| 1 | //===-- AMDGPU specific platform definitions for math support -------------===// |
|---|
| 2 | // |
|---|
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
|---|
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
|---|
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
|---|
| 6 | // |
|---|
| 7 | //===----------------------------------------------------------------------===// |
|---|
| 8 | |
|---|
| 9 | #ifndef LLVM_LIBC_SRC_MATH_AMDGPU_PLATFORM_H |
|---|
| 10 | #define LLVM_LIBC_SRC_MATH_AMDGPU_PLATFORM_H |
|---|
| 11 | |
|---|
| 12 | #include "src/__support/macros/attributes.h" |
|---|
| 13 | |
|---|
| 14 | #include <stdint.h> |
|---|
| 15 | |
|---|
| 16 | namespace LIBC_NAMESPACE { |
|---|
| 17 | |
|---|
| 18 | // The ROCm device library uses control globals to alter codegen for the |
|---|
| 19 | // different targets. To avoid needing to link them in manually we simply |
|---|
| 20 | // define them here. |
|---|
| 21 | extern "C" { |
|---|
| 22 | |
|---|
| 23 | // Disable unsafe math optimizations in the implementation. |
|---|
| 24 | extern const LIBC_INLINE_VAR uint8_t __oclc_unsafe_math_opt = 0; |
|---|
| 25 | |
|---|
| 26 | // Disable denormalization at zero optimizations in the implementation. |
|---|
| 27 | extern const LIBC_INLINE_VAR uint8_t __oclc_daz_opt = 0; |
|---|
| 28 | |
|---|
| 29 | // Disable rounding optimizations for 32-bit square roots. |
|---|
| 30 | extern const LIBC_INLINE_VAR uint8_t __oclc_correctly_rounded_sqrt32 = 1; |
|---|
| 31 | |
|---|
| 32 | // Disable finite math optimizations. |
|---|
| 33 | extern const LIBC_INLINE_VAR uint8_t __oclc_finite_only_opt = 0; |
|---|
| 34 | |
|---|
| 35 | // Set the ISA value to a high enough value that the ROCm device library math |
|---|
| 36 | // functions will assume we have fast FMA operations among other features. This |
|---|
| 37 | // is determined to be safe on all targets by looking at the source code. |
|---|
| 38 | // https://github.com/ROCm/ROCm-Device-Libs/blob/amd-stg-open/ocml/src/opts.h |
|---|
| 39 | extern const LIBC_INLINE_VAR uint32_t __oclc_ISA_version = 9000; |
|---|
| 40 | } |
|---|
| 41 | |
|---|
| 42 | // These aliases cause clang to emit the control constants with ODR linkage. |
|---|
| 43 | // This allows us to link against the symbols without preventing them from being |
|---|
| 44 | // optimized out or causing symbol collisions. |
|---|
| 45 | [[gnu::alias("__oclc_unsafe_math_opt" )]] const uint8_t __oclc_unsafe_math_opt__; |
|---|
| 46 | [[gnu::alias("__oclc_daz_opt" )]] const uint8_t __oclc_daz_opt__; |
|---|
| 47 | [[gnu::alias("__oclc_correctly_rounded_sqrt32" )]] const uint8_t |
|---|
| 48 | __oclc_correctly_rounded_sqrt32__; |
|---|
| 49 | [[gnu::alias("__oclc_finite_only_opt" )]] const uint8_t __oclc_finite_only_opt__; |
|---|
| 50 | [[gnu::alias("__oclc_ISA_version" )]] const uint32_t __oclc_ISA_version__; |
|---|
| 51 | |
|---|
| 52 | } // namespace LIBC_NAMESPACE |
|---|
| 53 | |
|---|
| 54 | #endif // LLVM_LIBC_SRC_MATH_AMDGPU_PLATFORM_H |
|---|
| 55 | |
|---|
