COMPMID-2608: Enable quantization with multiplier greater than 1 on NEON
Change-Id: Ib2b0c9ac88fc2b645f478c9981f71ee28f2c77fd
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/2425
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
diff --git a/src/core/NEON/kernels/NEGEMMLowpOffsetContributionOutputStageKernel.cpp b/src/core/NEON/kernels/NEGEMMLowpOffsetContributionOutputStageKernel.cpp
index 8418733..86abb2d 100644
--- a/src/core/NEON/kernels/NEGEMMLowpOffsetContributionOutputStageKernel.cpp
+++ b/src/core/NEON/kernels/NEGEMMLowpOffsetContributionOutputStageKernel.cpp
@@ -909,8 +909,12 @@
&& mm_result->info()->tensor_shape().y() != vector_sum_row->info()->tensor_shape().x();
// Check if we need to clamp the result using min and max
- const bool is_bounded_relu = ((output_stage.gemmlowp_min_bound != output_stage.gemmlowp_max_bound)
- && !(output_stage.gemmlowp_min_bound == 0 && output_stage.gemmlowp_max_bound == 255));
+ PixelValue type_min = 0;
+ PixelValue type_max = 0;
+ std::tie(type_min, type_max) = get_min_max(output->info()->data_type());
+ int type_min_int = type_min.get<int>();
+ int type_max_int = type_max.get<int>();
+ const bool is_bounded_relu = !(output_stage.gemmlowp_min_bound == type_min_int && output_stage.gemmlowp_max_bound == type_max_int);
// Check if we need to perform fixed point requantization
const bool is_fixed_point = output_stage.type != GEMMLowpOutputStageType::QUANTIZE_DOWN;