[IE CLDNN] Quantized deeplabv3 optimizations (#646)

Enabled dilation for imad dw fsv16 kernel
Added argmax and mutable_data to fsv16 white list
Enabled byxf input for quantize scale_shift kernel

inference-engine/thirdparty/clDNN/kernel_selector/core/actual_kernels/convolution/convolution_kernel_b_fs_yx_fsv_16_32_imad_dw.cpp

@@ -66,6 +66,7 @@ ParamsKey kernel_selector::ConvolutionKernel_b_fs_yx_fsv_16_32_imad_dw::GetSuppo
     k.EnableQuantization(QuantizationType::ASYMMETRIC_DATA_AND_WEIGHTS);
     k.EnableDepthwiseSeparableOpt();
     k.EnableGroupedConvolution();
+    k.EnableDilation();
     return k;
 }
 

inference-engine/thirdparty/clDNN/kernel_selector/core/actual_kernels/quantize/quantize_kernel_scale_shift_opt.cpp

@@ -33,6 +33,7 @@ ParamsKey QuantizeKernelScaleShift::GetSupportedKey() const {
     k.EnableOutputDataType(Datatype::INT8);
     k.EnableInputLayout(DataLayout::bfyx);
     k.EnableInputLayout(DataLayout::yxfb);
+    k.EnableInputLayout(DataLayout::byxf);
     k.EnableInputLayout(DataLayout::bfzyx);
     k.EnableInputLayout(DataLayout::b_fs_yx_fsv16);
     k.EnableInputLayout(DataLayout::b_fs_zyx_fsv16);

inference-engine/thirdparty/clDNN/src/layout_optimizer.cpp

@@ -367,8 +367,7 @@ bool layout_optimizer::convolution_b_fs_yx_fsv16_opt(layout const &input_layout,
             weights_layout.size.batch[0] >= 16 &&
             ((conv->groups == 1 && conv->split() == 1) ||
              conv->groups == static_cast<uint32_t>(input_layout.size.feature[0]) ||
-             conv->split() == static_cast<int32_t>(input_layout.size.feature[0])) &&
+             conv->split() == static_cast<int32_t>(input_layout.size.feature[0])))
-             conv->dilation == tensor{ 1 })
             return true;
         // Check for grouped convolution
         else if (input_layout.size.spatial[2] == 1 && input_layout.size.batch[0] < 16 &&

inference-engine/thirdparty/clDNN/src/program.cpp

@@ -46,6 +46,7 @@
 #include "deconvolution_inst.h"
 #include "detection_output_inst.h"
 #include "input_layout_inst.h"
+#include "arg_max_min_inst.h"
 #include "lstm_inst.h"
 #include "lstm_elt_inst.h"
 #include "lstm_gemm_inst.h"
@@ -1130,7 +1131,9 @@ void program_impl::set_layout_optimizer_attributes(layout_optimizer& lo) {
             (prim.type() != cldnn::mvn::type_id()
              || (prim.as<mvn>().input().get_output_layout().data_type != data_types::u8 &&
                  prim.as<mvn>().input().get_output_layout().data_type != data_types::i8)
-             || prim.as<mvn>().get_primitive()->across_channels))
+             || prim.as<mvn>().get_primitive()->across_channels) &&
+            prim.type() != cldnn::arg_max_min::type_id() &&
+            prim.type() != cldnn::mutable_data::type_id())
             can_use_fsv16 = false;
 
         // WA to keep bfyx_f16 layout disabled for some topologies where it leads to regressions.

inference-engine/thirdparty/clDNN/tests/test_cases/convolution_gpu_test.cpp

@@ -4746,7 +4746,6 @@ protected:
 
         build_options opts;
         opts.set_option(build_option::optimize_data(false));
-        opts.set_option(build_option::graph_dumps_dir("/tmp/cldnn_dumps/"));
 
         topology topology(input_layout("input", input.get_layout()),
                           data("weights", weights),
@@ -8214,6 +8213,11 @@ struct params_generator : std::vector<convolution_random_test_all_params> {
                 b, 32, 24, { 19, 19 }, { 3, 3 }, { 1, 1 }, { -1, -1 }, { 2, 2 }, true, 1, input_format, asymm_weights, asymm_data, padded_input });
             push_back(convolution_random_test_all_params{
                 b, 32, 24, { 19, 19 }, { 3, 3 }, { 2, 2 }, { -1, -1 }, { 2, 2 }, true, 1, input_format, asymm_weights, asymm_data, padded_input });
+            // depthwise + dilation
+            push_back(convolution_random_test_all_params{
+                b, 64, 64, { 19, 19 }, { 3, 3 }, { 1, 1 }, { -1, -1 }, { 2, 2 }, true, 64, input_format, asymm_weights, asymm_data, padded_input });
+            push_back(convolution_random_test_all_params{
+                b, 64, 64, { 19, 19 }, { 3, 3 }, { 2, 2 }, { -1, -1 }, { 2, 2 }, true, 64, input_format, asymm_weights, asymm_data, padded_input });
         }
         return *this;
     }