COMPMID-945: Fix GEMM CL FP32 mismatch - V2

The approach used in e415fc13 was not correct. Switch to allowing to use an
absolute tolerance if the relative tolerance comparison fails.

Change-Id: I8d94d2f8edd3e0eb7388d3d8ac3ebfc37790e267
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/122269
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>

2 files changed, 28 insertions, 9 deletions

diff --git a/tests/validation/CL/GEMM.cpp b/tests/validation/CL/GEMM.cpp
index 6539090fcf..217edf4438 100644
--- a/tests/validation/CL/GEMM.cpp
+++ b/tests/validation/CL/GEMM.cpp
@@ -49,7 +49,9 @@ namespace validation
 {
 namespace
 {
-RelativeTolerance<float>            tolerance_f32(0.001f);    /**< Tolerance value for comparing reference's output against implementation's output for floating point data types */
+RelativeTolerance<float> tolerance_f32(0.001f); /**< Tolerance value for comparing reference's output against implementation's output for floating point data types */
+constexpr float          abs_tolerance_f32(
+    0.0001f);                                                 /**< Absolute tolerance value for comparing reference's output against implementation's output for floating point data types in case using relative tolerance fails because of small values */
 RelativeTolerance<half_float::half> tolerance_f16(half(0.2)); /**< Tolerance value for comparing reference's output against implementation's output for floating point data types */
 constexpr AbsoluteTolerance<float>  tolerance_q(1.0f);        /**< Tolerance value for comparing reference's output against implementation's output for fixed point data types */
 constexpr float                     tolerance_num   = 0.02f;  /**< Tolerance number */
@@ -200,7 +202,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMFixture<float>, framework::DatasetMode::P
 FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMFixture<float>, framework::DatasetMode::NIGHTLY, combine(datasets::LargeGEMMDataset(), framework::dataset::make("DataType", DataType::F32)))
 {
     // Validate output
-    validate(CLAccessor(_target), _reference, tolerance_f32);
+    validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32);
 }
 TEST_SUITE_END()
 TEST_SUITE_END()
diff --git a/tests/validation/Validation.h b/tests/validation/Validation.h
index 651b7e5b3c..506318018d 100644
--- a/tests/validation/Validation.h
+++ b/tests/validation/Validation.h
@@ -180,7 +180,7 @@ void validate(const arm_compute::PaddingSize &padding, const arm_compute::Paddin
  * other test cases.
  */
 template <typename T, typename U = AbsoluteTolerance<T>>
-void validate(const IAccessor &tensor, const SimpleTensor<T> &reference, U tolerance_value = U(), float tolerance_number = 0.f);
+void validate(const IAccessor &tensor, const SimpleTensor<T> &reference, U tolerance_value = U(), float tolerance_number = 0.f, float absolute_tolerance_value = 0.f);
 
 /** Validate tensors with valid region.
  *
@@ -193,7 +193,7 @@ void validate(const IAccessor &tensor, const SimpleTensor<T> &reference, U toler
  * other test cases.
  */
 template <typename T, typename U = AbsoluteTolerance<T>>
-void validate(const IAccessor &tensor, const SimpleTensor<T> &reference, const ValidRegion &valid_region, U tolerance_value = U(), float tolerance_number = 0.f);
+void validate(const IAccessor &tensor, const SimpleTensor<T> &reference, const ValidRegion &valid_region, U tolerance_value = U(), float tolerance_number = 0.f, float absolute_tolerance_value = 0.f);
 
 /** Validate tensors with valid mask.
  *
@@ -206,7 +206,8 @@ void validate(const IAccessor &tensor, const SimpleTensor<T> &reference, const V
  * other test cases.
  */
 template <typename T, typename U = AbsoluteTolerance<T>>
-void validate(const IAccessor &tensor, const SimpleTensor<T> &reference, const SimpleTensor<T> &valid_mask, U tolerance_value = U(), float tolerance_number = 0.f);
+void validate(const IAccessor &tensor, const SimpleTensor<T> &reference, const SimpleTensor<T> &valid_mask, U tolerance_value = U(), float tolerance_number = 0.f,
+              float absolute_tolerance_value = 0.f);
 
 /** Validate tensors against constant value.
  *
@@ -317,10 +318,10 @@ struct compare<RelativeTolerance<U>> : public compare_base<RelativeTolerance<U>>
 };
 
 template <typename T, typename U>
-void validate(const IAccessor &tensor, const SimpleTensor<T> &reference, U tolerance_value, float tolerance_number)
+void validate(const IAccessor &tensor, const SimpleTensor<T> &reference, U tolerance_value, float tolerance_number, float absolute_tolerance_value)
 {
     // Validate with valid region covering the entire shape
-    validate(tensor, reference, shape_to_valid_region(tensor.shape()), tolerance_value, tolerance_number);
+    validate(tensor, reference, shape_to_valid_region(tensor.shape()), tolerance_value, tolerance_number, absolute_tolerance_value);
 }
 
 template <typename T, typename U, typename = typename std::enable_if<std::is_integral<T>::value>::type>
@@ -331,7 +332,7 @@ void validate_wrap(const IAccessor &tensor, const SimpleTensor<T> &reference, U
 }
 
 template <typename T, typename U>
-void validate(const IAccessor &tensor, const SimpleTensor<T> &reference, const ValidRegion &valid_region, U tolerance_value, float tolerance_number)
+void validate(const IAccessor &tensor, const SimpleTensor<T> &reference, const ValidRegion &valid_region, U tolerance_value, float tolerance_number, float absolute_tolerance_value)
 {
     int64_t num_mismatches = 0;
     int64_t num_elements   = 0;
@@ -365,6 +366,14 @@ void validate(const IAccessor &tensor, const SimpleTensor<T> &reference, const V
 
                 if(!compare<U>(target_value, reference_value, tolerance_value))
                 {
+                    if(absolute_tolerance_value != 0.f)
+                    {
+                        const AbsoluteTolerance<float> abs_tolerance(absolute_tolerance_value);
+                        if(compare<AbsoluteTolerance<float>>(target_value, reference_value, abs_tolerance))
+                        {
+                            continue;
+                        }
+                    }
                     ARM_COMPUTE_TEST_INFO("id = " << id);
                     ARM_COMPUTE_TEST_INFO("channel = " << c);
                     ARM_COMPUTE_TEST_INFO("target = " << std::setprecision(5) << framework::make_printable(target_value));
@@ -474,7 +483,7 @@ void validate_wrap(const IAccessor &tensor, const SimpleTensor<T> &reference, co
 }
 
 template <typename T, typename U>
-void validate(const IAccessor &tensor, const SimpleTensor<T> &reference, const SimpleTensor<T> &valid_mask, U tolerance_value, float tolerance_number)
+void validate(const IAccessor &tensor, const SimpleTensor<T> &reference, const SimpleTensor<T> &valid_mask, U tolerance_value, float tolerance_number, float absolute_tolerance_value)
 {
     int64_t num_mismatches = 0;
     int64_t num_elements   = 0;
@@ -508,6 +517,14 @@ void validate(const IAccessor &tensor, const SimpleTensor<T> &reference, const S
 
                 if(!compare<U>(target_value, reference_value, tolerance_value))
                 {
+                    if(absolute_tolerance_value != 0.f)
+                    {
+                        const AbsoluteTolerance<float> abs_tolerance(absolute_tolerance_value);
+                        if(compare<AbsoluteTolerance<float>>(target_value, reference_value, abs_tolerance))
+                        {
+                            continue;
+                        }
+                    }
                     ARM_COMPUTE_TEST_INFO("id = " << id);
                     ARM_COMPUTE_TEST_INFO("channel = " << c);
                     ARM_COMPUTE_TEST_INFO("target = " << std::setprecision(5) << framework::make_printable(target_value));