COMPMID-417: Port DepthConcatenate to QS8/QS16 for NEON/CL.

Change-Id: I3dddae63043c7aa18d908a4fc8abacf3c64f98ca
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/80081
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Steven Niu <steven.niu@arm.com>

2 files changed, 51 insertions, 21 deletions

diff --git a/src/core/CL/kernels/CLDepthConcatenateKernel.cpp b/src/core/CL/kernels/CLDepthConcatenateKernel.cpp
index 73f1ba15df..6a699ae710 100644
--- a/src/core/CL/kernels/CLDepthConcatenateKernel.cpp
+++ b/src/core/CL/kernels/CLDepthConcatenateKernel.cpp
@@ -35,6 +35,10 @@
 #include "arm_compute/core/Validate.h"
 #include "arm_compute/core/Window.h"
 
+#include "support/ToolchainSupport.h"
+
+#include <map>
+
 using namespace arm_compute;
 
 CLDepthConcatenateKernel::CLDepthConcatenateKernel()
@@ -49,12 +53,22 @@ BorderSize CLDepthConcatenateKernel::border_size() const
 
 void CLDepthConcatenateKernel::configure(const ICLTensor *input, unsigned int depth_offset, ICLTensor *output)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F32);
+    static std::map<int, std::pair<std::string, int>> configs_map =
+    {
+        { 1, { "uchar", 16 } },
+        { 2, { "ushort", 8 } },
+        { 4, { "uint", 4 } },
+        { 8, { "ulong", 2 } },
+    };
+
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT_POSITION(input, output);
     ARM_COMPUTE_ERROR_ON(input->info()->dimension(2) + depth_offset > output->info()->dimension(2));
     ARM_COMPUTE_ERROR_ON(input->info()->dimension(0) > output->info()->dimension(0));
     ARM_COMPUTE_ERROR_ON(input->info()->dimension(1) > output->info()->dimension(1));
     ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(3, input, output);
+    ARM_COMPUTE_ERROR_ON(configs_map.find(input->info()->element_size()) == configs_map.end());
 
     // The gaps between the two lowest dimensions of input and output need to be divisible by 2
     // Otherwise it is not clear how the padding should be added onto the input tensor
@@ -64,33 +78,44 @@ void CLDepthConcatenateKernel::configure(const ICLTensor *input, unsigned int de
     _input  = input;
     _output = output;
 
+    // Add build options
+    auto                  config = configs_map.find(static_cast<int>(input->info()->element_size()));
+    std::set<std::string> build_opts;
+    build_opts.emplace(("-DDATA_TYPE=" + config->second.first));
+    build_opts.emplace(("-DVEC_SIZE=" + support::cpp11::to_string(config->second.second)));
+
     // Create kernel
-    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("concatenate_depth"));
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("concatenate_depth", build_opts));
 
     // Configure kernel window
     _left_right = (output->info()->dimension(0) - input->info()->dimension(0)) / 2;
     _top_bottom = (output->info()->dimension(1) - input->info()->dimension(1)) / 2;
 
-    const unsigned int offset_to_first_elements_in_bytes = depth_offset * output->info()->strides_in_bytes()[2] + _left_right * output->info()->strides_in_bytes()[0] + _top_bottom *
-                                                           output->info()->strides_in_bytes()[1];
+    const int offset_to_first_elements_in_bytes = depth_offset * output->info()->strides_in_bytes()[2];
 
-    const unsigned int num_elems_processed_per_iteration = 4;
-    const unsigned int num_elems_read_per_iteration      = 4;
+    const unsigned int num_elems_processed_per_iteration = 16 / input->info()->element_size();
+    const unsigned int num_elems_read_per_iteration      = 16 / input->info()->element_size();
     const unsigned int num_rows_read_per_iteration       = 1;
 
     // The window needs to be based on input as we copy all the depths of input
-    Window win = calculate_max_enlarged_window(*input->info(), Steps(num_elems_processed_per_iteration), border_size());
+    Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration));
+    win.set(Window::DimZ, Window::Dimension(0, input->info()->tensor_shape().z(), 1));
 
+    AccessWindowRectangle  input_access(input->info(), -_left_right, -_top_bottom, num_elems_read_per_iteration, num_rows_read_per_iteration);
     AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration);
-
-    update_window_and_padding(win,
-                              AccessWindowRectangle(input->info(), -_left_right, -_top_bottom, num_elems_read_per_iteration, num_rows_read_per_iteration),
-                              output_access);
-
+    update_window_and_padding(win, input_access, output_access);
     output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), output->info()->tensor_shape()));
 
-    unsigned int idx = 2 * num_arguments_per_2D_tensor(); // Skip the input and output parameters
-    _kernel.setArg<unsigned int>(idx, offset_to_first_elements_in_bytes);
+    unsigned int  idx = 2 * num_arguments_per_3D_tensor(); // Skip the input and output parameters
+    const cl_int3 offsets =
+    {
+        {
+            static_cast<cl_int>(_left_right),
+            static_cast<cl_int>(_top_bottom),
+            static_cast<cl_int>(offset_to_first_elements_in_bytes),
+        }
+    };
+    _kernel.setArg<cl_int3>(idx, offsets);
 
     ICLKernel::configure(win);
 }
@@ -100,14 +125,14 @@ void CLDepthConcatenateKernel::run(const Window &window, cl::CommandQueue &queue
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
 
-    Window slice = window.first_slice_window_2D();
+    Window slice = window.first_slice_window_3D();
 
     do
     {
         unsigned int idx = 0;
-        add_2D_tensor_argument(idx, _input, slice);
-        add_2D_tensor_argument(idx, _output, slice);
+        add_3D_tensor_argument(idx, _input, slice);
+        add_3D_tensor_argument(idx, _output, slice);
         enqueue(queue, *this, slice);
     }
-    while(window.slide_window_slice_2D(slice));
+    while(window.slide_window_slice_3D(slice));
 }
diff --git a/src/core/CL/kernels/CLFillBorderKernel.cpp b/src/core/CL/kernels/CLFillBorderKernel.cpp
index 2c751a489c..7667491710 100644
--- a/src/core/CL/kernels/CLFillBorderKernel.cpp
+++ b/src/core/CL/kernels/CLFillBorderKernel.cpp
@@ -76,7 +76,7 @@ void CLFillBorderKernel::configure(ICLTensor *tensor, BorderSize border_size, Bo
 
     // Define select type required by replicate border > 1
     const DataType dt          = tensor->info()->data_type();
-    std::string    select_type = get_cl_type_from_data_type(dt);
+    std::string    select_type = get_underlying_cl_type_from_data_type(dt);
     if(is_data_type_float(dt))
     {
         select_type = (DataType::F32 == dt) ? "int" : "short";
@@ -84,7 +84,7 @@ void CLFillBorderKernel::configure(ICLTensor *tensor, BorderSize border_size, Bo
 
     // Define build options
     std::set<std::string> build_opts;
-    build_opts.emplace(("-DDATA_TYPE=" + get_cl_type_from_data_type(dt)));
+    build_opts.emplace(("-DDATA_TYPE=" + get_underlying_cl_type_from_data_type(dt)));
     build_opts.emplace(("-DSELECT_TYPE=" + select_type));
     build_opts.emplace(("-DBORDER_SIZE_TOP=" + support::cpp11::to_string(border_size.top)));
     build_opts.emplace(("-DBORDER_SIZE_BOTTOM=" + support::cpp11::to_string(border_size.bottom)));
@@ -119,9 +119,14 @@ void CLFillBorderKernel::configure(ICLTensor *tensor, BorderSize border_size, Bo
             case DataType::U8:
                 set_constant_border<uint8_t>(idx, constant_border_value);
                 break;
+            case DataType::QS8:
+            case DataType::S8:
+                set_constant_border<int8_t>(idx, constant_border_value);
+                break;
             case DataType::U16:
                 set_constant_border<uint16_t>(idx, constant_border_value);
                 break;
+            case DataType::QS16:
             case DataType::S16:
                 set_constant_border<int16_t>(idx, constant_border_value);
                 break;