Skip to content

Handle Byte scalar type in quantized conv2d NHWC (#18427) #18427

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

Open
khazaei wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Handle Byte scalar type in quantized conv2d NHWC (#18427) #18427

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
1f84837
Fix depthwise conv2d weight layout rearrangement for non-depthwise gr…
Mar 24, 2026
3757d54
Handle Byte scalar type in quantized conv2d NHWC (#18427)
Mar 24, 2026
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
27 changes: 25 additions & 2 deletions backends/cadence/hifi/operators/op_quantized_conv2d_nhwc_out.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,3 +1,3 @@
/*
* Copyright (c) Meta Platforms, Inc. and affiliates.
* All rights reserved.
Expand Down Expand Up @@ -166,7 +166,8 @@
bool conv1d = input.dim() == 3;
constexpr int kNnlibMaxDim = 4;

if (input.scalar_type() == ScalarType::Char) {
if (input.scalar_type() == ScalarType::Char ||
input.scalar_type() == ScalarType::Byte) {
WORD8* __restrict__ p_out =
(WORD8* __restrict__)out.mutable_data_ptr<int8_t>();
WORD8* __restrict__ p_inp =
Expand Down Expand Up @@ -303,6 +304,28 @@
if (groups == input_channels) {
WORD32 channels_multiplier = out_channels / input_channels;

// The NNLib depthwise function expects weights in [KH, KW, OC] layout.
// When is_depthwise is false, the weight is 4D [OC, KH, KW, IC/G] and
// needs to be rearranged from [OC, KH, KW] to [KH, KW, OC].
WORD8* p_kernel_dw = p_kernel;
if (!is_depthwise) {
WORD32 weight_size = kernel_height * kernel_width * out_channels;
p_kernel_dw =
(WORD8*)kernels::allocate_temp_memory(ctx, weight_size);
for (WORD32 oc = 0; oc < out_channels; oc++) {
for (WORD32 kh = 0; kh < kernel_height; kh++) {
for (WORD32 kw = 0; kw < kernel_width; kw++) {
p_kernel_dw
[kh * kernel_width * out_channels + kw * out_channels + oc] =
p_kernel
[oc * kernel_height * kernel_width * kernel_channels +
kh * kernel_width * kernel_channels +
kw * kernel_channels];
}
}
}
}

scratch_size = xa_nn_conv2d_depthwise_getsize(
input_height,
input_width,
Expand Down Expand Up @@ -332,7 +355,7 @@

xa_nn_conv2d_depthwise_per_chan_sym8sxasym8s(
out_batch,
p_kernel,
p_kernel_dw,
in_batch,
p_bias,
input_height,
Expand Down
Loading