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>

6 files changed, 172 insertions, 53 deletions

diff --git a/src/core/CL/CLHelpers.cpp b/src/core/CL/CLHelpers.cpp
index dd87e778d7..1073b39ca7 100644
--- a/src/core/CL/CLHelpers.cpp
+++ b/src/core/CL/CLHelpers.cpp
@@ -100,6 +100,19 @@ std::string get_cl_type_from_data_type(const DataType &dt)
     }
 }
 
+std::string get_underlying_cl_type_from_data_type(const DataType &dt)
+{
+    switch(dt)
+    {
+        case DataType::QS8:
+            return "char";
+        case DataType::QS16:
+            return "short";
+        default:
+            return get_cl_type_from_data_type(dt);
+    }
+}
+
 const std::string &string_from_target(GPUTarget target)
 {
     static std::map<GPUTarget, const std::string> gpu_target_map =
diff --git a/src/core/CL/cl_kernels/concatenate.cl b/src/core/CL/cl_kernels/concatenate.cl
index 00f5189508..a92ab5bdad 100644
--- a/src/core/CL/cl_kernels/concatenate.cl
+++ b/src/core/CL/cl_kernels/concatenate.cl
@@ -25,29 +25,35 @@
 
 /** This kernel concatenates the input tensor into the output tensor along the third dimension
  *
- * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: QS8, QS16, F16, F32
  * @param[in]  src_stride_x                      Stride of the source tensor in X dimension (in bytes)
  * @param[in]  src_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  src_stride_y                      Stride of the source tensor in Y dimension (in bytes)
  * @param[in]  src_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  src_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  src_step_z                        src_stride_z * number of elements along Z processed per workitem(in bytes)
  * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source tensor
- * @param[out] dst_ptr                           Pointer to the destination tensor. Supported data types: F32
+ * @param[out] dst_ptr                           Pointer to the destination tensor. Supported data types: same as @p src_ptr
  * @param[in]  dst_stride_x                      Stride of the destination tensor in X dimension (in bytes)
  * @param[in]  dst_step_x                        dst_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  dst_stride_y                      Stride of the destination tensor in Y dimension (in bytes)
  * @param[in]  dst_step_y                        dst_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  dst_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  dst_step_z                        dst_stride_z * number of elements along Z processed per workitem(in bytes)
  * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
- * @param[in]  offset                            The offset to the first valid element of the output tensor in bytes
+ * @param[in]  offsets                           The offsets to the first valid element of the output tensor in bytes
  */
 __kernel void concatenate_depth(
-    IMAGE_DECLARATION(src),
-    IMAGE_DECLARATION(dst),
-    unsigned int offset)
+    TENSOR3D_DECLARATION(src),
+    TENSOR3D_DECLARATION(dst),
+    int3 offsets)
 {
-    Image src = CONVERT_TO_IMAGE_STRUCT(src);
-    Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
+    Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
+    Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
 
-    float4 source_values = vload4(0, (__global float *)src.ptr);
+    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+    source_values = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)tensor3D_offset(&src, -offsets.x, -offsets.y, 0));
 
-    vstore4(source_values, 0, (__global float *)(dst.ptr + offset));
+    VSTORE(VEC_SIZE)
+    (source_values, 0, (__global DATA_TYPE *)(dst.ptr + offsets.z));
 }
diff --git a/src/core/CL/cl_kernels/helpers.h b/src/core/CL/cl_kernels/helpers.h
index 29a43f769b..0b6d92dfd0 100644
--- a/src/core/CL/cl_kernels/helpers.h
+++ b/src/core/CL/cl_kernels/helpers.h
@@ -30,6 +30,12 @@
 
 #define CLAMP(x, min_val, max_val) min(max(x, min_val), max_val)
 
+#define VLOAD_STR(size) vload##size
+#define VLOAD(size) VLOAD_STR(size)
+
+#define VSTORE_STR(size) vstore##size
+#define VSTORE(size) VSTORE_STR(size)
+
 #define VEC_DATA_TYPE_STR(type, size) type##size
 #define VEC_DATA_TYPE(type, size) VEC_DATA_TYPE_STR(type, size)
 
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;
diff --git a/src/core/NEON/kernels/NEDepthConcatenateKernel.cpp b/src/core/NEON/kernels/NEDepthConcatenateKernel.cpp
index 902490ec38..d58e4e0aa5 100644
--- a/src/core/NEON/kernels/NEDepthConcatenateKernel.cpp
+++ b/src/core/NEON/kernels/NEDepthConcatenateKernel.cpp
@@ -27,17 +27,76 @@
 #include "arm_compute/core/Helpers.h"
 #include "arm_compute/core/IAccessWindow.h"
 #include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/NEON/NEFixedPoint.h"
 #include "arm_compute/core/TensorInfo.h"
 #include "arm_compute/core/Utils.h"
 #include "arm_compute/core/Validate.h"
 #include "arm_compute/core/Window.h"
 
 #include <arm_neon.h>
+#include <cstdint>
 
 using namespace arm_compute;
 
+namespace
+{
+// Overloads of 128-bit vector loads
+uint8x16_t loadq(const uint8_t *ptr)
+{
+    return vld1q_u8(ptr);
+}
+uint16x8_t loadq(const uint16_t *ptr)
+{
+    return vld1q_u16(ptr);
+}
+uint32x4_t loadq(const uint32_t *ptr)
+{
+    return vld1q_u32(ptr);
+}
+// Overloads of 128-bit vector stores
+void storeq(uint8_t *ptr, uint8x16_t val)
+{
+    return vst1q_u8(ptr, val);
+}
+void storeq(uint16_t *ptr, uint16x8_t val)
+{
+    return vst1q_u16(ptr, val);
+}
+void storeq(uint32_t *ptr, uint32x4_t val)
+{
+    return vst1q_u32(ptr, val);
+}
+
+template <typename T>
+void depth_concat(const ITensor *in, ITensor *out, std::pair<int, int> start_xy, int depth_offset, const Window &window)
+{
+    const int start_x = start_xy.first;
+    const int start_y = start_xy.second;
+
+    // Offset input
+    const int input_offset_to_first_elements_in_bytes = in->info()->offset_first_element_in_bytes() - start_x * in->info()->strides_in_bytes()[0] - start_y * in->info()->strides_in_bytes()[1];
+    uint8_t *input_ptr                               = in->buffer() + input_offset_to_first_elements_in_bytes;
+
+    // Offset output
+    const unsigned int output_offset_to_first_elements_in_bytes = out->info()->offset_first_element_in_bytes() + depth_offset * out->info()->strides_in_bytes()[2];
+    uint8_t           *output_ptr                               = out->buffer() + output_offset_to_first_elements_in_bytes;
+
+    Iterator input(in, window);
+    Iterator output(out, window);
+
+    execute_window_loop(window, [&](const Coordinates & id)
+    {
+        const auto in_ptr  = reinterpret_cast<const T *>(input_ptr + input.offset());
+        const auto out_ptr = reinterpret_cast<T *>(output_ptr + output.offset());
+
+        storeq(out_ptr, loadq(in_ptr));
+    },
+    input, output);
+}
+} // namespace
+
 NEDepthConcatenateKernel::NEDepthConcatenateKernel()
-    : _input(nullptr), _output(nullptr), _top_bottom(0), _left_right(0), _depth_offset(0)
+    : _func(nullptr), _input(nullptr), _output(nullptr), _top_bottom(0), _left_right(0), _depth_offset(0)
 {
 }
 
@@ -48,8 +107,9 @@ BorderSize NEDepthConcatenateKernel::border_size() const
 
 void NEDepthConcatenateKernel::configure(const ITensor *input, unsigned int depth_offset, ITensor *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);
+    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));
@@ -60,18 +120,36 @@ void NEDepthConcatenateKernel::configure(const ITensor *input, unsigned int dept
     ARM_COMPUTE_ERROR_ON((output->info()->dimension(0) - input->info()->dimension(0)) % 2);
     ARM_COMPUTE_ERROR_ON((output->info()->dimension(1) - input->info()->dimension(1)) % 2);
 
+    _func         = nullptr;
     _input        = input;
     _output       = output;
     _depth_offset = depth_offset;
     _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 num_elems_processed_per_iteration = 4;
-    const unsigned int num_elems_read_per_iteration      = 4;
+    switch(input->info()->data_type())
+    {
+        case DataType::QS8:
+            _func = &depth_concat<uint8_t>;
+            break;
+        case DataType::QS16:
+        case DataType::F16:
+            _func = &depth_concat<uint16_t>;
+            break;
+        case DataType::F32:
+            _func = &depth_concat<uint32_t>;
+            break;
+        default:
+            ARM_COMPUTE_ERROR("Unsupported data type.");
+    }
+
+    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);
@@ -85,21 +163,7 @@ void NEDepthConcatenateKernel::run(const Window &window)
 {
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
+    ARM_COMPUTE_ERROR_ON(_func == nullptr);
 
-    // Offset output
-    const unsigned int offset_to_first_elements_in_bytes = _output->info()->offset_first_element_in_bytes() + _left_right * _output->info()->strides_in_bytes()[0] + _top_bottom *
-                                                           _output->info()->strides_in_bytes()[1] + _depth_offset * _output->info()->strides_in_bytes()[2];
-    uint8_t           *output_ptr                        = _output->buffer() + offset_to_first_elements_in_bytes;
-
-    Iterator input(_input, window);
-    Iterator output(_output, window);
-
-    execute_window_loop(window, [&](const Coordinates & id)
-    {
-        const auto in_ptr  = reinterpret_cast<const float *>(input.ptr());
-        const auto out_ptr = reinterpret_cast<float *>(output_ptr + output.offset());
-
-        vst1q_f32(out_ptr, vld1q_f32(in_ptr));
-    },
-    input, output);
+    (*_func)(_input, _output, std::make_pair(_left_right, _top_bottom), _depth_offset, window);
 }