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 /arm_compute/core/Types.h | |
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 'arm_compute/core/Types.h')
-rw-r--r-- | arm_compute/core/Types.h | 9 |
1 files changed, 9 insertions, 0 deletions
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. |