COMPMID-2417: NEDequantizationLayer support for QASYMM8_PER_CHANNEL

Change-Id: I1ef4ce8610e11e81702b0b7f0f7c437fed49833e
Signed-off-by: Michalis Spyrou <michalis.spyrou@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1795
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 46 insertions, 1 deletions

diff --git a/src/core/NEON/kernels/NEDequantizationLayerKernel.cpp b/src/core/NEON/kernels/NEDequantizationLayerKernel.cpp
index e52f53ea04..d880c80d82 100644
--- a/src/core/NEON/kernels/NEDequantizationLayerKernel.cpp
+++ b/src/core/NEON/kernels/NEDequantizationLayerKernel.cpp
@@ -43,7 +43,7 @@ namespace
 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QSYMM8, DataType::QSYMM16);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_PER_CHANNEL, DataType::QSYMM8, DataType::QSYMM16);
 
     if(output->tensor_shape().total_size() > 0)
     {
@@ -160,6 +160,48 @@ void run_dequantization_qasymm8(const ITensor *input, ITensor *output, const Win
 }
 
 template <typename T>
+void run_dequantization_qasymm8_per_channel(const ITensor *input, ITensor *output, const Window &window)
+{
+    const std::vector<float>   scale  = input->info()->quantization_info().scale();
+    const std::vector<int32_t> offset = input->info()->quantization_info().offset();
+
+    const int  window_step_x  = 16;
+    const auto window_start_x = static_cast<int>(window.x().start());
+    const auto window_end_x   = static_cast<int>(window.x().end());
+
+    // Reset first dimension to handle tail calculations manually
+    Window win(window);
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    // Create iterators
+    Iterator in(input, win);
+    Iterator out(output, win);
+
+    execute_window_loop(win, [&](const Coordinates & id)
+    {
+        const auto in_ptr  = reinterpret_cast<const uint8_t *>(in.ptr());
+        const auto out_ptr = reinterpret_cast<T *>(out.ptr());
+
+        int x = window_start_x;
+        for(; x <= (window_end_x - window_step_x); x += window_step_x)
+        {
+            const auto vin  = wrapper::vloadq(in_ptr + x);
+            const auto vdeq = vdequantize(vin, scale[id.z()], offset[id.z()]);
+
+            store_result<T>(reinterpret_cast<T *>(out_ptr + x), vdeq);
+        }
+
+        // Compute left-over elements
+        for(; x < window_end_x; ++x)
+        {
+            uint8_t val    = *(in_ptr + x);
+            *(out_ptr + x) = static_cast<T>(dequantize(val, scale[id.z()], offset[id.z()]));
+        }
+    },
+    in, out);
+}
+
+template <typename T>
 void run_dequantization_qsymm8(const ITensor *input, ITensor *output, const Window &window)
 {
     const UniformQuantizationInfo &qinfo = input->info()->quantization_info().uniform();
@@ -251,6 +293,9 @@ void run_dequantization_core(const ITensor *input, ITensor *output, const Window
         case DataType::QASYMM8:
             run_dequantization_qasymm8<T>(input, output, window);
             break;
+        case DataType::QASYMM8_PER_CHANNEL:
+            run_dequantization_qasymm8_per_channel<T>(input, output, window);
+            break;
         case DataType::QSYMM8:
             run_dequantization_qsymm8<T>(input, output, window);
             break;