Skip to content

Add circular tiling support to conv2d and pad, for Vulkan, CUDA, and CPU (used for making seamless textures) #16985

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
wants to merge 5 commits into
base: master
Choose a base branch
from
Open

Add circular tiling support to conv2d and pad, for Vulkan, CUDA, and CPU (used for making seamless textures) #16985

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
f6ac084
Feat: Added vulkan circular tiling support
Phylliida Nov 3, 2025
d7f5958
Feat: Added cpu circular
Phylliida Nov 3, 2025
1b62b49
Feat: Added cuda kernels
Phylliida Nov 3, 2025
60bed3b
Added tests
Phylliida Nov 3, 2025
5700a4e
Added tests
Phylliida Nov 3, 2025
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
82 changes: 82 additions & 0 deletions ggml/include/ggml.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1943,6 +1943,18 @@ extern "C" {
int d0, // dilation dimension 0
int d1); // dilation dimension 1


GGML_API struct ggml_tensor * ggml_conv_2d_circular(
Copy link
Collaborator

@jeffbolznv jeffbolznv Nov 4, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'd personally prefer the wrapping to be an option on existing commands (either add an optional parameter to existing functions, or do something like ggml_mul_mat_set_prec to modify it after it's created. But the core maintainers should decide. I just don't want to end up with 2^N different convolution functions as these additional options keep getting added.

Copy link
Author

@Phylliida Phylliida Nov 4, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I don't think an optional parameter is an option (iiuc) since they are c api? (could hack with macros, but not ideal) But I'm open to some state modifying thing if that's what they want

struct ggml_context * ctx,
struct ggml_tensor * a, // convolution kernel
struct ggml_tensor * b, // data
int s0, // stride dimension 0
int s1, // stride dimension 1
int p0, // padding dimension 0
int p1, // padding dimension 1
int d0, // dilation dimension 0
int d1); // dilation dimension 1

GGML_API struct ggml_tensor * ggml_im2col_3d(
struct ggml_context * ctx,
struct ggml_tensor * a,
Expand Down Expand Up @@ -2016,6 +2028,19 @@ extern "C" {
int d0, // dilation dimension 0
int d1); // dilation dimension 1


// depthwise (via im2col and mul_mat)
GGML_API struct ggml_tensor * ggml_conv_2d_dw_circular(
struct ggml_context * ctx,
struct ggml_tensor * a, // convolution kernel
struct ggml_tensor * b, // data
int s0, // stride dimension 0
int s1, // stride dimension 1
int p0, // padding dimension 0
int p1, // padding dimension 1
int d0, // dilation dimension 0
int d1); // dilation dimension 1

// Depthwise 2D convolution
// may be faster than ggml_conv_2d_dw, but not available in all backends
// a: KW KH 1 C convolution kernel
Expand All @@ -2032,12 +2057,35 @@ extern "C" {
int dilation0,
int dilation1);

// Depthwise 2D convolution (on a torus)
// may be faster than ggml_conv_2d_dw, but not available in all backends
// a: KW KH 1 C convolution kernel
// b: W H C N input data
// res: W_out H_out C N
GGML_API struct ggml_tensor * ggml_conv_2d_dw_direct_circular(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
int stride0,
int stride1,
int pad0,
int pad1,
int dilation0,
int dilation1);

GGML_API struct ggml_tensor * ggml_conv_transpose_2d_p0(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
int stride);

// circular (on a torus)
GGML_API struct ggml_tensor * ggml_conv_transpose_2d_p0_circular(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
int stride);

GGML_API struct ggml_tensor * ggml_conv_2d_direct(
struct ggml_context * ctx,
struct ggml_tensor * a, // convolution kernel [KW, KH, IC, OC]
Expand All @@ -2048,6 +2096,17 @@ extern "C" {
int p1, // padding dimension 1
int d0, // dilation dimension 0
int d1); // dilation dimension 1

GGML_API struct ggml_tensor * ggml_conv_2d_direct_circular(
struct ggml_context * ctx,
struct ggml_tensor * a, // convolution kernel [KW, KH, IC, OC]
struct ggml_tensor * b, // input data [W, H, C, N]
int s0, // stride dimension 0
int s1, // stride dimension 1
int p0, // padding dimension 0
int p1, // padding dimension 1
int d0, // dilation dimension 0
int d1); // dilation dimension 1

GGML_API struct ggml_tensor * ggml_conv_3d_direct(
struct ggml_context * ctx,
Expand Down Expand Up @@ -2156,6 +2215,15 @@ extern "C" {
int p2,
int p3);

// pad each dimension with values on the other side of the torus (looping around)
GGML_API struct ggml_tensor * ggml_pad_circular(
struct ggml_context * ctx,
struct ggml_tensor * a,
int p0,
int p1,
int p2,
int p3);

GGML_API struct ggml_tensor * ggml_pad_ext(
struct ggml_context * ctx,
struct ggml_tensor * a,
Expand All @@ -2169,6 +2237,20 @@ extern "C" {
int rp3
);

// circular padding
GGML_API struct ggml_tensor * ggml_pad_ext_circular(
struct ggml_context * ctx,
struct ggml_tensor * a,
int lp0,
int rp0,
int lp1,
int rp1,
int lp2,
int rp2,
int lp3,
int rp3
);

// pad each dimension with reflection: [a, b, c, d] -> [b, a, b, c, d, c]
GGML_API struct ggml_tensor * ggml_pad_reflect_1d(
struct ggml_context * ctx,
Expand Down
Loading