[ONCPUML-97]: Implement "int8" support for 2D decomposition at high core counts
Interleaved2d functionality was extended to uint8 and int8 kernels.
Change-Id: If78facbce56e9ec7b2f4c23436af0bd5db7f7b69
Signed-off-by: Aleksandr Nikolaev <aleksandr.nikolaev@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3467
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
diff --git a/src/runtime/NEON/functions/NEGEMMAssemblyDispatch.cpp b/src/runtime/NEON/functions/NEGEMMAssemblyDispatch.cpp
index b09ea2e..8e9f393 100644
--- a/src/runtime/NEON/functions/NEGEMMAssemblyDispatch.cpp
+++ b/src/runtime/NEON/functions/NEGEMMAssemblyDispatch.cpp
@@ -438,12 +438,18 @@
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 || _d->info()->data_type() == DataType::F16))
+ else if(_kernel_info.method == arm_gemm::GemmMethod::GEMM_INTERLEAVED_2D && (_d->info()->data_type() == DataType::F32 || _d->info()->data_type() == DataType::F16 || _d->info()->data_type() == DataType::U8 || _d->info()->data_type() == DataType::S8) )
{
//GEMM_INTERLEAVED supports 2D parallelism, IScheduler::split_dimensions_all signals to parallelise over all window dimensions
const int granule_threshold = 200;
scheduling_hint = IScheduler::Hints(IScheduler::split_dimensions_all, IScheduler::StrategyHint::STATIC, granule_threshold);
}
+ else if(_kernel_info.method == arm_gemm::GemmMethod::QUANTIZE_WRAPPER_2D && (_d->info()->data_type() == DataType::QASYMM8 || _d->info()->data_type() == DataType::QASYMM8_SIGNED))
+ {
+ //special case for QASYMM8 to support 2D parallelism, scheduler here may be tweaked differently compared to FP32 case
+ const int granule_threshold = 200;
+ scheduling_hint = IScheduler::Hints(IScheduler::split_dimensions_all, IScheduler::StrategyHint::STATIC, granule_threshold);
+ }
NEScheduler::get().schedule(_optimised_kernel.get(), scheduling_hint);
}