diff options
Diffstat (limited to 'src/runtime/NEON/functions/NELSTMLayerQuantized.cpp')
| -rw-r--r-- | src/runtime/NEON/functions/NELSTMLayerQuantized.cpp | 12 |
1 files changed, 6 insertions, 6 deletions
diff --git a/src/runtime/NEON/functions/NELSTMLayerQuantized.cpp b/src/runtime/NEON/functions/NELSTMLayerQuantized.cpp index e325619ae4..cfd996b538 100644 --- a/src/runtime/NEON/functions/NELSTMLayerQuantized.cpp +++ b/src/runtime/NEON/functions/NELSTMLayerQuantized.cpp @@ -138,8 +138,7 @@ void NELSTMLayerQuantized::configure(const ITensor *input, const float multiplier = 4096.f * qasymm.uniform().scale * qweights.uniform().scale; int output_multiplier = 0; int output_shift = 0; - - quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift); + quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift); _memory_group.manage(&_output_lowp); _output_stage.configure(&_output_highp, &_bias, &_output_lowp, output_multiplier, output_shift); @@ -340,12 +339,13 @@ Status NELSTMLayerQuantized::validate(const ITensorInfo *input, input_concatenated.set_quantization_info(QuantizationInfo(qasymm.uniform().scale, qasymm.uniform().offset)); weights_transposed.set_quantization_info(QuantizationInfo(qweights.uniform().scale, qweights.uniform().offset)); - // multiplier = (input_scale * weights_scale) / output_scale (2 ^ (-12)) const TensorInfo output_lowp(output_highp.tensor_shape(), 1, DataType::QSYMM16, qsymm_3); - const float multiplier = 4096.f * qasymm.uniform().scale * qweights.uniform().scale; - ARM_COMPUTE_UNUSED(multiplier); - ARM_COMPUTE_RETURN_ERROR_ON(multiplier > 1.0f); + const float multiplier = 4096.f * qasymm.uniform().scale * qweights.uniform().scale; + int output_multiplier = 0; + int output_shift = 0; + ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift)); + // _output_stage ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint::validate(&output_highp, &bias_concatenated, &output_lowp)); |