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>

1 files changed, 86 insertions, 22 deletions

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);
 }