Skip to content

[ET-VK] Using TmpTensor for width packed versions of q_linear op shader to reduce memory usage. #8009

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Jan 28, 2025
Merged

[ET-VK] Using TmpTensor for width packed versions of q_linear op shader to reduce memory usage. #8009

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 7 additions & 2 deletions backends/vulkan/runtime/graph/ops/impl/QuantizedLinear.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -73,13 +73,18 @@ void add_q_8w_linear_node(
auto viewFn = VK_GET_OP_FN("aten.view_copy.default");
ValueRef mat1_W_packed = mat1;
ValueRef out_W_packed = out;
// Create temporary tensors to store the width packed versions of mat1 and out
TmpTensor mat1_tmp(
&graph, graph.sizes_of(mat1), graph.dtype_of(mat1), utils::kWidthPacked);
TmpTensor out_tmp(
&graph, graph.sizes_of(out), graph.dtype_of(out), utils::kWidthPacked);
if (!graph.is_buffer_storage(out) &&
graph.packed_dim_of(mat1) != WHCN::kWidthDim) {
// Ensure mat1 is width packed
mat1_W_packed = graph.add_tensor_like(mat1, utils::kWidthPacked);
mat1_W_packed = mat1_tmp;
viewFn(graph, {mat1, graph.add_none(), mat1_W_packed});
// Ensure out is packed correctly
out_W_packed = graph.add_tensor_like(out, utils::kWidthPacked);
out_W_packed = out_tmp;
}
ValueRef q_mat2 = prepack_standard(
graph, q_mat2_data, graph.storage_type_of(out), utils::kWidthPacked);
Expand Down