From b0786ca8768fb59cb595ab7ba6b08a7f5b5a8361 Mon Sep 17 00:00:00 2001
From: pikaliov <pikaliov@gmail.com>
Date: Tue, 12 May 2026 12:41:36 +0300
Subject: [PATCH] Add fork README (mirror, contact, install) and ignore built
 DCNv4 .so.

Document the Gitea mirror for CVGL; extend .gitignore for extension artifacts.
Keep CUDA im2col/col2im and setup version tweaks in DCNv4_op.

Co-authored-by: Cursor <cursoragent@cursor.com>
---
 .gitignore                              |  4 +-
 DCNv4_op/setup.py                       |  2 +-
 DCNv4_op/src/cuda/dcnv4_col2im_cuda.cuh | 68 ++++++++++++++++++++++++-
 DCNv4_op/src/cuda/dcnv4_im2col_cuda.cuh | 30 +++++++++--
 README.md                               | 19 +++++++
 5 files changed, 116 insertions(+), 7 deletions(-)

diff --git a/.gitignore b/.gitignore
index 0ad7194..9225ddf 100644
--- a/.gitignore
+++ b/.gitignore
@@ -23,4 +23,6 @@ detection/ckpts
 segmentation/data
 segmentation/ckpts
 work_dirs/
-output
\ No newline at end of file
+output
+*.so
+DCNv4_op/DCNv4/*.so
\ No newline at end of file
diff --git a/DCNv4_op/setup.py b/DCNv4_op/setup.py
index 082dbeb..0ecf6ef 100644
--- a/DCNv4_op/setup.py
+++ b/DCNv4_op/setup.py
@@ -62,7 +62,7 @@ def get_extensions():
 
 setup(
     name="DCNv4",
-    version="1.0.0.post2",
+    version="1.0.0.post3+cvgl_k57",
     author="Yuwen Xiong, Feng Wang",
     url="",
     description="PyTorch Wrapper for CUDA Functions of DCNv4",
diff --git a/DCNv4_op/src/cuda/dcnv4_col2im_cuda.cuh b/DCNv4_op/src/cuda/dcnv4_col2im_cuda.cuh
index 78bb8dc..d29249e 100644
--- a/DCNv4_op/src/cuda/dcnv4_col2im_cuda.cuh
+++ b/DCNv4_op/src/cuda/dcnv4_col2im_cuda.cuh
@@ -410,8 +410,40 @@ void _dcnv4_col2im_cuda(
         kernel = backward_kernel_dcn<scalar_t, d_stride, stride_type, 1, 8, true>;
       }
       break;
+    case 25:
+      if(check_backward_warpp(d_stride, D)){
+        kernel = backward_kernel_dcn_warp_primitive<scalar_t, d_stride, stride_type, 1, 25, true>;
+      }
+      else {
+        kernel = backward_kernel_dcn<scalar_t, d_stride, stride_type, 1, 25, true>;
+      }
+      break;
+    case 24:
+      if(check_backward_warpp(d_stride, D)){
+        kernel = backward_kernel_dcn_warp_primitive<scalar_t, d_stride, stride_type, 1, 24, true>;
+      }
+      else {
+        kernel = backward_kernel_dcn<scalar_t, d_stride, stride_type, 1, 24, true>;
+      }
+      break;
+    case 49:
+      if(check_backward_warpp(d_stride, D)){
+        kernel = backward_kernel_dcn_warp_primitive<scalar_t, d_stride, stride_type, 1, 49, true>;
+      }
+      else {
+        kernel = backward_kernel_dcn<scalar_t, d_stride, stride_type, 1, 49, true>;
+      }
+      break;
+    case 48:
+      if(check_backward_warpp(d_stride, D)){
+        kernel = backward_kernel_dcn_warp_primitive<scalar_t, d_stride, stride_type, 1, 48, true>;
+      }
+      else {
+        kernel = backward_kernel_dcn<scalar_t, d_stride, stride_type, 1, 48, true>;
+      }
+      break;
     default:
-      printf("K=%ld\n", K);
+      printf("K=%d\n", K);
       throw std::invalid_argument("invalid kernel shape");
     }
   } else {
@@ -432,8 +464,40 @@ void _dcnv4_col2im_cuda(
         kernel = backward_kernel_dcn<scalar_t, d_stride, stride_type, 1, 8, false>;
       }
       break;
+    case 25:
+      if(check_backward_warpp(d_stride, D)){
+        kernel = backward_kernel_dcn_warp_primitive<scalar_t, d_stride, stride_type, 1, 25, false>;
+      }
+      else {
+        kernel = backward_kernel_dcn<scalar_t, d_stride, stride_type, 1, 25, false>;
+      }
+      break;
+    case 24:
+      if(check_backward_warpp(d_stride, D)){
+        kernel = backward_kernel_dcn_warp_primitive<scalar_t, d_stride, stride_type, 1, 24, false>;
+      }
+      else {
+        kernel = backward_kernel_dcn<scalar_t, d_stride, stride_type, 1, 24, false>;
+      }
+      break;
+    case 49:
+      if(check_backward_warpp(d_stride, D)){
+        kernel = backward_kernel_dcn_warp_primitive<scalar_t, d_stride, stride_type, 1, 49, false>;
+      }
+      else {
+        kernel = backward_kernel_dcn<scalar_t, d_stride, stride_type, 1, 49, false>;
+      }
+      break;
+    case 48:
+      if(check_backward_warpp(d_stride, D)){
+        kernel = backward_kernel_dcn_warp_primitive<scalar_t, d_stride, stride_type, 1, 48, false>;
+      }
+      else {
+        kernel = backward_kernel_dcn<scalar_t, d_stride, stride_type, 1, 48, false>;
+      }
+      break;
     default:
-      printf("K=%ld\n", K);
+      printf("K=%d\n", K);
       throw std::invalid_argument("invalid kernel shape");
     }
   }
diff --git a/DCNv4_op/src/cuda/dcnv4_im2col_cuda.cuh b/DCNv4_op/src/cuda/dcnv4_im2col_cuda.cuh
index 3f460f2..b09d337 100644
--- a/DCNv4_op/src/cuda/dcnv4_im2col_cuda.cuh
+++ b/DCNv4_op/src/cuda/dcnv4_im2col_cuda.cuh
@@ -279,8 +279,20 @@ void _dcnv4_im2col_cuda(cudaStream_t stream,
     case 8:
       kernel = forward_kernel_dcn_reg<scalar_t, d_stride, stride_type, 1, 8, true>;
       break;
+    case 25:
+      kernel = forward_kernel_dcn_reg<scalar_t, d_stride, stride_type, 1, 25, true>;
+      break;
+    case 24:
+      kernel = forward_kernel_dcn_reg<scalar_t, d_stride, stride_type, 1, 24, true>;
+      break;
+    case 49:
+      kernel = forward_kernel_dcn_reg<scalar_t, d_stride, stride_type, 1, 49, true>;
+      break;
+    case 48:
+      kernel = forward_kernel_dcn_reg<scalar_t, d_stride, stride_type, 1, 48, true>;
+      break;
     default:
-      printf("K=%ld\n", K);
+      printf("K=%d\n", K);
       throw std::invalid_argument("invalid kernel shape");
     }
   } else {
@@ -290,9 +302,21 @@ void _dcnv4_im2col_cuda(cudaStream_t stream,
       break;
     case 8:
       kernel = forward_kernel_dcn_reg<scalar_t, d_stride, stride_type, 1, 8, false>;
-    break;
+      break;
+    case 25:
+      kernel = forward_kernel_dcn_reg<scalar_t, d_stride, stride_type, 1, 25, false>;
+      break;
+    case 24:
+      kernel = forward_kernel_dcn_reg<scalar_t, d_stride, stride_type, 1, 24, false>;
+      break;
+    case 49:
+      kernel = forward_kernel_dcn_reg<scalar_t, d_stride, stride_type, 1, 49, false>;
+      break;
+    case 48:
+      kernel = forward_kernel_dcn_reg<scalar_t, d_stride, stride_type, 1, 48, false>;
+      break;
     default:
-      printf("K=%ld\n", K);
+      printf("K=%d\n", K);
       throw std::invalid_argument("invalid kernel shape");
     }
   }
diff --git a/README.md b/README.md
index ae7d1d1..1987f55 100644
--- a/README.md
+++ b/README.md
@@ -1,4 +1,23 @@
+## Зеркало и сборка (Pikaliov / CVGL)
 
+**Автор зеркала:** Пикалиов (Pikaliov) · `i@pikaliov.ru`  
+**Репозиторий (Gitea):** https://git.lissad.keenetic.name/Pikaliov/DCN_custom_op  
+**Апстрим OpenGVLab:** https://github.com/OpenGVLab/DCNv4
+
+Каталог **`DCNv4_op`** — PyTorch CUDA-расширение (модуль `DCNv4`). Сборка из исходников:
+
+```bash
+cd DCNv4_op
+pip install -e .
+```
+
+Нужны **CUDA** и согласованные **PyTorch / nvcc** (в `setup.py` при отсутствии CUDA сборка не выполняется). Пример импорта: `from DCNv4 import DCNv4`.
+
+Лицензия исходного кода — **Apache-2.0**, см. файл `LICENSE`.
+
+Ниже — оригинальный README проекта OpenGVLab (модели, таблицы, ссылки на веса).
+
+---
 
 # [DCNv4](https://arxiv.org/pdf/2401.06197.pdf)