| Commit message (Collapse) | Author | Age | Files | Lines |
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
sub-packages
The 2026.0 oneAPI MKL component grew to 1.9 GB on disk, dominated by
~900 MB of static archives, ~330 MB of SYCL/GPU offload libraries and
~80 MB of headers / cmake metadata / sample sources. The IPP component
similarly carries ~280 MB of static archives on top of a 537 MB
runtime payload. Consumers that only need a CPU BLAS/FFT backend
(PyTorch CPU, NumPy, image-processing pipelines) were paying for all
of it whenever they installed intel-oneapi-toolkit-mkl or
intel-oneapi-toolkit-ipp.
Split the two heaviest components along the usual runtime/dev/static
seams:
intel-oneapi-toolkit-mkl CPU shared libs + dispatch
libraries + env/etc/bin glue
(~640 MB)
intel-oneapi-toolkit-mkl-sycl SYCL / GPU offload shared libs
(~330 MB, RDEPENDS -runtime)
intel-oneapi-toolkit-mkl-staticdev Static archives, lib/*.a
(~900 MB)
intel-oneapi-toolkit-mkl-dev Headers, cmake modules,
pkgconfig files, share/ examples
(~85 MB)
intel-oneapi-toolkit-ipp ipp + ippcp runtime shared libs
(~540 MB)
intel-oneapi-toolkit-ipp-staticdev Static archives (~280 MB)
intel-oneapi-toolkit-ipp-dev Headers, cmake, pkgconfig, share
PACKAGES ordering places the specific sub-sub-packages before the
catch-all component package so first-match-wins puts the heavyweight
bits in their dedicated package. FILES patterns use [0-9]* to match
the versioned directory only (the sibling 'latest' compatibility
symlink is captured by the parent component package). RDEPENDS form
a closed chain: -dev depends on the runtime package, -staticdev on
-dev, -sycl on -runtime + the CPU runtime, so users that explicitly
install a sub-sub-package still get a working environment.
The top-level meta-package RDEPENDS:${PN} continues to pull all of
the new sub-packages, so an image that asks for the unqualified
intel-oneapi-toolkit recipe is unchanged.
Signed-off-by: Yogesh Tyagi <yogesh.tyagi@intel.com>
Tested-by: Zoltán Böszörményi <zboszor@gmail.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Remove the per-component apt-package recipes superseded by the
unified intel-oneapi-toolkit_2026.0.0.198.bb recipe, which packages
the same components as sub-packages of a single offline installer
(intel-oneapi-toolkit-{compiler,runtime,mkl,tbb,dpl,ipp,...}).
Consumers should reference the canonical toolkit sub-package names
directly.
Recipes removed:
* intel-oneapi-dpcpp-cpp_2024.0.0-49819.bb
* intel-oneapi-dpcpp-cpp-runtime_2024.0.0-49819.bb
* intel-oneapi-mkl_2024.0.0-49656.bb
* intel-oneapi-ipp_2021.10.0-653.bb
* setup-intel-oneapi-env_2023.0.0-25370.bb (and its
intel-oneapi-runtime.conf helper)
The setup-intel-oneapi-env helper installed a single ld.so.conf.d
snippet pointing the dynamic loader at /opt/intel/oneapi/lib (and
lib/intel64, lib/ia32, lib/emu). It dates back to the 2023.0
oneAPI layout. The unified toolkit installs libs under
per-component directories such as /opt/intel/oneapi/{compiler,mkl,
tbb,...}/2026.0/lib, not the flat /opt/intel/oneapi/lib that the
conf file targets, so the ld.so.conf entries no longer point at
real directories. Loadability of oneAPI binaries on the target is
provided by:
* setvars.sh / per-component env/vars.sh sourcing (the canonical
Intel-supported entry point that every test and example uses)
* $ORIGIN-relative RUNPATH baked into the shipped shared objects
by the Intel installer
* /etc/OpenCL/vendors/intel-cpu.icd written by the unified toolkit
recipe's do_install for ICD loader discovery
* virtual-opencl-icd / level-zero-loader RDEPENDS on
intel-oneapi-toolkit-runtime for the OpenCL / Level Zero loaders
The setup-intel-oneapi-env filename version (2023.0.0-25370) was
a stale carry-over from when the recipe was first added; nothing
in the recipe ever referenced ${PV}.
Signed-off-by: Yogesh Tyagi <yogesh.tyagi@intel.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Intel oneAPI 2026.0 ships the former Base and HPC toolkits as a
single offline self-extracting installer instead of as the per-
component apt packages used in the 2021/2024 layout. Add a recipe
that consumes intel-oneapi-toolkit-2026.0.0.198_offline.sh via a
pseudo-free do_unpack_vendor task and stages /opt/intel/oneapi
into the rootfs.
The recipe partitions the installed tree into sub-packages along
component boundaries: intel-oneapi-toolkit-{compiler,runtime,mkl,
tbb,dpl,debugger,ipp,ccl,vtune,common,licensing}. Consumers should
reference these canonical sub-package names directly rather than
the per-component apt names that the 2021/2024 standalone recipes
exposed.
The runtime sub-package additionally:
* registers /opt/intel/oneapi/compiler/2026.0/lib/libintelocl.so as
/etc/OpenCL/vendors/intel-cpu.icd so systems without an Intel GPU
expose an Intel OpenCL platform via the standard ICD mechanism;
* installs a libxml2.so.2 -> libxml2.so.16 SONAME compat symlink,
because the prebuilt libcommon_clang.so / libintelocl.so are
linked against the libxml2 2.x SONAME while current oe-core ships
libxml2 3.x as libxml2.so.16. libxml2 3.x is API-compatible with
the option-parsing surface the binaries use.
The bundle is x86_64-only so COMPATIBLE_HOST is restricted.
Signed-off-by: Yogesh Tyagi <yogesh.tyagi@intel.com>
|
| |
|
|
| |
Signed-off-by: Yogesh Tyagi <yogesh.tyagi@intel.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Add INSANE_SKIP for buildpaths QA check in RenderKit recipes that use
ISPC for code generation.
ISPC (Intel SPMD Program Compiler) embeds absolute build paths in
generated binaries and debug information by design. This is necessary
for:
1. Debug information accuracy - ISPC generates C++ wrapper code with
#line directives pointing to original .ispc source files
2. Error reporting during kernel execution - backtraces reference
original source locations
3. Profiling support - performance tools need source file locations
Affected packages:
- embree: ISPC kernels for ray tracing primitives
- openvkl: ISPC kernels for volume rendering
- ospray: ISPC kernels for rendering algorithms
This resolves QA errors that prevent SDK image builds when RenderKit
stack is included.
Signed-off-by: Yogesh Tyagi <yogesh.tyagi@intel.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Add Intel Open Image Denoise (OIDN) 2.4.1, an open-source library for
high-performance, high-quality denoising of images rendered with ray tracing.
Key features:
- Machine learning-based denoising filters optimized for Intel CPUs
- ISPC-accelerated kernels for vectorization
- Supports HDR and LDR images
- Multiple filter types: RT, RTLightmap
Dependencies:
- ISPC 1.25.0+ for kernel compilation
- TBB for task parallelism
- CMake 3.1+
Build configuration:
- GPU support disabled (OIDN_DEVICE_GPU=OFF)
- Apps disabled for minimal footprint
- Static TBB linkage
Part of Intel RenderKit stack upgrade coordinated with:
- ISPC 1.29.1
- embree 4.4.0
- openvkl 2.0.1
- ospray 3.2.0
- rkcommon 1.15.2
Signed-off-by: Yogesh Tyagi <yogesh.tyagi@intel.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Add patch to update ospray's dependency version requirements:
- rkcommon: 1.14.0 -> 1.15.2 (remove EXACT matching)
- embree: 4.3.3 -> 4.4.0
Removing EXACT version matching for rkcommon is safe because:
1. rkcommon declares COMPATIBILITY SameMajorVersion in CMake
2. Changes from 1.14.0 to 1.15.2 are backward compatible:
- Added shutdownTaskingSystem() function (new, non-breaking)
- Optimized frame() math algorithm (same result, better performance)
- Bug fixes in tasking system (improved reliability)
- Compiler compatibility improvements (icx, ARM)
3. No API breakage: zero functions removed or signatures changed
The 1.15.2 version provides performance improvements and bug fixes
while maintaining full API compatibility with 1.14.0.
Signed-off-by: Yogesh Tyagi <yogesh.tyagi@intel.com>
|
| |
|
|
|
|
|
| |
Release Notes:
https://github.com/embree/embree/releases/tag/v4.4.0
Signed-off-by: Yogesh Tyagi <yogesh.tyagi@intel.com>
|
| |
|
|
|
|
|
|
|
|
| |
Dropped 0001-use-fully-qualified-rkcommon-math-rsqrt-to-avoid-ove.patch
as it has been merged upstream.
Release Notes:
https://github.com/ospray/rkcommon/releases/tag/v1.15.2
Signed-off-by: Yogesh Tyagi <yogesh.tyagi@intel.com>
|
| |
|
|
|
|
|
| |
Release Notes:
https://github.com/uxlfoundation/oneDPL/releases/tag/oneDPL-2022.11.0-release
Signed-off-by: Yogesh Tyagi <yogesh.tyagi@intel.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Reorganize meta-intel recipes to follow Yocto best practices and
clear semantic boundaries between clang/SYCL components, BSP enablement,
core OS policy, and generic support libraries.
Key changes:
- Move oneDPL into clang dynamic layer (SYCL / Clang-semantic dependency)
- Move linux-npu-driver from recipes-core to recipes-bsp (platform enablement)
- Move formfactor from recipes-bsp to recipes-core (system-wide OS policy)
- Move intel-oneapi-ipp out of clang layer into recipes-oneapi (CPU-only, compiler-agnostic)
- Move intel-crypto-mb from recipes-oneapi to recipes-support/crypto
- Move intel-cmt-cat and metee from recipes-bsp to recipes-support
This refactor improves layer clarity, avoids unnecessary clang coupling,
and aligns recipe placement with Yocto conventions.
Signed-off-by: Yogesh Tyagi <yogesh.tyagi@intel.com>
|
|
|
Yocto 5.3 merged most of meta-clang.
Fix recipes structure in dynamic-layers/clang-layer
clang recipes now depend on meta-clang-revival layer
Signed-off-by: Yogesh Tyagi <yogesh.tyagi@intel.com>
|