diff options
author | cfRod <crefeda.rodrigues@arm.com> | 2021-11-05 11:29:53 +0000 |
---|---|---|
committer | Gian Marco Iodice <gianmarco.iodice@arm.com> | 2021-11-09 10:19:03 +0000 |
commit | f2c022e54ae65467cca4bc698b5b94e5b3c62c47 (patch) | |
tree | dcf3f38e85ff9761573d2e7e37cc746291f828af /src | |
parent | f349655cf75a686ed97e90d193a45602d6e82e6f (diff) | |
download | ComputeLibrary-f2c022e54ae65467cca4bc698b5b94e5b3c62c47.tar.gz |
Enable fast_math in CpuFullyConnected
ONCPUML-529
* Add support for passing fast_math for fullyconnected layers via fc_info.
* Add support for passing fast_math to run ACL benchmark graphs.
* Add validation test and accuracy tests (updated fixtures).
Note: abs and rel. tolerance for fast math mode are set based on experimental data.
Signed-off-by: cfRod <crefeda.rodrigues@arm.com>
change-Id: Ib107d6264d3ae5e36555334f39a13e678f8618df
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6521
Reviewed-by: SiCong Li <sicong.li@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src')
-rw-r--r-- | src/cpu/operators/CpuFullyConnected.cpp | 15 | ||||
-rw-r--r-- | src/cpu/operators/CpuFullyConnected.h | 1 | ||||
-rw-r--r-- | src/graph/GraphBuilder.cpp | 8 | ||||
-rw-r--r-- | src/graph/nodes/FullyConnectedLayer.cpp | 15 |
4 files changed, 28 insertions, 11 deletions
diff --git a/src/cpu/operators/CpuFullyConnected.cpp b/src/cpu/operators/CpuFullyConnected.cpp index 03c53b001d..6d77c614f7 100644 --- a/src/cpu/operators/CpuFullyConnected.cpp +++ b/src/cpu/operators/CpuFullyConnected.cpp @@ -109,7 +109,7 @@ Status get_gemmlowp_output_stage_info(const ITensorInfo *src, const ITensorInfo return Status{}; } -Status validate_mm(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const ActivationLayerInfo &act) +Status validate_mm(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const ActivationLayerInfo &act, bool enable_fast_math) { if(is_data_type_quantized_asymmetric(src->data_type())) { @@ -123,6 +123,7 @@ Status validate_mm(const ITensorInfo *src, const ITensorInfo *weights, const ITe GEMMInfo gemm_info; gemm_info.set_gemmlowp_output_stage(gemmlowp_output_stage_info); + gemm_info.set_fast_math(enable_fast_math); // Validate gemmlowp function TensorInfo src_info = src->clone()->set_quantization_info(src_quantization_info); @@ -135,7 +136,9 @@ Status validate_mm(const ITensorInfo *src, const ITensorInfo *weights, const ITe } else { - ARM_COMPUTE_RETURN_ON_ERROR(CpuGemm::validate(src, weights, biases, dst, 1.f, 1.0f, GEMMInfo(false, false, true /* Reshape weights only for the first run */))); + GEMMInfo gemm_info(false, false, true /* Reshape weights only for the first run */); + gemm_info.set_fast_math(enable_fast_math); + ARM_COMPUTE_RETURN_ON_ERROR(CpuGemm::validate(src, weights, biases, dst, 1.f, 1.0f, gemm_info)); } return Status{}; @@ -158,7 +161,8 @@ CpuFullyConnected::CpuFullyConnected() _needs_weights_reshape(false), _is_fc_after_conv(false), _is_quantized_asymmetric(false), - _is_prepared(false) + _is_prepared(false), + _enable_fast_math(false) { } @@ -185,6 +189,7 @@ void CpuFullyConnected::configure_mm(const ITensorInfo *src, const ITensorInfo * GEMMInfo gemm_info; gemm_info.set_gemmlowp_output_stage(gemmlowp_output_stage_info); gemm_info.set_activation_info(act); + gemm_info.set_fast_math(_enable_fast_math); _mm_gemmlowp = std::make_unique<CpuGemmLowpMatrixMultiplyCore>(); _mm_gemmlowp->configure(&src_info, &weights_info, biases, dst, gemm_info); } @@ -193,6 +198,7 @@ void CpuFullyConnected::configure_mm(const ITensorInfo *src, const ITensorInfo * // Configure matrix multiply kernel GEMMInfo gemm_info(false, false, true /* Reshape weights only for the first run */); gemm_info.set_activation_info(act); + gemm_info.set_fast_math(_enable_fast_math); _mm_gemm = std::make_unique<CpuGemm>(); _mm_gemm->configure(src, weights, biases, dst, 1.f, 1.0f, gemm_info); } @@ -241,6 +247,7 @@ void CpuFullyConnected::configure(const ITensorInfo *src, const ITensorInfo *wei _is_quantized_asymmetric = is_data_type_quantized_asymmetric(src->data_type()); _is_prepared = false; _trans_weights_idx = AuxTensorIdx::Count; + _enable_fast_math = fc_info.enable_fast_math; // With the Fully Connected layer we can have 4 different cases: // 1) Convolution layer -> Fully Connected layer without batches @@ -418,7 +425,7 @@ Status CpuFullyConnected::validate(const ITensorInfo *src, const ITensorInfo *we ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(0) != weights_to_use->dimension(1)); } // Validate matrix multiply kernel - ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(src_to_use, weights_to_use, biases, dst, fc_info.activation_info)); + ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(src_to_use, weights_to_use, biases, dst, fc_info.activation_info, fc_info.enable_fast_math)); return Status{}; } diff --git a/src/cpu/operators/CpuFullyConnected.h b/src/cpu/operators/CpuFullyConnected.h index 304ea3c62b..44fa21f9f8 100644 --- a/src/cpu/operators/CpuFullyConnected.h +++ b/src/cpu/operators/CpuFullyConnected.h @@ -141,6 +141,7 @@ private: bool _is_fc_after_conv; bool _is_quantized_asymmetric; bool _is_prepared; + bool _enable_fast_math; }; } // namespace cpu } // namespace arm_compute diff --git a/src/graph/GraphBuilder.cpp b/src/graph/GraphBuilder.cpp index 01d35a15b9..15abf3738a 100644 --- a/src/graph/GraphBuilder.cpp +++ b/src/graph/GraphBuilder.cpp @@ -467,7 +467,7 @@ NodeID GraphBuilder::add_flatten_node(Graph &g, NodeParams params, NodeIdxPair i NodeID GraphBuilder::add_fully_connected_layer(Graph &g, NodeParams params, NodeIdxPair input, unsigned int num_outputs, NodeID weights_nid, NodeID bias_nid, - const FullyConnectedLayerInfo fc_info, const QuantizationInfo &out_quant_info) + const FullyConnectedLayerInfo fc_info, const QuantizationInfo &out_quant_info, FastMathHint fast_math_hint) { check_nodeidx_pair(input, g); ARM_COMPUTE_ERROR_ON(num_outputs == 0); @@ -479,7 +479,7 @@ NodeID GraphBuilder::add_fully_connected_layer(Graph &g, NodeParams params, Node const TensorDescriptor input_tensor_desc = get_tensor_descriptor(g, g.node(input.node_id)->outputs()[0]); // Create fully connected node and connect - NodeID fc_nid = g.add_node<FullyConnectedLayerNode>(num_outputs, out_quant_info, fc_info); + NodeID fc_nid = g.add_node<FullyConnectedLayerNode>(num_outputs, out_quant_info, fc_info, fast_math_hint); g.add_connection(input.node_id, input.index, fc_nid, 0); g.add_connection(weights_nid, 0, fc_nid, 1); if(has_bias) @@ -495,7 +495,7 @@ NodeID GraphBuilder::add_fully_connected_layer(Graph &g, NodeParams params, Node NodeID GraphBuilder::add_fully_connected_layer(Graph &g, NodeParams params, NodeIdxPair input, unsigned int num_outputs, ITensorAccessorUPtr weights_accessor, ITensorAccessorUPtr bias_accessor, const FullyConnectedLayerInfo fc_info, - const QuantizationInfo &weights_quant_info, const QuantizationInfo &out_quant_info) + const QuantizationInfo &weights_quant_info, const QuantizationInfo &out_quant_info, FastMathHint fast_math_hint) { check_nodeidx_pair(input, g); ARM_COMPUTE_ERROR_ON(num_outputs == 0); @@ -523,7 +523,7 @@ NodeID GraphBuilder::add_fully_connected_layer(Graph &g, NodeParams params, Node } // Create fully connected node and connect - NodeID fc_nid = g.add_node<FullyConnectedLayerNode>(num_outputs, out_quant_info, fc_info); + NodeID fc_nid = g.add_node<FullyConnectedLayerNode>(num_outputs, out_quant_info, fc_info, fast_math_hint); g.add_connection(input.node_id, input.index, fc_nid, 0); g.add_connection(w_nid, 0, fc_nid, 1); if(has_bias) diff --git a/src/graph/nodes/FullyConnectedLayer.cpp b/src/graph/nodes/FullyConnectedLayer.cpp index 442f636b61..6278227878 100644 --- a/src/graph/nodes/FullyConnectedLayer.cpp +++ b/src/graph/nodes/FullyConnectedLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2020 Arm Limited. + * Copyright (c) 2018-2021 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -31,12 +31,21 @@ namespace arm_compute { namespace graph { -FullyConnectedLayerNode::FullyConnectedLayerNode(unsigned int num_outputs, QuantizationInfo out_quant_info, FullyConnectedLayerInfo fc_info) - : _num_outputs(num_outputs), _out_quant_info(std::move(out_quant_info)), _info(fc_info) +FullyConnectedLayerNode::FullyConnectedLayerNode(unsigned int num_outputs, QuantizationInfo out_quant_info, FullyConnectedLayerInfo fc_info, FastMathHint fast_math_hint) + : _num_outputs(num_outputs), _out_quant_info(std::move(out_quant_info)), _info(fc_info), _fast_math_hint(fast_math_hint) { _input_edges.resize(3, EmptyEdgeID); _outputs.resize(1, NullTensorID); } +void FullyConnectedLayerNode::set_fast_math_hint(FastMathHint hint) +{ + _fast_math_hint = hint; +} + +FastMathHint FullyConnectedLayerNode::fast_math_hint() const +{ + return _fast_math_hint; +} void FullyConnectedLayerNode::set_fused_activation(ActivationLayerInfo fused_activation) { |