From f29d1b7d8bf2d1619554eb3443556b44d4aa1a4c Mon Sep 17 00:00:00 2001 From: Michele Di Giorgio Date: Tue, 29 Oct 2019 10:58:13 +0000 Subject: COMPMID-2608: Enable quantization with multiplier greater than 1 on NEON Change-Id: Ib2b0c9ac88fc2b645f478c9981f71ee28f2c77fd Signed-off-by: Michele Di Giorgio Reviewed-on: https://review.mlplatform.org/c/2425 Comments-Addressed: Arm Jenkins Reviewed-by: Georgios Pinitas Tested-by: Arm Jenkins --- src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp | 12 +----------- 1 file changed, 1 insertion(+), 11 deletions(-) (limited to 'src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp') diff --git a/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp b/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp index ca4fe732a7..ddcc71f466 100644 --- a/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp +++ b/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp @@ -60,16 +60,6 @@ Status validate_arguments_optimized(const ITensorInfo *input, const ITensorInfo const bool is_quantized = (!is_data_type_quantized_per_channel(weights->data_type())) && is_data_type_quantized_asymmetric(input->data_type()); - if(is_quantized) - { - const UniformQuantizationInfo iq_info = input->quantization_info().uniform(); - const UniformQuantizationInfo wq_info = weights->quantization_info().uniform(); - const UniformQuantizationInfo oq_info = output->quantization_info().uniform(); - - float multiplier = (iq_info.scale * wq_info.scale) / oq_info.scale; - ARM_COMPUTE_UNUSED(multiplier); - ARM_COMPUTE_RETURN_ERROR_ON(multiplier > 1.0f); - } if(!NEDepthwiseConvolutionAssemblyDispatch::is_optimized_supported(input, weights, conv_info, depth_multiplier, dilation)) { TensorInfo accumulator = TensorInfo(output->clone()->set_is_resizable(true).reset_padding().set_data_type(DataType::S32)); @@ -205,7 +195,7 @@ void NEDepthwiseConvolutionLayer::NEDepthwiseConvolutionLayerOptimizedInternal:: float multiplier = (iq_info.scale * wq_info.scale) / oq_info.scale; int32_t output_multiplier; int32_t output_shift; - quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift); + quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift); _output_stage_kernel.configure(&_accumulator, biases, _is_nchw ? output : &_permuted_output, output_multiplier, output_shift, oq_info.offset); _accumulator.allocator()->allocate(); } -- cgit v1.2.1