From f2c022e54ae65467cca4bc698b5b94e5b3c62c47 Mon Sep 17 00:00:00 2001 From: cfRod Date: Fri, 5 Nov 2021 11:29:53 +0000 Subject: 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 change-Id: Ib107d6264d3ae5e36555334f39a13e678f8618df Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6521 Reviewed-by: SiCong Li Reviewed-by: Gian Marco Iodice Tested-by: Arm Jenkins Comments-Addressed: Arm Jenkins --- arm_compute/core/Types.h | 9 +++++++++ 1 file changed, 9 insertions(+) (limited to 'arm_compute/core/Types.h') diff --git a/arm_compute/core/Types.h b/arm_compute/core/Types.h index bff672c361..47df44cb67 100644 --- a/arm_compute/core/Types.h +++ b/arm_compute/core/Types.h @@ -1578,6 +1578,7 @@ struct FullyConnectedLayerInfo 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 @@ public: { 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. -- cgit v1.2.1