COMPMID-791: Adds support of QASYMM8 in NEDepthwiseConvolution3x3

Change-Id: I1a9ed6c3420ddf8978aeaad48d9915333b006b49
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/116374
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

1 files changed, 4 insertions, 55 deletions

diff --git a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp b/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp
index 8b3f2383ab..5e14e1a95d 100644
--- a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp
+++ b/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -96,57 +96,11 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
     Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
     return std::make_pair(err, win);
 }
+} // namespace
 
-template <bool    is_bounded_relu>
-inline uint8x16_t finalize_quantization(int32x4x4_t &in_s32, int result_fixedpoint_multiplier, int32_t result_shift, int32x4_t result_offset_after_shift_s32, uint8x16_t min_u8,
-                                        uint8x16_t max_u8)
+namespace arm_compute
 {
-    const static int32x4_t zero_s32 = vdupq_n_s32(0);
-
-    // Fixed point multiplication with vector saturating rounding doubling multiply high with scalar
-    in_s32.val[0] = vqrdmulhq_n_s32(in_s32.val[0], result_fixedpoint_multiplier);
-    in_s32.val[1] = vqrdmulhq_n_s32(in_s32.val[1], result_fixedpoint_multiplier);
-    in_s32.val[2] = vqrdmulhq_n_s32(in_s32.val[2], result_fixedpoint_multiplier);
-    in_s32.val[3] = vqrdmulhq_n_s32(in_s32.val[3], result_fixedpoint_multiplier);
-
-    // Round to the nearest division by a power-of-two using result_shift_s32
-    in_s32.val[0] = rounding_divide_by_pow2(in_s32.val[0], result_shift);
-    in_s32.val[1] = rounding_divide_by_pow2(in_s32.val[1], result_shift);
-    in_s32.val[2] = rounding_divide_by_pow2(in_s32.val[2], result_shift);
-    in_s32.val[3] = rounding_divide_by_pow2(in_s32.val[3], result_shift);
-
-    // Add the offset terms
-    in_s32.val[0] = vaddq_s32(in_s32.val[0], result_offset_after_shift_s32);
-    in_s32.val[1] = vaddq_s32(in_s32.val[1], result_offset_after_shift_s32);
-    in_s32.val[2] = vaddq_s32(in_s32.val[2], result_offset_after_shift_s32);
-    in_s32.val[3] = vaddq_s32(in_s32.val[3], result_offset_after_shift_s32);
-
-    // Saturate negative values
-    in_s32.val[0] = vmaxq_s32(in_s32.val[0], zero_s32);
-    in_s32.val[1] = vmaxq_s32(in_s32.val[1], zero_s32);
-    in_s32.val[2] = vmaxq_s32(in_s32.val[2], zero_s32);
-    in_s32.val[3] = vmaxq_s32(in_s32.val[3], zero_s32);
-
-    // Convert S32 to S16
-    const int16x8x2_t in_s16 =
-    {
-        {
-            vcombine_s16(vqmovn_s32(in_s32.val[0]), vqmovn_s32(in_s32.val[1])),
-            vcombine_s16(vqmovn_s32(in_s32.val[2]), vqmovn_s32(in_s32.val[3]))
-        }
-    };
-
-    // Convert S16 to U8
-    uint8x16_t out_u8 = vcombine_u8(vqmovun_s16(in_s16.val[0]), vqmovun_s16(in_s16.val[1]));
-
-    if(is_bounded_relu)
-    {
-        out_u8 = vmaxq_u8(out_u8, min_u8);
-        out_u8 = vminq_u8(out_u8, max_u8);
-    }
-
-    return out_u8;
-}
+class Coordinates;
 
 /* Function used by the left-over for loop to perform the quantization */
 template <bool is_bounded_relu>
@@ -178,11 +132,6 @@ inline uint8_t finalize_quantization(int32x4_t in_s32, int result_fixedpoint_mul
 
     return out_u8;
 }
-} // namespace
-
-namespace arm_compute
-{
-class Coordinates;
 } // namespace arm_compute
 
 template <bool is_bounded_relu>