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/NEDepthwiseConvolutionLayerNativeKernel.cpp b/src/core/NEON/kernels/NEDepthwiseConvolutionLayerNativeKernel.cpp
index a9a3183..aee13ee 100644
--- a/src/core/NEON/kernels/NEDepthwiseConvolutionLayerNativeKernel.cpp
+++ b/src/core/NEON/kernels/NEDepthwiseConvolutionLayerNativeKernel.cpp
@@ -289,7 +289,16 @@
acc.at(i) += *reinterpret_cast<int32_t *>(biases_it.ptr() + i * sizeof(int32_t));
}
- acc.at(i) = rounding_divide_by_exp2(saturating_doubling_high_mul(acc.at(i), output_multiplier.at(id.x() + i)), output_shift.at(id.x() + i)) + output_qoffset;
+ const int out_mul = output_multiplier.at(id.x() + i);
+ const int out_shift = output_shift.at(id.x() + i);
+ if(out_shift < 0)
+ {
+ acc.at(i) = saturating_doubling_high_mul(acc.at(i) * (1 << (-out_shift)), out_mul) + output_qoffset;
+ }
+ else
+ {
+ acc.at(i) = rounding_divide_by_exp2(saturating_doubling_high_mul(acc.at(i), out_mul), out_shift) + output_qoffset;
+ }
out_vals[i] = static_cast<T>(utility::clamp<int32_t, uint8_t>(acc.at(i)));
}
@@ -381,21 +390,20 @@
if(has_biases)
{
- const auto biases_val = *(reinterpret_cast<int32_t *>(biases_it.ptr() + m * sizeof(int32_t)));
+ acc.at(m) += *(reinterpret_cast<int32_t *>(biases_it.ptr() + m * sizeof(int32_t)));
+ }
- int32_t out_val = acc.at(m) + biases_val;
- out_val = rounding_divide_by_exp2(saturating_doubling_high_mul(out_val, output_multiplier.at(id.x() + m)),
- output_shift.at(id.x() + m))
- + output_qoffset;
- *(reinterpret_cast<T *>(output_it.ptr() + m * sizeof(T))) = static_cast<T>(utility::clamp<int32_t, uint8_t>(out_val));
+ const int out_mul = output_multiplier.at(id.x() + m);
+ const int out_shift = output_shift.at(id.x() + m);
+ if(out_shift < 0)
+ {
+ acc.at(m) = saturating_doubling_high_mul(acc.at(m) * (1 << (-out_shift)), out_mul) + output_qoffset;
}
else
{
- int32_t out_val = rounding_divide_by_exp2(saturating_doubling_high_mul(acc.at(m), output_multiplier.at(id.x() + m)),
- output_shift.at(id.x() + m))
- + output_qoffset;
- *(reinterpret_cast<T *>(output_it.ptr() + m * sizeof(T))) = static_cast<T>(utility::clamp<int32_t, uint8_t>(out_val));
+ acc.at(m) = rounding_divide_by_exp2(saturating_doubling_high_mul(acc.at(m), out_mul), out_shift) + output_qoffset;
}
+ *(reinterpret_cast<T *>(output_it.ptr() + m * sizeof(T))) = static_cast<T>(utility::clamp<int32_t, uint8_t>(acc.at(m)));
}
},
input_it, weights_it, biases_it, output_it);
@@ -531,8 +539,7 @@
int32_t out_mult = 0;
int32_t out_shift = 0;
const float multiplier = input_scale * weights_scale.at(i) / output_scale;
- ARM_COMPUTE_ERROR_ON(multiplier > 1.f);
- arm_compute::quantization::calculate_quantized_multiplier_less_than_one(multiplier, &out_mult, &out_shift);
+ arm_compute::quantization::calculate_quantized_multiplier(multiplier, &out_mult, &out_shift);
_output_multiplier.push_back(out_mult);
_output_shift.push_back(out_shift);
diff --git a/src/core/NEON/kernels/NEDirectConvolutionLayerOutputStageKernel.cpp b/src/core/NEON/kernels/NEDirectConvolutionLayerOutputStageKernel.cpp
index 4313a5e..8834d97 100644
--- a/src/core/NEON/kernels/NEDirectConvolutionLayerOutputStageKernel.cpp
+++ b/src/core/NEON/kernels/NEDirectConvolutionLayerOutputStageKernel.cpp
@@ -38,14 +38,15 @@
#include <cstddef>
#include <cstdint>
-using namespace arm_compute;
-
+namespace arm_compute
+{
namespace
{
Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output,
int result_fixedpoint_multiplier, int result_shift, int result_offset_after_shift)
{
ARM_COMPUTE_UNUSED(result_fixedpoint_multiplier);
+ ARM_COMPUTE_UNUSED(result_shift);
ARM_COMPUTE_UNUSED(result_offset_after_shift);
ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input);
ARM_COMPUTE_RETURN_ERROR_ON(input->data_layout() == DataLayout::UNKNOWN);
@@ -53,7 +54,6 @@
DataType::F16,
DataType::S32, DataType::F32);
- ARM_COMPUTE_RETURN_ERROR_ON_MSG(result_shift < 0, "Result shift must be a non negative integer");
if(bias != nullptr)
{
ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(bias, 1, DataType::F16, DataType::S32, DataType::F32);
@@ -596,3 +596,4 @@
(*_func)(_input, _bias, window, _output, _result_fixedpoint_multiplier, _result_shift, _result_offset_after_shift);
}
+} // namespace arm_compute
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;
diff --git a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp b/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp
index 4906e6a..bb0b864 100644
--- a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp
+++ b/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp
@@ -39,8 +39,8 @@
#include <cstddef>
#include <cstdint>
-using namespace arm_compute;
-
+namespace arm_compute
+{
namespace
{
Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
@@ -244,4 +244,5 @@
ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
(this->*_func)(window);
-}
\ No newline at end of file
+}
+} // namespace arm_compute
\ No newline at end of file