COMPMID-2796: Add support for QASYMM8_SIGNED in NEActivationLayer and NEPReluLayer

Change-Id: I089fd19a6beab7779d690bc9ace327f661c2753d
Signed-off-by: Michalis Spyrou <michalis.spyrou@arm.com>
Reviewed-on: https://review.mlplatform.org/c/2407
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>

2 files changed, 366 insertions, 6 deletions

diff --git a/src/core/NEON/kernels/NEActivationLayerKernel.cpp b/src/core/NEON/kernels/NEActivationLayerKernel.cpp
index c338ef09c7..44f76f6e22 100644
--- a/src/core/NEON/kernels/NEActivationLayerKernel.cpp
+++ b/src/core/NEON/kernels/NEActivationLayerKernel.cpp
@@ -48,7 +48,7 @@ namespace
 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const ActivationLayerInfo &activation_info)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::QASYMM8, DataType::QSYMM16, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::QASYMM8_SIGNED, DataType::QASYMM8, DataType::QSYMM16, DataType::F16, DataType::F32);
 
     static std::set<ActivationLayerInfo::ActivationFunction> qasymm8_supported_activations =
     {
@@ -72,8 +72,13 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, c
 
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(is_data_type_quantized_symmetric(data_type) && (qsymm16_supported_activations.count(f_act) == 0),
                                     "For QSYMM16 only tanh and logistic are supported");
-    ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized_asymmetric(data_type) && (f_act == ActivationLayerInfo::ActivationFunction::TANH) && (oq_info != QuantizationInfo(1.f / 128.f, 128)));
-    ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized_asymmetric(data_type) && (f_act == ActivationLayerInfo::ActivationFunction::LOGISTIC) && (oq_info != QuantizationInfo(1.f / 256.f, 0)));
+    ARM_COMPUTE_RETURN_ERROR_ON((data_type == DataType::QASYMM8 || data_type == DataType::QASYMM16) && (f_act == ActivationLayerInfo::ActivationFunction::TANH)
+                                && (oq_info != QuantizationInfo(1.f / 128.f, 128)));
+    ARM_COMPUTE_RETURN_ERROR_ON((data_type == DataType::QASYMM8 || data_type == DataType::QASYMM16) && (f_act == ActivationLayerInfo::ActivationFunction::LOGISTIC)
+                                && (oq_info != QuantizationInfo(1.f / 256.f, 0)));
+
+    ARM_COMPUTE_RETURN_ERROR_ON(data_type == DataType::QASYMM8_SIGNED && (f_act == ActivationLayerInfo::ActivationFunction::TANH) && (oq_info != QuantizationInfo(1.f / 128.f, 0)));
+    ARM_COMPUTE_RETURN_ERROR_ON(data_type == DataType::QASYMM8_SIGNED && (f_act == ActivationLayerInfo::ActivationFunction::LOGISTIC) && (oq_info != QuantizationInfo(1.f / 256.f, -128)));
 
     ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized_symmetric(data_type) && (f_act == ActivationLayerInfo::ActivationFunction::TANH) && (oq_info != QuantizationInfo(1.f / 32768.f, 0)));
     ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized_symmetric(data_type) && (f_act == ActivationLayerInfo::ActivationFunction::LOGISTIC) && (oq_info != QuantizationInfo(1.f / 32768.f, 0)));
@@ -173,6 +178,17 @@ void NEActivationLayerKernel::configure(ITensor *input, ITensor *output, Activat
     };
 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC*/
 
+    // Activation functions : QASYMM8_SIGNED
+    static std::map<ActivationFunction, ActivationFunctionExecutorPtr> act_map_qasymm8_signed =
+    {
+        { ActivationFunction::LOGISTIC, &NEActivationLayerKernel::activation<ActivationFunction::LOGISTIC, qasymm8_signed_t> },
+        { ActivationFunction::BOUNDED_RELU, &NEActivationLayerKernel::activation<ActivationFunction::BOUNDED_RELU, qasymm8_signed_t> },
+        { ActivationFunction::LU_BOUNDED_RELU, &NEActivationLayerKernel::activation<ActivationFunction::LU_BOUNDED_RELU, qasymm8_signed_t> },
+        { ActivationFunction::RELU, &NEActivationLayerKernel::activation<ActivationFunction::RELU, qasymm8_signed_t> },
+        { ActivationFunction::TANH, &NEActivationLayerKernel::activation<ActivationFunction::TANH, qasymm8_signed_t> },
+        { ActivationFunction::IDENTITY, &NEActivationLayerKernel::activation<ActivationFunction::IDENTITY, qasymm8_signed_t> },
+    };
+
     // Activation functions : QASYMM8
     static std::map<ActivationFunction, ActivationFunctionExecutorPtr> act_map_qasymm8 =
     {
@@ -193,6 +209,9 @@ void NEActivationLayerKernel::configure(ITensor *input, ITensor *output, Activat
 
     switch(input->info()->data_type())
     {
+        case DataType::QASYMM8_SIGNED:
+            _func = act_map_qasymm8_signed[activation_info.activation()];
+            break;
         case DataType::QASYMM8:
             _func = act_map_qasymm8[activation_info.activation()];
             break;
@@ -508,6 +527,156 @@ typename std::enable_if<std::is_same<T, qasymm8_t>::value, void>::type NEActivat
 }
 
 template <ActivationLayerInfo::ActivationFunction F, typename T>
+typename std::enable_if<std::is_same<T, qasymm8_signed_t>::value, void>::type NEActivationLayerKernel::activation(const Window &window)
+{
+    const int                window_step_x  = 16 / sizeof(T);
+    const auto               window_start_x = static_cast<int>(window.x().start());
+    const auto               window_end_x   = static_cast<int>(window.x().end());
+    const ActivationFunction act            = F;
+
+    Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
+    win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    Iterator input(_input, win_collapsed);
+    Iterator output(_output, win_collapsed);
+
+    const UniformQuantizationInfo qi_in    = _input->info()->quantization_info().uniform();
+    const UniformQuantizationInfo qi_out   = _output->info()->quantization_info().uniform();
+    const qasymm8x16_signed_t     va       = vdupq_n_s8(quantize_qasymm8_signed(_act_info.a(), qi_in));
+    const qasymm8x16_signed_t     vb       = vdupq_n_s8(quantize_qasymm8_signed(_act_info.b(), qi_in));
+    const qasymm8_signed_t        a        = quantize_qasymm8_signed(_act_info.a(), qi_in);
+    const qasymm8_signed_t        b        = quantize_qasymm8_signed(_act_info.b(), qi_in);
+    const qasymm8_signed_t        const_0  = quantize_qasymm8_signed(0.f, qi_in);
+    const qasymm8x16_signed_t     vconst_0 = vdupq_n_s8(const_0);
+    const auto                    vconst_1 = vdupq_n_f32(1.f);
+    const float32x4_t             va_f32   = vdupq_n_f32(_act_info.a());
+    const float32x4_t             vb_f32   = vdupq_n_f32(_act_info.b());
+    const float                   a_f32    = _act_info.a();
+    const float                   b_f32    = _act_info.b();
+
+    // Initialise scale/offset for re-quantization
+    float       s  = qi_in.scale / qi_out.scale;
+    float       o  = -qi_in.offset * s + qi_out.offset;
+    float32x4_t vs = vdupq_n_f32(s);
+    float32x4_t vo = vdupq_n_f32(o);
+
+    execute_window_loop(win_collapsed, [&](const Coordinates &)
+    {
+        const auto input_ptr  = reinterpret_cast<const T *>(input.ptr());
+        const auto output_ptr = reinterpret_cast<T *>(output.ptr());
+
+        wrapper::traits::neon_bitvector_t<T, wrapper::traits::BitWidth::W128> tmp;
+
+        // Compute S elements per iteration
+        int x = window_start_x;
+        for(; x <= (window_end_x - window_step_x); x += window_step_x)
+        {
+            const auto vin = wrapper::vloadq(input_ptr + x);
+            if(act == ActivationFunction::RELU)
+            {
+                // Perform activation
+                tmp = vmaxq_s8(vconst_0, vin);
+                // Re-quantize to new output space
+                tmp = vmlaq_qasymm8_signed(tmp, vs, vo);
+            }
+            else if(act == ActivationFunction::BOUNDED_RELU)
+            {
+                // Perform activation
+                tmp = vminq_s8(va, vmaxq_s8(vconst_0, vin));
+                // Re-quantize to new output space
+                tmp = vmlaq_qasymm8_signed(tmp, vs, vo);
+            }
+            else if(act == ActivationFunction::LU_BOUNDED_RELU)
+            {
+                // Perform activation
+                tmp = vminq_s8(va, vmaxq_s8(vb, vin));
+                // Re-quantize to new output space
+                tmp = vmlaq_qasymm8_signed(tmp, vs, vo);
+            }
+            else if(act == ActivationFunction::LOGISTIC)
+            {
+                // De-quantize
+                const auto vin_deq = vdequantize(vin, qi_in);
+                // Perform activation
+                const float32x4x4_t tmp_dep =
+                {
+                    {
+                        wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[0])))),
+                        wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[1])))),
+                        wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[2])))),
+                        wrapper::vdiv(vconst_1, wrapper::vadd(vconst_1, wrapper::vexpq(wrapper::vneg(vin_deq.val[3])))),
+                    }
+                };
+                // Re-quantize to new output space
+                tmp = vquantize_signed(tmp_dep, qi_out);
+            }
+            else if(act == ActivationFunction::TANH)
+            {
+                // De-quantize
+                const auto vin_deq = vdequantize(vin, qi_in);
+                // Perform activation
+                const float32x4x4_t tmp_dep =
+                {
+                    {
+                        wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[0], vb_f32))),
+                        wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[1], vb_f32))),
+                        wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[2], vb_f32))),
+                        wrapper::vmul(va_f32, wrapper::vtanh(wrapper::vmul(vin_deq.val[3], vb_f32))),
+                    }
+                };
+                // Re-quantize to new output space
+                tmp = vquantize_signed(tmp_dep, qi_out);
+            }
+            else
+            {
+                ARM_COMPUTE_ERROR("Unsupported activation function");
+            }
+            wrapper::vstore(output_ptr + x, tmp);
+        }
+
+        // Compute left-over elements
+        for(; x < window_end_x; ++x)
+        {
+            T in = *(reinterpret_cast<const T *>(input_ptr + x));
+            T tmp;
+            if(act == ActivationFunction::RELU)
+            {
+                tmp = std::max(const_0, in);
+                tmp = std::max<int32_t>(0, std::min<int32_t>(tmp * s + o, 255));
+            }
+            else if(act == ActivationFunction::BOUNDED_RELU)
+            {
+                tmp = std::min(a, std::max(const_0, in));
+                tmp = std::max<int32_t>(0, std::min<int32_t>(tmp * s + o, 255));
+            }
+            else if(act == ActivationFunction::LU_BOUNDED_RELU)
+            {
+                tmp = std::min(a, std::max(b, in));
+                tmp = std::max<int32_t>(0, std::min<int32_t>(tmp * s + o, 255));
+            }
+            else if(act == ActivationFunction::LOGISTIC)
+            {
+                float tmp_f = dequantize_qasymm8_signed(in, qi_in);
+                tmp_f       = 1.f / (1.f + std::exp(-tmp_f));
+                tmp         = quantize_qasymm8_signed(tmp_f, qi_out);
+            }
+            else if(act == ActivationFunction::TANH)
+            {
+                float tmp_f = dequantize_qasymm8_signed(in, qi_in);
+                tmp_f       = a_f32 * std::tanh(b_f32 * tmp_f);
+                tmp         = quantize_qasymm8_signed(tmp_f, qi_out);
+            }
+            else
+            {
+                ARM_COMPUTE_ERROR("Unsupported activation function");
+            }
+            *(output_ptr + x) = tmp;
+        }
+    },
+    input, output);
+}
+
+template <ActivationLayerInfo::ActivationFunction F, typename T>
 typename std::enable_if<std::is_same<T, qsymm16_t>::value, void>::type NEActivationLayerKernel::activation(const Window &window)
 {
     const int                window_step_x  = 16 / sizeof(T);
diff --git a/src/core/NEON/kernels/NEElementwiseOperationKernel.cpp b/src/core/NEON/kernels/NEElementwiseOperationKernel.cpp
index 9bd080983c..4928ae9bdd 100644
--- a/src/core/NEON/kernels/NEElementwiseOperationKernel.cpp
+++ b/src/core/NEON/kernels/NEElementwiseOperationKernel.cpp
@@ -61,6 +61,21 @@ float32x4x4_t load_quantized(const uint8_t *input1_ptr, const int32x4_t &offset,
     return out;
 }
 
+float32x4x4_t load_quantized_signed(const int8_t *input1_ptr, const int32x4_t &offset, const float32x4_t &scale)
+{
+    qasymm8x16_signed_t x = vld1q_s8(input1_ptr);
+    const float32x4x4_t out =
+    {
+        {
+            vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_low_s8(x)))), offset)), scale),
+            vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_low_s8(x)))), offset)), scale),
+            vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_high_s8(x)))), offset)), scale),
+            vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_high_s8(x)))), offset)), scale),
+        }
+    };
+    return out;
+}
+
 void store_quantized(uint8_t *output_ptr, const uint32x4x4_t &out)
 {
     const uint8x8_t pa = vqmovn_u16(vcombine_u16(vqmovn_u32(out.val[0]), vqmovn_u32(out.val[1])));
@@ -89,6 +104,27 @@ void store_quantized(uint8_t *output_ptr, const float32x4x4_t &rf, const float32
     store_quantized(output_ptr, out);
 }
 
+void store_quantized_signed(int8_t *output_ptr, const int32x4x4_t &out)
+{
+    const int8x8_t pa = vqmovn_s16(vcombine_s16(vqmovn_s32(out.val[0]), vqmovn_s32(out.val[1])));
+    const int8x8_t pb = vqmovn_s16(vcombine_s16(vqmovn_s32(out.val[2]), vqmovn_s32(out.val[3])));
+    vst1q_s8(output_ptr, vcombine_s8(pa, pb));
+}
+
+void store_quantized_signed(int8_t *output_ptr, const float32x4x4_t &rf, const float32x4_t &offset, const float32x4_t &invscale)
+{
+    int32x4x4_t out =
+    {
+        {
+            vcvtq_s32_f32(vmlaq_f32(offset, rf.val[0], invscale)),
+            vcvtq_s32_f32(vmlaq_f32(offset, rf.val[1], invscale)),
+            vcvtq_s32_f32(vmlaq_f32(offset, rf.val[2], invscale)),
+            vcvtq_s32_f32(vmlaq_f32(offset, rf.val[3], invscale)),
+        }
+    };
+    store_quantized_signed(output_ptr, out);
+}
+
 float32x4x4_t dup_quantized(qasymm8_t broadcast_value, int offset, float scale)
 {
     const qasymm8x16_t broadcast_value_vec = vdupq_n_u8(broadcast_value);
@@ -152,6 +188,12 @@ inline uint8_t elementwise_arithm_op_quantized_scalar(const float &a, const floa
     return quantize_qasymm8(elementwise_arithm_op_scalar<op>(a, b), qinfo);
 }
 
+template <ArithmeticOperation op>
+inline int8_t elementwise_arithm_op_quantized_signed_scalar(const float &a, const float &b, UniformQuantizationInfo qinfo)
+{
+    return quantize_qasymm8_signed(elementwise_arithm_op_scalar<op>(a, b), qinfo);
+}
+
 template <ArithmeticOperation    op, typename VectorType>
 inline typename VectorType::type elementwise_arithm_op(const typename VectorType::type &a, const typename VectorType::type &b)
 {
@@ -368,6 +410,24 @@ inline int elementwise_arithm_op_quantized_loop(int window_start_x, int window_e
     return x;
 }
 
+template <ArithmeticOperation op>
+inline int elementwise_arithm_op_quantized_singed_loop(int window_start_x, int window_end_x, int window_step_x,
+                                                       const int8_t *input1_ptr, const int8_t *input2_ptr, int8_t *output_ptr,
+                                                       int32x4_t voffset1, int32x4_t voffset2, float32x4_t vscale1, float32x4_t vscale2,
+                                                       float32x4_t voffseto, float32x4_t invvscaleo)
+{
+    int x = window_start_x;
+    for(; x <= (window_end_x - window_step_x); x += window_step_x)
+    {
+        // Get inputs and compute output
+        const float32x4x4_t af = load_quantized_signed(input1_ptr + x, voffset1, vscale1);
+        const float32x4x4_t bf = load_quantized_signed(input2_ptr + x, voffset2, vscale2);
+        const float32x4x4_t rf = elementwise_arithm_op<op>(af, bf);
+        store_quantized_signed(output_ptr + x, rf, voffseto, invvscaleo);
+    }
+    return x;
+}
+
 template <ArithmeticOperation op, typename ScalarType, typename VectorType>
 inline int elementwise_arithm_op_broadcast_loop(int window_start_x, int window_end_x, int window_step_x,
                                                 const ScalarType *non_broadcast_input_ptr, const ScalarType &broadcast_value, ScalarType *output_ptr, const bool reorder)
@@ -396,6 +456,21 @@ inline int elementwise_arithm_op_quantized_broadcast_loop(int window_start_x, in
     }
     return x;
 }
+template <ArithmeticOperation op>
+inline int elementwise_arithm_op_quantized_signed_broadcast_loop(int window_start_x, int window_end_x, int window_step_x,
+                                                                 const int8_t *non_broadcast_input_ptr, float32x4x4_t broadcast_vector, int8_t *output_ptr,
+                                                                 int32x4_t voffset_non_broadcast, float32x4_t vscale_non_broadcast,
+                                                                 float32x4_t voffseto, float32x4_t invvscaleo, bool reorder)
+{
+    int x = window_start_x;
+    for(; x <= (window_end_x - window_step_x); x += window_step_x)
+    {
+        const float32x4x4_t af = load_quantized_signed(non_broadcast_input_ptr + x, voffset_non_broadcast, vscale_non_broadcast);
+        const float32x4x4_t rf = elementwise_arithm_op<op>(reorder ? broadcast_vector : af, reorder ? af : broadcast_vector);
+        store_quantized_signed(output_ptr + x, rf, voffseto, invvscaleo);
+    }
+    return x;
+}
 
 template <ComparisonOperation op, typename InputScalarType, typename InputVectorType>
 inline int elementwise_comp_op_16_loop(int window_start_x, int window_end_x, int window_step_x,
@@ -697,6 +772,114 @@ void elementwise_op_quantized(const ITensor *in1, const ITensor *in2, ITensor *o
     }
 }
 
+void elementwise_op_quantized_signed(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window,
+                                     int8_t (*scalar_func)(const float &, const float &, UniformQuantizationInfo),
+                                     int (*broadcast_func)(int, int, int, const int8_t *, float32x4x4_t, int8_t *, int32x4_t, float32x4_t,
+                                                           float32x4_t, float32x4_t, const bool),
+                                     int (*neon_func)(int, int, int, const int8_t *, const int8_t *, int8_t *,
+                                                      int32x4_t, int32x4_t, float32x4_t, float32x4_t,
+                                                      float32x4_t, float32x4_t))
+{
+    // Create input windows
+    Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape());
+    Window input2_win = window.broadcast_if_dimension_le_one(in2->info()->tensor_shape());
+
+    // Clear X Dimension on execution window as we handle manually
+    Window win = window;
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    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());
+    const bool is_broadcast_across_x = (input1_win.x().step() == 0) || (input2_win.x().step() == 0);
+
+    const UniformQuantizationInfo output_qinfo = out->info()->quantization_info().uniform();
+
+    // Output quantization info (add 0.5 to round toward the nearest integer - 0.5 rounds away from zero)
+    const float32x4_t voffseto   = vdupq_n_f32(output_qinfo.offset + 0.5f);
+    const float32x4_t invvscaleo = vdupq_n_f32(1.f / output_qinfo.scale);
+
+    if(is_broadcast_across_x)
+    {
+        // Select the broadcast input on the X axis
+        const bool     is_broadcast_input_2 = input2_win.x().step() == 0;
+        Window         broadcast_win        = is_broadcast_input_2 ? input2_win : input1_win;
+        Window         non_broadcast_win    = !is_broadcast_input_2 ? input2_win : input1_win;
+        const ITensor *broadcast_tensor     = is_broadcast_input_2 ? in2 : in1;
+        const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1;
+
+        const UniformQuantizationInfo broadcast_qinfo     = broadcast_tensor->info()->quantization_info().uniform();
+        const UniformQuantizationInfo non_broadcast_qinfo = non_broadcast_tensor->info()->quantization_info().uniform();
+
+        const int32x4_t   voffset_non_broadcast = vdupq_n_s32(non_broadcast_qinfo.offset);
+        const float32x4_t vscale_non_broadcast  = vdupq_n_f32(non_broadcast_qinfo.scale);
+
+        // Clear X Dimension on execution window as we handle manually
+        non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator broadcast_input(broadcast_tensor, broadcast_win);
+        Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
+        Iterator output(out, win);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            const auto non_broadcast_input_ptr = reinterpret_cast<const int8_t *>(non_broadcast_input.ptr());
+            const auto output_ptr              = reinterpret_cast<int8_t *>(output.ptr());
+
+            const int8_t        broadcast_value  = *reinterpret_cast<const int8_t *>(broadcast_input.ptr());
+            const float32x4x4_t broadcast_vector = dup_quantized(broadcast_value, broadcast_qinfo.offset, broadcast_qinfo.scale);
+
+            int x = (*broadcast_func)(window_start_x, window_end_x, window_step_x, non_broadcast_input_ptr, broadcast_vector, output_ptr,
+                                      voffset_non_broadcast, vscale_non_broadcast, voffseto, invvscaleo, !is_broadcast_input_2);
+            for(; x < window_end_x; ++x)
+            {
+                const float afs   = dequantize_qasymm8_signed(*(non_broadcast_input_ptr + x), non_broadcast_qinfo);
+                const float bfs   = dequantize_qasymm8_signed(broadcast_value, broadcast_qinfo);
+                *(output_ptr + x) = (*scalar_func)(!is_broadcast_input_2 ? bfs : afs, !is_broadcast_input_2 ? afs : bfs, output_qinfo);
+            }
+        },
+        broadcast_input, non_broadcast_input, output);
+    }
+    else
+    {
+        const UniformQuantizationInfo input1_qinfo = in1->info()->quantization_info().uniform();
+        const UniformQuantizationInfo input2_qinfo = in2->info()->quantization_info().uniform();
+
+        // Input1 quantization info
+        const int32x4_t   voffset1 = vdupq_n_s32(input1_qinfo.offset);
+        const float32x4_t vscale1  = vdupq_n_f32(input1_qinfo.scale);
+
+        // Input2 quantization info
+        const int32x4_t   voffset2 = vdupq_n_s32(input2_qinfo.offset);
+        const float32x4_t vscale2  = vdupq_n_f32(input2_qinfo.scale);
+
+        // Clear X Dimension on execution window as we handle manually
+        input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+        input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+        Iterator input1(in1, input1_win);
+        Iterator input2(in2, input2_win);
+        Iterator output(out, win);
+
+        execute_window_loop(win, [&](const Coordinates &)
+        {
+            const auto input1_ptr = reinterpret_cast<const int8_t *>(input1.ptr());
+            const auto input2_ptr = reinterpret_cast<const int8_t *>(input2.ptr());
+            const auto output_ptr = reinterpret_cast<int8_t *>(output.ptr());
+
+            int x = (*neon_func)(window_start_x, window_end_x, window_step_x, input1_ptr, input2_ptr, output_ptr, voffset1, voffset2,
+                                 vscale1, vscale2, voffseto, invvscaleo);
+            for(; x < window_end_x; ++x)
+            {
+                const float afs   = dequantize_qasymm8_signed(*(input1_ptr + x), input1_qinfo);
+                const float bfs   = dequantize_qasymm8_signed(*(input2_ptr + x), input2_qinfo);
+                *(output_ptr + x) = (*scalar_func)(afs, bfs, output_qinfo);
+            }
+        },
+        input1, input2, output);
+    }
+}
+
 template <ComparisonOperation op, typename InputScalarType, typename InputVectorType>
 void elementwise_comp_op_16(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
 {
@@ -733,6 +916,13 @@ void elementwise_arithm_op_quantized(const ITensor *in1, const ITensor *in2, ITe
                              &elementwise_arithm_op_quantized_broadcast_loop<op>,
                              &elementwise_arithm_op_quantized_loop<op>);
 }
+template <ArithmeticOperation op>
+void elementwise_arithm_op_quantized_signed(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
+{
+    elementwise_op_quantized_signed(in1, in2, out, window, &elementwise_arithm_op_quantized_signed_scalar<op>,
+                                    &elementwise_arithm_op_quantized_signed_broadcast_loop<op>,
+                                    &elementwise_arithm_op_quantized_singed_loop<op>);
+}
 
 template <ComparisonOperation op>
 void elementwise_comp_op_quantized(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
@@ -773,7 +963,8 @@ configure_arithm_func(const ITensor *input1, const ITensor *input2, ITensor *out
         { "op_F32_F32_F32", &elementwise_arithm_op<op, typename wrapper::traits::neon_vector<float, 4>> },
         { "op_S16_S16_S16", &elementwise_arithm_op<op, typename wrapper::traits::neon_vector<int16_t, 8>> },
         { "op_S32_S32_S32", &elementwise_arithm_op<op, typename wrapper::traits::neon_vector<int32_t, 4>> },
-        { "op_QASYMM8_QASYMM8_QASYMM8", &elementwise_arithm_op_quantized<op> }
+        { "op_QASYMM8_QASYMM8_QASYMM8", &elementwise_arithm_op_quantized<op> },
+        { "op_QASYMM8_SIGNED_QASYMM8_SIGNED_QASYMM8_SIGNED", &elementwise_arithm_op_quantized_signed<op> }
     };
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
     map_function["op_F16_F16_F16"] = &elementwise_arithm_op<op, typename wrapper::traits::neon_vector<float16_t, 8>>;
@@ -808,8 +999,8 @@ NEElementwiseOperationKernel::NEElementwiseOperationKernel()
 
 Status NEElementwiseOperationKernel::validate_arguments_common(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output)
 {
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input1, 1, DataType::QASYMM8, DataType::S16, DataType::F16, DataType::S32, DataType::F32);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input2, 1, DataType::QASYMM8, DataType::S16, DataType::F16, DataType::S32, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input1, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::S16, DataType::F16, DataType::S32, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input2, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::S16, DataType::F16, DataType::S32, DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(&input1);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&input1, &input2);