diff options
author | Giorgio Arena <giorgio.arena@arm.com> | 2021-09-24 14:04:27 +0100 |
---|---|---|
committer | Giorgio Arena <giorgio.arena@arm.com> | 2021-09-29 10:31:08 +0000 |
commit | 63e0beb9fb9646407d123e830165546e9129e95d (patch) | |
tree | 9bfe80e8d853327a82f9f622d89c3b43df0400f4 /src/cpu/operators/CpuFullyConnected.cpp | |
parent | b1ba1e33f2b03b211f561123559c24517c0e5865 (diff) | |
download | ComputeLibrary-63e0beb9fb9646407d123e830165546e9129e95d.tar.gz |
Add support for non-constant weights and biases in CpuFullyConnected
Changing the approach for specifying that weights and biases tensors are
non-constant by making it a member of TensorInfo rather than an option
of the functions.
Resolves: COMPMID-4222, COMPMID-4811
Signed-off-by: Giorgio Arena <giorgio.arena@arm.com>
Change-Id: I9b0081ccbcf8271ce029ba6755563d64c59e1d32
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6313
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Pablo Marquez Tello <pablo.tello@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/cpu/operators/CpuFullyConnected.cpp')
-rw-r--r-- | src/cpu/operators/CpuFullyConnected.cpp | 27 |
1 files changed, 22 insertions, 5 deletions
diff --git a/src/cpu/operators/CpuFullyConnected.cpp b/src/cpu/operators/CpuFullyConnected.cpp index 4133d9e8ca..03c53b001d 100644 --- a/src/cpu/operators/CpuFullyConnected.cpp +++ b/src/cpu/operators/CpuFullyConnected.cpp @@ -314,9 +314,14 @@ void CpuFullyConnected::configure(const ITensorInfo *src, const ITensorInfo *wei if(_aux_mem[Pretranspose].size > 0) { - // Release permuted weights at the of prepare as they are further transposed by the assembly dispatch - _aux_mem[TransposedWeights] = MemoryInfo(offset_int_vec(TransposedWeights), MemoryLifetime::Prepare, _reshaped_weights.total_size()); - _aux_mem[ConvertedWeights] = MemoryInfo(offset_int_vec(ConvertedWeights), MemoryLifetime::Prepare, _converted_weights.total_size()); + // Release permuted weights at the end of prepare as they are further transposed by the assembly dispatch + // Do not release them if biases are dynamic and data type is quantized, since the weights tensor will be used for biases offset calculation + _aux_mem[TransposedWeights] = MemoryInfo(offset_int_vec(TransposedWeights), (_is_quantized_asymmetric + && biases && !(biases->are_values_constant())) ? + MemoryLifetime::Persistent : + MemoryLifetime::Prepare, + _reshaped_weights.total_size()); + _aux_mem[ConvertedWeights] = MemoryInfo(offset_int_vec(ConvertedWeights), MemoryLifetime::Prepare, _converted_weights.total_size()); } else { @@ -334,10 +339,9 @@ Status CpuFullyConnected::validate(const ITensorInfo *src, const ITensorInfo *we ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, weights, dst); ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 2); - ARM_COMPUTE_RETURN_ERROR_ON(biases != nullptr && biases->num_dimensions() > 1); ARM_COMPUTE_RETURN_ERROR_ON(fc_info.activation_info.enabled() && is_data_type_quantized(src->data_type()) && fc_info.activation_info.activation() != ActivationLayerInfo::ActivationFunction::RELU && fc_info.activation_info.activation() != ActivationLayerInfo::ActivationFunction::BOUNDED_RELU && fc_info.activation_info.activation() != ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(!fc_info.constant_weights, "Non-constant weights are currently not supported"); + ARM_COMPUTE_RETURN_ERROR_ON(!weights->are_values_constant() && (!fc_info.are_weights_reshaped || fc_info.transpose_weights)); bool weights_reshaped = fc_info.transpose_weights ? fc_info.are_weights_reshaped : true; bool is_fc_after_conv = true; @@ -358,6 +362,19 @@ Status CpuFullyConnected::validate(const ITensorInfo *src, const ITensorInfo *we // Check if we have a fully connected layer with batches const bool is_batched_fc_layer = dst->dimension(1) > 1; + if(biases != nullptr) + { + ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); + if(is_data_type_quantized(src->data_type())) + { + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32); + } + else + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, biases); + } + } + if(is_batched_fc_layer) { is_fc_after_conv = (TensorShape::num_max_dimensions >= 4) && (std::equal(src->tensor_shape().cbegin() + 3, |