Feature: Add initial SYCL backend support for gsplat

gsplat/__init__.py

@@ -1,25 +1,72 @@
+import os
+import sys
+import torch
 import warnings
 
+BACKEND: str = ""
+
+FORCE_BACKEND = os.getenv("GSPLAT_BACKEND", "").lower()
+
+if FORCE_BACKEND == "cuda" or (FORCE_BACKEND == "" and torch.cuda.is_available()):
+    try:
+        BACKEND = "cuda"
+        from .cuda._wrapper import (
+            RollingShutterType,
+            fully_fused_projection,
+            fully_fused_projection_2dgs,
+            fully_fused_projection_with_ut,
+            isect_offset_encode,
+            isect_tiles,
+            proj,
+            quat_scale_to_covar_preci,
+            rasterize_to_indices_in_range,
+            rasterize_to_indices_in_range_2dgs,
+            rasterize_to_pixels,
+            rasterize_to_pixels_2dgs,
+            rasterize_to_pixels_eval3d,
+            spherical_harmonics,
+            world_to_cam,
+        )
+        print("gsplat: CUDA backend successfully loaded.", file=sys.stderr)
+    except ImportError:
+        if FORCE_BACKEND == "cuda":
+            print("gsplat: Error! GSPLAT_BACKEND=cuda was set but CUDA backend failed to load.", file=sys.stderr)
+        pass
+
+if not BACKEND and (FORCE_BACKEND == "sycl" or FORCE_BACKEND == ""):
+    try:
+        BACKEND = "sycl"
+        from .sycl._wrapper import (
+            RollingShutterType,
+            fully_fused_projection,
+            fully_fused_projection_2dgs,
+            fully_fused_projection_with_ut,
+            isect_offset_encode,
+            isect_tiles,
+            proj,
+            quat_scale_to_covar_preci,
+            rasterize_to_indices_in_range,
+            rasterize_to_indices_in_range_2dgs,
+            rasterize_to_pixels,
+            rasterize_to_pixels_2dgs,
+            rasterize_to_pixels_eval3d,
+            spherical_harmonics,
+            world_to_cam,
+        )
+        print("gsplat: SYCL backend successfully loaded.", file=sys.stderr)
+    except ImportError as e:
+        if FORCE_BACKEND == "sycl":
+            print(f"gsplat: Error! GSPLAT_BACKEND=sycl was set but SYCL backend failed to load: {e}", file=sys.stderr)
+        pass
+
+if not BACKEND:
+    print(
+        "gsplat: Warning! No high-performance backend (CUDA or SYCL) found.",
+        file=sys.stderr,
+    )
+
+
 from .compression import PngCompression
-from .cuda._torch_impl import accumulate
-from .cuda._torch_impl_2dgs import accumulate_2dgs
-from .cuda._wrapper import (
-    RollingShutterType,
-    fully_fused_projection,
-    fully_fused_projection_2dgs,
-    fully_fused_projection_with_ut,
-    isect_offset_encode,
-    isect_tiles,
-    proj,
-    quat_scale_to_covar_preci,
-    rasterize_to_indices_in_range,
-    rasterize_to_indices_in_range_2dgs,
-    rasterize_to_pixels,
-    rasterize_to_pixels_2dgs,
-    rasterize_to_pixels_eval3d,
-    spherical_harmonics,
-    world_to_cam,
-)
 from .exporter import export_splats
 from .optimizers import SelectiveAdam
 from .rendering import (
@@ -31,7 +78,9 @@
 from .strategy import DefaultStrategy, MCMCStrategy, Strategy
 from .version import __version__
 
-all = [
+
+__all__ = [
+    "BACKEND",
     "PngCompression",
     "DefaultStrategy",
     "MCMCStrategy",
@@ -47,16 +96,16 @@
     "quat_scale_to_covar_preci",
     "rasterize_to_pixels",
     "world_to_cam",
-    "accumulate",
     "rasterize_to_indices_in_range",
     "fully_fused_projection_2dgs",
     "rasterize_to_pixels_2dgs",
     "rasterize_to_indices_in_range_2dgs",
-    "accumulate_2dgs",
     "rasterization_2dgs_inria_wrapper",
     "RollingShutterType",
     "fully_fused_projection_with_ut",
     "rasterize_to_pixels_eval3d",
     "export_splats",
     "__version__",
-]
+    "SelectiveAdam",
+    # Note: accumulate and accumulate_2dgs are not typically part of the public API
+]

gsplat/cuda/_torch_impl.py → gsplat/_torch_impl.py

@@ -639,7 +639,7 @@ def _rasterize_to_pixels(
         This function requires the `nerfacc` package to be installed. Please install it
         using the following command `pip install nerfacc`.
     """
-    from ._wrapper import rasterize_to_indices_in_range
+    from .cuda._wrapper import rasterize_to_indices_in_range
 
     image_dims = means2d.shape[:-2]
     channels = colors.shape[-1]

gsplat/cuda/_torch_impl_2dgs.py → gsplat/_torch_impl_2dgs.py

@@ -4,7 +4,7 @@
 import torch
 from torch import Tensor
 
-from gsplat.cuda._torch_impl import _quat_scale_to_matrix
+from ._torch_impl import _quat_scale_to_matrix
 
 
 def _fully_fused_projection_2dgs(
@@ -231,7 +231,7 @@ def _rasterize_to_pixels_2dgs(
         This function requires the `nerfacc` package to be installed. Please install it
         using the following command `pip install nerfacc`.
     """
-    from ._wrapper import rasterize_to_indices_in_range_2dgs
+    from .cuda._wrapper import rasterize_to_indices_in_range_2dgs
 
     image_dims = means2d.shape[:-2]
     channels = colors.shape[-1]

gsplat/cuda/_wrapper.py

@@ -112,7 +112,7 @@ def world_to_cam(
         - **Gaussian means in camera coordinate system**. [..., C, N, 3]
         - **Gaussian covariances in camera coordinate system**. [..., C, N, 3, 3]
     """
-    from ._torch_impl import _world_to_cam
+    from .._torch_impl import _world_to_cam
 
     warnings.warn(
         "world_to_cam() is removed from the CUDA backend as it's relatively easy to "

gsplat/strategy/ops.py

@@ -5,7 +5,18 @@
 import torch.nn.functional as F
 from torch import Tensor
 
-from gsplat import quat_scale_to_covar_preci
+from gsplat import BACKEND
+if BACKEND == "cuda":
+    from gsplat.cuda._wrapper import (
+        quat_scale_to_covar_preci,
+    )
+elif BACKEND == "sycl":
+    from gsplat.sycl._wrapper import (
+        quat_scale_to_covar_preci,
+    )
+else:
+    raise ImportError("gsplat: No backend loaded, cannot import strategy ops.")
+
 from gsplat.relocation import compute_relocation
 from gsplat.utils import normalized_quat_to_rotmat
 

gsplat/sycl/CMakeLists.txt

@@ -0,0 +1,119 @@
+cmake_minimum_required(VERSION 3.23...4.0)  # Need min 3.23 on Windows
+
+set(CMAKE_C_COMPILER icx)
+set(CMAKE_CXX_COMPILER icx)
+
+project(gsplat_sycl)
+
+if (NOT CMAKE_BUILD_TYPE AND NOT CMAKE_CONFIGURATION_TYPES)
+  set(CMAKE_BUILD_TYPE Release CACHE STRING "Choose the type of build." FORCE)
+  set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS "Debug" "Release" "MinSizeRel" "RelWithDebInfo")
+endif()
+
+if (NOT SYCL_AOT_TARGETS)
+        set (SYCL_AOT_TARGETS "spir64" CACHE STRING "Comma separated list of SYCL targets for ahead of time compilation. See https://github.com/intel/llvm/blob/sycl/sycl/doc/UsersManual.md for a full list.")
+endif()
+
+find_package(Python 3.8 COMPONENTS Interpreter Development.Module REQUIRED)
+
+
+execute_process(
+  COMMAND "${Python_EXECUTABLE}" -c "import torch.utils; print(torch.utils.cmake_prefix_path)"
+  OUTPUT_STRIP_TRAILING_WHITESPACE
+  OUTPUT_VARIABLE Torch_DIR_From_Python 
+  RESULT_VARIABLE _torch_path_result    
+  ERROR_QUIET                           
+)
+if(NOT _torch_path_result EQUAL 0)
+  message(WARNING "Failed to get Torch CMake path from Python. "
+                  "Make sure PyTorch is installed in the Python environment: ${Python_EXECUTABLE}")
+  set(Torch_DIR_From_Python "")
+endif()
+
+if(Torch_DIR_From_Python AND IS_DIRECTORY "${Torch_DIR_From_Python}")
+  set(Torch_DIR ${Torch_DIR_From_Python})
+  message(STATUS "Found Torch CMake directory via Python: ${Torch_DIR}")
+  find_package(Torch REQUIRED HINTS ${Torch_DIR_From_Python})
+else()
+  message(FATAL_ERROR "Could not find Torch via Python introspection. "
+                      "Please ensure PyTorch is installed or set CMAKE_PREFIX_PATH/Torch_DIR manually.")
+endif()
+execute_process(
+  COMMAND "${Python_EXECUTABLE}" -c "import os; from torch.utils import cpp_extension; print(os.path.join(cpp_extension.library_paths(True)[0], 'libtorch_python.so'))"
+  OUTPUT_STRIP_TRAILING_WHITESPACE
+  OUTPUT_VARIABLE TORCH_PYTHON_LIB
+)
+
+if (NOT EXISTS "${TORCH_PYTHON_LIB}")
+    message(FATAL_ERROR "Could not find libtorch_python.so at ${TORCH_PYTHON_LIB}. Please check your PyTorch installation.")
+else()
+    message(STATUS "Found torch_python library at: ${TORCH_PYTHON_LIB}")
+endif()
+
+
+set(PYBIND11_FINDPYTHON ON)
+find_package(pybind11 CONFIG REQUIRED)
+
+set( SYCL_SOURCES
+  ${CMAKE_CURRENT_SOURCE_DIR}/ext.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/src/adam.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/src/intersect_offset.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/src/intersect_tile.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/src/null.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/src/projection_2dgs_fused_bwd.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/src/projection_2dgs_fused_fwd.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/src/projection_2dgs_packed_bwd.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/src/projection_2dgs_packed_fwd.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/src/projection_ewa_3dgs_fused_bwd.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/src/projection_ewa_3dgs_fused_fwd.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/src/projection_ewa_3dgs_packed_bwd.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/src/projection_ewa_3dgs_packed_fwd.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/src/projection_ewa_simple_bwd.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/src/projection_ewa_simple_fwd.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/src/projection_ut_3dgs_fused.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/src/quat_scale_to_covar_preci_bwd.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/src/quat_scale_to_covar_preci_fwd.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/src/rasterize_to_indices_2dgs.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/src/rasterize_to_indices_3dgs.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/src/rasterize_to_pixels_2dgs_bwd.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/src/rasterize_to_pixels_2dgs_fwd.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/src/rasterize_to_pixels_3dgs_bwd.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/src/rasterize_to_pixels_3dgs_fwd.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/src/rasterize_to_pixels_from_world_3dgs_bwd.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/src/rasterize_to_pixels_from_world_3dgs_fwd.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/src/relocation.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/src/spherical_harmonics_bwd.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/src/spherical_harmonics_fwd.cpp
+)
+
+set(SYCL_MODULE_NAME gsplat_sycl_kernels)
+
+pybind11_add_module(${SYCL_MODULE_NAME} MODULE ${SYCL_SOURCES})
+
+target_include_directories( ${SYCL_MODULE_NAME} SYSTEM PRIVATE
+    ${CMAKE_CURRENT_SOURCE_DIR}/include 
+    ${CMAKE_CURRENT_SOURCE_DIR}/../cuda/csrc/third_party/glm 
+)
+
+target_compile_definitions(${SYCL_MODULE_NAME} PRIVATE 
+    TORCH_EXTENSION_NAME=${SYCL_MODULE_NAME}
+)
+
+target_compile_options(${SYCL_MODULE_NAME} PRIVATE -fsycl)
+target_compile_features(${SYCL_MODULE_NAME} PUBLIC cxx_std_17)
+
+target_link_options(${SYCL_MODULE_NAME} PRIVATE -fsycl -fsycl-targets=${SYCL_AOT_TARGETS})
+target_link_libraries(${SYCL_MODULE_NAME} PRIVATE torch)
+
+
+# Fix for icx: error: '-MP' is not supported with offloading enabled
+if (WIN32 AND NOT UNIX)
+    get_target_property(CURRENT_OPTIONS ${SYCL_MODULE_NAME} COMPILE_OPTIONS)
+    string(REPLACE "/MP" "" MODIFIED_OPTIONS "${CURRENT_OPTIONS}")
+    set_target_properties(${SYCL_MODULE_NAME} PROPERTIES COMPILE_OPTIONS "${MODIFIED_OPTIONS}")
+endif ()
+if (UNIX AND NOT APPLE)
+   # Find libtorch_xpu and libsycl at runtime in Python environment
+   set_target_properties(${SYCL_MODULE_NAME} PROPERTIES INSTALL_RPATH 
+           "$ORIGIN/../../torch/lib/;$ORIGIN/../../../../")
+endif()

gsplat/sycl/_backend.py

@@ -0,0 +1,7 @@
+_C = None
+
+try:
+    # Try to import the compiled module (via setup.py or pre-built .so)
+    from gsplat import gsplat_sycl_kernels as _C
+except ImportError: 
+    raise ImportError("Unable to find compiled sycl kernels package")