Add openmp library from llvm

system/lib/llvm-openmp/include/ompx.h

@@ -0,0 +1,240 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef __OMPX_H
+#define __OMPX_H
+
+#if (defined(__NVPTX__) || defined(__AMDGPU__) || defined(__SPIRV__))
+#include <gpuintrin.h>
+#define __OMPX_TARGET_IS_GPU
+#endif
+
+typedef unsigned long uint64_t;
+typedef unsigned int uint32_t;
+
+static inline uint32_t __warpSize(void) {
+#ifdef __OMPX_TARGET_IS_GPU
+  return __gpu_num_lanes();
+#else
+  __builtin_trap();
+#endif
+}
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+int omp_get_ancestor_thread_num(int);
+int omp_get_team_size(int);
+
+#ifdef __cplusplus
+}
+#endif
+
+/// Target kernel language extensions
+///
+/// These extensions exist for the host to allow fallback implementations,
+/// however, they cannot be arbitrarily composed with OpenMP. If the rules of
+/// the kernel language are followed, the host fallbacks should behave as
+/// expected since the kernel is represented as 3 sequential outer loops, one
+/// for each grid dimension, and three (nested) parallel loops, one for each
+/// block dimension. This fallback is not supposed to be optimal and should be
+/// configurable by the user.
+///
+///{
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+enum {
+  ompx_relaxed = __ATOMIC_RELAXED,
+  ompx_aquire = __ATOMIC_ACQUIRE,
+  ompx_release = __ATOMIC_RELEASE,
+  ompx_acq_rel = __ATOMIC_ACQ_REL,
+  ompx_seq_cst = __ATOMIC_SEQ_CST,
+};
+
+enum {
+  ompx_dim_x = 0,
+  ompx_dim_y = 1,
+  ompx_dim_z = 2,
+};
+
+// TODO: The following implementation is for host fallback. We need to disable
+// generation of host fallback in kernel language mode.
+#pragma omp begin declare variant match(device = {kind(cpu)})
+
+/// ompx_{thread,block}_{id,dim}
+///{
+#define _TGT_KERNEL_LANGUAGE_HOST_IMPL_GRID_C(NAME, VALUE)                     \
+  static inline int ompx_##NAME(int Dim) { return VALUE; }
+
+_TGT_KERNEL_LANGUAGE_HOST_IMPL_GRID_C(thread_id,
+                                      omp_get_ancestor_thread_num(Dim + 1))
+_TGT_KERNEL_LANGUAGE_HOST_IMPL_GRID_C(block_dim, omp_get_team_size(Dim + 1))
+_TGT_KERNEL_LANGUAGE_HOST_IMPL_GRID_C(block_id, 0)
+_TGT_KERNEL_LANGUAGE_HOST_IMPL_GRID_C(grid_dim, 1)
+#undef _TGT_KERNEL_LANGUAGE_HOST_IMPL_GRID_C
+///}
+
+/// ompx_{sync_block}_{,divergent}
+///{
+#define _TGT_KERNEL_LANGUAGE_HOST_IMPL_SYNC_C(RETTY, NAME, ARGS, BODY)         \
+  static inline RETTY ompx_##NAME(ARGS) { BODY; }
+
+_TGT_KERNEL_LANGUAGE_HOST_IMPL_SYNC_C(void, sync_block, int Ordering,
+                                      _Pragma("omp barrier"))
+_TGT_KERNEL_LANGUAGE_HOST_IMPL_SYNC_C(void, sync_block_acq_rel, void,
+                                      ompx_sync_block(ompx_acq_rel))
+_TGT_KERNEL_LANGUAGE_HOST_IMPL_SYNC_C(void, sync_block_divergent, int Ordering,
+                                      ompx_sync_block(Ordering))
+#undef _TGT_KERNEL_LANGUAGE_HOST_IMPL_SYNC_C
+///}
+
+static inline uint64_t ompx_ballot_sync(uint64_t mask, int pred) {
+  __builtin_trap();
+}
+
+/// ompx_shfl_down_sync_{i,f,l,d}
+///{
+#define _TGT_KERNEL_LANGUAGE_SHFL_DOWN_SYNC_HOST_IMPL(TYPE, TY)                \
+  static inline TYPE ompx_shfl_down_sync_##TY(uint64_t mask, TYPE var,         \
+                                              unsigned delta, int width) {     \
+    __builtin_trap();                                                          \
+  }
+
+_TGT_KERNEL_LANGUAGE_SHFL_DOWN_SYNC_HOST_IMPL(int, i)
+_TGT_KERNEL_LANGUAGE_SHFL_DOWN_SYNC_HOST_IMPL(float, f)
+_TGT_KERNEL_LANGUAGE_SHFL_DOWN_SYNC_HOST_IMPL(long, l)
+_TGT_KERNEL_LANGUAGE_SHFL_DOWN_SYNC_HOST_IMPL(double, d)
+
+#undef _TGT_KERNEL_LANGUAGE_SHFL_DOWN_SYNC_HOST_IMPL
+///}
+
+#pragma omp end declare variant
+
+/// ompx_{sync_block}_{,divergent}
+///{
+#define _TGT_KERNEL_LANGUAGE_DECL_SYNC_C(RETTY, NAME, ARGS)         \
+  RETTY ompx_##NAME(ARGS);
+
+_TGT_KERNEL_LANGUAGE_DECL_SYNC_C(void, sync_block, int Ordering)
+_TGT_KERNEL_LANGUAGE_DECL_SYNC_C(void, sync_block_acq_rel, void)
+_TGT_KERNEL_LANGUAGE_DECL_SYNC_C(void, sync_block_divergent, int Ordering)
+#undef _TGT_KERNEL_LANGUAGE_DECL_SYNC_C
+///}
+
+/// ompx_{thread,block}_{id,dim}_{x,y,z}
+///{
+#define _TGT_KERNEL_LANGUAGE_DECL_GRID_C(NAME)                                 \
+  int ompx_##NAME(int Dim);                                                    \
+  static inline int ompx_##NAME##_x() { return ompx_##NAME(ompx_dim_x); }      \
+  static inline int ompx_##NAME##_y() { return ompx_##NAME(ompx_dim_y); }      \
+  static inline int ompx_##NAME##_z() { return ompx_##NAME(ompx_dim_z); }
+
+_TGT_KERNEL_LANGUAGE_DECL_GRID_C(thread_id)
+_TGT_KERNEL_LANGUAGE_DECL_GRID_C(block_dim)
+_TGT_KERNEL_LANGUAGE_DECL_GRID_C(block_id)
+_TGT_KERNEL_LANGUAGE_DECL_GRID_C(grid_dim)
+#undef _TGT_KERNEL_LANGUAGE_DECL_GRID_C
+///}
+
+uint64_t ompx_ballot_sync(uint64_t mask, int pred);
+
+/// ompx_shfl_down_sync_{i,f,l,d}
+///{
+#define _TGT_KERNEL_LANGUAGE_SHFL_DOWN_SYNC(TYPE, TY)                          \
+  TYPE ompx_shfl_down_sync_##TY(uint64_t mask, TYPE var, unsigned delta,       \
+                                int width);
+
+_TGT_KERNEL_LANGUAGE_SHFL_DOWN_SYNC(int, i)
+_TGT_KERNEL_LANGUAGE_SHFL_DOWN_SYNC(float, f)
+_TGT_KERNEL_LANGUAGE_SHFL_DOWN_SYNC(long, l)
+_TGT_KERNEL_LANGUAGE_SHFL_DOWN_SYNC(double, d)
+
+#undef _TGT_KERNEL_LANGUAGE_SHFL_DOWN_SYNC
+///}
+
+#ifdef __cplusplus
+}
+#endif
+
+#ifdef __cplusplus
+
+namespace ompx {
+
+enum {
+  dim_x = ompx_dim_x,
+  dim_y = ompx_dim_y,
+  dim_z = ompx_dim_z,
+};
+
+enum {
+  relaxed = ompx_relaxed ,
+  aquire = ompx_aquire,
+  release = ompx_release,
+  acc_rel = ompx_acq_rel,
+  seq_cst = ompx_seq_cst,
+};
+
+/// ompx::{thread,block}_{id,dim}_{,x,y,z}
+///{
+#define _TGT_KERNEL_LANGUAGE_HOST_IMPL_GRID_CXX(NAME)                          \
+  static inline int NAME(int Dim) noexcept { return ompx_##NAME(Dim); }        \
+  static inline int NAME##_x() noexcept { return NAME(ompx_dim_x); }           \
+  static inline int NAME##_y() noexcept { return NAME(ompx_dim_y); }           \
+  static inline int NAME##_z() noexcept { return NAME(ompx_dim_z); }
+
+_TGT_KERNEL_LANGUAGE_HOST_IMPL_GRID_CXX(thread_id)
+_TGT_KERNEL_LANGUAGE_HOST_IMPL_GRID_CXX(block_dim)
+_TGT_KERNEL_LANGUAGE_HOST_IMPL_GRID_CXX(block_id)
+_TGT_KERNEL_LANGUAGE_HOST_IMPL_GRID_CXX(grid_dim)
+#undef _TGT_KERNEL_LANGUAGE_HOST_IMPL_GRID_CXX
+///}
+
+/// ompx_{sync_block}_{,divergent}
+///{
+#define _TGT_KERNEL_LANGUAGE_HOST_IMPL_SYNC_CXX(RETTY, NAME, ARGS, CALL_ARGS)  \
+  static inline RETTY NAME(ARGS) {               \
+    return ompx_##NAME(CALL_ARGS);                                             \
+  }
+
+_TGT_KERNEL_LANGUAGE_HOST_IMPL_SYNC_CXX(void, sync_block, int Ordering = acc_rel,
+                                        Ordering)
+_TGT_KERNEL_LANGUAGE_HOST_IMPL_SYNC_CXX(void, sync_block_divergent,
+                                        int Ordering = acc_rel, Ordering)
+#undef _TGT_KERNEL_LANGUAGE_HOST_IMPL_SYNC_CXX
+///}
+
+static inline uint64_t ballot_sync(uint64_t mask, int pred) {
+  return ompx_ballot_sync(mask, pred);
+}
+
+/// shfl_down_sync
+///{
+#define _TGT_KERNEL_LANGUAGE_SHFL_DOWN_SYNC(TYPE, TY)                          \
+  static inline TYPE shfl_down_sync(uint64_t mask, TYPE var, unsigned delta,   \
+                                    int width = __warpSize()) {                \
+    return ompx_shfl_down_sync_##TY(mask, var, delta, width);                  \
+  }
+
+_TGT_KERNEL_LANGUAGE_SHFL_DOWN_SYNC(int, i)
+_TGT_KERNEL_LANGUAGE_SHFL_DOWN_SYNC(float, f)
+_TGT_KERNEL_LANGUAGE_SHFL_DOWN_SYNC(long, l)
+_TGT_KERNEL_LANGUAGE_SHFL_DOWN_SYNC(double, d)
+
+#undef _TGT_KERNEL_LANGUAGE_SHFL_DOWN_SYNC
+///}
+
+} // namespace ompx
+#endif
+
+///}
+
+#endif /* __OMPX_H */

system/lib/llvm-openmp/prebuilt/kmp_config.h

@@ -0,0 +1,149 @@
+/*
+ * kmp_config.h -- Feature macros
+ */
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+#ifndef KMP_CONFIG_H
+#define KMP_CONFIG_H
+
+#include "kmp_platform.h"
+
+// cmakedefine01 MACRO will define MACRO as either 0 or 1
+// cmakedefine MACRO 1 will define MACRO as 1 or leave undefined
+#define DEBUG_BUILD 0
+#define RELWITHDEBINFO_BUILD 0
+#define LIBOMP_USE_ITT_NOTIFY 0
+#define USE_ITT_NOTIFY LIBOMP_USE_ITT_NOTIFY
+#if ! LIBOMP_USE_ITT_NOTIFY
+# define INTEL_NO_ITTNOTIFY_API
+#endif
+#define LIBOMP_USE_VERSION_SYMBOLS 0
+#if LIBOMP_USE_VERSION_SYMBOLS
+# define KMP_USE_VERSION_SYMBOLS
+#endif
+#define LIBOMP_HAVE_WEAK_ATTRIBUTE 1
+#define KMP_HAVE_WEAK_ATTRIBUTE LIBOMP_HAVE_WEAK_ATTRIBUTE
+#define LIBOMP_HAVE_PSAPI 0
+#define KMP_HAVE_PSAPI LIBOMP_HAVE_PSAPI
+#define LIBOMP_STATS 0
+#define KMP_STATS_ENABLED LIBOMP_STATS
+#define LIBOMP_HAVE_X86INTRIN_H 0
+#define KMP_HAVE_X86INTRIN_H LIBOMP_HAVE_X86INTRIN_H
+#define LIBOMP_HAVE___BUILTIN_READCYCLECOUNTER 0
+#define KMP_HAVE___BUILTIN_READCYCLECOUNTER LIBOMP_HAVE___BUILTIN_READCYCLECOUNTER
+#define LIBOMP_HAVE___RDTSC 0
+#define KMP_HAVE___RDTSC LIBOMP_HAVE___RDTSC
+#define LIBOMP_USE_DEBUGGER 0
+#define USE_DEBUGGER LIBOMP_USE_DEBUGGER
+#define LIBOMP_OMPT_DEBUG 0
+#define OMPT_DEBUG LIBOMP_OMPT_DEBUG
+#define LIBOMP_OMPT_SUPPORT 0
+#define OMPT_SUPPORT LIBOMP_OMPT_SUPPORT
+#define LIBOMP_OMPD_SUPPORT 0
+#define OMPD_SUPPORT LIBOMP_OMPD_SUPPORT
+#define LIBOMP_TASKGRAPH_EXPERIMENTAL 0
+#define OMP_TASKGRAPH_EXPERIMENTAL LIBOMP_TASKGRAPH_EXPERIMENTAL
+#define LIBOMP_PROFILING_SUPPORT 0
+#define OMP_PROFILING_SUPPORT LIBOMP_PROFILING_SUPPORT
+#define LIBOMP_OMPT_OPTIONAL 1
+#define OMPT_OPTIONAL LIBOMP_OMPT_OPTIONAL
+#define LIBOMP_USE_ADAPTIVE_LOCKS 0
+#define KMP_USE_ADAPTIVE_LOCKS LIBOMP_USE_ADAPTIVE_LOCKS
+#define KMP_DEBUG_ADAPTIVE_LOCKS 0
+#define LIBOMP_USE_INTERNODE_ALIGNMENT 0
+#define KMP_USE_INTERNODE_ALIGNMENT LIBOMP_USE_INTERNODE_ALIGNMENT
+#define LIBOMP_ENABLE_ASSERTIONS 0
+#define KMP_USE_ASSERT LIBOMP_ENABLE_ASSERTIONS
+#define LIBOMP_USE_HIER_SCHED 0
+#define KMP_USE_HIER_SCHED LIBOMP_USE_HIER_SCHED
+#define STUBS_LIBRARY 0
+#define LIBOMP_USE_HWLOC 0
+#define KMP_USE_HWLOC LIBOMP_USE_HWLOC
+#define LIBOMP_ENABLE_SHARED 0
+#define KMP_DYNAMIC_LIB LIBOMP_ENABLE_SHARED
+#define KMP_ARCH_STR "wasm32"
+#define KMP_LIBRARY_FILE "libomp.a"
+#define KMP_VERSION_MAJOR 5
+#define KMP_VERSION_MINOR 0
+#define MSVC 0
+#define KMP_MSVC_COMPAT MSVC
+#define LIBOMP_HAVE_WAITPKG_INTRINSICS 0
+#define KMP_HAVE_WAITPKG_INTRINSICS LIBOMP_HAVE_WAITPKG_INTRINSICS
+#define LIBOMP_HAVE_RTM_INTRINSICS 0
+#define KMP_HAVE_RTM_INTRINSICS LIBOMP_HAVE_RTM_INTRINSICS
+#define LIBOMP_HAVE_IMMINTRIN_H 0
+#define KMP_HAVE_IMMINTRIN_H LIBOMP_HAVE_IMMINTRIN_H
+#define LIBOMP_HAVE_INTRIN_H 0
+#define KMP_HAVE_INTRIN_H LIBOMP_HAVE_INTRIN_H
+#define LIBOMP_HAVE_ATTRIBUTE_WAITPKG 0
+#define KMP_HAVE_ATTRIBUTE_WAITPKG LIBOMP_HAVE_ATTRIBUTE_WAITPKG
+#define LIBOMP_HAVE_ATTRIBUTE_RTM 0
+#define KMP_HAVE_ATTRIBUTE_RTM LIBOMP_HAVE_ATTRIBUTE_RTM
+#define LIBOMP_ARCH_AARCH64_A64FX 0
+#define KMP_ARCH_AARCH64_A64FX LIBOMP_ARCH_AARCH64_A64FX
+#define LIBOMP_HAVE_XMMINTRIN_H 0
+#define KMP_HAVE_XMMINTRIN_H LIBOMP_HAVE_XMMINTRIN_H
+#define LIBOMP_HAVE__MM_MALLOC 0
+#define KMP_HAVE__MM_MALLOC LIBOMP_HAVE__MM_MALLOC
+#define LIBOMP_HAVE_ALIGNED_ALLOC 1
+#define KMP_HAVE_ALIGNED_ALLOC LIBOMP_HAVE_ALIGNED_ALLOC
+#define LIBOMP_HAVE_POSIX_MEMALIGN 1
+#define KMP_HAVE_POSIX_MEMALIGN LIBOMP_HAVE_POSIX_MEMALIGN
+#define LIBOMP_HAVE__ALIGNED_MALLOC 0
+#define KMP_HAVE__ALIGNED_MALLOC LIBOMP_HAVE__ALIGNED_MALLOC
+#define LIBOMP_USE_CANCEL_THREADS 1
+#define KMP_CANCEL_THREADS LIBOMP_USE_CANCEL_THREADS
+#define OPENMP_ENABLE_LIBOMPTARGET 0
+#define ENABLE_LIBOMPTARGET OPENMP_ENABLE_LIBOMPTARGET
+
+// Configured cache line based on architecture
+#if KMP_ARCH_PPC64 || KMP_ARCH_PPC
+# define CACHE_LINE 128
+#elif KMP_ARCH_AARCH64_A64FX
+# define CACHE_LINE 256
+#elif KMP_ARCH_S390X
+# define CACHE_LINE 256
+#else
+# define CACHE_LINE 64
+#endif
+
+#if ! KMP_32_BIT_ARCH
+# define BUILD_I8 1
+#endif
+
+#define KMP_ADJUST_BLOCKTIME 1
+#define BUILD_PARALLEL_ORDERED 1
+#define KMP_ASM_INTRINS 1
+#define USE_ITT_BUILD LIBOMP_USE_ITT_NOTIFY
+#define INTEL_ITTNOTIFY_PREFIX __kmp_itt_
+#if ! (KMP_MIC || KMP_OS_HAIKU)
+# define USE_LOAD_BALANCE 1
+#endif
+#if ! (KMP_OS_WINDOWS || KMP_OS_DARWIN)
+# define KMP_TDATA_GTID 1
+#endif
+#if STUBS_LIBRARY
+# define KMP_STUB 1
+#endif
+#if DEBUG_BUILD || RELWITHDEBINFO_BUILD
+# define KMP_DEBUG 1
+#endif
+
+#if KMP_OS_WINDOWS
+# define KMP_WIN_CDECL
+#else
+# define BUILD_TV
+# define KMP_GOMP_COMPAT
+#endif
+
+// use shared memory with dynamic library (except Android, where shm_*
+// functions don't exist).
+#if KMP_OS_UNIX && KMP_DYNAMIC_LIB && !__ANDROID__
+#define KMP_USE_SHM
+#endif
+#endif // KMP_CONFIG_H