diff options
author | Georgios Pinitas <georgios.pinitas@arm.com> | 2019-05-21 13:32:43 +0100 |
---|---|---|
committer | Georgios Pinitas <georgios.pinitas@arm.com> | 2019-06-03 14:51:29 +0000 |
commit | 4c5469b192665c94118a8a558787cb9cec2d0765 (patch) | |
tree | 168aa969de8243bdbb1f25247dd9f54d037ae32c /src/runtime/NEON/functions/NEFullyConnectedLayer.cpp | |
parent | 43a129e94df41f9ac8bc78b702da5a387ada0494 (diff) | |
download | ComputeLibrary-4c5469b192665c94118a8a558787cb9cec2d0765.tar.gz |
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 <georgios.pinitas@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1236
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/runtime/NEON/functions/NEFullyConnectedLayer.cpp')
-rw-r--r-- | src/runtime/NEON/functions/NEFullyConnectedLayer.cpp | 16 |
1 files changed, 10 insertions, 6 deletions
diff --git a/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp b/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp index e1a17db6d4..7a74a7ea90 100644 --- a/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp +++ b/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp @@ -44,8 +44,8 @@ Status validate_mm(const ITensorInfo &input, const ITensorInfo &weights, const I { // 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(input.quantization_info().uniform().scale, -input.quantization_info().uniform().offset); + const QuantizationInfo weights_quantization_info(weights.quantization_info().uniform().scale, -weights.quantization_info().uniform().offset); // Validate gemmlowp function ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpMatrixMultiplyCore::validate(&input.clone()->set_quantization_info(input_quantization_info), @@ -90,8 +90,8 @@ void NEFullyConnectedLayer::configure_mm(const ITensor *input, const ITensor *we 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); @@ -227,11 +227,15 @@ void NEFullyConnectedLayer::configure(const ITensor *input, const ITensor *weigh // 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(); } |