Release Notes for PoCL 7.2¶

Release highlights¶

Conformance results were submitted for OpenCL 3.0 conformance with the CPU x86-64 (AVX512) , CPU RISC-V RV64GC and CPU RISC-V RVA22+RVV1.0 devices, using OpenCL-CTS tag v2026-03-25-00. Conformance testing via OpenCL-CTS passed with both OpenCL C and SPIR-V compilation modes, for all of these devices.
Support for LLVM version 22

PoCL now supports configuring LLVM via both llvm-config binary (-DWITH_LLVM_CONFIG=<path>), and LLVMConfig.cmake (-DLLVM_DIR=<path>). Support for cross-compiling was also improved, and requires using both these options, pointing to target & host LLVMs.
LLVM_SPIRV CMake option was replaced by HOST_LLVM_SPIRV
Added POCL_DEBUG_LLVM_OPTS environment variable. This allows directly passing multiple LLVM options for debugging purposes.

OpenCL-CTS updated to upstream tag v2026-03-25-00 and fixed related bugs:
- clCreateContext: return error on invalid CL_CONTEXT_INTEROP_USER_SYNC flag
- clSetKernelArg: for buffer argument, accept arg_value NULL or pointer to NULL
- clSetKernelExecInfo: fix out-of-bounds read of the USM indirect pointer list.
- clCreateContextFromType: return error on unknown values of device_type bitfield
- clGetDeviceIDs: return error on unknown values of device_type bitfield
- clSetKernelArgSVMPointer: return CL_INVALID_ARG_INDEX for some error conditions
- clEnqueueNDRange: accept local_size == NULL if required-wg-size is nonzero
- clCreateProgramWithIL: return CL_INVALID_OPERATION if no devices support SPIR-V
Fixed an out-of-bounds read in clSetKernelExecInfo when registering indirect USM pointers (CL_KERNEL_EXEC_INFO_USM_PTRS_INTEL): the size of the pointer list was passed on in bytes where a pointer count was expected, causing the runtime to read past the end of the user-provided list.
clGetKernelArgInfo(CL_KERNEL_ARG_NAME) now returns the source argument names for SPIR/SPIR-V binaries on the CPU driver (previously returned CL_KERNEL_ARG_INFO_NOT_AVAILABLE even when the binary contained the names). The llvm-spirv translator preserves SPIR-V OpName decorations on llvm::Argument`s rather than as `!kernel_arg_name metadata, and PoCL now reads them from there.

Implemented sub_group_{reduce,scan_exclusive,scan_inclusive}_* and sub_group_{all,any,broadcast}. (reduce is only available for PTX6.0+)
Note that CUDA driver does not support LLVM 21, due to a bug in upstream Clang code. Users must use LLVM 18,19,20 or 22 with CUDA. For details see https://github.com/llvm/llvm-project/issues/154772
Fixed pre-mangled subgroup intrinsic names
Implemented support for additional parameters to clGetKernelSubgroupInfo(): CL_KERNEL_MAX_SUB_GROUP_SIZE_FOR_NDRANGE, CL_KERNEL_SUB_GROUP_COUNT_FOR_NDRANGE, CL_KERNEL_LOCAL_SIZE_FOR_SUB_GROUP_COUNT
Fixed printf to work correctly when using LLVM 19/20 (issue 2021)

The atomic operations on floats (cl_ext_float_atomic) are now disabled when -DENABLE_CONFORMANCE=ON. When -DENABLE_CONFORMANCE=OFF, cl_ext_float_atomic is enabled. This is due to issues in OpenCL-CTS that prevent passing the testsuite when cl_ext_float_atomic is enabled.
Implemented new extensions: cl_khr_extended_bit_ops, cl_khr_device_uuid, cl_khr_suggested_local_work_size, cl_khr_integer_dot_product
Improved support for FP16 math builtins. Not all of the math builtins are supported yet; approx 15 are still missing.
When using SPIR-V input, extensions cl_khr_spirv_linkonce_odr and cl_khr_spirv_no_integer_wrap_decoration are now supported
Fixed missing support of a few FP16 overloads of vload/vstore/convert_type builtins in SPIRV input
Reported sizes for INTEL_SUB_GROUP_SIZES were reverted to earlier behavior, PoCL now only reports power-of-2 sizes.
Implemented support for cl_khr_kernel_clock (requires Clang support for __builtin_readcyclecounter builtin function)
Re-enabled support for __opencl_c_program_scope_global_variables. This had a problem with SPIR-V input when using certain versions of llvm-spirv translator, however this has now been fixed in the translator.
Mac OS X: statically-linked LLVM symbols are now hidden also when using the CMake-package setup
Fixed kernel-side sub-group queries for non-divisible work-group sizes
Honor intel_reqd_sub_group_size attribute in clGetKernelSubGroupInfo()
SIGFPE handler is now fixed to enable only on x86-64, not all x86

Added a work-in-progress MLIR-based OpenCL C kernel compilation flow for CPUs. The flow supports both Polygeist and ClangIR front-ends. Support for basic features such as local memory and some built-ins is included, but the majority of built-in functions or barriers are not yet supported. Contributions are welcome to increase the test coverage.