COMPMID-2235: Extend type support for CL/NEON DequantizationLayer.

Adds support for:
- QSYMM8

Change-Id: Ia0b839fc844ce0f968dad1b69a001f9a660dbcd5
Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1378
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>

2 files changed, 48 insertions, 2 deletions

diff --git a/arm_compute/core/NEON/NEAsymm.h b/arm_compute/core/NEON/NEAsymm.h
index 2347c468ab..4c8f797360 100644
--- a/arm_compute/core/NEON/NEAsymm.h
+++ b/arm_compute/core/NEON/NEAsymm.h
@@ -223,6 +223,52 @@ inline float32x4x4_t vdequantize(const uint8x16_t &qv, const UniformQuantization
     return vdequantized_input;
 }
 
+/** Dequantize following an asymmetric quantization scheme a neon vector holding 16 quantized values.
+ *
+ * @param[in] qv     Input values to be dequantized.
+ * @param[in] scale  Quantization scaling factor.
+ * @param[in] offset Zero quantization offset.
+ *
+ * @return Dequantized values in a neon vector
+ */
+inline float32x4x4_t vdequantize(const uint8x16_t &qv, float scale, int32_t offset)
+{
+    const int32x4_t     voffset = vdupq_n_s32(offset);
+    const float32x4_t   vscale  = vdupq_n_f32(scale);
+    const float32x4x4_t vdequantized_input =
+    {
+        {
+            vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_low_u8(qv))))), voffset)), vscale),
+            vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_low_u8(qv))))), voffset)), vscale),
+            vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_high_u8(qv))))), voffset)), vscale),
+            vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_high_u8(qv))))), voffset)), vscale),
+        }
+    };
+    return vdequantized_input;
+}
+
+/** Dequantize following a symmetric quantization scheme a neon vector holding 16 quantized values.
+ *
+ * @param[in] qv    Input values to be dequantized.
+ * @param[in] scale Quantization scaling factor.
+ *
+ * @return Dequantized values in a neon vector
+ */
+inline float32x4x4_t vdequantize(const int8x16_t &qv, float scale)
+{
+    const float32x4_t   vscale = vdupq_n_f32(scale);
+    const float32x4x4_t vdequantized_input =
+    {
+        {
+            vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_low_s8(qv))))), vscale),
+            vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_low_s8(qv))))), vscale),
+            vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_high_s8(qv))))), vscale),
+            vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_high_s8(qv))))), vscale),
+        }
+    };
+    return vdequantized_input;
+}
+
 /** Quantize a neon vector holding 8 floating point values.
  *
  * @param[in] qv Input values to be quantized.
diff --git a/arm_compute/core/NEON/kernels/NEDequantizationLayerKernel.h b/arm_compute/core/NEON/kernels/NEDequantizationLayerKernel.h
index 7d215f5f7b..3320ba6889 100644
--- a/arm_compute/core/NEON/kernels/NEDequantizationLayerKernel.h
+++ b/arm_compute/core/NEON/kernels/NEDequantizationLayerKernel.h
@@ -52,13 +52,13 @@ public:
     ~NEDequantizationLayerKernel() = default;
     /** Set input, output tensors.
      *
-     * @param[in]  input  Source tensor. Data type supported: QASYMM8.
+     * @param[in]  input  Source tensor. Data type supported: QASYMM8/QSYMM8.
      * @param[out] output Destination tensor with the same dimensions of input. Data type supported: F16/F32.
      */
     void configure(const ITensor *input, ITensor *output);
     /** Static function to check if given info will lead to a valid configuration of @ref NEDequantizationLayerKernel
      *
-     * @param[in] input  Input tensor info. Data types supported: QASYMM8.
+     * @param[in] input  Input tensor info. Data types supported: QASYMM8/QSYMM8.
      * @param[in] output Output tensor info. Data types supported: F16/F32.
      *
      * @return a status