IntenseWebs/openvino
3
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[GPU] Update tuning params of shape agnostic version of fully_connected_bf_tiled kernel for dGPUs (#16482)

Browse Source
This commit is contained in:
Sergey Shlyapnikov 2023-03-24 09:08:09 +04:00 committed by GitHub
parent 7601e8a874
commit e434c320f5
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 11 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
src/plugins/intel_gpu/src/kernel_selector/kernels/fully_connected/fully_connected_kernel_bf_tiled.cpp
View File

@ -201,12 +201,21 @@ FullyConnected_bf_tiled::GetAutoTuneParams(const fully_connected_params& params,
max_tile_ofm *= 2;
if (params.is_shape_agnostic) {
// Use special tuning params for Gen12HP dGPUs, since these parameters demonstrate higher performance
// due to better HW utilization (reduced TILE_OFM parameter) and better assembler kernel's code
// generation (extended TILE_K parameter) for both FP16 and FP32 data types
if (dtype == Datatype::F16) {
// tune_params(tile_b, tile_ofm, tile_ifm, tile_k, dispatch_bsv, dispatch_fsv, exec_options)
selector.Case(tune_params(8, std::min(max_tile_ofm, 2u), 1, 2, 1, 1, EXE_MODE_AGE_BASED));
if (params.engineInfo.supports_immad)
selector.Case(tune_params(8, 1, 1, 4, 1, 1, EXE_MODE_AGE_BASED));
else
selector.Case(tune_params(8, std::min(max_tile_ofm, 2u), 1, 2, 1, 1, EXE_MODE_AGE_BASED));
} else if (dtype == Datatype::F32) {
// tune_params(tile_b, tile_ofm, tile_ifm, tile_k, dispatch_bsv, dispatch_fsv, exec_options)
selector.Case(tune_params(8, std::min(max_tile_ofm, 2u), 1, 1, 1, 1, EXE_MODE_AGE_BASED));
if (params.engineInfo.supports_immad)
selector.Case(tune_params(8, 1, 1, 4, 1, 1, EXE_MODE_AGE_BASED));
else
selector.Case(tune_params(8, std::min(max_tile_ofm, 2u), 1, 1, 1, 1, EXE_MODE_AGE_BASED));
}
} else {
if (dtype == Datatype::F16) {