diff options
author | cfRod <crefeda.rodrigues@arm.com> | 2020-06-25 18:12:25 +0100 |
---|---|---|
committer | Georgios Pinitas <georgios.pinitas@arm.com> | 2020-07-15 15:17:51 +0000 |
commit | 534fdeaabbf3052a133e366f17634e5dd8c26498 (patch) | |
tree | 01c5ad1086674fffe31f3bf5ab89f4efe7193f95 /src/runtime/NEON | |
parent | ed5fe69b6612a5cf0dd52340f6781885d77afbc9 (diff) | |
download | ComputeLibrary-534fdeaabbf3052a133e366f17634e5dd8c26498.tar.gz |
[ONCPUML-96] FP16 support for 2D decomposition at high core counts.
Added changes to gemm_fp16 to pick gemm interleaved pretransposed 2D for hgemm_24x8 and sgemm_12x8. Also added the change for scheduling hints based on datatype F16.
Signed-off-by: cfRod <crefeda.rodrigues@arm.com>
Change-Id: Idd754cf14b47d00a70eab79bbb5ee3ecaf77450f
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3477
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/runtime/NEON')
-rw-r--r-- | src/runtime/NEON/functions/NEGEMMAssemblyDispatch.cpp | 2 |
1 files changed, 1 insertions, 1 deletions
diff --git a/src/runtime/NEON/functions/NEGEMMAssemblyDispatch.cpp b/src/runtime/NEON/functions/NEGEMMAssemblyDispatch.cpp index 19a8374bc9..b09ea2ec57 100644 --- a/src/runtime/NEON/functions/NEGEMMAssemblyDispatch.cpp +++ b/src/runtime/NEON/functions/NEGEMMAssemblyDispatch.cpp @@ -438,7 +438,7 @@ void Fallback<TypeInput, TypeOutput, OutputStage>::run() const int granule_threshold = 200; scheduling_hint = IScheduler::Hints(Window::DimX, IScheduler::StrategyHint::DYNAMIC, granule_threshold); } - else if(_kernel_info.method == arm_gemm::GemmMethod::GEMM_INTERLEAVED_2D && _d->info()->data_type() == DataType::F32) + else if(_kernel_info.method == arm_gemm::GemmMethod::GEMM_INTERLEAVED_2D && (_d->info()->data_type() == DataType::F32 || _d->info()->data_type() == DataType::F16)) { //GEMM_INTERLEAVED supports 2D parallelism, IScheduler::split_dimensions_all signals to parallelise over all window dimensions const int granule_threshold = 200; |