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>
diff --git a/arm_compute/core/Types.h b/arm_compute/core/Types.h
index bff672c..47df44c 100644
--- a/arm_compute/core/Types.h
+++ b/arm_compute/core/Types.h
@@ -1578,6 +1578,7 @@
bool transpose_weights{ true }; /**< Transpose weights if true. */
bool are_weights_reshaped{ false }; /**< Reshape the weights tensor if false. */
bool retain_internal_weights{ false }; /**< Retain internal reshaped weights. */
+ bool enable_fast_math{ false }; /**< Enable fast math computation. */
/* Other parameters */
bool fp_mixed_precision{ false }; /**< Use wider accumulators (32 bit instead of 16 for FP16) to improve accuracy. */
@@ -2107,6 +2108,14 @@
{
return _fast_math;
};
+ /** Set fast math flag
+ *
+ * @param[in] fast_math Flag to set
+ */
+ void set_fast_math(bool fast_math)
+ {
+ _fast_math = fast_math;
+ }
/** Flag which specifies whether to broadcast the shape of the bias tensor.
*
* @return True if the shape of the bias tensor is to be broadcasted.
diff --git a/arm_compute/graph/GraphBuilder.h b/arm_compute/graph/GraphBuilder.h
index 14ad057..cb88c0e 100644
--- a/arm_compute/graph/GraphBuilder.h
+++ b/arm_compute/graph/GraphBuilder.h
@@ -295,13 +295,15 @@
* @param[in] bias_nid (Optional) Node ID of the bias node data. Defaults to EmptyNodeID
* @param[in] fc_info (Optional) Fully connected layer metadata
* @param[in] out_quant_info (Optional) Output quantization info
+ * @param[in] fast_math_hint (Optional) Fast math hint
*
* @return Node ID of the created node, EmptyNodeID in case of error
*/
static NodeID add_fully_connected_layer(Graph &g, NodeParams params, NodeIdxPair input, unsigned int num_outputs,
NodeID weights_nid, NodeID bias_nid = EmptyNodeID,
const FullyConnectedLayerInfo fc_info = FullyConnectedLayerInfo(),
- const QuantizationInfo &out_quant_info = QuantizationInfo());
+ const QuantizationInfo &out_quant_info = QuantizationInfo(),
+ FastMathHint fast_math_hint = FastMathHint::Disabled);
/** Adds a fully connected layer node to the graph
*
* @param[in] g Graph to add the layer to
@@ -313,6 +315,7 @@
* @param[in] fc_info (Optional) Fully connected layer metadata
* @param[in] weights_quant_info (Optional) Weights quantization info
* @param[in] out_quant_info (Optional) Output quantization info
+ * @param[in] fast_math_hint (Optional) Fast math hint
*
* @return Node ID of the created node, EmptyNodeID in case of error
*/
@@ -320,7 +323,8 @@
ITensorAccessorUPtr weights_accessor = nullptr, ITensorAccessorUPtr bias_accessor = nullptr,
const FullyConnectedLayerInfo fc_info = FullyConnectedLayerInfo(),
const QuantizationInfo &weights_quant_info = QuantizationInfo(),
- const QuantizationInfo &out_quant_info = QuantizationInfo());
+ const QuantizationInfo &out_quant_info = QuantizationInfo(),
+ FastMathHint fast_math_hint = FastMathHint::Disabled);
/** Adds a generate proposals layer node to the graph
*
* @param[in] g Graph to add the layer to
diff --git a/arm_compute/graph/backends/FunctionHelpers.h b/arm_compute/graph/backends/FunctionHelpers.h
index 6aec3f6..55af056 100644
--- a/arm_compute/graph/backends/FunctionHelpers.h
+++ b/arm_compute/graph/backends/FunctionHelpers.h
@@ -1096,7 +1096,8 @@
typename TargetInfo::TensorType *weights = get_backing_tensor<TargetInfo>(node.input(1));
typename TargetInfo::TensorType *biases = get_backing_tensor<TargetInfo>(node.input(2));
typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
- const FullyConnectedLayerInfo fc_info = node.info();
+ FullyConnectedLayerInfo fc_info = node.info();
+ fc_info.enable_fast_math = (node.fast_math_hint() == FastMathHint::Enabled);
ARM_COMPUTE_ERROR_ON(input == nullptr);
ARM_COMPUTE_ERROR_ON(weights == nullptr);
diff --git a/arm_compute/graph/frontend/Layers.h b/arm_compute/graph/frontend/Layers.h
index bf68b26..fe0539b 100644
--- a/arm_compute/graph/frontend/Layers.h
+++ b/arm_compute/graph/frontend/Layers.h
@@ -776,7 +776,7 @@
{
return GraphBuilder::add_fully_connected_layer(s.graph(), common_params, input, _num_outputs,
std::move(_weights), std::move(_bias), _fc_info,
- std::move(_weights_quant_info), std::move(_out_quant_info));
+ std::move(_weights_quant_info), std::move(_out_quant_info), s.hints().fast_math_hint);
}
else
{
@@ -785,7 +785,7 @@
NodeID bias_nid = (_bias_ss == nullptr) ? EmptyNodeID : _bias_ss->tail_node();
return GraphBuilder::add_fully_connected_layer(s.graph(), common_params, input, _num_outputs,
_weights_ss->tail_node(), bias_nid, _fc_info,
- std::move(_out_quant_info));
+ std::move(_out_quant_info), s.hints().fast_math_hint);
}
}
diff --git a/arm_compute/graph/nodes/FullyConnectedLayerNode.h b/arm_compute/graph/nodes/FullyConnectedLayerNode.h
index a7712f4..9ade62b 100644
--- a/arm_compute/graph/nodes/FullyConnectedLayerNode.h
+++ b/arm_compute/graph/nodes/FullyConnectedLayerNode.h
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2018-2020 Arm Limited.
+ * Copyright (c) 2018-2021 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -39,10 +39,22 @@
* @param[in] num_outputs Number of neurons in the layer
* @param[in] out_quant_info (Optional) Output quantization info
* @param[in] fc_info (Optional) Additional information about the fully connected layer
+ * @param[in] fast_math_hint (Optional) Fast math hint
*/
FullyConnectedLayerNode(unsigned int num_outputs,
QuantizationInfo out_quant_info = QuantizationInfo(),
- FullyConnectedLayerInfo fc_info = FullyConnectedLayerInfo());
+ FullyConnectedLayerInfo fc_info = FullyConnectedLayerInfo(),
+ FastMathHint fast_math_hint = FastMathHint::Disabled);
+ /** Sets the fast math fast hint
+ *
+ * @param[in] hint Hint to use for fullyconnected layer
+ */
+ void set_fast_math_hint(FastMathHint hint);
+ /** Fast math hint accessor
+ *
+ * @return Fast math hint to be used by the node
+ */
+ FastMathHint fast_math_hint() const;
/** Sets fused activation
*
* @param[in] fused_activation Fused activation to set
@@ -94,6 +106,7 @@
unsigned int _num_outputs;
QuantizationInfo _out_quant_info;
FullyConnectedLayerInfo _info;
+ FastMathHint _fast_math_hint;
};
} // namespace graph
} // namespace arm_compute