diff options
author | Georgios Pinitas <georgios.pinitas@arm.com> | 2021-07-16 16:16:43 +0100 |
---|---|---|
committer | Georgios Pinitas <georgios.pinitas@arm.com> | 2021-07-22 02:25:50 +0000 |
commit | 4ee8b1599dbaf7634d25607fa5ac96ba3dc6b0f2 (patch) | |
tree | 2f8362d33cdad4212f4b96995681c68184c759e1 /src/runtime/CL | |
parent | 59fd7a722e5bc7e85309d6200bc37a772721a719 (diff) | |
download | ComputeLibrary-4ee8b1599dbaf7634d25607fa5ac96ba3dc6b0f2.tar.gz |
Update GEMM assembly kernels
- Introduce Fp32 kernels with internal calculations in Bfloat16 when
fast_mode is enabled
- Improve kernel selection heuristics
Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Change-Id: I68a9e7e862b6fd2721b46e0d7cc791091c4ab279
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5965
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/runtime/CL')
-rw-r--r-- | src/runtime/CL/functions/CLFullyConnectedLayer.cpp | 2 | ||||
-rw-r--r-- | src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp | 2 |
2 files changed, 4 insertions, 0 deletions
diff --git a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp index bc9a3056e8..0647a473e2 100644 --- a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp +++ b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp @@ -101,6 +101,7 @@ Status validate_mm(const ITensorInfo &input, const ITensorInfo &weights, const I fc_info.retain_internal_weights, // retain_internal_weights gemmlowp_output_stage, // gemmlowp_output_stage fc_info.fp_mixed_precision, // fp_mixed_precision + false, // fast_math true, // broadcast_bias ActivationLayerInfo()); // activation_info @@ -151,6 +152,7 @@ void CLFullyConnectedLayer::configure_mm(const CLCompileContext &compile_context fc_info.retain_internal_weights, // retain_internal_weights gemmlowp_output_stage, // gemmlowp_output_stage 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 diff --git a/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp b/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp index f926b1d0a6..16735dde0e 100644 --- a/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp +++ b/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp @@ -128,6 +128,7 @@ void CLGEMMConvolutionLayer::configure_mm(const CLCompileContext &compile_contex false, // retain_internal_weights gemmlowp_output_stage, // gemmlowp_output_stage false, // fp_mixed_precision + false, // fast_math true, // broadcast_bias act_info); // activation_info @@ -167,6 +168,7 @@ Status CLGEMMConvolutionLayer::validate_mm(const ITensorInfo *input, const ITens false, // retain_internal_weights gemmlowp_output_stage, // gemmlowp_output_stage false, // fp_mixed_precision + false, // fast_math true, // broadcast_bias act_info); // activation_info |