pierotofy · pierotofy · Mar 20, 2024 · Mar 11, 2024 · Mar 11, 2024 · Mar 12, 2024
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -6,17 +6,17 @@ set(GPU_RUNTIME "CUDA" CACHE STRING "HIP or CUDA")
 set(OPENCV_DIR "OPENCV_DIR-NOTFOUND" CACHE PATH "Path to the OPENCV installation directory")
 
 if(NOT CMAKE_BUILD_TYPE)
-set(CMAKE_BUILD_TYPE "Release" CACHE STRING "Choose the type of build, options are: Debug Release RelWithDebInfo MinSizeRel." FORCE)
+    set(CMAKE_BUILD_TYPE "Release" CACHE STRING "Choose the type of build, options are: Debug Release RelWithDebInfo MinSizeRel." FORCE)
 endif()
 
-enable_language(${GPU_RUNTIME})
-set(CMAKE_${GPU_RUNTIME}_STANDARD 17)
-set(${GPU_RUNTIME}_STANDARD 17)
-
 if(GPU_RUNTIME STREQUAL "CUDA")
     set(CMAKE_CUDA_ARCHITECTURES 70 75)
-    find_package(CUDAToolkit REQUIRED)
-else()
+    find_package(CUDAToolkit)
+    if (CUDAToolkit-NOTFOUND)
+        message(WARNING "CUDA toolkit not found, building with CPU support only")
+        set(GPU_RUNTIME "CPU")
+    endif()
+elseif(GPU_RUNTIME STREQUAL "HIP")
     set(USE_HIP ON CACHE BOOL "Use HIP for GPU acceleration")
 
     if(NOT DEFINED HIP_PATH)
@@ -40,6 +40,13 @@ else()
     list(APPEND CMAKE_PREFIX_PATH "${ROCM_ROOT}")
 endif()
 
+set(CMAKE_CXX_STANDARD 17)
+if((GPU_RUNTIME STREQUAL "CUDA") OR (GPU_RUNTIME STREQUAL "HIP"))
+    enable_language(${GPU_RUNTIME})
+    set(CMAKE_${GPU_RUNTIME}_STANDARD 17)
+    set(${GPU_RUNTIME}_STANDARD 17)
+endif()
+
 if (NOT WIN32 AND NOT APPLE)
     set(STDPPFS_LIBRARY stdc++fs)
 endif()
@@ -52,38 +59,48 @@ if (NOT WIN32 AND NOT APPLE)
 endif()
 set(OpenCV_LIBS opencv_core opencv_imgproc opencv_highgui opencv_calib3d)
 
-add_library(gsplat vendor/gsplat/forward.cu vendor/gsplat/backward.cu vendor/gsplat/bindings.cu vendor/gsplat/helpers.cuh)
-if(GPU_RUNTIME STREQUAL "CUDA")
-    set(GPU_LIBRARIES "cuda")
-    target_link_libraries(gsplat PUBLIC cuda)
-else(GPU_RUNTIME STREQUAL "HIP")
-    set(GPU_INCLUDE_DIRS "${ROCM_ROOT}/include")
-    target_compile_definitions(gsplat PRIVATE USE_HIP __HIP_PLATFORM_AMD__)
+set(GSPLAT_LIBS gsplat_cpu)
+if((GPU_RUNTIME STREQUAL "CUDA") OR (GPU_RUNTIME STREQUAL "HIP"))
+    add_library(gsplat vendor/gsplat/forward.cu vendor/gsplat/backward.cu vendor/gsplat/bindings.cu vendor/gsplat/helpers.cuh)
+    list(APPEND GSPLAT_LIBS gsplat)
+    if(GPU_RUNTIME STREQUAL "CUDA")
+        set(GPU_LIBRARIES "cuda")
+        target_link_libraries(gsplat PUBLIC cuda)
+        set_target_properties(gsplat PROPERTIES CUDA_ARCHITECTURES "70;75")
+    else(GPU_RUNTIME STREQUAL "HIP")
+        set(GPU_INCLUDE_DIRS "${ROCM_ROOT}/include")
+        target_compile_definitions(gsplat PRIVATE USE_HIP __HIP_PLATFORM_AMD__)
+    endif()
+    target_include_directories(gsplat PRIVATE
+        ${PROJECT_SOURCE_DIR}/vendor/glm        
+        ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES}
+        ${TORCH_INCLUDE_DIRS}
+    )
+    set_target_properties(gsplat PROPERTIES LINKER_LANGUAGE CXX)
 endif()
 
-target_include_directories(gsplat PRIVATE
-    ${PROJECT_SOURCE_DIR}/vendor/glm        
-    ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES}
-    ${TORCH_INCLUDE_DIRS}
-)
-set_target_properties(gsplat PROPERTIES LINKER_LANGUAGE CXX)
-set_target_properties(gsplat PROPERTIES CUDA_ARCHITECTURES "70;75")
+add_library(gsplat_cpu vendor/gsplat-cpu/gsplat_cpu.cpp)
+target_include_directories(gsplat_cpu PRIVATE ${TORCH_INCLUDE_DIRS})
 
 add_executable(opensplat opensplat.cpp point_io.cpp nerfstudio.cpp model.cpp kdtree_tensor.cpp spherical_harmonics.cpp cv_utils.cpp utils.cpp project_gaussians.cpp rasterize_gaussians.cpp ssim.cpp optim_scheduler.cpp colmap.cpp input_data.cpp tensor_math.cpp)
 set_property(TARGET opensplat PROPERTY CXX_STANDARD 17)
 target_include_directories(opensplat PRIVATE ${PROJECT_SOURCE_DIR}/vendor/glm ${GPU_INCLUDE_DIRS})
-target_link_libraries(opensplat PUBLIC ${STDPPFS_LIBRARY} ${GPU_LIBRARIES} gsplat ${TORCH_LIBRARIES} ${OpenCV_LIBS})
+target_link_libraries(opensplat PUBLIC ${STDPPFS_LIBRARY} ${GPU_LIBRARIES} ${GSPLAT_LIBS} ${TORCH_LIBRARIES} ${OpenCV_LIBS})
 if(GPU_RUNTIME STREQUAL "HIP")
     target_compile_definitions(opensplat PRIVATE USE_HIP __HIP_PLATFORM_AMD__)
+elseif(GPU_RUNTIME STREQUAL "CUDA")
+    target_compile_definitions(opensplat PRIVATE USE_CUDA)
 endif()
 
 if(OPENSPLAT_BUILD_SIMPLE_TRAINER)
-    add_executable(simple_trainer simple_trainer.cpp project_gaussians.cpp rasterize_gaussians.cpp)
+    add_executable(simple_trainer simple_trainer.cpp project_gaussians.cpp rasterize_gaussians.cpp cv_utils.cpp)
     target_include_directories(simple_trainer PRIVATE ${PROJECT_SOURCE_DIR}/vendor/glm ${GPU_INCLUDE_DIRS})
-    target_link_libraries(simple_trainer PUBLIC ${GPU_LIBRARIES} gsplat ${TORCH_LIBRARIES} ${OpenCV_LIBS})
+    target_link_libraries(simple_trainer PUBLIC ${GPU_LIBRARIES} ${GSPLAT_LIBS} ${TORCH_LIBRARIES} ${OpenCV_LIBS})
     set_property(TARGET simple_trainer PROPERTY CXX_STANDARD 17)
     if(GPU_RUNTIME STREQUAL "HIP")
         target_compile_definitions(simple_trainer PRIVATE USE_HIP __HIP_PLATFORM_AMD__)
+    elseif(GPU_RUNTIME STREQUAL "CUDA")
+        target_compile_definitions(simple_trainer PRIVATE USE_CUDA)
     endif()
 endif()
 

diff --git a/README.md b/README.md
@@ -6,15 +6,37 @@ A free and open source implementation of 3D gaussian splatting written in C++, f
 
 OpenSplat takes camera poses + sparse points in [COLMAP](https://colmap.github.io/) or [nerfstudio](https://docs.nerf.studio/quickstart/custom_dataset.html) project format and computes a [scene file](https://drive.google.com/file/d/1w-CBxyWNXF3omA8B_IeOsRmSJel3iwyr/view?usp=sharing) (.ply) that can be later imported for viewing, editing and rendering in other [software](https://github.com/MrNeRF/awesome-3D-gaussian-splatting?tab=readme-ov-file#open-source-implementations).
 
+Graphics card recommended, but not required! OpenSplat runs the fastest on NVIDIA and AMD GPUs, but can also run entirely on CPU power (~100x slower).
+
 Commercial use allowed and encouraged under the terms of the [AGPLv3](https://www.tldrlegal.com/license/gnu-affero-general-public-license-v3-agpl-3-0). ✅
 
-## Build (CUDA)
+## Build
 
 Requirements:
 
- * **CUDA**: Make sure you have the CUDA compiler (`nvcc`) in your PATH and that `nvidia-smi` is working. https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html 
- * **libtorch**: Visit https://pytorch.org/get-started/locally/ and select your OS, for package select "LibTorch". Make sure to match your version of CUDA if you want to leverage GPU support in libtorch.
  * **OpenCV**: `sudo apt install libopencv-dev` should do it.
+ * **libtorch**: See instructions below
+
+### CPU
+
+ **libtorch**: Visit https://pytorch.org/get-started/locally/ and select your OS, for package select "LibTorch". For compute platform you can select "CPU".
+
+ Then:
+
+ ```bash
+ git clone https://github.com/pierotofy/OpenSplat OpenSplat
+ cd OpenSplat
+ mkdir build && cd build
+ cmake -DCMAKE_PREFIX_PATH=/path/to/libtorch/ .. && make -j$(nproc)
+ ```
+
+### CUDA
+
+Additional requirement:
+
+ * **CUDA**: Make sure you have the CUDA compiler (`nvcc`) in your PATH and that `nvidia-smi` is working. https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html 
+
+ **libtorch**: Visit https://pytorch.org/get-started/locally/ and select your OS, for package select "LibTorch". Make sure to match your version of CUDA if you want to leverage GPU support in libtorch.
 
  Then:
 
@@ -27,12 +49,13 @@ Requirements:
 
  The software has been tested on Ubuntu 20.04 and Windows. With some changes it could run on macOS (help us by opening a PR?).
 
-## Build (ROCm via HIP)
-Requirements:
+### ROCm via HIP
+
+Additional requirement:
 
 * **ROCm**: Make sure you have the ROCm installed at /opt/rocm. https://rocm.docs.amd.com/projects/install-on-linux/en/latest/tutorial/quick-start.html
-* **libtorch**: Visit https://pytorch.org/get-started/locally/ and select your OS, for package select "LibTorch". Make sure to match your version of ROCm (5.7) if you want to leverage AMD GPU support in libtorch.
-* **OpenCV**: `sudo apt install libopencv-dev` should do it.
+
+**libtorch**: Visit https://pytorch.org/get-started/locally/ and select your OS, for package select "LibTorch". Make sure to match your version of ROCm (5.7) if you want to leverage AMD GPU support in libtorch.
 
 Then:
 
@@ -44,13 +67,18 @@ Then:
  cmake -DCMAKE_PREFIX_PATH=/path/to/libtorch/ -DGPU_RUNTIME="HIP" -DHIP_ROOT_DIR=/opt/rocm -DOPENSPLAT_BUILD_SIMPLE_TRAINER=ON ..
  make
  ```
+
 In addition, you can leverage Jinja to build the project
- ```
- cmake -GNinja -DCMAKE_PREFIX_PATH=/path/to/libtorch/ -DGPU_RUNTIME="HIP" -DHIP_ROOT_DIR=/opt/rocm -DOPENSPLAT_BUILD_SIMPLE_TRAINER=ON ..
- jinja
- ```
 
-## Docker Build (CUDA)
+```bash
+cmake -GNinja -DCMAKE_PREFIX_PATH=/path/to/libtorch/ -DGPU_RUNTIME="HIP" -DHIP_ROOT_DIR=/opt/rocm -DOPENSPLAT_BUILD_SIMPLE_TRAINER=ON ..
+jinja
+```
+
+## Docker Build
+
+### CUDA
+
 Navigate to the root directory of OpenSplat repo that has Dockerfile and run the following command to build the Docker image:
 
 ```bash
@@ -70,7 +98,8 @@ docker build \
   --build-arg CMAKE_BUILD_TYPE=Release .
 ```
 
-## Docker Build (ROCm via HIP)
+### ROCm via HIP
+
 Navigate to the root directory of OpenSplat repo that has Dockerfile and run the following command to build the Docker image:
 ```bash
 docker build \
@@ -138,11 +167,11 @@ cd /code/build
 We recently released OpenSplat, so there's lots of work to do.
 
  * Support for running on AMD cards (more testing needed)
- * Support for running on CPU-only
  * Improve speed / reduce memory usage
  * Distributed computation using multiple machines
  * Real-time training viewer output
  * Compressed scene outputs
+ * Automatic filtering
  * Your ideas?
 
  https://github.com/pierotofy/OpenSplat/issues?q=is%3Aopen+is%3Aissue+label%3Aenhancement

diff --git a/gsplat.hpp b/gsplat.hpp
@@ -0,0 +1,12 @@
+#ifndef GSPLAT_H
+#define GSPLAT_H
+
+#include "vendor/gsplat/config.h"
+
+#if defined(USE_HIP) || defined(USE_CUDA)
+#include "vendor/gsplat/bindings.h"
+#endif
+
+#include "vendor/gsplat-cpu/bindings.h"
+
+#endif