COMPMID-3441: Nightly failed due to NEScale QASYMM8_SIGNED output wrong result

Change-Id: I4a97523408f44c509ba0f6bf180c56ea33b964cf
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3138
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Sheri Zhang <sheri.zhang@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

2 files changed, 36 insertions, 10 deletions

diff --git a/arm_compute/core/Helpers.h b/arm_compute/core/Helpers.h
index 34e4e6f885..09c672ecfa 100644
--- a/arm_compute/core/Helpers.h
+++ b/arm_compute/core/Helpers.h
@@ -146,7 +146,7 @@ inline T delta_bilinear_c1(const T *pixel_ptr, size_t stride, float dx, float dy
 }
 
 /** Computes bilinear interpolation for quantized input and output, using the pointer to the top-left pixel and the pixel's distance between
- * the real coordinates and the smallest following integer coordinates. Input must be quantized and in single channel format.
+ * the real coordinates and the smallest following integer coordinates. Input must be QASYMM8 and in single channel format.
  *
  * @param[in] pixel_ptr Pointer to the top-left pixel value of a single channel input.
  * @param[in] stride    Stride to access the bottom-left and bottom-right pixel values
@@ -179,6 +179,40 @@ inline uint8_t delta_bilinear_c1_quantized(const uint8_t *pixel_ptr, size_t stri
     return static_cast<uint8_t>(quantize_qasymm8(res, oq_info));
 }
 
+/** Computes bilinear interpolation for quantized input and output, using the pointer to the top-left pixel and the pixel's distance between
+ * the real coordinates and the smallest following integer coordinates. Input must be QASYMM8_SIGNED and in single channel format.
+ *
+ * @param[in] pixel_ptr Pointer to the top-left pixel value of a single channel input.
+ * @param[in] stride    Stride to access the bottom-left and bottom-right pixel values
+ * @param[in] dx        Pixel's distance between the X real coordinate and the smallest X following integer
+ * @param[in] dy        Pixel's distance between the Y real coordinate and the smallest Y following integer
+ * @param[in] iq_info   Input QuantizationInfo
+ * @param[in] oq_info   Output QuantizationInfo
+ *
+ * @note dx and dy must be in the range [0, 1.0]
+ *
+ * @return The bilinear interpolated pixel value
+ */
+inline int8_t delta_bilinear_c1_quantized(const int8_t *pixel_ptr, size_t stride, float dx, float dy, UniformQuantizationInfo iq_info, UniformQuantizationInfo oq_info)
+{
+    ARM_COMPUTE_ERROR_ON(pixel_ptr == nullptr);
+
+    const float dx1 = 1.0f - dx;
+    const float dy1 = 1.0f - dy;
+
+    const float a00 = dequantize_qasymm8_signed(*pixel_ptr, iq_info);
+    const float a01 = dequantize_qasymm8_signed(*(pixel_ptr + 1), iq_info);
+    const float a10 = dequantize_qasymm8_signed(*(pixel_ptr + stride), iq_info);
+    const float a11 = dequantize_qasymm8_signed(*(pixel_ptr + stride + 1), iq_info);
+
+    const float w1  = dx1 * dy1;
+    const float w2  = dx * dy1;
+    const float w3  = dx1 * dy;
+    const float w4  = dx * dy;
+    float       res = a00 * w1 + a01 * w2 + a10 * w3 + a11 * w4;
+    return static_cast<int8_t>(quantize_qasymm8_signed(res, oq_info));
+}
+
 /** Computes linear interpolation using the pointer to the top pixel and the pixel's distance between
  * the real coordinates and the smallest following integer coordinates. Input must be in single channel format.
  *
diff --git a/src/core/NEON/kernels/NEScaleKernel.cpp b/src/core/NEON/kernels/NEScaleKernel.cpp
index 47a02d77c5..4f2f925c3c 100644
--- a/src/core/NEON/kernels/NEScaleKernel.cpp
+++ b/src/core/NEON/kernels/NEScaleKernel.cpp
@@ -25,20 +25,12 @@
 
 #include "arm_compute/core/AccessWindowStatic.h"
 #include "arm_compute/core/CPP/Validate.h"
-#include "arm_compute/core/Coordinates.h"
-#include "arm_compute/core/Error.h"
 #include "arm_compute/core/Helpers.h"
-#include "arm_compute/core/ITensor.h"
 #include "arm_compute/core/NEON/wrapper/wrapper.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_compute/core/utils/misc/Utility.h"
 
 #include <arm_neon.h>
-#include <cstddef>
-#include <cstdint>
 
 namespace arm_compute
 {
@@ -724,7 +716,7 @@ void NEScaleKernel::scale_bilinear_nchw(const Window &window)
                 const auto offsets_ptr = reinterpret_cast<const int32_t *>(offsets.ptr());
                 const auto dx_ptr      = reinterpret_cast<const float *>(dx.ptr());
                 const auto dy_ptr      = reinterpret_cast<const float *>(dy.ptr());
-                const auto in_ptr      = reinterpret_cast<const uint8_t *>(in.ptr());
+                const auto in_ptr      = reinterpret_cast<const int8_t *>(in.ptr());
 
                 const int in_yi      = std::floor((id.y() + _sampling_offset) * hr - _sampling_offset);
                 const int offset_row = in_yi * in_stide_in_bytes;