COMPMID-447: Support scaling factors different than 1 for QS8/QS16 NEPixelWiseMultiplication.

Change-Id: I6d90a18df861d53546bdca982192b4ffc0dbb3c2
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/80794
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Pablo Tello <pablo.tello@arm.com>

2 files changed, 101 insertions, 9 deletions

diff --git a/tests/validation/NEON/PixelWiseMultiplication.cpp b/tests/validation/NEON/PixelWiseMultiplication.cpp
index c46ad6a9d7..f80944821c 100644
--- a/tests/validation/NEON/PixelWiseMultiplication.cpp
+++ b/tests/validation/NEON/PixelWiseMultiplication.cpp
@@ -442,6 +442,36 @@ BOOST_DATA_TEST_CASE(RunSmall, SmallShapes() * DataType::QS8 *ConvertPolicies()
     validate(Accessor(dst), ref_dst);
 }
 
+BOOST_TEST_DECORATOR(*boost::unit_test::label("precommit"))
+BOOST_DATA_TEST_CASE(RunSmallScale255, SmallShapes() * DataType::QS8 * (1.f / 255.f) * ConvertPolicies() * RoundingPolicy::TO_NEAREST_UP * boost::unit_test::data::xrange(1, 7),
+                     shape, dt, scale, convert_policy, rounding_policy, fixed_point_position)
+{
+    // Compute function
+    Tensor dst = compute_pixel_wise_multiplication(shape, dt, dt, dt, scale, convert_policy, rounding_policy, fixed_point_position);
+
+    // Compute reference
+    RawTensor ref_dst = Reference::compute_reference_fixed_point_pixel_wise_multiplication(shape, dt, dt, dt, scale, fixed_point_position, convert_policy, rounding_policy);
+
+    // Validate output
+    validate(Accessor(dst), ref_dst);
+}
+
+BOOST_TEST_DECORATOR(*boost::unit_test::label("precommit"))
+BOOST_DATA_TEST_CASE(RunSmallScaleOther, SmallShapes() * DataType::QS8 *ConvertPolicies() * RoundingPolicy::TO_ZERO * boost::unit_test::data::xrange(1, 7),
+                     shape, dt, convert_policy, rounding_policy, fixed_point_position)
+{
+    const float scale = 1.f / static_cast<float>(1 << fixed_point_position);
+
+    // Compute function
+    Tensor dst = compute_pixel_wise_multiplication(shape, dt, dt, dt, scale, convert_policy, rounding_policy, fixed_point_position);
+
+    // Compute reference
+    RawTensor ref_dst = Reference::compute_reference_fixed_point_pixel_wise_multiplication(shape, dt, dt, dt, scale, fixed_point_position, convert_policy, rounding_policy);
+
+    // Validate output
+    validate(Accessor(dst), ref_dst, 1.f);
+}
+
 BOOST_TEST_DECORATOR(*boost::unit_test::label("nightly"))
 BOOST_DATA_TEST_CASE(RunLarge, LargeShapes() * DataType::QS8 *ConvertPolicies() * RoundingPolicy::TO_ZERO * boost::unit_test::data::xrange<int>(1, 7),
                      shape, dt, convert_policy, rounding_policy, fixed_point_position)
@@ -455,6 +485,36 @@ BOOST_DATA_TEST_CASE(RunLarge, LargeShapes() * DataType::QS8 *ConvertPolicies()
     // Validate output
     validate(Accessor(dst), ref_dst);
 }
+
+BOOST_TEST_DECORATOR(*boost::unit_test::label("nightly"))
+BOOST_DATA_TEST_CASE(RunLargeScale255, LargeShapes() * DataType::QS8 * (1.f / 255.f) * ConvertPolicies() * RoundingPolicy::TO_ZERO * boost::unit_test::data::xrange(1, 7),
+                     shape, dt, scale, convert_policy, rounding_policy, fixed_point_position)
+{
+    // Compute function
+    Tensor dst = compute_pixel_wise_multiplication(shape, dt, dt, dt, scale, convert_policy, rounding_policy, fixed_point_position);
+
+    // Compute reference
+    RawTensor ref_dst = Reference::compute_reference_fixed_point_pixel_wise_multiplication(shape, dt, dt, dt, scale, fixed_point_position, convert_policy, rounding_policy);
+
+    // Validate output
+    validate(Accessor(dst), ref_dst);
+}
+
+BOOST_TEST_DECORATOR(*boost::unit_test::label("nightly"))
+BOOST_DATA_TEST_CASE(RunLargeScaleOther, LargeShapes() * DataType::QS8 *ConvertPolicies() * RoundingPolicy::TO_ZERO * boost::unit_test::data::xrange(1, 7),
+                     shape, dt, convert_policy, rounding_policy, fixed_point_position)
+{
+    const float scale = 1.f / static_cast<float>(1 << fixed_point_position);
+
+    // Compute function
+    Tensor dst = compute_pixel_wise_multiplication(shape, dt, dt, dt, scale, convert_policy, rounding_policy, fixed_point_position);
+
+    // Compute reference
+    RawTensor ref_dst = Reference::compute_reference_fixed_point_pixel_wise_multiplication(shape, dt, dt, dt, scale, fixed_point_position, convert_policy, rounding_policy);
+
+    // Validate output
+    validate(Accessor(dst), ref_dst, 1.f);
+}
 BOOST_AUTO_TEST_SUITE_END()
 
 BOOST_AUTO_TEST_SUITE(QS16)
@@ -472,6 +532,36 @@ BOOST_DATA_TEST_CASE(RunSmall, SmallShapes() * DataType::QS16 *ConvertPolicies()
     validate(Accessor(dst), ref_dst);
 }
 
+BOOST_TEST_DECORATOR(*boost::unit_test::label("precommit"))
+BOOST_DATA_TEST_CASE(RunSmallScale255, SmallShapes() * DataType::QS16 * (1.f / 255.f) * ConvertPolicies() * RoundingPolicy::TO_NEAREST_UP * boost::unit_test::data::xrange(1, 15),
+                     shape, dt, scale, convert_policy, rounding_policy, fixed_point_position)
+{
+    // Compute function
+    Tensor dst = compute_pixel_wise_multiplication(shape, dt, dt, dt, scale, convert_policy, rounding_policy, fixed_point_position);
+
+    // Compute reference
+    RawTensor ref_dst = Reference::compute_reference_fixed_point_pixel_wise_multiplication(shape, dt, dt, dt, scale, fixed_point_position, convert_policy, rounding_policy);
+
+    // Validate output
+    validate(Accessor(dst), ref_dst);
+}
+
+BOOST_TEST_DECORATOR(*boost::unit_test::label("precommit"))
+BOOST_DATA_TEST_CASE(RunSmallScaleOther, SmallShapes() * DataType::QS16 *ConvertPolicies() * RoundingPolicy::TO_ZERO * boost::unit_test::data::xrange(1, 15),
+                     shape, dt, convert_policy, rounding_policy, fixed_point_position)
+{
+    const float scale = 1.f / static_cast<float>(1 << fixed_point_position);
+
+    // Compute function
+    Tensor dst = compute_pixel_wise_multiplication(shape, dt, dt, dt, scale, convert_policy, rounding_policy, fixed_point_position);
+
+    // Compute reference
+    RawTensor ref_dst = Reference::compute_reference_fixed_point_pixel_wise_multiplication(shape, dt, dt, dt, scale, fixed_point_position, convert_policy, rounding_policy);
+
+    // Validate output
+    validate(Accessor(dst), ref_dst, 1.f);
+}
+
 BOOST_TEST_DECORATOR(*boost::unit_test::label("nightly"))
 BOOST_DATA_TEST_CASE(RunLarge, LargeShapes() * DataType::QS16 *ConvertPolicies() * RoundingPolicy::TO_ZERO * boost::unit_test::data::xrange<int>(1, 15),
                      shape, dt, convert_policy, rounding_policy, fixed_point_position)
diff --git a/tests/validation/TensorOperations.h b/tests/validation/TensorOperations.h
index 4d067ac748..319047816c 100644
--- a/tests/validation/TensorOperations.h
+++ b/tests/validation/TensorOperations.h
@@ -866,7 +866,7 @@ void pixel_wise_multiplication(const Tensor<T1> &in1, const Tensor<T2> &in2, Ten
 
 // Fixed-point Pixel-wise Multiplication
 template <typename T, typename = typename std::enable_if<std::is_integral<T>::value>::type>
-void fixed_point_pixel_wise_multiplication(const Tensor<T> &in1, const Tensor<T> &in2, Tensor<T> &out, int scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy)
+void fixed_point_pixel_wise_multiplication(const Tensor<T> &in1, const Tensor<T> &in2, Tensor<T> &out, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy)
 {
     using namespace fixed_point_arithmetic;
 
@@ -881,18 +881,20 @@ void fixed_point_pixel_wise_multiplication(const Tensor<T> &in1, const Tensor<T>
     ARM_COMPUTE_ERROR_ON((in1.data_type() == DataType::QS8) && (fixed_point_position == 0 || fixed_point_position > 7));
     ARM_COMPUTE_ERROR_ON((in1.data_type() == DataType::QS16) && (fixed_point_position == 0 || fixed_point_position > 15));
 
-    fixed_point<T> fp_scale(scale, fixed_point_position);
-    const bool     is_sat     = convert_policy == ConvertPolicy::SATURATE;
-    const bool     do_scaling = scale != 1;
+    const fixed_point<T> fp_scale(scale, fixed_point_position);
+    const bool           is_sat = convert_policy == ConvertPolicy::SATURATE;
 
     for(int i = 0; i < in1.num_elements(); ++i)
     {
-        fixed_point<T> val1(in1[i], fixed_point_position, true);
-        fixed_point<T> val2(in2[i], fixed_point_position, true);
-        fixed_point<T> res = (is_sat) ? val1 * val2 : mul<OverflowPolicy::WRAP>(val1, val2);
-        if(do_scaling)
+        const fixed_point<T> val1(in1[i], fixed_point_position, true);
+        fixed_point<T>       res(in2[i], fixed_point_position, true);
+        if(is_sat)
         {
-            res = (is_sat) ? res * fp_scale : mul<OverflowPolicy::WRAP>(res, fp_scale);
+            res = mul(mul(res, val1), fp_scale);
+        }
+        else
+        {
+            res = mul<OverflowPolicy::WRAP>(mul<OverflowPolicy::WRAP>(res, val1), fp_scale);
         }
         out[i] = res.raw();
     }