Enable quantized data types for CpuElementwiseUnary on Armv7a
- Adding fallback functions neon_qasymm8_signed_elementwise_unary() and
neon_qasymm8_elementwise_unary()
- They would be called in case target is not aarch64
Resolves: COMPMID-5994
Change-Id: Id0db1e7cb0fe92f1eaef0b3a9ed2bea01b3f2a15
Signed-off-by: Ramy Elgammal <ramy.elgammal@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9416
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
diff --git a/tests/validation/NEON/ElementwiseAbsoluteValue.cpp b/tests/validation/NEON/ElementwiseAbsoluteValue.cpp
index 7f6a6a5..0667ac7 100644
--- a/tests/validation/NEON/ElementwiseAbsoluteValue.cpp
+++ b/tests/validation/NEON/ElementwiseAbsoluteValue.cpp
@@ -46,8 +46,13 @@
#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
RelativeTolerance<float> tolerance_fp16(0.01f);
#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+#if defined(__aarch64__)
constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(0);
constexpr AbsoluteTolerance<int8_t> tolerance_qasymm8_signed(0);
+#else // #if !defined(__aarch64__)
+constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(1); // There is difference of 1, because quantizing in reference uses round policy "TO_NEAREST_UP", where the armv7a neon kernel uses "TO_ZERO"
+constexpr AbsoluteTolerance<int8_t> tolerance_qasymm8_signed(1);
+#endif // #if !defined(__aarch64__)
} // namespace
TEST_SUITE(NEON)
@@ -136,8 +141,8 @@
validate(Accessor(_target), _reference, tolerance_qasymm8_signed);
}
TEST_SUITE_END() // QASYMM8_SIGNED
-
TEST_SUITE_END() // Quantized
+
TEST_SUITE_END() // AbsLayer
TEST_SUITE_END() // Neon
} // namespace validation
diff --git a/tests/validation/NEON/ElementwiseExpLayer.cpp b/tests/validation/NEON/ElementwiseExpLayer.cpp
index e8940c5..31cd786 100644
--- a/tests/validation/NEON/ElementwiseExpLayer.cpp
+++ b/tests/validation/NEON/ElementwiseExpLayer.cpp
@@ -46,8 +46,15 @@
#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
RelativeTolerance<float> tolerance_fp16(0.01f);
#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+
+#if defined(__aarch64__)
constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(0);
constexpr AbsoluteTolerance<int8_t> tolerance_qasymm8_signed(0);
+#else // #if !defined(__aarch64__)
+constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(1); // There is difference of 1, because quantizing in reference uses round policy "TO_NEAREST_UP", where the armv7a neon kernel uses "TO_ZERO"
+constexpr AbsoluteTolerance<int8_t> tolerance_qasymm8_signed(1);
+#endif // #if !defined(__aarch64__)
+
} // namespace
TEST_SUITE(NEON)
TEST_SUITE(ExpLayer)
@@ -111,7 +118,6 @@
validate(Accessor(_target), _reference, tolerance_qasymm8_signed);
}
TEST_SUITE_END() // QASYMM8_SIGNED
-
TEST_SUITE_END() // Quantized
TEST_SUITE_END() // ExpLayer
diff --git a/tests/validation/NEON/ElementwiseLog.cpp b/tests/validation/NEON/ElementwiseLog.cpp
index 49a88ce..320ece2 100644
--- a/tests/validation/NEON/ElementwiseLog.cpp
+++ b/tests/validation/NEON/ElementwiseLog.cpp
@@ -46,8 +46,15 @@
#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
RelativeTolerance<float> tolerance_fp16(0.01f);
#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+
+#if defined(__aarch64__)
constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(0);
constexpr AbsoluteTolerance<int8_t> tolerance_qasymm8_signed(0);
+#else // #if !defined(__aarch64__)
+constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(1); // There is difference of 1, because quantizing in reference uses round policy "TO_NEAREST_UP", where the armv7a neon kernel uses "TO_ZERO"
+constexpr AbsoluteTolerance<int8_t> tolerance_qasymm8_signed(1);
+#endif // #if !defined(__aarch64__)
+
} // namespace
TEST_SUITE(NEON)
TEST_SUITE(LogLayer)
@@ -118,8 +125,8 @@
validate(Accessor(_target), _reference, tolerance_qasymm8_signed);
}
TEST_SUITE_END() // QASYMM8_SIGNED
-
TEST_SUITE_END() // Quantized
+
TEST_SUITE_END() // LogLayer
TEST_SUITE_END() // Neon
} // namespace validation
diff --git a/tests/validation/NEON/ElementwiseNegation.cpp b/tests/validation/NEON/ElementwiseNegation.cpp
index 038058c..5b8ae8f 100644
--- a/tests/validation/NEON/ElementwiseNegation.cpp
+++ b/tests/validation/NEON/ElementwiseNegation.cpp
@@ -46,8 +46,13 @@
#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
RelativeTolerance<float> tolerance_fp16(0.01f);
#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+#if defined(__aarch64__)
constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(0);
constexpr AbsoluteTolerance<int8_t> tolerance_qasymm8_signed(0);
+#else // #if !defined(__aarch64__)
+constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(1); // There is difference of 1, because quantizing in reference uses round policy "TO_NEAREST_UP", where the armv7a neon kernel uses "TO_ZERO"
+constexpr AbsoluteTolerance<int8_t> tolerance_qasymm8_signed(1);
+#endif // #if !defined(__aarch64__)
} // namespace
TEST_SUITE(NEON)
TEST_SUITE(NegLayer)
@@ -142,8 +147,8 @@
validate(Accessor(_target), _reference, tolerance_qasymm8_signed);
}
TEST_SUITE_END() // QASYMM8_SIGNED
-
TEST_SUITE_END() // Quantized
+
TEST_SUITE_END() // NegLayer
TEST_SUITE_END() // Neon
} // namespace validation
diff --git a/tests/validation/NEON/ElementwiseRound.cpp b/tests/validation/NEON/ElementwiseRound.cpp
index a6ff47c..620618c 100644
--- a/tests/validation/NEON/ElementwiseRound.cpp
+++ b/tests/validation/NEON/ElementwiseRound.cpp
@@ -114,8 +114,8 @@
validate(Accessor(_target), _reference, tolerance_qasymm8_signed);
}
TEST_SUITE_END() // QASYMM8_SIGNED
-
TEST_SUITE_END() // Quantized
+
TEST_SUITE_END() // RoundLayer
TEST_SUITE_END() // Neon
} // namespace validation
diff --git a/tests/validation/NEON/ElementwiseRsqrtLayer.cpp b/tests/validation/NEON/ElementwiseRsqrtLayer.cpp
index 1d291ac..80788c8 100644
--- a/tests/validation/NEON/ElementwiseRsqrtLayer.cpp
+++ b/tests/validation/NEON/ElementwiseRsqrtLayer.cpp
@@ -46,8 +46,13 @@
#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
RelativeTolerance<float> tolerance_fp16(0.01f);
#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+#if defined(__aarch64__)
constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(0);
constexpr AbsoluteTolerance<int8_t> tolerance_qasymm8_signed(0);
+#else // #if !defined(__aarch64__)
+constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(1); // There is difference of 1, because quantizing in reference uses round policy "TO_NEAREST_UP", where the armv7a neon kernel uses "TO_ZERO"
+constexpr AbsoluteTolerance<int8_t> tolerance_qasymm8_signed(1);
+#endif // #if !defined(__aarch64__)
} // namespace
TEST_SUITE(NEON)
TEST_SUITE(RsqrtLayer)
@@ -131,8 +136,8 @@
validate(Accessor(_target), _reference, tolerance_qasymm8_signed);
}
TEST_SUITE_END() // QASYMM8_SIGNED
-
TEST_SUITE_END() // Quantized
+
TEST_SUITE_END() // RsqrtLayer
TEST_SUITE_END() // Neon
} // namespace validation
diff --git a/tests/validation/NEON/ElementwiseSin.cpp b/tests/validation/NEON/ElementwiseSin.cpp
index 76f4c50..9c2d7ae 100644
--- a/tests/validation/NEON/ElementwiseSin.cpp
+++ b/tests/validation/NEON/ElementwiseSin.cpp
@@ -120,7 +120,6 @@
TEST_SUITE_END() // QASYMM8_SIGNED
TEST_SUITE_END() // Quantized
-
TEST_SUITE_END() // SinLayer
TEST_SUITE_END() // Neon
} // namespace validation