Complete to run benchmark_app f32 for resample cpu impl

kelvinchoi-intel · kelvinchoi-intel · commit 931097edfb67 · 2024-12-27T00:56:46.000+09:00
diff --git a/src/plugins/intel_gpu/src/graph/impls/cpu/resample.cpp b/src/plugins/intel_gpu/src/graph/impls/cpu/resample.cpp
@@ -106,34 +106,32 @@ struct resample_impl : public typed_primitive_impl<resample> {
         auto input_mem_ptr = instance.input_memory_ptr();
         cldnn::mem_lock<uint8_t, mem_lock_type::read> input_lock(input_mem_ptr, stream);
         
+        auto input_rank = params->input_layouts[0].get_rank();
         for (size_t i = 0; i < params->input_layouts.size(); i++) {
             auto input_tensor = make_tensor(params->input_layouts[0], input_lock.data());
             input_host_tensors.push_back(input_tensor);
         }
 
+        if (scales.size() < input_rank)
+            scales.insert(scales.begin(), input_rank - scales.size(), 1.f);
+        
+        for(size_t i = (input_rank - axes.size()); i > 0; i--)
+            axes.insert(axes.begin(), 1, (i - 1));
+
         if (input_host_tensors.size() == 1) {
             auto target_shape_sizes = params->output_layouts[0].get_tensor().sizes();
             std::vector<int64_t> target_shape_ps;
-            for (size_t i = 0; i < axes.size(); i++)
+            for (size_t i = 0; i < input_rank; i++)
                 target_shape_ps.push_back(target_shape_sizes[i]);
 
             auto target_shape_tensor = ov::Tensor(ov::element::i32, {target_shape_ps.size()}, target_shape_ps.data());
             input_host_tensors.push_back(target_shape_tensor);
 
-            if (shape_calc_mode == ov::op::util::InterpolateBase::ShapeCalcMode::SIZES) {
-                auto new_scales = scales;
-                auto input_shape_sizes = params->input_layouts[0].get_tensor().sizes();
-                for (size_t i = 0; i < sizes.size(); i++)
-                    new_scales[i] = sizes[i] / input_shape_sizes[i];
-
-                auto scales_tensor = ov::Tensor(ov::element::f32, {new_scales.size()}, new_scales.data());
-                input_host_tensors.push_back(scales_tensor);
-                shape_calc_mode = ov::op::util::InterpolateBase::ShapeCalcMode::SCALES;
-            } else if (shape_calc_mode == ov::op::util::InterpolateBase::ShapeCalcMode::SCALES) {
+            if (shape_calc_mode == ov::op::util::InterpolateBase::ShapeCalcMode::SCALES) {
                 auto scales_tensor = ov::Tensor(ov::element::f32, {scales.size()}, scales.data());
                 input_host_tensors.push_back(scales_tensor);
             } else {
-                OPENVINO_ASSERT(false, "[GPU] Not supported Interpolate ShapeCalcMode", instance.id());
+                OPENVINO_ASSERT(false, "[GPU] Not supported Interpolate ShapeCalcMode of CPU impl", instance.id());
             }
             
             auto axes_tensor = ov::Tensor(ov::element::i64, {axes.size()}, axes.data());
@@ -189,19 +187,6 @@ struct resample_impl : public typed_primitive_impl<resample> {
 namespace detail {
 
 attach_resample_impl::attach_resample_impl() {
-    // auto formats = {
-    //     format::bfyx,
-    // };
-
-    // auto types = {
-    //     data_types::f32,
-    // };
-
-    // implementation_map<resample>::add(impl_types::cpu, shape_types::static_shape, resample_impl::create, types, formats);
-    // implementation_map<resample>::add(impl_types::cpu, shape_types::dynamic_shape, resample_impl::create, types, formats);
-
-    //std::set<implementation_map<resample>::key_type> keys;
-
     const auto types = {data_types::f32, data_types::i32};
     const auto formats = {
         format::bfyx,
@@ -219,13 +204,6 @@ attach_resample_impl::attach_resample_impl() {
         format::bs_fs_zyx_bsv32_fsv32,
         format::bs_fs_zyx_bsv32_fsv16,
     };
-    // for (const auto type : types) {
-    //     for (const auto format : formats) {
-    //         keys.emplace(type, format);
-    //     }
-    // }
-
-    // keys.emplace(data_types::f32, format::yxfb);
 
     implementation_map<resample>::add(impl_types::cpu, shape_types::static_shape, resample_impl::create, types, formats);
     implementation_map<resample>::add(impl_types::cpu, shape_types::dynamic_shape, resample_impl::create, types, formats);
diff --git a/src/plugins/intel_gpu/src/graph/impls/ocl/register.cpp b/src/plugins/intel_gpu/src/graph/impls/ocl/register.cpp
@@ -72,7 +72,6 @@ void register_implementations() {
     REGISTER_OCL(swiglu);
     REGISTER_OCL(tile);
     REGISTER_OCL(gather_tree);
-    //REGISTER_OCL(resample);
     REGISTER_OCL(grn);
     REGISTER_OCL(ctc_greedy_decoder);
     REGISTER_OCL(ctc_loss);
diff --git a/src/plugins/intel_gpu/src/graph/impls/ocl/register.hpp b/src/plugins/intel_gpu/src/graph/impls/ocl/register.hpp
@@ -47,7 +47,6 @@
 #include "intel_gpu/primitives/reduce.hpp"
 #include "intel_gpu/primitives/region_yolo.hpp"
 #include "intel_gpu/primitives/reorg_yolo.hpp"
-//#include "intel_gpu/primitives/resample.hpp"
 #include "intel_gpu/primitives/reshape.hpp"
 #include "intel_gpu/primitives/reverse_sequence.hpp"
 #include "intel_gpu/primitives/rms.hpp"
diff --git a/src/plugins/intel_gpu/src/graph/impls/ocl/resample.cpp b/src/plugins/intel_gpu/src/graph/impls/ocl/resample.cpp
@@ -175,42 +175,6 @@ struct resample_impl : typed_primitive_impl_ocl<resample> {
     }
 };
 
-// namespace detail {
-
-// attach_resample_impl::attach_resample_impl() {
-//     std::set<implementation_map<resample>::key_type> keys;
-
-//     const auto types = {data_types::f16, data_types::f32, data_types::i8, data_types::u8, data_types::i32};
-//     const auto formats = {
-//         format::bfyx,
-//         format::b_fs_yx_fsv16,
-//         format::b_fs_yx_fsv32,
-//         format::bs_fs_yx_bsv16_fsv16,
-//         format::bs_fs_yx_bsv32_fsv16,
-//         format::bs_fs_yx_bsv32_fsv32,
-
-//         format::bfzyx,
-//         format::b_fs_zyx_fsv16,
-//         format::b_fs_zyx_fsv32,
-//         format::bs_fs_zyx_bsv16_fsv32,
-//         format::bs_fs_zyx_bsv16_fsv16,
-//         format::bs_fs_zyx_bsv32_fsv32,
-//         format::bs_fs_zyx_bsv32_fsv16,
-//     };
-//     for (const auto type : types) {
-//         for (const auto format : formats) {
-//             keys.emplace(type, format);
-//         }
-//     }
-
-//     keys.emplace(data_types::f32, format::yxfb);
-//     keys.emplace(data_types::f16, format::yxfb);
-//     keys.emplace(data_types::f16, format::fs_b_yx_fsv32);
-
-//     implementation_map<resample>::add(impl_types::ocl, typed_primitive_impl_ocl<resample>::create<resample_impl>, keys);
-// }
-
-// }  // namespace detail
 std::unique_ptr<primitive_impl> ResampleImplementationManager::create_impl(const program_node& node, const kernel_impl_params& params) const {
     assert(node.is_type<resample>());
     return typed_primitive_impl_ocl<resample>::create<resample_impl>(static_cast<const resample_node&>(node), params);
diff --git a/src/plugins/intel_gpu/src/graph/impls/ocl/resample.hpp b/src/plugins/intel_gpu/src/graph/impls/ocl/resample.hpp
@@ -4,7 +4,7 @@
 
 #include "impls/registry/implementation_manager.hpp"
 #include "program_node.h"
-//#include "intel_gpu/primitives/resample.hpp"
+#include "resample_inst.h"
 
 #include <memory>
 namespace cldnn {
@@ -15,16 +15,6 @@ struct ResampleImplementationManager : public ImplementationManager {
     ResampleImplementationManager(shape_types shape_type, ValidateFunc vf = nullptr) : ImplementationManager(impl_types::ocl, shape_type, vf) {}
     std::unique_ptr<primitive_impl> create_impl(const program_node& node, const kernel_impl_params& params) const override;
     bool validate_impl(const program_node& node) const override {
-        // auto prim = node.as<resample>().get_primitive();
-        // const auto& in0_layout = node.get_input_layout(0);
-
-        // if (in0_layout.data_type == ov::element::f32 &&
-        //     prim->operation_type == ov::op::util::InterpolateBase::InterpolateMode::LINEAR_ONNX &&
-        //     prim->coord_trans_mode == ov::op::util::InterpolateBase::CoordinateTransformMode::ALIGN_CORNERS &&
-        //     prim->shape_calc_mode == ov::op::util::InterpolateBase::ShapeCalcMode::SCALES) {
-        //     return false;
-        // }
-
         return true;
     }
 };
diff --git a/src/plugins/intel_gpu/src/graph/impls/registry/registry.hpp b/src/plugins/intel_gpu/src/graph/impls/registry/registry.hpp
@@ -201,7 +201,6 @@ REGISTER_DEFAULT_IMPLS(space_to_batch, OCL_S);
 REGISTER_DEFAULT_IMPLS(space_to_depth, OCL_S);
 REGISTER_DEFAULT_IMPLS(swiglu, OCL_S, OCL_D);
 REGISTER_DEFAULT_IMPLS(gather_tree, OCL_S);
-//REGISTER_DEFAULT_IMPLS(resample, CPU_S, OCL_S);
 REGISTER_DEFAULT_IMPLS(grn, OCL_S);
 REGISTER_DEFAULT_IMPLS(ctc_greedy_decoder, OCL_S);
 REGISTER_DEFAULT_IMPLS(ctc_loss, OCL_S);
diff --git a/src/plugins/intel_gpu/src/graph/impls/registry/resample_impls.cpp b/src/plugins/intel_gpu/src/graph/impls/registry/resample_impls.cpp
@@ -33,19 +33,10 @@ const std::vector<std::shared_ptr<cldnn::ImplementationManager>>& Registry<resam
 
                 return true;
         })
-        // OV_GPU_CREATE_INSTANCE_OCL(ocl::ResampleImplementationManager, shape_types::dynamic_shape,
-        //     [](const cldnn::program_node& node){
-        //         return false;
-        // })
-
         OV_GPU_GET_INSTANCE_CPU(resample, shape_types::static_shape,
             [](const cldnn::program_node& node){
                 return true;
         })
-        // OV_GPU_GET_INSTANCE_CPU(resample, shape_types::dynamic_shape,
-        //     [](const cldnn::program_node& node){
-        //         return false;
-        // })
     };
 
     return impls;
diff --git a/src/plugins/intel_gpu/tests/functional/single_layer_tests/dynamic/interpolate.cpp b/src/plugins/intel_gpu/tests/functional/single_layer_tests/dynamic/interpolate.cpp
@@ -549,13 +549,20 @@ INSTANTIATE_TEST_SUITE_P(InterpolateLinear_Layout_Test, InterpolateLayerGPUTest,
 const std::vector<ShapeParams> shapeParams4D_LargeShape = {
     ShapeParams{
         ov::op::v4::Interpolate::ShapeCalcMode::SCALES,
-        //InputShape{{-1, {2, 100}, -1, -1}, {{1, 64, 148, 148}}},
-        InputShape{{-1, -1, -1, -1}, {{1, 3, 48, 48}}}, // min shape for failure
+        InputShape{{-1, {2, 100}, -1, -1}, {{1, 64, 148, 148}}},
         ov::test::utils::InputLayerType::CONSTANT,
         ov::test::utils::InputLayerType::CONSTANT,
         {{1.f, 1.f, 2.f, 2.f}},
         defaultAxes4D.front()
     },
+    ShapeParams{
+        ov::op::v4::Interpolate::ShapeCalcMode::SCALES,
+        InputShape{{-1, -1, -1, -1}, {{1, 3, 48, 48}}},
+        ov::test::utils::InputLayerType::CONSTANT,
+        ov::test::utils::InputLayerType::CONSTANT,
+        {{2.f, 2.f}},
+        reducedAxes4D.front()
+    },
     ShapeParams{
         ov::op::v4::Interpolate::ShapeCalcMode::SIZES,
         InputShape{{-1, -1, -1, -1}, {{1, 3, 48, 48}}},
