From 4c5469b192665c94118a8a558787cb9cec2d0765 Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Tue, 21 May 2019 13:32:43 +0100 Subject: COMPMID-2225: Add interface support for new quantized data types. Add support for: -QSYMM8, 8-bit quantized symmetric -QSYMM8_PER_CHANNEL, 8-bit quantized symmetric with per channel quantization Change-Id: I00c4ff98e44af37419470af61419ee95d0de2463 Signed-off-by: Georgios Pinitas Reviewed-on: https://review.mlplatform.org/c/1236 Tested-by: Arm Jenkins Reviewed-by: Gian Marco Iodice Comments-Addressed: Arm Jenkins --- src/runtime/CL/functions/CLFullyConnectedLayer.cpp | 19 +++++++++++++------ 1 file changed, 13 insertions(+), 6 deletions(-) (limited to 'src/runtime/CL/functions/CLFullyConnectedLayer.cpp') diff --git a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp index 7b9229c4ae..87d4c56a0e 100644 --- a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp +++ b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp @@ -41,10 +41,13 @@ Status validate_mm(const ITensorInfo &input, const ITensorInfo &weights, const I { if(is_data_type_quantized_asymmetric(input.data_type())) { + const UniformQuantizationInfo iq_info = input.quantization_info().uniform(); + const UniformQuantizationInfo wq_info = weights.quantization_info().uniform(); + // Since we need negative offsets for computing convolution, we need to change QuantizationInfo() // Extract and negate input and weights offset - const QuantizationInfo input_quantization_info(input.quantization_info().scale, -input.quantization_info().offset); - const QuantizationInfo weights_quantization_info(weights.quantization_info().scale, -weights.quantization_info().offset); + const QuantizationInfo input_quantization_info(iq_info.scale, -iq_info.offset); + const QuantizationInfo weights_quantization_info(wq_info.scale, -wq_info.offset); // Validate gemmlowp function ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyCore::validate(&input.clone()->set_quantization_info(input_quantization_info), @@ -88,8 +91,8 @@ void CLFullyConnectedLayer::configure_mm(const ICLTensor *input, const ICLTensor const QuantizationInfo input_quantization_info = input->info()->quantization_info(); const QuantizationInfo weights_quantization_info = weights->info()->quantization_info(); - input->info()->set_quantization_info(QuantizationInfo(input_quantization_info.scale, -input_quantization_info.offset)); - weights->info()->set_quantization_info(QuantizationInfo(weights_quantization_info.scale, -weights_quantization_info.offset)); + input->info()->set_quantization_info(QuantizationInfo(input_quantization_info.uniform().scale, -input_quantization_info.uniform().offset)); + weights->info()->set_quantization_info(QuantizationInfo(weights_quantization_info.uniform().scale, -weights_quantization_info.uniform().offset)); // Configure gemmlowp function _mm_gemmlowp.configure(input, weights, nullptr, output); @@ -230,11 +233,15 @@ void CLFullyConnectedLayer::configure(const ICLTensor *input, const ICLTensor *w // Configure output stage for asymmetric quantized types if(_is_quantized) { - float multiplier = input->info()->quantization_info().scale * weights->info()->quantization_info().scale / output->info()->quantization_info().scale; + const UniformQuantizationInfo iq_info = input->info()->quantization_info().uniform(); + const UniformQuantizationInfo wq_info = weights->info()->quantization_info().uniform(); + const UniformQuantizationInfo oq_info = output->info()->quantization_info().uniform(); + + float multiplier = iq_info.scale * wq_info.scale / oq_info.scale; int output_multiplier; int output_shift; quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift); - _gemmlowp_output_stage.configure(&_gemmlowp_output, biases, output, output_multiplier, output_shift, output->info()->quantization_info().offset); + _gemmlowp_output_stage.configure(&_gemmlowp_output, biases, output, output_multiplier, output_shift, oq_info.offset); _gemmlowp_output.allocator()->allocate(); } } -- cgit v1.2.1