[IE CLDNN] Use single primitive reduce when axis=1(feature) (openvinotoolkit#4548)

Author: hyunback

inference-engine/src/cldnn_engine/cldnn_engine.cpp

@@ -134,11 +134,8 @@ cldnn::device_info clDNNEngine::GetDeviceInfo(const std::map<std::string, std::s
 template<typename T>
 static bool disableReduceDecomposition(const std::shared_ptr<const ngraph::Node> node) {
     if (auto op = std::dynamic_pointer_cast<const T>(node)) {
-        auto reduction_axes = op->get_reduction_axes().to_vector();
-        bool reduce_along_f = op->get_reduction_axes().size() == 1 && std::count(reduction_axes.begin(), reduction_axes.end(), 1) != 0;
         bool fp16_batch_not_1 = op->get_element_type() == ngraph::element::f16 && op->input(0).get_shape()[0] != 1;
-        bool can_use_reduce = !reduce_along_f && !fp16_batch_not_1;
-        return can_use_reduce;
+        return !fp16_batch_not_1;
     }
     return false;
 }

inference-engine/thirdparty/clDNN/kernel_selector/core/actual_kernels/resample/resample_kernel_opt.cpp

@@ -22,6 +22,8 @@ ParamsKey ResampleKernelOpt::GetSupportedKey() const {
     ParamsKey k;
     k.EnableInputDataType(Datatype::F16);
     k.EnableInputDataType(Datatype::F32);
+    k.EnableInputDataType(Datatype::UINT8);
+    k.EnableInputDataType(Datatype::INT8);
     k.EnableOutputDataType(Datatype::F16);
     k.EnableOutputDataType(Datatype::F32);
     k.EnableOutputDataType(Datatype::UINT8);
@@ -84,6 +86,11 @@ bool ResampleKernelOpt::Validate(const Params& p, const optional_params& o) cons
 
     const auto& input = params.inputs[0];
 
+    if ((input.GetDType() == Datatype::UINT8 || input.GetDType() == Datatype::INT8) &&
+        params.resampleType != ResampleType::NEAREST_NEIGHBOR &&
+        params.resampleType != ResampleType::BILINEAR_INTERP)
+        return false;
+
     if (input.GetLayout() != DataLayout::fs_b_yx_fsv32 && input.GetLayout() != DataLayout::b_fs_yx_fsv16)
         return false;
 

inference-engine/thirdparty/clDNN/kernel_selector/core/cl_kernels/reduce_gpu_b_fs_yx_fsv16.cl

@@ -256,7 +256,7 @@ uint offset = batch_out * input_batch_pitch + ((feature_out + FSV - 1) / FSV) *
             for (uint yi = y_out; yi < y_max_val; ++yi) {
                 for (uint xi = x_out; xi < x_max_val; ++xi) {
                     INPUT_VEC input = (INPUT_VEC)(INPUT_INIT_VAL);
-                    #if (REDUCE_MAX_MODE || REDUCE_MIN_MODE || REDUCE_PROD_MODE || REDUCE_AND_MODE || REDUCE_LOG_SUM_EXP_MODE) && REDUCE_FEATURE && (INPUT0_FEATURE_NUM % FSV != 0)
+                    #if REDUCE_FEATURE && (INPUT0_FEATURE_NUM % FSV != 0)
                         if (fi + FSV <= INPUT0_FEATURE_NUM)
                             input = BLOCK_READ(data, offset);
                         else
@@ -273,7 +273,7 @@ uint offset = batch_out * input_batch_pitch + ((feature_out + FSV - 1) / FSV) *
                 #if INPUT0_SIZE_X % READ_OFFSET != 0
                     for (uint xi = x_leftover_start; xi < x_leftover_end; ++xi) {
                         INPUT0_TYPE leftovers = INIT_VAL;
-                        #if (REDUCE_MAX_MODE || REDUCE_MIN_MODE || REDUCE_PROD_MODE || REDUCE_AND_MODE || REDUCE_LOG_SUM_EXP_MODE) && REDUCE_FEATURE && (INPUT0_FEATURE_NUM % FSV != 0)
+                        #if REDUCE_FEATURE && (INPUT0_FEATURE_NUM % FSV != 0)
                             if (fi + FSV <= INPUT0_FEATURE_NUM)
                                 leftovers = DT_INPUT_BLOCK_READ(data, offset);
                             else
@@ -341,7 +341,7 @@ uint offset = batch_out * input_batch_pitch + ((feature_out + FSV - 1) / FSV) *
             for (uint yi = y_out; yi < y_max_val; ++yi) {
                 for (uint xi = x_out; xi < x_max_val; ++xi) {
                     INPUT_VEC input = (INPUT_VEC)(INPUT_INIT_VAL);
-                    #if (REDUCE_MAX_MODE || REDUCE_MIN_MODE || REDUCE_PROD_MODE || REDUCE_AND_MODE || REDUCE_LOG_SUM_EXP_MODE) && REDUCE_FEATURE && (INPUT0_FEATURE_NUM % FSV != 0)
+                    #if REDUCE_FEATURE && (INPUT0_FEATURE_NUM % FSV != 0)
                         if (fi + FSV <= INPUT0_FEATURE_NUM)
                             input = BLOCK_READ(data, offset);
                         else