Add SYCL support for geometry and pipelines kernels, Transform ops, Nearest neighbor search and HashMap

.gitignore

@@ -3,9 +3,6 @@
 # Docker build artifacts (wheel, dev package, ccache)
 docker/*.tar.gz
 docker/*.tar.xz
-open3d-artifacts/*.tar.xz
-open3d-artifacts/*.deb
-open3d-artifacts/*.whl
 
 # Windows build temp cmake file
 CMakeSettings.json
@@ -41,6 +38,8 @@ _vimrc_local.vim
 .vscode/
 .clangd/
 .ccls-cache/
+AGENTS.local.md
+CLAUDE.local.md
 
 # Build artifacts
 **/package-lock.json

3rdparty/curl/curl.cmake

@@ -14,6 +14,8 @@ endif()
 
 if (APPLE)  # homebrew does not package libidn2
     set(curl_cmake_extra_args -DUSE_APPLE_IDN=ON -DUSE_LIBIDN2=OFF -DUSE_NGHTTP2=OFF)
+else()
+    set(curl_cmake_extra_args -DUSE_NGHTTP2=OFF)
 endif()
 
 ExternalProject_Add(

3rdparty/find_dependencies.cmake

@@ -960,17 +960,27 @@ endif()
 # flattened by CMake (e.g. when Open3D itself is a shared library and must
 # fully resolve all symbols at build time), they can end up in an order where
 # ld's single left-to-right archive scan fails to resolve symbols. Wrap them
-# in --start-group and --end-group on UNIX (non-Apple) to ensure GNU ld
-# rescans this group of libraries until all symbols resolve, regardless of
-# order. We avoid CMake's modern LINK_GROUP RESCAN generator expression here
-# because it produces developer warnings when applied to INTERFACE library
-# targets (which Open3D::3rdparty_curl/openssl are).
+# in a GNU ld archive group on UNIX (non-Apple) to ensure GNU ld rescans this
+# group of libraries until all symbols resolve, regardless of order. Use the
+# short "-Wl,-(" / "-Wl,-)" alias rather than "-Wl,--start-group" /
+# "-Wl,--end-group": the latter are also used verbatim by MKL's GROUPED
+# import (see open3d_import_3rdparty_library's GROUPED option below), and
+# CMake's link-line deduplication collapses repeated identical literal flag
+# strings across the whole dependency graph, emptying both groups and
+# stranding MKL's libraries outside their own grouping (undefined mkl_*
+# symbols at link time). We avoid CMake's modern LINK_GROUP RESCAN generator
+# expression here because it produces developer warnings when applied to
+# INTERFACE library targets (which Open3D::3rdparty_curl/openssl are).
+# The parentheses are backslash-escaped because CMake/Ninja invokes the link
+# command via "/bin/sh -c" without shell-quoting raw linker flag strings, and
+# bare "(" / ")" are shell metacharacters that would otherwise cause a shell
+# syntax error at link time.
 if(UNIX AND NOT APPLE)
     list(APPEND Open3D_3RDPARTY_PRIVATE_TARGETS_FROM_CUSTOM
-        "-Wl,-("
+        "-Wl,-\\("
         Open3D::3rdparty_curl
         Open3D::3rdparty_openssl
-        "-Wl,-)")
+        "-Wl,-\\)")
 else()
     list(APPEND Open3D_3RDPARTY_PRIVATE_TARGETS_FROM_CUSTOM Open3D::3rdparty_curl Open3D::3rdparty_openssl)
 endif()

CHANGELOG.md

@@ -1,4 +1,5 @@
 ## Main
+-   Add SYCL tensor backends for HashMap, nearest-neighbor search (KNN, fixed-radius, hybrid), geometry transforms, and registration / odometry / feature pipelines (parity with CUDA paths where applicable).
 -   Add compressed SPZ file I/O for tensor-based Gaussian splats, with zstd dependency integration, round-trip tests, and notebook samples.
 -   Add Windows shared-library CUDA and SYCL Python wheels (`open3d-cuda`, `open3d-xpu`) built against the installed devel package; ship NVIDIA CUDA 12.6 runtime pip dependencies (`python/requirements_win_cuda.txt`) since CUDA is linked dynamically on Windows
 -   Fix WebRTC prebuilt packaging and CI workflow across Linux, macOS arm64, and Windows runtime variants (PR #7515)

CMakeLists.txt

@@ -85,6 +85,9 @@ if(BUILD_SYCL_MODULE)
 endif()
 option(GLIBCXX_USE_CXX11_ABI      "Set -D_GLIBCXX_USE_CXX11_ABI=1"           ON )
 option(ENABLE_SYCL_UNIFIED_SHARED_MEMORY "Enable SYCL unified shared memory" OFF)
+set(ENABLE_SANITIZER OFF CACHE STRING "Enable sanitizer.")
+set_property(CACHE ENABLE_SANITIZER PROPERTY STRINGS
+                    "OFF" "address" "memory" "thread" "undefined")
 if(BUILD_GUI AND (WIN32 OR UNIX AND NOT LINUX_AARCH64 AND NOT APPLE_AARCH64))
     option(BUILD_WEBRTC           "Build WebRTC visualizer"                  ON )
 else()

cmake/Open3DAddEncodedShader.cmake

@@ -69,11 +69,11 @@ endfunction()
 if (NOT TARGET ShaderEncoder)
     add_executable(ShaderEncoder EXCLUDE_FROM_ALL)
     target_sources(ShaderEncoder PRIVATE ${CMAKE_CURRENT_LIST_DIR}/ShaderEncoder.cpp)
-    target_compile_features(ShaderEncoder PRIVATE cxx_std_14)
+    open3d_set_global_properties(ShaderEncoder)
 endif()
 
 if (NOT TARGET ShaderLinker)
     add_executable(ShaderLinker EXCLUDE_FROM_ALL)
     target_sources(ShaderLinker PRIVATE ${CMAKE_CURRENT_LIST_DIR}/ShaderLinker.cpp)
-    target_compile_features(ShaderLinker PRIVATE cxx_std_14)
+    open3d_set_global_properties(ShaderLinker)
 endif()

cmake/Open3DSetGlobalProperties.cmake

@@ -47,8 +47,10 @@ function(open3d_set_global_properties target)
     target_compile_definitions(${target} PRIVATE OPEN3D_CXX_STANDARD="${CMAKE_CXX_STANDARD}")
     target_compile_definitions(${target} PRIVATE OPEN3D_CXX_COMPILER_ID="${CMAKE_CXX_COMPILER_ID}")
     target_compile_definitions(${target} PRIVATE OPEN3D_CXX_COMPILER_VERSION="${CMAKE_CXX_COMPILER_VERSION}")
-    target_compile_definitions(${target} PRIVATE OPEN3D_CUDA_COMPILER_ID="${CMAKE_CUDA_COMPILER_ID}")
-    target_compile_definitions(${target} PRIVATE OPEN3D_CUDA_COMPILER_VERSION="${CMAKE_CUDA_COMPILER_VERSION}")
+    if (BUILD_CUDA_MODULE)
+        target_compile_definitions(${target} PRIVATE OPEN3D_CUDA_COMPILER_ID="${CMAKE_CUDA_COMPILER_ID}")
+        target_compile_definitions(${target} PRIVATE OPEN3D_CUDA_COMPILER_VERSION="${CMAKE_CUDA_COMPILER_VERSION}")
+    endif()
 
     # std::filesystem (C++17) or std::experimental::filesystem (C++14)
     #
@@ -187,7 +189,7 @@ function(open3d_set_global_properties target)
     endif()
 
     # IEEE-compliant FP for IntelLLVM / icx (SYCL on Windows and Linux).
-    # -fno-fast-math: preserve NaN handling. -no-ftz / -mno-daz-ftz: icx
+    # -fno-fast-math: preserve NaN handling. -no-ftz : icx
     # enables FTZ/DAZ at -O1+, which zeroes denormals (e.g. RGB packed into a
     # float in PCD). Windows icx is clang-cl; prefix with /clang: so the
     # GNU-style flags take effect.
@@ -199,7 +201,8 @@ function(open3d_set_global_properties target)
     target_compile_options(${target} PRIVATE
         $<$<AND:$<CXX_COMPILER_ID:IntelLLVM>,$<NOT:$<COMPILE_LANGUAGE:ISPC>>>:${opt-prefix}-fno-fast-math>
         $<$<AND:$<CXX_COMPILER_ID:IntelLLVM>,$<NOT:$<COMPILE_LANGUAGE:ISPC>>>:${opt-prefix}-no-ftz>
-        $<$<AND:$<CXX_COMPILER_ID:IntelLLVM>,$<NOT:$<COMPILE_LANGUAGE:ISPC>>>:${opt-prefix}-mno-daz-ftz>)
+    )
+    # should be /Qftz- for WIN32?
 
     # Enable strip
     open3d_enable_strip(${target})

cpp/benchmarks/CMakeLists.txt

@@ -14,10 +14,6 @@ if (BUILD_CUDA_MODULE)
     find_package(CUDAToolkit REQUIRED)
     target_link_libraries(benchmarks PRIVATE CUDA::cudart)
 endif()
-if (BUILD_SYCL_MODULE)
-    find_package(IntelSYCL REQUIRED)   # requires cmake>=3.25 on Windows
-    add_sycl_to_target(TARGET benchmarks)
-endif()
 
 open3d_show_and_abort_on_warning(benchmarks)
 open3d_set_global_properties(benchmarks)
@@ -38,3 +34,9 @@ target_sources(benchmarks PRIVATE
 )
 target_link_libraries(benchmarks PRIVATE benchmark::benchmark)
 target_include_directories(benchmarks PRIVATE ${CMAKE_CURRENT_SOURCE_DIR})
+
+if (BUILD_SYCL_MODULE)
+    # benchmarks compiles its own SYCL device code above, so it needs the
+    # -fsycl link step.
+    target_link_libraries(benchmarks PRIVATE Open3D::3rdparty_sycl)
+endif()

cpp/benchmarks/core/CMakeLists.txt

@@ -1,6 +1,7 @@
 target_sources(benchmarks PRIVATE
     BinaryEW.cpp
     HashMap.cpp
+    NearestNeighborSearch.cpp
     Linalg.cpp
     MemoryManager.cpp
     ParallelFor.cpp