[clang] [Clang] Implement resource directory headers for common GPU intrinsics (PR #110179)

[Johannes Doerfert via cfe-commits]

================
@@ -0,0 +1,187 @@
+//===-- amdgpuintrin.h - AMDPGU intrinsic functions -----------------------===//
+//
+// 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 __AMDGPUINTRIN_H
+#define __AMDGPUINTRIN_H
+
+#ifndef __AMDGPU__
+#error "This file is intended for AMDGPU targets or offloading to AMDGPU
+#endif
+
+#include <stdbool.h>
+#include <stdint.h>
+
+#if defined(__HIP__) || defined(__CUDA__)
+#define _DEFAULT_ATTRS __attribute__((device)) __attribute__((always_inline))
+#else
+#define _DEFAULT_ATTRS __attribute__((always_inline))
+#endif
+
+#pragma omp begin declare target device_type(nohost)
+#pragma omp begin declare variant match(device = {arch(amdgcn)})
+
+// Type aliases to the address spaces used by the AMDGPU backend.
+#define _private __attribute__((opencl_private))
+#define _constant __attribute__((opencl_constant))
+#define _local __attribute__((opencl_local))
+#define _global __attribute__((opencl_global))
+
+// Attribute to declare a function as a kernel.
+#define _kernel __attribute__((amdgpu_kernel, visibility("protected")))
+
+// Returns the number of workgroups in the 'x' dimension of the grid.
+_DEFAULT_ATTRS static inline uint32_t _get_num_blocks_x() {
+  return __builtin_amdgcn_grid_size_x() / __builtin_amdgcn_workgroup_size_x();
+}
+
+// Returns the number of workgroups in the 'y' dimension of the grid.
+_DEFAULT_ATTRS static inline uint32_t _get_num_blocks_y() {
+  return __builtin_amdgcn_grid_size_y() / __builtin_amdgcn_workgroup_size_y();
+}
+
+// Returns the number of workgroups in the 'z' dimension of the grid.
+_DEFAULT_ATTRS static inline uint32_t _get_num_blocks_z() {
+  return __builtin_amdgcn_grid_size_z() / __builtin_amdgcn_workgroup_size_z();
+}
+
+// Returns the total number of workgruops in the grid.
+_DEFAULT_ATTRS static inline uint64_t _get_num_blocks() {
+  return _get_num_blocks_x() * _get_num_blocks_y() * _get_num_blocks_z();
+}
+
+// Returns the 'x' dimension of the current AMD workgroup's id.
+_DEFAULT_ATTRS static inline uint32_t _get_block_id_x() {
+  return __builtin_amdgcn_workgroup_id_x();
+}
+
+// Returns the 'y' dimension of the current AMD workgroup's id.
+_DEFAULT_ATTRS static inline uint32_t _get_block_id_y() {
+  return __builtin_amdgcn_workgroup_id_y();
+}
+
+// Returns the 'z' dimension of the current AMD workgroup's id.
+_DEFAULT_ATTRS static inline uint32_t _get_block_id_z() {
+  return __builtin_amdgcn_workgroup_id_z();
+}
+
+// Returns the absolute id of the AMD workgroup.
+_DEFAULT_ATTRS static inline uint64_t _get_block_id() {
+  return _get_block_id_x() + _get_num_blocks_x() * _get_block_id_y() +
+         _get_num_blocks_x() * _get_num_blocks_y() * _get_block_id_z();
+}
+
+// Returns the number of workitems in the 'x' dimension.
+_DEFAULT_ATTRS static inline uint32_t _get_num_threads_x() {
+  return __builtin_amdgcn_workgroup_size_x();
+}
+
+// Returns the number of workitems in the 'y' dimension.
+_DEFAULT_ATTRS static inline uint32_t _get_num_threads_y() {
+  return __builtin_amdgcn_workgroup_size_y();
+}
+
+// Returns the number of workitems in the 'z' dimension.
+_DEFAULT_ATTRS static inline uint32_t _get_num_threads_z() {
+  return __builtin_amdgcn_workgroup_size_z();
+}
+
+// Returns the total number of workitems in the workgroup.
+_DEFAULT_ATTRS static inline uint64_t _get_num_threads() {
+  return _get_num_threads_x() * _get_num_threads_y() * _get_num_threads_z();
+}
+
+// Returns the 'x' dimension id of the workitem in the current AMD workgroup.
+_DEFAULT_ATTRS static inline uint32_t _get_thread_id_x() {
+  return __builtin_amdgcn_workitem_id_x();
+}
+
+// Returns the 'y' dimension id of the workitem in the current AMD workgroup.
+_DEFAULT_ATTRS static inline uint32_t _get_thread_id_y() {
+  return __builtin_amdgcn_workitem_id_y();
+}
+
+// Returns the 'z' dimension id of the workitem in the current AMD workgroup.
+_DEFAULT_ATTRS static inline uint32_t _get_thread_id_z() {
+  return __builtin_amdgcn_workitem_id_z();
+}
+
+// Returns the absolute id of the thread in the current AMD workgroup.
+_DEFAULT_ATTRS static inline uint64_t _get_thread_id() {
+  return _get_thread_id_x() + _get_num_threads_x() * _get_thread_id_y() +
+         _get_num_threads_x() * _get_num_threads_y() * _get_thread_id_z();
+}
+
+// Returns the size of an AMD wavefront, either 32 or 64 depending on hardware
+// and compilation options.
+_DEFAULT_ATTRS static inline uint32_t _get_lane_size() {
+  return __builtin_amdgcn_wavefrontsize();
+}
+
+// Returns the id of the thread inside of an AMD wavefront executing together.
+_DEFAULT_ATTRS [[clang::convergent]] static inline uint32_t _get_lane_id() {
+  return __builtin_amdgcn_mbcnt_hi(~0u, __builtin_amdgcn_mbcnt_lo(~0u, 0u));
+}
+
+// Returns the bit-mask of active threads in the current wavefront.
+_DEFAULT_ATTRS [[clang::convergent]] static inline uint64_t _get_lane_mask() {
+  return __builtin_amdgcn_read_exec();
+}
+
+// Copies the value from the first active thread in the wavefront to the rest.
+_DEFAULT_ATTRS [[clang::convergent]] static inline uint32_t
+_broadcast_value(uint64_t, uint32_t x) {
+  return __builtin_amdgcn_readfirstlane(x);
+}
+
+// Returns a bitmask of threads in the current lane for which \p x is true.
+_DEFAULT_ATTRS [[clang::convergent]] static inline uint64_t
+_ballot(uint64_t lane_mask, bool x) {
+  // The lane_mask & gives the nvptx semantics when lane_mask is a subset of
+  // the active threads
+  return lane_mask & __builtin_amdgcn_ballot_w64(x);
+}
+
+// Waits for all the threads in the block to converge and issues a fence.
+_DEFAULT_ATTRS [[clang::convergent]] static inline void _sync_threads() {
+  __builtin_amdgcn_s_barrier();
+  __builtin_amdgcn_fence(__ATOMIC_ACQUIRE, "workgroup");
+}
+
+// Wait for all threads in the wavefront to converge, this is a noop on AMDGPU.
+_DEFAULT_ATTRS [[clang::convergent]] static inline void _sync_lane(uint64_t) {
----------------
jdoerfert wrote:

why `_sync_lane`? `_sync_warp` would math `_sync_threads` wrt. CUDA, `_sync_lanes` would match the plural. If you don't keep the CUDA matching, it's weird since `_sync_threads` then arbitrarily syncs over the workgroup while `_sync_lanes` does not, even though "threads" and "lanes" are the same thing here (i.a., "Wait for all threads in the wavefront to converge"). I would stick to CUDA names, *or* pick descriptive names. That is orthogonal to the memory fence issues partially mentioned above.

https://github.com/llvm/llvm-project/pull/110179