[libclc] Produce subnormal results for exp/exp2 on FP32 (#212696) The FP32 __clc_exp/__clc_exp2 implementations scale the reduced result by 2^p using integer bit manipulation (as_int(y) + (p << 23)). That representation cannot encode a subnormal, so inputs whose result falls below the smallest normal were unconditionally flushed to zero via "x < llim ? 0.0f". On devices that report CL_FP_DENORM this loses the correct subnormal result (e.g. erfc(), which is built on exp(), failed CTS math_brute_force with the reference subnormal vs a returned 0). Recompute the subnormal range with __clc_ldexp, which does produce subnormals, gated on __clc_fp32_subnormals_supported() so flush-to-zero targets keep the previous fast path. The ldexp path is restricted to [subnorm_llim, llim); inputs below subnorm_llim (including -inf and large-magnitude negatives) still flush to zero, since their reduced exponent argument would be ill-formed for ldexp. GitOrigin-RevId: 7cec6725695788dc9a82aa4464611a9dacde2549
libclc is an open source implementation of the library requirements of the OpenCL C programming language, as specified by the OpenCL 1.1 Specification. The following sections of the specification impose library requirements:
libclc is intended to be used with the Clang compiler's OpenCL frontend.
libclc is designed to be portable and extensible. To this end, it provides generic implementations of most library requirements, allowing the target to override the generic implementation at the granularity of individual functions.
libclc currently supports PTX, AMDGPU, SPIRV and CLSPV targets, but support for more targets is welcome.
libclc is built as part of an LLVM runtimes build.
Select the targets to build with LLVM_RUNTIME_TARGETS, and enable libclc for each selected target with the matching RUNTIMES_<target-triple>_LLVM_ENABLE_RUNTIMES cache entry.
CMake cache files for target build are provided under libclc/cmake/caches/. Pass one with -C to avoid setting each variable manually:
| Cache file | Targets |
|---|---|
amdgcn-amd-amdhsa-llvm.cmake | AMDGPU |
nvptx64-nvidia-cuda.cmake | NVPTX64 |
spirv.cmake | SPIR-V |
spirv-vulkan.cmake | Vulkan (clspv) |
all-targets.cmake | All of the above |
For example, from the root of the llvm-project repository:
cmake -C libclc/cmake/caches/amdgcn-amd-amdhsa-llvm.cmake -G Ninja -B build llvm
The cache file sets LLVM_ENABLE_PROJECTS, LLVM_RUNTIME_TARGETS, and all required RUNTIMES_* entries. Additional -D flags can still be appended to override individual settings.
cd llvm-project mkdir build cd build cmake ../llvm -G Ninja -DLLVM_ENABLE_PROJECTS="clang" -DCMAKE_BUILD_TYPE=Release \ -DRUNTIMES_amdgcn-amd-amdhsa-llvm_LLVM_ENABLE_RUNTIMES=libclc \ -DLLVM_RUNTIME_TARGETS="amdgcn-amd-amdhsa-llvm"
cmake ../llvm -G Ninja -DLLVM_ENABLE_PROJECTS="clang" -DCMAKE_BUILD_TYPE=Release \ -DRUNTIMES_nvptx64-nvidia-cuda_LLVM_ENABLE_RUNTIMES=libclc \ -DLLVM_RUNTIME_TARGETS="nvptx64-nvidia-cuda"
cmake ../llvm -G Ninja -DLLVM_ENABLE_PROJECTS="clang" -DCMAKE_BUILD_TYPE=Release \ -DRUNTIMES_spirv32-unknown-vulkan_LLVM_ENABLE_RUNTIMES=libclc \ -DRUNTIMES_spirv64-unknown-vulkan_LLVM_ENABLE_RUNTIMES=libclc \ -DLLVM_RUNTIME_TARGETS="spirv32-unknown-vulkan;spirv64-unknown-vulkan"
cmake ../llvm -G Ninja -DLLVM_ENABLE_PROJECTS="clang" -DCMAKE_BUILD_TYPE=Release \ -DRUNTIMES_spirv32-unknown-unknown_LLVM_ENABLE_RUNTIMES=libclc \ -DRUNTIMES_spirv64-unknown-unknown_LLVM_ENABLE_RUNTIMES=libclc \ -DLLVM_RUNTIME_TARGETS="spirv32-unknown-unknown;spirv64-unknown-unknown"
To build multiple targets, pass them as a semicolon-separated list in LLVM_RUNTIME_TARGETS and provide a matching RUNTIMES_<target-triple>_LLVM_ENABLE_RUNTIMES=libclc entry for each target.
ninja
ninja install
Note you can use the DESTDIR Makefile variable to do staged installs.
DESTDIR=/path/for/staged/install ninja install
libclc utilizes the LLVM testing infrastructure.
To execute all per-target tests for libclc.
ninja check-libclc
check-libclc is a top-level target that aggregates all per-target tests.
If you are working on a specific target, you can run tests for just that target triple:
ninja check-libclc-<target-triple>
Alternatively, you can run target-specific tests via the runtimes build by pointing to the target-specific build directory:
ninja -C runtimes/runtimes-<target-triple>-bins check-libclc
To build out of tree, or in other words, against an existing LLVM build or install:
CC=$(<path-to>/llvm-config --bindir)/clang cmake \ <path-to>/llvm-project/libclc/CMakeLists.txt -DCMAKE_BUILD_TYPE=Release \ -G Ninja -DLLVM_DIR=$(<path-to>/llvm-config --cmakedir) \ -DLLVM_RUNTIMES_TARGET=<target-triple> $ ninja
Then install as before.
In both cases, the LLVM used must include the targets you want libclc support for (AMDGPU and NVPTX are enabled in LLVM by default). Apart from SPIRV where you do not need an LLVM target but you do need the llvm-spirv tool available. Either build this in-tree, or place it in the directory pointed to by LLVM_TOOLS_BINARY_DIR.