Fix fully connected and matmul mismatches
* There is an issue with quantized fully connected and matmul
when the lower bound of bounded ReLU is negative.
* Use int32_t for the calculation of min/max quantized value
rather than PixelValue to avoid this issue.
Partially resolves: COMPMID-5996
Signed-off-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Change-Id: I7b22e9d56a2441fc6a4c5c4e627f57d6e00d6ff1
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9502
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Jakub Sujak <jakub.sujak@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
diff --git a/src/core/utils/quantization/AsymmHelpers.cpp b/src/core/utils/quantization/AsymmHelpers.cpp
index ba9a97a..f5b69c7 100644
--- a/src/core/utils/quantization/AsymmHelpers.cpp
+++ b/src/core/utils/quantization/AsymmHelpers.cpp
@@ -24,6 +24,7 @@
#include "arm_compute/core/utils/quantization/AsymmHelpers.h"
#include "arm_compute/core/Helpers.h"
#include "support/ToolchainSupport.h"
+#include "src/core/utils/quantization/AsymmHelpers.h"
#include <cmath>
#include <limits>
@@ -177,11 +178,15 @@
return std::make_pair(min_quant_val, max_quant_val);
}
-std::tuple<PixelValue, PixelValue> get_quantized_asymmetric_output_min_max(const QuantizationInfo &q_info, const ActivationLayerInfo &act_info, DataType data_type)
+std::tuple<int32_t, int32_t> get_quantized_asymmetric_output_min_max(const QuantizationInfo &q_info, const ActivationLayerInfo &act_info, DataType data_type)
{
- PixelValue type_min{};
- PixelValue type_max{};
- std::tie(type_min, type_max) = get_min_max(data_type);
+ ARM_COMPUTE_ERROR_ON(data_type != DataType::QASYMM8 && data_type != DataType::QASYMM8_SIGNED);
+
+ const auto min_max = get_min_max(data_type);
+
+ int32_t type_min = std::get<0>(min_max).get<int32_t>();
+ int32_t type_max = std::get<1>(min_max).get<int32_t>();
+
const UniformQuantizationInfo q_unif = q_info.uniform();
if(act_info.enabled())
@@ -189,15 +194,15 @@
switch(act_info.activation())
{
case ActivationLayerInfo::ActivationFunction::RELU:
- type_min = PixelValue(q_unif.offset);
+ type_min = q_unif.offset;
break;
case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU:
- type_min = PixelValue(q_unif.offset);
- type_max = PixelValue(act_info.a(), data_type, q_info);
+ type_min = q_unif.offset;
+ type_max = (data_type == DataType::QASYMM8) ? quantize_qasymm8(act_info.a(), q_info) : quantize_qasymm8_signed(act_info.a(), q_info);
break;
case ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU:
- type_min = PixelValue(act_info.b(), data_type, q_info);
- type_max = PixelValue(act_info.a(), data_type, q_info);
+ type_min = (data_type == DataType::QASYMM8) ? quantize_qasymm8(act_info.b(), q_info) : quantize_qasymm8_signed(act_info.b(), q_info);
+ type_max = (data_type == DataType::QASYMM8) ? quantize_qasymm8(act_info.a(), q_info) : quantize_qasymm8_signed(act_info.a(), q_info);
break;
default:
ARM_COMPUTE_ERROR("Activation function not supported.");
@@ -205,7 +210,7 @@
}
}
- return std::make_pair(type_min, type_max);
+ return std::make_tuple(type_min, type_max);
}
void compute_quantized_multipliers_and_shifts(const ITensorInfo *input,