorderParams> get_weights_reorder(const kernel_impl_params& impl_params, const dnnl::primitive_desc& pd) {
         auto cldnn_prim = impl_params.typed_desc<deconvolution>();
-        auto weights_layout = impl_params.get_input_layout(1);
-        auto grouped_weights = format::is_grouped(weights_layout.format) || cldnn_prim->grouped_weights_shape;
-        cldnn::format out_fmt = onednn::find_format(pd.weights_desc(0), grouped_weights);
-        kernel_selector::WeightsLayout reqLayout = to_weights_layout(out_fmt, cldnn_prim->grouped_weights_shape);
-
-        weights_reorder_params.src = convert_weights_tensor(weights_layout, cldnn_prim->grouped_weights_shape);
-        weights_reorder_params.dest = weights_reorder_params.src.TransformIgnorePadding(reqLayout,
-                                                                                        weights_reorder_params.src.GetDType(),
-                                                                                        cldnn_prim->groups,
-                                                                                        false);
-        weights_reorder_params.rotate = false;
-        weights_reorder_params.is_initialized = true;
-
-        return weights_reorder_params;
+
+        auto input_weights_layout = impl_params.get_input_layout(1);
+        auto grouped_weights = format::is_grouped(input_weights_layout.format) || cldnn_prim->grouped_weights_shape;
+        format out_fmt = onednn::find_format(pd.weights_desc(0), grouped_weights);
+
+        auto output_weights_layout = input_weights_layout;
+        output_weights_layout.format = out_fmt;
+
+        return std::make_shared<WeightsReorderParams>(input_weights_layout, output_weights_layout, false);
     }
 
 public:

diff --git a/src/plugins/intel_gpu/src/graph/impls/onednn/fully_connected_onednn.cpp b/src/plugins/intel_gpu/src/graph/impls/onednn/fully_connected_onednn.cpp
@@ -51,7 +51,7 @@ struct fully_connected_onednn : typed_primitive_onednn_impl<fully_connected> {
         return args;
     }
 
-    static kernel_selector::WeightsReorderParams get_weights_reorder(const kernel_impl_params& impl_params, const dnnl::primitive_desc& pd) {
+    static std::shared_ptr<WeightsReorderParams> get_weights_reorder(const kernel_impl_params& impl_params, const dnnl::primitive_desc& pd) {
         auto input_layout = impl_params.get_input_layout(0);
         auto weights_layout = impl_params.get_input_layout(1);
         auto cldnn_prim = impl_params.typed_desc<fully_connected>();
@@ -66,20 +66,12 @@ struct fully_connected_onednn : typed_primitive_onednn_impl<fully_connected> {
             weights_layout.set_partial_shape(reshape_to_2d(weights_pshape, feature));
         }
 
-        kernel_selector::WeightsReorderParams weights_reorder_params;
+        format out_fmt = onednn::find_format(pd.weights_desc(0));
 
-        cldnn::format out_fmt = onednn::find_format(pd.weights_desc(0));
-        kernel_selector::WeightsLayout req_layout = to_weights_layout(out_fmt, false);
+        auto output_weights_layout = weights_layout;
+        output_weights_layout.format = out_fmt;
 
-        weights_reorder_params.src =  convert_weights_tensor(weights_layout, false);
-        weights_reorder_params.dest = weights_reorder_params.src.TransformIgnorePadding(req_layout,
-                                                                                        weights_reorder_params.src.GetDType(),
-                                                                                        1,
-                                                                                        false);
-        weights_reorder_params.rotate = false;
-        weights_reorder_params.is_initialized = true;
-
-        return weights_reorder_params;
+        return std::make_shared<WeightsReorderParams>(weights_layout, output_weights_layout, false);
     }
 
     static std::shared_ptr<dnnl::inner_product_forward::primitive_desc> get_fully_connected_primitive_descriptor(const kernel_impl_params& impl_params,

diff --git a/src/plugins/intel_gpu/src/graph/impls/onednn/primitive_onednn_base.h b/src/plugins/intel_gpu/src/graph/impls/onednn/primitive_onednn_base.h
@@ -46,8 +46,8 @@ struct typed_primitive_onednn_impl : public typed_primitive_impl<PType> {
             const ExecutionConfig& config,
             std::shared_ptr<dnnl::primitive_attr> attrs,
             const PrimDescType& pd,
-            kernel_selector::WeightsReorderParams weights_reorder = {})
-        : typed_primitive_impl<PType>(create_weights_reorder_params(weights_reorder), pd.impl_info_str()),
+            std::shared_ptr<WeightsReorderParams> weights_reorder = {})
+        : typed_primitive_impl<PType>(weights_reorder, pd.impl_info_str()),
         _engine(&engine),
         _attrs(attrs),
         _pd(pd) {

diff --git a/src/plugins/intel_gpu/src/graph/include/primitive_inst.h b/src/plugins/intel_gpu/src/graph/include/primitive_inst.h
@@ -98,7 +98,6 @@ struct primitive_impl {
 
     bool need_weights_reorder() const { return _weights_reorder_params != nullptr; }
     std::shared_ptr<WeightsReorderParams> get_weights_reorder_params() const { return _weights_reorder_params; }
-    void reset_weights_reorder_params() { _weights_reorder_params = nullptr; }
 
     std::shared_ptr<kernel_impl_params> get_weights_reorder_kernel_params() const;