Skip to content
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

Jump to bottom

[Don't merge] Refactor grouped convolution #1255

Open
wants to merge 2 commits into
base: develop
Choose a base branch
from
+510 −5,993
Open

[Don't merge] Refactor grouped convolution #1255

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
74c655f
Refactor grouped convolution
bartekxk Apr 22, 2024
3635e7a
Merge branch 'develop' into barkocot/lwpck-1572
bartekxk Apr 22, 2024
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
34 changes: 20 additions & 14 deletions client_example/07_grouped_convnd_fwd/grouped_conv1d_fwd.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,18 +17,24 @@ using InDataType = ck::half_t;
using WeiDataType = ck::half_t;
using OutDataType = ck::half_t;

using InLayout = ck::tensor_layout::convolution::GNWC;
using WeiLayout = ck::tensor_layout::convolution::GKXC;
using OutLayout = ck::tensor_layout::convolution::GNWK;
using InLayout = ck::tensor_layout::convolution::GNDHWC;
using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
using OutLayout = ck::tensor_layout::convolution::GNDHWK;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;

static constexpr ck::index_t NumDimSpatial = 1;
static constexpr ck::index_t NumDimSpatial = 3;
static constexpr ck::index_t G = 32;
static constexpr ck::index_t N = 256;
static constexpr ck::index_t K = 192;
static constexpr ck::index_t C = 192;
static constexpr ck::index_t Z = 1; // Z == 1 for 1d
static constexpr ck::index_t Y = 1; // Y == 1 for 1d
static constexpr ck::index_t X = 3;
static constexpr ck::index_t Di = 1; // Di = 1 for 1d
static constexpr ck::index_t Hi = 1; // Hi = 1 for 1d
static constexpr ck::index_t Wi = 28;
static constexpr ck::index_t Do = 1; // Do = 1 for 1d
static constexpr ck::index_t Ho = 1; // Ho = 1 for 1d
static constexpr ck::index_t Wo = 28;

struct SimpleDeviceMem
Expand All @@ -49,14 +55,14 @@ struct SimpleDeviceMem

int main()
{
std::array<ck::index_t, NumDimSpatial + 3> in_lengths{G, N, Wi, C};
std::array<ck::index_t, NumDimSpatial + 3> in_strides{0, 0, 0, 1};
std::array<ck::index_t, NumDimSpatial + 3> in_lengths{G, N, Di, Hi, Wi, C};
std::array<ck::index_t, NumDimSpatial + 3> in_strides{0, 0, 0, 0, 0, 1};

std::array<ck::index_t, NumDimSpatial + 3> wei_lengths{G, K, X, C};
std::array<ck::index_t, NumDimSpatial + 3> wei_strides{0, 0, 0, 1};
std::array<ck::index_t, NumDimSpatial + 3> wei_lengths{G, K, Z, Y, X, C};
std::array<ck::index_t, NumDimSpatial + 3> wei_strides{0, 0, 0, 0, 0, 1};

std::array<ck::index_t, NumDimSpatial + 3> out_lengths{G, N, Wo, K};
std::array<ck::index_t, NumDimSpatial + 3> out_strides{0, 0, 0, 1};
std::array<ck::index_t, NumDimSpatial + 3> out_lengths{G, N, Do, Ho, Wo, K};
std::array<ck::index_t, NumDimSpatial + 3> out_strides{0, 0, 0, 0, 0, 1};

std::partial_sum(rbegin(in_lengths),
std::prev(rend(in_lengths)),
Expand Down Expand Up @@ -91,10 +97,10 @@ int main()
std::next(rbegin(out_strides)),
std::next(rbegin(out_strides), NumDimSpatial + 1));

std::array<ck::index_t, NumDimSpatial> filter_strides{1};
std::array<ck::index_t, NumDimSpatial> filter_dilations{1};
std::array<ck::index_t, NumDimSpatial> input_left_pads{1};
std::array<ck::index_t, NumDimSpatial> input_right_pads{1};
std::array<ck::index_t, NumDimSpatial> filter_strides{1, 1, 1};
std::array<ck::index_t, NumDimSpatial> filter_dilations{1, 1, 1};
std::array<ck::index_t, NumDimSpatial> input_left_pads{0, 0, 1};
std::array<ck::index_t, NumDimSpatial> input_right_pads{0, 0, 1};

SimpleDeviceMem in(sizeof(InDataType) * G * N * Wi * C);
SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * X * C);
Expand Down
35 changes: 20 additions & 15 deletions client_example/07_grouped_convnd_fwd/grouped_conv2d_fwd.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,20 +17,23 @@ using InDataType = ck::half_t;
using WeiDataType = ck::half_t;
using OutDataType = ck::half_t;

using InLayout = ck::tensor_layout::convolution::NHWGC;
using WeiLayout = ck::tensor_layout::convolution::GKYXC;
using OutLayout = ck::tensor_layout::convolution::NHWGK;
using InLayout = ck::tensor_layout::convolution::NDHWGC;
using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
using OutLayout = ck::tensor_layout::convolution::NDHWGK;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;

static constexpr ck::index_t NumDimSpatial = 2;
static constexpr ck::index_t NumDimSpatial = 3;
static constexpr ck::index_t G = 32;
static constexpr ck::index_t N = 256; // batch size
static constexpr ck::index_t K = 64; // output channel
static constexpr ck::index_t C = 32; // input channel (per group)
static constexpr ck::index_t Z = 1; // filter D == 1 for 2d
static constexpr ck::index_t Y = 3; // filter H
static constexpr ck::index_t X = 3; // filter W
static constexpr ck::index_t Di = 1; // input D = 1 for 2d
static constexpr ck::index_t Hi = 28; // input H
static constexpr ck::index_t Wi = 28; // input W
static constexpr ck::index_t Do = 1; // output D = 1 for 2d
static constexpr ck::index_t Ho = 28; // output H
static constexpr ck::index_t Wo = 28; // output W

Expand All @@ -55,17 +58,19 @@ int main()
// We have NHWGC/GKYXC/NHWGK (x, weight, y) in memory space
// However, CK's API only accept length and stride with order of GNCHW/GKCYX/GNCHW
// Hence, we need to adjust the order of stride
std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
std::array<ck::index_t, 5> in_strides{C, Hi * Wi * G * C, 1, Wi * G * C, G * C};
std::array<ck::index_t, 5> wei_lengths{G, K, C, Y, X};
std::array<ck::index_t, 5> wei_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
std::array<ck::index_t, 5> out_strides{C, Ho * Wo * G * C, 1, Wo * G * C, G * C};

std::array<ck::index_t, NumDimSpatial> filter_strides{1, 1};
std::array<ck::index_t, NumDimSpatial> filter_dilations{1, 1};
std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1};
std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1};
std::array<ck::index_t, 6> in_lengths{G, N, C, Di, Hi, Wi};
std::array<ck::index_t, 6> in_strides{
C, Hi * Wi * G * C, 1, Hi * Wi * G * C, Wi * G * C, G * C};
std::array<ck::index_t, 6> wei_lengths{G, K, C, Z, Y, X};
std::array<ck::index_t, 6> wei_strides{K * Y * X * C, Y * X * C, 1, Y * X * C, X * C, C};
std::array<ck::index_t, 6> out_lengths{G, N, K, Do, Ho, Wo};
std::array<ck::index_t, 6> out_strides{
C, Ho * Wo * G * C, 1, Ho * Wo * G * C, Wo * G * C, G * C};

std::array<ck::index_t, NumDimSpatial> filter_strides{1, 1, 1};
std::array<ck::index_t, NumDimSpatial> filter_dilations{1, 1, 1};
std::array<ck::index_t, NumDimSpatial> input_left_pads{0, 1, 1};
std::array<ck::index_t, NumDimSpatial> input_right_pads{0, 1, 1};

SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
Expand Down
83 changes: 31 additions & 52 deletions client_example/10_grouped_convnd_bwd_data/grouped_conv2d_bwd_data.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,22 +18,25 @@ using InDataType = ck::half_t;
using WeiDataType = ck::half_t;
using OutDataType = ck::half_t;

using InLayout = ck::tensor_layout::convolution::GNHWC;
using WeiLayout = ck::tensor_layout::convolution::GKYXC;
using OutLayout = ck::tensor_layout::convolution::GNHWK;
using InLayout = ck::tensor_layout::convolution::NDHWGC;
using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
using OutLayout = ck::tensor_layout::convolution::NDHWGK;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;

static constexpr ck::index_t NumDimSpatial = 2;
static constexpr ck::index_t G = 32;
static constexpr ck::index_t N = 256;
static constexpr ck::index_t K = 192;
static constexpr ck::index_t C = 192;
static constexpr ck::index_t NumDimSpatial = 3;
static constexpr ck::index_t G = 2;
static constexpr ck::index_t N = 16;
static constexpr ck::index_t K = 16;
static constexpr ck::index_t C = 16;
static constexpr ck::index_t Z = 1; // Z == 1 for 2d
static constexpr ck::index_t Y = 3;
static constexpr ck::index_t X = 3;
static constexpr ck::index_t Hi = 28;
static constexpr ck::index_t Wi = 28;
static constexpr ck::index_t Ho = 28;
static constexpr ck::index_t Wo = 28;
static constexpr ck::index_t Di = 1; // Di == 1 for 2d
static constexpr ck::index_t Hi = 14;
static constexpr ck::index_t Wi = 14;
static constexpr ck::index_t Do = 1; // Do == 1 for 2d
static constexpr ck::index_t Ho = 14;
static constexpr ck::index_t Wo = 14;

struct SimpleDeviceMem
{
Expand All @@ -53,46 +56,22 @@ struct SimpleDeviceMem

int main()
{
std::array<ck::index_t, NumDimSpatial + 3> in_lengths{G, N, Hi, Wi, C};
std::array<ck::index_t, NumDimSpatial + 3> in_strides{0, 0, 0, 0, 1};

std::array<ck::index_t, NumDimSpatial + 3> wei_lengths{G, K, Y, X, C};
std::array<ck::index_t, NumDimSpatial + 3> wei_strides{0, 0, 0, 0, 1};

std::array<ck::index_t, NumDimSpatial + 3> out_lengths{G, N, Ho, Wo, K};
std::array<ck::index_t, NumDimSpatial + 3> out_strides{0, 0, 0, 0, 1};

std::partial_sum(rbegin(in_lengths),
std::prev(rend(in_lengths)),
std::next(rbegin(in_strides)),
std::multiplies<>{});
std::partial_sum(rbegin(wei_lengths),
std::prev(rend(wei_lengths)),
std::next(rbegin(wei_strides)),
std::multiplies<>{});
std::partial_sum(rbegin(out_lengths),
std::prev(rend(out_lengths)),
std::next(rbegin(out_strides)),
std::multiplies<>{});

// transpose GNHWC/GKYXC/GNHWK to GNCHW/GKCYX/GNCHW
std::rotate(
rbegin(in_lengths), std::next(rbegin(in_lengths)), std::next(rbegin(in_lengths), 3));
std::rotate(
rbegin(in_strides), std::next(rbegin(in_strides)), std::next(rbegin(in_strides), 3));
std::rotate(
rbegin(wei_lengths), std::next(rbegin(wei_lengths)), std::next(rbegin(wei_lengths), 3));
std::rotate(
rbegin(wei_strides), std::next(rbegin(wei_strides)), std::next(rbegin(wei_strides), 3));
std::rotate(
rbegin(out_lengths), std::next(rbegin(out_lengths)), std::next(rbegin(out_lengths), 3));
std::rotate(
rbegin(out_strides), std::next(rbegin(out_strides)), std::next(rbegin(out_strides), 3));

std::array<ck::index_t, NumDimSpatial> filter_strides{1, 1};
std::array<ck::index_t, NumDimSpatial> filter_dilations{1, 1};
std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1};
std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1};
std::array<ck::index_t, NumDimSpatial + 3> in_lengths{G, N, C, Di, Hi, Wi};
std::array<ck::index_t, NumDimSpatial + 3> in_strides{
C, Di * Hi * Wi * G * C, 1, Hi * Wi * G * C, Wi * G * C, G * C};

std::array<ck::index_t, NumDimSpatial + 3> wei_lengths{G, K, C, Z, Y, X};
std::array<ck::index_t, NumDimSpatial + 3> wei_strides{
K * Z * Y * X * C, Z * Y * X * C, 1, Y * X * C, X * C, C};

std::array<ck::index_t, NumDimSpatial + 3> out_lengths{G, N, K, Do, Ho, Wo};
std::array<ck::index_t, NumDimSpatial + 3> out_strides{
K, Do * Ho * Wo * G * K, 1, Ho * Wo * G * K, Wo * G * K, G * K};

std::array<ck::index_t, NumDimSpatial> filter_strides{1, 1, 1};
std::array<ck::index_t, NumDimSpatial> filter_dilations{1, 1, 1};
std::array<ck::index_t, NumDimSpatial> input_left_pads{0, 1, 1};
std::array<ck::index_t, NumDimSpatial> input_right_pads{0, 1, 1};

SimpleDeviceMem in(sizeof(InDataType) * G * N * Hi * Wi * C);
SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
Expand Down
45 changes: 27 additions & 18 deletions client_example/11_grouped_conv_bwd_weight/grouped_conv1d_bwd_weight_fp16.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,28 +10,37 @@ using InDataType = ck::half_t;
using WeiDataType = ck::half_t;
using OutDataType = ck::half_t;

using InLayout = ck::tensor_layout::convolution::GNWC;
using WeiLayout = ck::tensor_layout::convolution::GKXC;
using OutLayout = ck::tensor_layout::convolution::GNWK;
using InLayout = ck::tensor_layout::convolution::GNDHWC;
using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
using OutLayout = ck::tensor_layout::convolution::GNDHWK;

static constexpr ck::index_t NumDimSpatial = 1;
static constexpr ck::index_t G = 32;
static constexpr ck::index_t N = 256;
static constexpr ck::index_t K = 192;
static constexpr ck::index_t C = 192;
static constexpr ck::index_t NumDimSpatial = 3;
static constexpr ck::index_t G = 8;
static constexpr ck::index_t N = 64;
static constexpr ck::index_t K = 128;
static constexpr ck::index_t C = 128;
static constexpr ck::index_t Z = 1; // Z == 1 for 1d
static constexpr ck::index_t Y = 1; // Y == 1 for 1d
static constexpr ck::index_t X = 3;
static constexpr ck::index_t Di = 1; // Di == 1 for 1d
static constexpr ck::index_t Hi = 1; // Hi == 1 for 1d
static constexpr ck::index_t Wi = 28;
static constexpr ck::index_t Do = 1; // Do == 1 for 1d
static constexpr ck::index_t Ho = 1; // Ho == 1 for 1d
static constexpr ck::index_t Wo = 28;
static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_lengths{G, N, C, Wi};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> filter_lengths{G, K, C, X};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_lengths{G, N, K, Wo};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_strides{N * Wi * C, Wi* C, 1, C};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> weights_strides{K * X * C, X* C, 1, C};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_strides{N * Wo * K, Wo* K, 1, K};
static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_strides{1};
static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_dilations{1};
static constexpr std::array<ck::index_t, NumDimSpatial> input_left_pads{1};
static constexpr std::array<ck::index_t, NumDimSpatial> input_right_pads{1};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_lengths{G, N, C, Di, Hi, Wi};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> filter_lengths{G, K, C, Z, Y, X};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_lengths{G, N, K, Do, Ho, Wo};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_strides{
N * Di * Hi * Wi * C, Di* Hi* Wi* C, 1, Hi* Wi* C, Wi* C, C};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> weights_strides{
K * Z * Y * X * C, Z* Y* X* C, 1, Y* X* C, X* C, C};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_strides{
N * Do * Ho * Wo * K, Do* Ho* Wo* K, 1, Ho* Wo* K, Wo* K, K};
static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_strides{1, 1, 1};
static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_dilations{1, 1, 1};
static constexpr std::array<ck::index_t, NumDimSpatial> input_left_pads{0, 0, 1};
static constexpr std::array<ck::index_t, NumDimSpatial> input_right_pads{0, 0, 1};

int main()
{
Expand Down
39 changes: 21 additions & 18 deletions client_example/11_grouped_conv_bwd_weight/grouped_conv2d_bwd_weight_fp16.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,34 +10,37 @@ using InDataType = ck::half_t;
using WeiDataType = ck::half_t;
using OutDataType = ck::half_t;

using InLayout = ck::tensor_layout::convolution::GNHWC;
using WeiLayout = ck::tensor_layout::convolution::GKYXC;
using OutLayout = ck::tensor_layout::convolution::GNHWK;
using InLayout = ck::tensor_layout::convolution::GNDHWC;
using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
using OutLayout = ck::tensor_layout::convolution::GNDHWK;

static constexpr ck::index_t NumDimSpatial = 2;
static constexpr ck::index_t G = 32;
static constexpr ck::index_t N = 256;
static constexpr ck::index_t K = 192;
static constexpr ck::index_t C = 192;
static constexpr ck::index_t NumDimSpatial = 3;
static constexpr ck::index_t G = 8;
static constexpr ck::index_t N = 64;
static constexpr ck::index_t K = 128;
static constexpr ck::index_t C = 128;
static constexpr ck::index_t Z = 1; // Z == 1 for 1d
static constexpr ck::index_t Y = 3;
static constexpr ck::index_t X = 3;
static constexpr ck::index_t Di = 1; // Di == 1 for 1d
static constexpr ck::index_t Hi = 28;
static constexpr ck::index_t Wi = 28;
static constexpr ck::index_t Do = 1; // Do == 1 for 1d
static constexpr ck::index_t Ho = 28;
static constexpr ck::index_t Wo = 28;
static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_lengths{G, N, C, Hi, Wi};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> filter_lengths{G, K, C, Y, X};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_lengths{G, N, K, Ho, Wo};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_lengths{G, N, C, Di, Hi, Wi};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> filter_lengths{G, K, C, Z, Y, X};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_lengths{G, N, K, Do, Ho, Wo};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_strides{
N * Hi * Wi * C, Hi* Wi* C, 1, Wi* C, C};
N * Di * Hi * Wi * C, Di* Hi* Wi* C, 1, Hi* Wi* C, Wi* C, C};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> weights_strides{
K * Y * X * C, Y* X* C, 1, X* C, C};
K * Z * Y * X * C, Z* Y* X* C, 1, Y* X* C, X* C, C};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_strides{
N * Ho * Wo * K, Ho* Wo* K, 1, Wo* K, K};
static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_strides{1, 1};
static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_dilations{1, 1};
static constexpr std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1};
static constexpr std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1};
N * Do * Ho * Wo * K, Do* Ho* Wo* K, 1, Ho* Wo* K, Wo* K, K};
static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_strides{1, 1, 1};
static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_dilations{1, 1, 1};
static constexpr std::array<ck::index_t, NumDimSpatial> input_left_pads{0, 1, 1};
static constexpr std::array<ck::index_t, NumDimSpatial> input_right_pads{0, 1, 1};

int main()
{
Expand Down
9 changes: 7 additions & 2 deletions include/ck/tensor_operation/gpu/device/impl/device_column_to_image_impl.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.

#pragma once

Expand Down Expand Up @@ -66,7 +66,12 @@ struct DeviceColumnToImageImpl
static constexpr auto spatial_offset = Number<3>{};

static constexpr auto conv_to_gemm_transformer =
TransformConvFwdToGemm<NDimSpatial, ConvolutionForwardSpecialization::Default>{};
TransformConvFwdToGemm<NDimSpatial,
MPerBlock,
0 /* NPerBlock*/,
KPerBlock,
GemmSpecialization::MKPadding,
ConvolutionForwardSpecialization::Default>{};
static constexpr auto matrix_padder =
MatrixPadder<GemmSpecialization::MKPadding, index_t, index_t, index_t>{
MPerBlock, 0 /* NPerBlock*/, KPerBlock};
Expand Down
Loading