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/gpu | |
| 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/gpu')
| -rw-r--r-- | src/gpu/cl/operators/ClFullyConnected.cpp | 18 | ||||
| -rw-r--r-- | src/gpu/cl/operators/ClGemm.cpp | 4 |
2 files changed, 17 insertions, 5 deletions
diff --git a/src/gpu/cl/operators/ClFullyConnected.cpp b/src/gpu/cl/operators/ClFullyConnected.cpp index 165ffe9a47..8afd036e7c 100644 --- a/src/gpu/cl/operators/ClFullyConnected.cpp +++ b/src/gpu/cl/operators/ClFullyConnected.cpp @@ -170,8 +170,7 @@ void ClFullyConnected::configure_mm(const CLCompileContext &compile_context, ITe fc_info.fp_mixed_precision, // fp_mixed_precision false, // fast_math true, // broadcast_bias - fc_info.activation_info, // activation_info - fc_info.constant_weights); // constant_weights + fc_info.activation_info); // activation_info if(_is_quantized) { @@ -335,7 +334,7 @@ Status ClFullyConnected::validate(const ITensorInfo *src, const ITensorInfo *wei ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 2); 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(!fc_info.constant_weights && (!fc_info.are_weights_reshaped || fc_info.transpose_weights)); + 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; @@ -353,6 +352,19 @@ Status ClFullyConnected::validate(const ITensorInfo *src, const ITensorInfo *wei const ITensorInfo *src_to_use = src; const ITensorInfo *weights_to_use = weights; + 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); + } + } + // Check if we have a fully connected layer with batches const bool is_batched_fc_layer = dst->dimension(1) > 1; if(is_batched_fc_layer) diff --git a/src/gpu/cl/operators/ClGemm.cpp b/src/gpu/cl/operators/ClGemm.cpp index e955ae3d65..4cd5237b11 100644 --- a/src/gpu/cl/operators/ClGemm.cpp +++ b/src/gpu/cl/operators/ClGemm.cpp @@ -575,7 +575,7 @@ void ClGemm::configure(const CLCompileContext &compile_context, ITensorInfo *a, // Select GEMMType _gemm_kernel_type = auto_select_gemm_kernel(auto_heuristics::CommonQuery{ CLScheduler::get().target(), a->data_type(), m, n, k, batch_size }, _reshape_b_only_on_first_run, - gemm_info.constant_weights()); + b->are_values_constant()); const bool fuse_add_c = (!(helpers::float_ops::is_zero(beta)) && c != nullptr); @@ -624,7 +624,7 @@ Status ClGemm::validate(const ITensorInfo *a, const ITensorInfo *b, const ITenso { CLScheduler::get().target(), a->data_type(), m, n, k, batch_size, }, - gemm_info.reshape_b_only_on_first_run(), gemm_info.constant_weights()); + gemm_info.reshape_b_only_on_first_run(), b->are_values_constant()); const bool fuse_add_c = (!(helpers::float_ops::is_zero(beta)) && c != nullptr); |