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

Fix softmax perf of stable diffusion (#16869)

Browse Source
This commit is contained in:
Taylor Yeonbok Lee
2023-04-12 12:01:31 -07:00
committed by GitHub
parent 298bf15a1b
commit bee357bcf8
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
src/plugins/intel_gpu/src/kernel_selector/kernels/softmax/softmax_kernel_bf.cpp
View File

@@ -54,9 +54,9 @@ SoftmaxKernel_bf::Parent::DispatchData SoftmaxKernel_bf::SetDefault(const softma
}
dispatchData.leftovers = dispatchData.dataSetSize % dispatchData.lws[0];
//if (dispatchData.leftovers % subgroup_size) {
// To use subgroup read/write, the starting address should be aligned to 128 bit
if ((dispatchData.dataSetSize * params.inputs[0].ElementSize()) >> 4) {
size_t dataSetSizeInByte = dispatchData.dataSetSize * params.inputs[0].ElementSize();
if ((dispatchData.dataSetsCount > 1) && ((dataSetSizeInByte - ((dataSetSizeInByte >> 4) << 4)))) {
dispatchData.subgroupBlockSize = 1;
} else {
if (dispatchData.itemsNum >> 3)