Skip to content

Use fp32 FMA in channelwise conv #4809

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

Draft
klin2024 wants to merge 1 commit into
base: develop
Choose a base branch
from
+4 −4
Draft

Use fp32 FMA in channelwise conv #4809

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
8 changes: 4 additions & 4 deletions src/targets/gpu/kernels/include/migraphx/kernels/channelwise_conv.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -53,19 +53,19 @@ channelwise_conv(TileLens, Padding, F f, Output output, Input x, Weights w, Inpu
auto xs_pack = pack(tiler.slice(inputs)...);

using type = typename Output::type;
array<type, decltype(w_ch.get_shape().elements()){}> wregs_arr;
array<float, decltype(w_ch.get_shape().elements()){}> wregs_arr;
auto wregs = make_tensor_view(wregs_arr.begin(), make_packed_shape(w_ch.get_shape()));
copy(w_ch.begin(), w_ch.end(), wregs.begin());

__syncthreads();

tiler.for_each([&](auto out_pos, auto out_multi) {
type acc = 0;
float acc = 0;
repeat(wregs.get_shape().elements(), [&](auto ki) {
auto k_multi = wregs.get_shape().multi(ki);
acc += x_ch[out_multi + k_multi] * wregs[k_multi];
acc += static_cast<float>(x_ch[out_multi + k_multi]) * wregs[k_multi];
});
xs_pack([&](auto... xs) { out_ch[out_pos] = f(acc, xs[out_pos]...); });
xs_pack([&](auto... xs) { out_ch[out_pos] = f(static_cast<type>(acc), xs[out_pos]...); });
});
}

Expand Down
Loading