Skip to content

Fix NHWC depthwise conv weight layout and dimension indexing for HiFi operators #17423

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
JakeStevens wants to merge 1 commit into
base: main
Choose a base branch
from
+66 −25
Open

Fix NHWC depthwise conv weight layout and dimension indexing for HiFi operators #17423

Show file tree
Hide file tree
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
29 changes: 21 additions & 8 deletions .../hifi/operators/op_quantized_conv2d_nhwc_depthwise_asym8sxsym8s_asym8s_per_tensor_out.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -48,14 +48,15 @@ void xa_opt_quantized_conv2d_nhwc_depthwise_asym8sxsym8s_asym8s(
WORD32* __restrict__ p_bias =
(WORD32* __restrict__)bias.const_data_ptr<int32_t>();

WORD32 input_height = conv1d ? 1 : input.size(2);
WORD32 input_width = conv1d ? input.size(2) : input.size(3);
WORD32 input_channels = input.size(1);
WORD32 kernel_height = conv1d ? 1 : weight.size(2);
WORD32 kernel_width = conv1d ? weight.size(2) : weight.size(3);
// NHWC layout: 4D=[N,H,W,C], 3D=[N,W,C]
WORD32 input_height = conv1d ? 1 : input.size(1);
WORD32 input_width = conv1d ? input.size(1) : input.size(2);
WORD32 input_channels = conv1d ? input.size(2) : input.size(3);
WORD32 kernel_height = conv1d ? 1 : weight.size(1);
WORD32 kernel_width = conv1d ? weight.size(1) : weight.size(2);
WORD32 out_channels = weight.size(0);
WORD32 out_height = conv1d ? 1 : out.size(2);
WORD32 out_width = conv1d ? out.size(2) : out.size(3);
WORD32 out_height = conv1d ? 1 : out.size(1);
WORD32 out_width = conv1d ? out.size(1) : out.size(2);
WORD32 batches = input.size(0);

WORD32 x_stride = stride[1];
Expand All @@ -79,6 +80,18 @@ void xa_opt_quantized_conv2d_nhwc_depthwise_asym8sxsym8s_asym8s(
out_shift32[i] = 0;
}

// Rearrange weight from [OC, KH, KW, IC/G] (graph NHWC format) to
// [KH, KW, OC] (NNLib HWC format expected for inp_data_format=0).
// For depthwise IC/G=1, so this is a transpose of [OC, KH*KW] to [KH*KW, OC].
WORD32 kernel_size = kernel_height * kernel_width;
WORD32 weight_size = out_channels * kernel_size;
WORD8* p_kernel_hwc = (WORD8*)kernels::allocate_temp_memory(ctx, weight_size);
for (int oc = 0; oc < out_channels; oc++) {
for (int k = 0; k < kernel_size; k++) {
p_kernel_hwc[k * out_channels + oc] = p_kernel[oc * kernel_size + k];
}
}

WORD32 scratch_size = xa_nn_conv2d_depthwise_getsize(
input_height,
input_width,
Expand Down Expand Up @@ -107,7 +120,7 @@ void xa_opt_quantized_conv2d_nhwc_depthwise_asym8sxsym8s_asym8s(

xa_nn_conv2d_depthwise_per_chan_sym8sxasym8s(
out_batch,
p_kernel,
p_kernel_hwc,
in_batch,
p_bias,
input_height,
Expand Down
30 changes: 22 additions & 8 deletions .../hifi/operators/op_quantized_conv2d_nhwc_depthwise_asym8uxsym8u_asym8u_per_tensor_out.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -48,14 +48,15 @@ void xa_opt_quantized_conv2d_nhwc_depthwise_asym8uxsym8u_asym8u(
WORD32* __restrict__ p_bias =
(WORD32* __restrict__)bias.const_data_ptr<int32_t>();

WORD32 input_height = conv1d ? 1 : input.size(2);
WORD32 input_width = conv1d ? input.size(2) : input.size(3);
WORD32 input_channels = input.size(1);
WORD32 kernel_height = conv1d ? 1 : weight.size(2);
WORD32 kernel_width = conv1d ? weight.size(2) : weight.size(3);
// NHWC layout: 4D=[N,H,W,C], 3D=[N,W,C]
WORD32 input_height = conv1d ? 1 : input.size(1);
WORD32 input_width = conv1d ? input.size(1) : input.size(2);
WORD32 input_channels = conv1d ? input.size(2) : input.size(3);
WORD32 kernel_height = conv1d ? 1 : weight.size(1);
WORD32 kernel_width = conv1d ? weight.size(1) : weight.size(2);
WORD32 out_channels = weight.size(0);
WORD32 out_height = conv1d ? 1 : out.size(2);
WORD32 out_width = conv1d ? out.size(2) : out.size(3);
WORD32 out_height = conv1d ? 1 : out.size(1);
WORD32 out_width = conv1d ? out.size(1) : out.size(2);
WORD32 batches = input.size(0);

WORD32 x_stride = stride[1];
Expand All @@ -79,6 +80,19 @@ void xa_opt_quantized_conv2d_nhwc_depthwise_asym8uxsym8u_asym8u(
out_shift32[i] = 0;
}

// Rearrange weight from [OC, KH, KW, IC/G] (graph NHWC format) to
// [KH, KW, OC] (NNLib HWC format expected for inp_data_format=0).
// For depthwise IC/G=1, so this is a transpose of [OC, KH*KW] to [KH*KW, OC].
WORD32 kernel_size = kernel_height * kernel_width;
WORD32 weight_size = out_channels * kernel_size;
UWORD8* p_kernel_hwc =
(UWORD8*)kernels::allocate_temp_memory(ctx, weight_size);
for (int oc = 0; oc < out_channels; oc++) {
for (int k = 0; k < kernel_size; k++) {
p_kernel_hwc[k * out_channels + oc] = p_kernel[oc * kernel_size + k];
}
}

WORD32 scratch_size = xa_nn_conv2d_depthwise_getsize(
input_height,
input_width,
Expand Down Expand Up @@ -107,7 +121,7 @@ void xa_opt_quantized_conv2d_nhwc_depthwise_asym8uxsym8u_asym8u(

xa_nn_conv2d_depthwise_per_chan_sym8sxasym8s(
(WORD8*)out_batch,
(WORD8*)p_kernel,
(WORD8*)p_kernel_hwc,
(WORD8*)in_batch,
p_bias,
input_height,
Expand Down
32 changes: 23 additions & 9 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
Expand Up @@ -176,15 +176,16 @@ void xa_opt_quantized_conv2d_nhwc(
WORD32* __restrict__ p_bias =
(WORD32* __restrict__)bias.const_data_ptr<int32_t>();

WORD32 input_height = conv1d ? 1 : input.size(2);
WORD32 input_width = conv1d ? input.size(2) : input.size(3);
WORD32 input_channels = input.size(1);
WORD32 kernel_height = conv1d ? 1 : weight.size(2);
WORD32 kernel_width = conv1d ? weight.size(2) : weight.size(3);
WORD32 kernel_channels = weight.size(1);
// NHWC layout: 4D=[N,H,W,C], 3D=[N,W,C]
WORD32 input_height = conv1d ? 1 : input.size(1);
WORD32 input_width = conv1d ? input.size(1) : input.size(2);
WORD32 input_channels = conv1d ? input.size(2) : input.size(3);
WORD32 kernel_height = conv1d ? 1 : weight.size(1);
WORD32 kernel_width = conv1d ? weight.size(1) : weight.size(2);
WORD32 kernel_channels = conv1d ? weight.size(2) : weight.size(3);
WORD32 out_channels = weight.size(0);
WORD32 out_height = conv1d ? 1 : out.size(2);
WORD32 out_width = conv1d ? out.size(2) : out.size(3);
WORD32 out_height = conv1d ? 1 : out.size(1);
WORD32 out_width = conv1d ? out.size(1) : out.size(2);
WORD32 batches = input.size(0);

WORD32 x_stride = stride[1];
Expand Down Expand Up @@ -285,6 +286,19 @@ void xa_opt_quantized_conv2d_nhwc(
if (groups == input_channels) {
WORD32 channels_multiplier = out_channels / input_channels;

// Rearrange weight from [OC, KH, KW, IC/G] (graph NHWC format) to
// [KH, KW, OC] (NNLib HWC format expected for inp_data_format=0).
WORD32 kernel_size_dw = kernel_height * kernel_width;
WORD32 weight_size_dw = out_channels * kernel_size_dw;
WORD8* p_kernel_hwc =
(WORD8*)kernels::allocate_temp_memory(ctx, weight_size_dw);
for (int oc = 0; oc < out_channels; oc++) {
for (int k = 0; k < kernel_size_dw; k++) {
p_kernel_hwc[k * out_channels + oc] =
p_kernel[oc * kernel_size_dw + k];
}
}

scratch_size = xa_nn_conv2d_depthwise_getsize(
input_height,
input_width,
Expand Down Expand Up @@ -322,7 +336,7 @@ void xa_opt_quantized_conv2d_nhwc(

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