Move s32 to f32 conversion in reference layers from quantization to dequantization
Signed-off-by: Radu Salavat <radu.salavat@arm.com>
Change-Id: Ib17946b526d35deeca94b5d2f163b92101e313c4
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/11420
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
diff --git a/tests/validation/fixtures/GEMMLowpFixture.h b/tests/validation/fixtures/GEMMLowpFixture.h
index 11a491f..6b7cbba 100644
--- a/tests/validation/fixtures/GEMMLowpFixture.h
+++ b/tests/validation/fixtures/GEMMLowpFixture.h
@@ -31,7 +31,7 @@
#include "tests/validation/Validation.h"
#include "tests/validation/reference/GEMMLowp.h"
#include "tests/validation/reference/ArithmeticOperations.h"
-#include "tests/validation/reference/QuantizationLayer.h"
+#include "tests/validation/reference/DequantizationLayer.h"
#include <cstdint>
#include <vector>
@@ -485,7 +485,7 @@
const auto b_qinfo = QuantizationInfo(5.0f / 255, b_offset);
TensorFillInfo finfo;
_target = compute_target(shape_a, shape_b, shape_output, a_qinfo, b_qinfo, finfo, accumulate, dynamic_qinfo);
- _reference = compute_reference(shape_a, shape_b, shape_output, a_qinfo, b_qinfo, finfo, accumulate);
+ _reference = compute_reference(shape_a, shape_b, shape_output, a_qinfo, b_qinfo, finfo, accumulate, dynamic_qinfo);
}
protected:
@@ -495,14 +495,16 @@
return compute_gemmlowp_target<TensorType, AccessorType, FunctionType, reinterpret_input_as_3d, reinterpret_output_as_3d, int32_t, false, run_twice>(shape_a, shape_b, shape_output, a_qinfo, b_qinfo, output_qinfo, DataType::QASYMM8_SIGNED, DataType::QASYMM8_SIGNED, GEMMLowpOutputStageInfo(), false, finfo, accumulate, dynamic_qinfo, DataType::F32);
}
- SimpleTensor<float> compute_reference(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_output, const QuantizationInfo& a_qinfo, const QuantizationInfo& b_qinfo, const TensorFillInfo& finfo, bool accumulate)
+ SimpleTensor<float> compute_reference(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_output, const QuantizationInfo& a_qinfo, const QuantizationInfo& b_qinfo, const TensorFillInfo& finfo, bool accumulate, const bool dynamic_qinfo)
{
+ QuantizationInfo s32_ref_output_quant_info = QuantizationInfo(a_qinfo.uniform().scale * b_qinfo.uniform().scale, 0, dynamic_qinfo);
+
SimpleTensor<int32_t> s32_ref_output = compute_gemmlowp_reference<reinterpret_input_as_3d, int8_t, int8_t, false, false, run_twice>(shape_a, shape_b, shape_output, a_qinfo, b_qinfo,
DataType::QASYMM8_SIGNED, DataType::QASYMM8_SIGNED, finfo);
+ s32_ref_output.quantization_info(s32_ref_output_quant_info);
SimpleTensor<float> f32_ref_output(s32_ref_output.shape(), DataType::F32);
- QuantizationInfo dst_quant_info = QuantizationInfo(a_qinfo.uniform().scale * b_qinfo.uniform().scale, 0);
- f32_ref_output = reference::quantization_layer<int32_t, float>(s32_ref_output, DataType::F32, dst_quant_info);
+ f32_ref_output = reference::dequantization_layer<float, int32_t>(s32_ref_output);
if (accumulate)
{
diff --git a/tests/validation/reference/DequantizationLayer.cpp b/tests/validation/reference/DequantizationLayer.cpp
index 64a89aa..67d69c2 100644
--- a/tests/validation/reference/DequantizationLayer.cpp
+++ b/tests/validation/reference/DequantizationLayer.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2017-2020 Arm Limited.
+ * Copyright (c) 2017-2020, 2024 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -59,6 +59,12 @@
ARM_COMPUTE_UNUSED(dt);
return static_cast<TOut>(dequantize_qsymm16(val, qinfo));
}
+template <typename TOut>
+TOut dequantize(int32_t val, const UniformQuantizationInfo qinfo, DataType dt)
+{
+ ARM_COMPUTE_UNUSED(dt);
+ return static_cast<TOut>(dequantize_s32(val, qinfo));
+}
} // namespace
template <typename TOut, typename TIn>
SimpleTensor<TOut> dequantization_layer(const SimpleTensor<TIn> &src)
@@ -115,6 +121,7 @@
template SimpleTensor<float> dequantization_layer(const SimpleTensor<int8_t> &src);
template SimpleTensor<half> dequantization_layer(const SimpleTensor<int16_t> &src);
template SimpleTensor<float> dequantization_layer(const SimpleTensor<int16_t> &src);
+template SimpleTensor<float> dequantization_layer(const SimpleTensor<int32_t> &src);
} // namespace reference
} // namespace validation
} // namespace test
diff --git a/tests/validation/reference/QuantizationLayer.cpp b/tests/validation/reference/QuantizationLayer.cpp
index b76263b..ad7ba7a 100644
--- a/tests/validation/reference/QuantizationLayer.cpp
+++ b/tests/validation/reference/QuantizationLayer.cpp
@@ -80,15 +80,6 @@
dst[i] = quantize_qasymm16((src[i]), qinfo, rounding_policy);
}
break;
- case DataType::F32:
-#if defined(_OPENMP)
- #pragma omp parallel for
-#endif /* _OPENMP */
- for(int i = 0; i < src.num_elements(); ++i)
- {
- dst[i] = dequantize_s32((src[i]), qinfo);
- }
- break;
default:
ARM_COMPUTE_ERROR("Unsupported output data type");
}
@@ -136,7 +127,6 @@
template SimpleTensor<uint8_t> quantization_layer(const SimpleTensor<float> &src, DataType output_data_type, const QuantizationInfo &quantization_info);
template SimpleTensor<uint16_t> quantization_layer(const SimpleTensor<half> &src, DataType output_data_type, const QuantizationInfo &quantization_info);
template SimpleTensor<uint16_t> quantization_layer(const SimpleTensor<float> &src, DataType output_data_type, const QuantizationInfo &quantization_info);
-template SimpleTensor<float> quantization_layer(const SimpleTensor<int32_t> &src, DataType output_data_type, const QuantizationInfo &quantization_info);
} // namespace reference
} // namespace validation
} // namespace test