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

[GPU] apply LoadType change condition when input1 fs != output fs(broadcast) in eltwise fusing (#12710)

Browse Source 
* apply condition when input1 fs != output fs(broadcast)
* modify condition of LT unaligned read transition
This commit is contained in:
Wilson Seok 2022-08-30 15:19:52 +09:00 committed by GitHub
parent 2d04f93486
commit cf5d93b273
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 4 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
src/plugins/intel_gpu/src/kernel_selector/core/common/jitter.cpp
View File

@ -1877,10 +1877,13 @@ std::string FusedOpsCodeGenerator::GetJitLoad(const FusedOpsConfiguration& conf,
bool safe_load = conf.boundary_check == FusedOpsConfiguration::BoundaryCheck::ENABLED;
// Fsv16 Eltwise whcih requires f axis broadcast such as input[1,1,z,1,1], output[b,f,z,y,x] need to use LT unligned read.
// In this case, intel_sub_group_block_read() introduces increasing index in feature block.
bool f_axis_broadcast = ((input_tensor.Feature().v != prim_output.Feature().v) && (input_tensor.Feature().v == 1) && (vec_size == 1));
// Change JitLoad to ignore LT_ALIGNED_READ LoadType if this input tensor has a planar format(SimpleLayout)
if (desc.GetType() == KernelType::ELTWISE && input_tensor.SimpleLayout() && input_tensor.GetLayout() != orig_output_layout &&
conf.load_type == FusedOpsConfiguration::LoadType::LT_ALIGNED_READ &&
input_tensor.SameDimsSizes(prim_output) && input_tensor.LogicalSize() != 1) {
(input_tensor.SameDimsSizes(prim_output) || f_axis_broadcast) && input_tensor.LogicalSize() != 1) {
std::string sub_group_local_id_str = "get_sub_group_local_id";
size_t found_sub = conf.bfzyx_idx_order[1].rfind(sub_group_local_id_str);
if (found_sub != std::string::npos) {