COMPMID-1564: Add QASYMM8 on CLPixelwiseMultiplication

Change-Id: I5f719f5b2915c18cd0ca6271db401152112863a6
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/148982
Tested-by: bsgcomp <bsgcomp@arm.com>
Reviewed-by: Isabella Gottardi <isabella.gottardi@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Reviewed-by: Giuseppe Rossini <giuseppe.rossini@arm.com>

8 files changed, 211 insertions, 51 deletions

diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index a7a95b8b94..391a0bb6d5 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -332,6 +332,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "permute_3201", "permute.cl" },
     { "pixelwise_mul_float", "pixelwise_mul_float.cl" },
     { "pixelwise_mul_int", "pixelwise_mul_int.cl" },
+    { "pixelwise_mul_quantized", "pixelwise_mul_int.cl" },
     { "pooling_layer_2", "pooling_layer.cl" },
     { "pooling_layer_3", "pooling_layer.cl" },
     { "pooling_layer_optimized_3", "pooling_layer.cl" },
diff --git a/src/core/CL/cl_kernels/pixelwise_mul_float.cl b/src/core/CL/cl_kernels/pixelwise_mul_float.cl
index f4f36a02af..9fa540e946 100644
--- a/src/core/CL/cl_kernels/pixelwise_mul_float.cl
+++ b/src/core/CL/cl_kernels/pixelwise_mul_float.cl
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016, 2017 ARM Limited.
+ * Copyright (c) 2016-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -30,6 +30,7 @@
 #endif /* SATURATE */
 #define CONVERT_OP_FLOAT(x, type, round) CONVERT_OP_FLOAT_STR(x, type, round)
 
+#if defined(DATA_TYPE_IN1) && defined(DATA_TYPE_IN2) && defined(DATA_TYPE_RES) && defined(DATA_TYPE_OUT)
 /** Performs a pixelwise multiplication with float scale of either integer or float inputs.
  *
  * @attention The inputs and output data types need to be passed at compile time using -DDATA_TYPE_IN1, -DDATA_TYPE_IN2 and -DDATA_TYPE_OUT:
@@ -93,3 +94,4 @@ __kernel void pixelwise_mul_float(
     // Store result
     vstore16(res, 0, (__global DATA_TYPE_OUT *)out.ptr);
 }
+#endif /* defined(DATA_TYPE_IN1) && defined(DATA_TYPE_IN2) && defined(DATA_TYPE_RES) && defined(DATA_TYPE_OUT) */
\ No newline at end of file
diff --git a/src/core/CL/cl_kernels/pixelwise_mul_int.cl b/src/core/CL/cl_kernels/pixelwise_mul_int.cl
index c99a08a583..5b3acb7ae6 100644
--- a/src/core/CL/cl_kernels/pixelwise_mul_int.cl
+++ b/src/core/CL/cl_kernels/pixelwise_mul_int.cl
@@ -32,6 +32,7 @@
 
 #define MUL_OP(x, y, scale, type, size) CONVERT_OP_INT((x) * (y) >> scale, type, size)
 
+#if defined(DATA_TYPE_IN1) && defined(DATA_TYPE_IN2) && defined(DATA_TYPE_RES) && defined(DATA_TYPE_OUT)
 /** Performs a pixelwise multiplication with integer scale of integer inputs.
  *
  * @attention The inputs and output data types need to be passed at compile time using -DDATA_TYPE_IN1, -DDATA_TYPE_IN2 and -DDATA_TYPE_OUT:
@@ -85,3 +86,70 @@ __kernel void pixelwise_mul_int(
     // Perform multiplication and store result
     vstore16(MUL_OP(in1_data, in2_data, scale, DATA_TYPE_OUT, 16), 0, (__global DATA_TYPE_OUT *)out.ptr);
 }
+#endif /* defined(DATA_TYPE_IN1) && defined(DATA_TYPE_IN2) && defined(DATA_TYPE_RES) && defined(DATA_TYPE_OUT) */
+
+#if defined(OFFSET_IN1) && defined(OFFSET_IN2) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_IN2) && defined(SCALE_OUT)
+/** Performs a pixelwise multiplication with float scale of quantized inputs.
+ *
+ * @note The quantization offset of the first operand must be passed at compile time using -DOFFSET_IN1, e.g. -DOFFSET_IN1=10
+ * @note The quantization offset of the second operand must be passed at compile time using -DOFFSET_IN2, e.g. -DOFFSET_IN2=10
+ * @note The quantization offset of the output must be passed at compile time using -DOFFSET_OUT, e.g. -DOFFSET_OUT=10
+ * @note The quantization scale of the first operand must be passed at compile time using -DSCALE_IN1, e.g. -DSCALE_IN1=10
+ * @note The quantization scale of the second operand must be passed at compile time using -DSCALE_IN2, e.g. -DSCALE_IN2=10
+ * @note The quantization scale of the output must be passed at compile time using -DSCALE_OUT, e.g. -DSCALE_OUT=10
+ * @note To perform saturating operation -DSATURATE has to be passed to the compiler otherwise wrapping policy will be used.
+ *
+ * @param[in]  in1_ptr                           Pointer to the source image. Supported data types: U8, S16, F16, F32
+ * @param[in]  in1_stride_x                      Stride of the source image in X dimension (in bytes)
+ * @param[in]  in1_step_x                        in1_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  in1_stride_y                      Stride of the source image in Y dimension (in bytes)
+ * @param[in]  in1_step_y                        in1_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  in1_stride_z                      Stride of the source image in Y dimension (in bytes)
+ * @param[in]  in1_step_z                        in1_stride_z * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  in1_offset_first_element_in_bytes The offset of the first element in the source image
+ * @param[in]  in2_ptr                           Pointer to the source image. Supported data types: U8, S16, F16, F32
+ * @param[in]  in2_stride_x                      Stride of the source image in X dimension (in bytes)
+ * @param[in]  in2_step_x                        in2_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  in2_stride_y                      Stride of the source image in Y dimension (in bytes)
+ * @param[in]  in2_step_y                        in2_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  in2_stride_z                      Stride of the source image in Y dimension (in bytes)
+ * @param[in]  in2_step_z                        in2_stride_z * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  in2_offset_first_element_in_bytes The offset of the first element in the source image
+ * @param[out] out_ptr                           Pointer to the destination image. Supported data types: U8, S16, F16, F32
+ * @param[in]  out_stride_x                      Stride of the destination image in X dimension (in bytes)
+ * @param[in]  out_step_x                        out_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  out_stride_y                      Stride of the destination image in Y dimension (in bytes)
+ * @param[in]  out_step_y                        out_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  out_stride_z                      Stride of the destination image in Y dimension (in bytes)
+ * @param[in]  out_step_z                        out_stride_z * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  out_offset_first_element_in_bytes The offset of the first element in the destination image
+ * @param[in]  scale                             Float scaling factor. Supported data types: F32
+ */
+__kernel void pixelwise_mul_quantized(
+    TENSOR3D_DECLARATION(in1),
+    TENSOR3D_DECLARATION(in2),
+    TENSOR3D_DECLARATION(out),
+    const float scale)
+{
+    // Get pixels pointer
+    Tensor3D in1 = CONVERT_TO_TENSOR3D_STRUCT(in1);
+    Tensor3D in2 = CONVERT_TO_TENSOR3D_STRUCT(in2);
+    Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(out);
+
+    // Load data
+    int16 in_a = CONVERT(vload16(0, (__global uchar *)in1.ptr), int16);
+    int16 in_b = CONVERT(vload16(0, (__global uchar *)in2.ptr), int16);
+
+    // Dequantize
+    in_a -= (int16)(int)OFFSET_IN1;
+    in_b -= (int16)(int)OFFSET_IN2;
+    const float16 in1f32 = convert_float16(in_a) * (float16)(float)SCALE_IN1;
+    const float16 in2f32 = convert_float16(in_b) * (float16)(float)SCALE_IN2;
+
+    const float16 qresf32 = (in1f32 * in2f32 * scale) / ((float16)(float)SCALE_OUT) + ((float16)((float16)OFFSET_OUT));
+    const uchar16 res     = convert_uchar16_sat(convert_int16_rte(qresf32));
+
+    // Store result
+    vstore16(res, 0, (__global uchar *)out.ptr);
+}
+#endif /* defined(OFFSET_IN1) && defined(OFFSET_IN2) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_IN2) && defined(SCALE_OUT) */
\ No newline at end of file
diff --git a/src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp b/src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp
index 4ca2ef8aa3..286b94ebdc 100644
--- a/src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp
+++ b/src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp
@@ -51,9 +51,9 @@ Status validate_arguments(const ITensorInfo *input1, const ITensorInfo *input2,
     ARM_COMPUTE_UNUSED(rounding_policy);
 
     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input1);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::U8, DataType::QASYMM8, DataType::S16, DataType::F16, DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input2);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input2, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input2, 1, DataType::U8, DataType::QASYMM8, DataType::S16, DataType::F16, DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(scale < 0, "Scale cannot be negative.");
 
     const TensorShape &out_shape = TensorShape::broadcast_shape(input1->tensor_shape(), input2->tensor_shape());
@@ -64,7 +64,7 @@ Status validate_arguments(const ITensorInfo *input1, const ITensorInfo *input2,
     if(output->total_size() > 0)
     {
         ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(output);
-        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32);
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8, DataType::QASYMM8, DataType::S16, DataType::F16, DataType::F32);
         ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->data_type() == DataType::U8 && (input1->data_type() != DataType::U8 || input2->data_type() != DataType::U8),
                                         "Output can only be U8 if both inputs are U8");
         ARM_COMPUTE_RETURN_ERROR_ON_MSG(detail::have_different_dimensions(out_shape, output->tensor_shape(), 0), "Wrong shape for output");
@@ -168,27 +168,44 @@ void CLPixelWiseMultiplicationKernel::configure(const ICLTensor *input1, const I
         data_type = "DATA_TYPE_INT";
     }
 
+    const bool is_quantized = is_data_type_quantized_asymmetric(input1->info()->data_type());
+
     // Construct kernel name
     std::string kernel_name = "pixelwise_mul";
-    kernel_name += (scale_int >= 0) ? "_int" : "_float";
+    if(!is_data_type_quantized(output->info()->data_type()))
+    {
+        kernel_name += (scale_int >= 0) ? "_int" : "_float";
+    }
 
     // Set kernel build options
-    std::set<std::string> build_opts;
-    build_opts.emplace((overflow_policy == ConvertPolicy::WRAP || is_data_type_float(output->info()->data_type())) ? "-DWRAP" : "-DSATURATE");
-    build_opts.emplace((rounding_policy == RoundingPolicy::TO_ZERO) ? "-DROUND=_rtz" : "-DROUND=_rte");
-    build_opts.emplace("-DDATA_TYPE_IN1=" + get_cl_type_from_data_type(input1->info()->data_type()));
-    build_opts.emplace("-DDATA_TYPE_IN2=" + get_cl_type_from_data_type(input2->info()->data_type()));
-    build_opts.emplace("-DDATA_TYPE_OUT=" + get_cl_type_from_data_type(output->info()->data_type()));
-    build_opts.emplace("-DDATA_TYPE_RES=" + compute_type);
-    build_opts.emplace("-D" + data_type);
+    CLBuildOptions build_opts;
+    if(is_quantized)
+    {
+        build_opts.add_option("-DOFFSET_IN1=" + support::cpp11::to_string(input1->info()->quantization_info().offset));
+        build_opts.add_option("-DOFFSET_IN2=" + support::cpp11::to_string(input2->info()->quantization_info().offset));
+        build_opts.add_option("-DOFFSET_OUT=" + support::cpp11::to_string(output->info()->quantization_info().offset));
+        build_opts.add_option("-DSCALE_IN1=" + support::cpp11::to_string(input1->info()->quantization_info().scale));
+        build_opts.add_option("-DSCALE_IN2=" + support::cpp11::to_string(input2->info()->quantization_info().scale));
+        build_opts.add_option("-DSCALE_OUT=" + support::cpp11::to_string(output->info()->quantization_info().scale));
+        kernel_name += "_quantized";
+    }
+    else
+    {
+        build_opts.add_option_if_else(overflow_policy == ConvertPolicy::WRAP || is_data_type_float(output->info()->data_type()), "-DWRAP", "-DSATURATE");
+        build_opts.add_option_if_else(rounding_policy == RoundingPolicy::TO_ZERO, "-DROUND=_rtz", "-DROUND=_rte");
+        build_opts.add_option("-DDATA_TYPE_IN1=" + get_cl_type_from_data_type(input1->info()->data_type()));
+        build_opts.add_option("-DDATA_TYPE_IN2=" + get_cl_type_from_data_type(input2->info()->data_type()));
+        build_opts.add_option("-DDATA_TYPE_OUT=" + get_cl_type_from_data_type(output->info()->data_type()));
+        build_opts.add_option("-DDATA_TYPE_RES=" + compute_type);
+    }
 
     // Create kernel
-    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts));
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
 
     // Set scale argument
     unsigned int idx = 3 * num_arguments_per_3D_tensor(); //Skip the inputs and output parameters
 
-    if(scale_int >= 0)
+    if(scale_int >= 0 && !is_quantized)
     {
         _kernel.setArg(idx++, scale_int);
     }
diff --git a/tests/validation/CL/PixelWiseMultiplication.cpp b/tests/validation/CL/PixelWiseMultiplication.cpp
index 8aa19dc105..b61ec394ea 100644
--- a/tests/validation/CL/PixelWiseMultiplication.cpp
+++ b/tests/validation/CL/PixelWiseMultiplication.cpp
@@ -38,8 +38,11 @@ namespace validation
 {
 namespace
 {
-const float scale_255 = 1.f / 255.f;
-
+namespace
+{
+const float                        scale_255 = 1.f / 255.f;
+constexpr AbsoluteTolerance<float> tolerance_qasymm8(1); /**< Tolerance value for comparing reference's output against implementation's output for quantized data types */
+} //namespace
 // *INDENT-OFF*
 // clang-format off
 #define VALIDATE(TYPE, TOLERANCE) validate(CLAccessor(_target), _reference, AbsoluteTolerance<TYPE>(TOLERANCE), 0.f);
@@ -105,31 +108,41 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(
 // *INDENT-ON*
 
 TEST_SUITE(F16toF16)
-
 TEST_SUITE(Scale255)
 PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE(RunSmall, ToF16Fixture<half_float::half>, PRECOMMIT, SmallShapes(), F16, F16, scale_255, TO_NEAREST_UP, VALIDATE(float, 1.f))
 TEST_SUITE_END() // Scale255
-
 TEST_SUITE_END() // F16toF16
 
 TEST_SUITE(F32toF32)
-
 TEST_SUITE(Scale255)
 PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE(RunSmall, ToF32Fixture<float>, PRECOMMIT, SmallShapes(), F32, F32, scale_255, TO_NEAREST_UP, VALIDATE(float, 1.f))
 TEST_SUITE_END() // Scale255
-
 TEST_SUITE_END() // F32toF32
 
-TEST_SUITE_END() // PixelWiseMultiplication
-
-TEST_SUITE(FixedPointPixelWiseMultiplication)
+PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, BroadcastFixture<float>, PRECOMMIT, SmallShapesBroadcast(), F32, F32, scale_255, TO_NEAREST_UP, VALIDATE(float, 1.f))
 
-TEST_SUITE(Broadcast)
-PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE(RunSmall, BroadcastFixture<float>, PRECOMMIT, SmallShapesBroadcast(), F32, F32, scale_255, TO_NEAREST_UP, VALIDATE(float, 1.f))
-TEST_SUITE_END() // Broadcast
+template <typename T>
+using CLPixelWiseMultiplicationQuantizedFixture = PixelWiseMultiplicationValidationQuantizedFixture<CLTensor, CLAccessor, CLPixelWiseMultiplication, T, T>;
+
+TEST_SUITE(Quantized)
+TEST_SUITE(QASYMM8)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLPixelWiseMultiplicationQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(combine(combine(datasets::SmallShapes(),
+                       framework::dataset::make("DataType", DataType::QASYMM8)),
+                       framework::dataset::make("Scale", { 1.f, 2.f })),
+                       framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })),
+                       framework::dataset::make("RoundingPolicy", RoundingPolicy::TO_NEAREST_EVEN)),
+                       framework::dataset::make("QuantizationInfo", { QuantizationInfo(5.f / 255.f, 20) })),
+                       framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 10) })),
+                       framework::dataset::make("QuantizationInfo", { QuantizationInfo(1.f / 255.f, 5) })))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_qasymm8);
+}
+TEST_SUITE_END() // QASYMM8
+TEST_SUITE_END() // Quantized
 
-TEST_SUITE_END() // FixedPointPixelWiseMultiplication
-TEST_SUITE_END()
+TEST_SUITE_END() // PixelWiseMultiplication
+TEST_SUITE_END() // CL
 } // namespace validation
 } // namespace test
 } // namespace arm_compute
diff --git a/tests/validation/fixtures/PixelWiseMultiplicationFixture.h b/tests/validation/fixtures/PixelWiseMultiplicationFixture.h
index b9f19f3e77..9927b75032 100644
--- a/tests/validation/fixtures/PixelWiseMultiplicationFixture.h
+++ b/tests/validation/fixtures/PixelWiseMultiplicationFixture.h
@@ -40,7 +40,7 @@ namespace test
 namespace validation
 {
 template <typename TensorType, typename AccessorType, typename FunctionType, typename T1, typename T2>
-class PixelWiseMultiplicationBroadcastValidationFixture : public framework::Fixture
+class PixelWiseMultiplicationGenericValidationFixture : public framework::Fixture
 {
 public:
     template <typename...>
@@ -50,10 +50,13 @@ public:
                DataType           dt_in2,
                float              scale,
                ConvertPolicy      convert_policy,
-               RoundingPolicy     rounding_policy)
+               RoundingPolicy     rounding_policy,
+               QuantizationInfo   qinfo0,
+               QuantizationInfo   qinfo1,
+               QuantizationInfo   qinfo_out)
     {
-        _target    = compute_target(shape0, shape1, dt_in1, dt_in2, scale, convert_policy, rounding_policy);
-        _reference = compute_reference(shape0, shape1, dt_in1, dt_in2, scale, convert_policy, rounding_policy);
+        _target    = compute_target(shape0, shape1, dt_in1, dt_in2, scale, convert_policy, rounding_policy, qinfo0, qinfo1, qinfo_out);
+        _reference = compute_reference(shape0, shape1, dt_in1, dt_in2, scale, convert_policy, rounding_policy, qinfo0, qinfo1, qinfo_out);
     }
 
 protected:
@@ -64,12 +67,13 @@ protected:
     }
 
     TensorType compute_target(const TensorShape &shape0, const TensorShape &shape1, DataType dt_in1, DataType dt_in2,
-                              float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy)
+                              float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy,
+                              QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out)
     {
         // Create tensors
-        TensorType src1 = create_tensor<TensorType>(shape0, dt_in1);
-        TensorType src2 = create_tensor<TensorType>(shape1, dt_in2);
-        TensorType dst  = create_tensor<TensorType>(TensorShape::broadcast_shape(shape0, shape1), dt_in2);
+        TensorType src1 = create_tensor<TensorType>(shape0, dt_in1, 1, qinfo0);
+        TensorType src2 = create_tensor<TensorType>(shape1, dt_in2, 1, qinfo1);
+        TensorType dst  = create_tensor<TensorType>(TensorShape::broadcast_shape(shape0, shape1), dt_in2, 1, qinfo_out);
 
         // Create and configure function
         FunctionType multiply;
@@ -99,17 +103,18 @@ protected:
     }
 
     SimpleTensor<T2> compute_reference(const TensorShape &shape0, const TensorShape &shape1, DataType dt_in1, DataType dt_in2,
-                                       float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy)
+                                       float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy,
+                                       QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out)
     {
         // Create reference
-        SimpleTensor<T1> src1{ shape0, dt_in1 };
-        SimpleTensor<T2> src2{ shape1, dt_in2 };
+        SimpleTensor<T1> src1{ shape0, dt_in1, 1, qinfo0 };
+        SimpleTensor<T2> src2{ shape1, dt_in2, 1, qinfo1 };
 
         // Fill reference
         fill(src1, 0);
         fill(src2, 1);
 
-        return reference::pixel_wise_multiplication<T1, T2>(src1, src2, scale, convert_policy, rounding_policy);
+        return reference::pixel_wise_multiplication<T1, T2>(src1, src2, scale, convert_policy, rounding_policy, qinfo_out);
     }
 
     TensorType       _target{};
@@ -117,16 +122,41 @@ protected:
 };
 
 template <typename TensorType, typename AccessorType, typename FunctionType, typename T1, typename T2>
-class PixelWiseMultiplicationValidationFixture : public PixelWiseMultiplicationBroadcastValidationFixture<TensorType, AccessorType, FunctionType, T1, T2>
+class PixelWiseMultiplicationValidationFixture : public PixelWiseMultiplicationGenericValidationFixture<TensorType, AccessorType, FunctionType, T1, T2>
 {
 public:
     template <typename...>
     void setup(const TensorShape &shape, DataType dt_in1, DataType dt_in2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy)
     {
-        PixelWiseMultiplicationBroadcastValidationFixture<TensorType, AccessorType, FunctionType, T1, T2>::setup(shape, shape, dt_in1, dt_in2, scale, convert_policy, rounding_policy);
+        PixelWiseMultiplicationGenericValidationFixture<TensorType, AccessorType, FunctionType, T1, T2>::setup(shape, shape, dt_in1, dt_in2, scale, convert_policy, rounding_policy,
+                                                                                                               QuantizationInfo(), QuantizationInfo(), QuantizationInfo());
     }
 };
 
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T1, typename T2>
+class PixelWiseMultiplicationBroadcastValidationFixture : public PixelWiseMultiplicationGenericValidationFixture<TensorType, AccessorType, FunctionType, T1, T2>
+{
+public:
+    template <typename...>
+    void setup(const TensorShape &shape0, const TensorShape &shape1, DataType dt_in1, DataType dt_in2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy)
+    {
+        PixelWiseMultiplicationGenericValidationFixture<TensorType, AccessorType, FunctionType, T1, T2>::setup(shape0, shape1, dt_in1, dt_in2, scale, convert_policy, rounding_policy,
+                                                                                                               QuantizationInfo(), QuantizationInfo(), QuantizationInfo());
+    }
+};
+
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T1, typename T2>
+class PixelWiseMultiplicationValidationQuantizedFixture : public PixelWiseMultiplicationGenericValidationFixture<TensorType, AccessorType, FunctionType, T1, T2>
+{
+public:
+    template <typename...>
+    void setup(const TensorShape &shape, DataType dt, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy,
+               QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out)
+    {
+        PixelWiseMultiplicationGenericValidationFixture<TensorType, AccessorType, FunctionType, T1, T2>::setup(shape, shape, dt, dt, scale, convert_policy, rounding_policy,
+                                                                                                               qinfo0, qinfo1, qinfo_out);
+    }
+};
 } // namespace validation
 } // namespace test
 } // namespace arm_compute
diff --git a/tests/validation/reference/PixelWiseMultiplication.cpp b/tests/validation/reference/PixelWiseMultiplication.cpp
index 859da5ce59..d86f8aae0d 100644
--- a/tests/validation/reference/PixelWiseMultiplication.cpp
+++ b/tests/validation/reference/PixelWiseMultiplication.cpp
@@ -23,6 +23,8 @@
  */
 #include "PixelWiseMultiplication.h"
 
+#include "tests/validation/Helpers.h"
+
 namespace arm_compute
 {
 namespace test
@@ -126,8 +128,10 @@ struct BroadcastUnroll<0>
 } // namespace
 
 template <typename T1, typename T2>
-SimpleTensor<T2> pixel_wise_multiplication(const SimpleTensor<T1> &src1, const SimpleTensor<T2> &src2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy)
+SimpleTensor<T2> pixel_wise_multiplication(const SimpleTensor<T1> &src1, const SimpleTensor<T2> &src2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy, QuantizationInfo qout)
 {
+    ARM_COMPUTE_UNUSED(qout);
+
     SimpleTensor<T2> dst(TensorShape::broadcast_shape(src1.shape(), src2.shape()), src2.data_type());
 
     if(scale < 0)
@@ -142,13 +146,37 @@ SimpleTensor<T2> pixel_wise_multiplication(const SimpleTensor<T1> &src1, const S
     return dst;
 }
 
+template <>
+SimpleTensor<uint8_t> pixel_wise_multiplication(const SimpleTensor<uint8_t> &src1, const SimpleTensor<uint8_t> &src2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy,
+                                                QuantizationInfo qout)
+{
+    SimpleTensor<uint8_t> dst(TensorShape::broadcast_shape(src1.shape(), src2.shape()), src2.data_type(), 1, qout);
+
+    if(src1.data_type() == DataType::QASYMM8 && src2.data_type() == DataType::QASYMM8)
+    {
+        SimpleTensor<float> src1_tmp = convert_from_asymmetric(src1);
+        SimpleTensor<float> src2_tmp = convert_from_asymmetric(src2);
+        SimpleTensor<float> dst_tmp  = pixel_wise_multiplication<float>(src1_tmp, src2_tmp, scale, convert_policy, rounding_policy, qout);
+        dst                          = convert_to_asymmetric(dst_tmp, qout);
+    }
+    else
+    {
+        if(scale < 0)
+        {
+            ARM_COMPUTE_ERROR("Scale of pixel-wise multiplication must be non-negative");
+        }
+
+        Coordinates id_src1, id_src2, id_dst;
+        BroadcastUnroll<Coordinates::num_max_dimensions>::unroll(src1, src2, dst, scale, convert_policy, rounding_policy, id_src1, id_src2, id_dst);
+    }
+    return dst;
+}
 // *INDENT-OFF*
 // clang-format off
-template SimpleTensor<uint8_t> pixel_wise_multiplication(const SimpleTensor<uint8_t> &src1, const SimpleTensor<uint8_t> &src2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy);
-template SimpleTensor<int16_t> pixel_wise_multiplication(const SimpleTensor<uint8_t> &src1, const SimpleTensor<int16_t> &src2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy);
-template SimpleTensor<int16_t> pixel_wise_multiplication(const SimpleTensor<int16_t> &src1, const SimpleTensor<int16_t> &src2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy);
-template SimpleTensor<float> pixel_wise_multiplication(const SimpleTensor<float> &src1, const SimpleTensor<float> &src2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy);
-template SimpleTensor<half_float::half> pixel_wise_multiplication(const SimpleTensor<half_float::half> &src1, const SimpleTensor<half_float::half> &src2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy);
+template SimpleTensor<int16_t> pixel_wise_multiplication(const SimpleTensor<uint8_t> &src1, const SimpleTensor<int16_t> &src2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy, QuantizationInfo qout);
+template SimpleTensor<int16_t> pixel_wise_multiplication(const SimpleTensor<int16_t> &src1, const SimpleTensor<int16_t> &src2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy, QuantizationInfo qout);
+template SimpleTensor<float> pixel_wise_multiplication(const SimpleTensor<float> &src1, const SimpleTensor<float> &src2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy, QuantizationInfo qout);
+template SimpleTensor<half_float::half> pixel_wise_multiplication(const SimpleTensor<half_float::half> &src1, const SimpleTensor<half_float::half> &src2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy, QuantizationInfo qout);
 // clang-format on
 // *INDENT-ON*
 } // namespace reference
diff --git a/tests/validation/reference/PixelWiseMultiplication.h b/tests/validation/reference/PixelWiseMultiplication.h
index 1dce154719..787a7b23e2 100644
--- a/tests/validation/reference/PixelWiseMultiplication.h
+++ b/tests/validation/reference/PixelWiseMultiplication.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -35,7 +35,8 @@ namespace validation
 namespace reference
 {
 template <typename T1, typename T2>
-SimpleTensor<T2> pixel_wise_multiplication(const SimpleTensor<T1> &src1, const SimpleTensor<T2> &src2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy);
+SimpleTensor<T2> pixel_wise_multiplication(const SimpleTensor<T1> &src1, const SimpleTensor<T2> &src2, float scale,
+                                           ConvertPolicy convert_policy, RoundingPolicy rounding_policy, QuantizationInfo qout = QuantizationInfo());
 } // namespace reference
 } // namespace validation
 } // namespace test