clang 22.0.0git
Macros | Functions
nvptxintrin.h File Reference

Go to the source code of this file.

Macros

#define __CUDA_ARCH__   0
#define __gpu_private   __attribute__((address_space(5)))
#define __gpu_constant   __attribute__((address_space(4)))
#define __gpu_local   __attribute__((address_space(3)))
#define __gpu_global   __attribute__((address_space(1)))
#define __gpu_generic   __attribute__((address_space(0)))
#define __gpu_kernel   __attribute__((nvptx_kernel, visibility("protected")))

Functions

 _Pragma ("omp begin declare target device_type(nohost)")
 _Pragma ("omp begin declare variant match(device = {arch(nvptx64)})")
static _DEFAULT_FN_ATTRS __inline__ uint32_t __gpu_num_blocks_x (void)
static _DEFAULT_FN_ATTRS __inline__ uint32_t __gpu_num_blocks_y (void)
static _DEFAULT_FN_ATTRS __inline__ uint32_t __gpu_num_blocks_z (void)
static _DEFAULT_FN_ATTRS __inline__ uint32_t __gpu_block_id_x (void)
static _DEFAULT_FN_ATTRS __inline__ uint32_t __gpu_block_id_y (void)
static _DEFAULT_FN_ATTRS __inline__ uint32_t __gpu_block_id_z (void)
static _DEFAULT_FN_ATTRS __inline__ uint32_t __gpu_num_threads_x (void)
static _DEFAULT_FN_ATTRS __inline__ uint32_t __gpu_num_threads_y (void)
static _DEFAULT_FN_ATTRS __inline__ uint32_t __gpu_num_threads_z (void)
static _DEFAULT_FN_ATTRS __inline__ uint32_t __gpu_thread_id_x (void)
static _DEFAULT_FN_ATTRS __inline__ uint32_t __gpu_thread_id_y (void)
static _DEFAULT_FN_ATTRS __inline__ uint32_t __gpu_thread_id_z (void)
static _DEFAULT_FN_ATTRS __inline__ uint32_t __gpu_num_lanes (void)
static _DEFAULT_FN_ATTRS __inline__ uint32_t __gpu_lane_id (void)
static _DEFAULT_FN_ATTRS __inline__ uint64_t __gpu_lane_mask (void)
static _DEFAULT_FN_ATTRS __inline__ uint32_t __gpu_read_first_lane_u32 (uint64_t __lane_mask, uint32_t __x)
static _DEFAULT_FN_ATTRS __inline__ uint64_t __gpu_ballot (uint64_t __lane_mask, bool __x)
static _DEFAULT_FN_ATTRS __inline__ void __gpu_sync_threads (void)
static _DEFAULT_FN_ATTRS __inline__ void __gpu_sync_lane (uint64_t __lane_mask)
static _DEFAULT_FN_ATTRS __inline__ uint32_t __gpu_shuffle_idx_u32 (uint64_t __lane_mask, uint32_t __idx, uint32_t __x, uint32_t __width)
static _DEFAULT_FN_ATTRS __inline__ uint64_t __gpu_match_any_u32 (uint64_t __lane_mask, uint32_t __x)
static _DEFAULT_FN_ATTRS __inline__ uint64_t __gpu_match_any_u64 (uint64_t __lane_mask, uint64_t __x)
static _DEFAULT_FN_ATTRS __inline__ uint64_t __gpu_match_all_u32 (uint64_t __lane_mask, uint32_t __x)
static _DEFAULT_FN_ATTRS __inline__ uint64_t __gpu_match_all_u64 (uint64_t __lane_mask, uint64_t __x)
static _DEFAULT_FN_ATTRS __inline__ bool __gpu_is_ptr_local (void *ptr)
static _DEFAULT_FN_ATTRS __inline__ bool __gpu_is_ptr_private (void *ptr)
static _DEFAULT_FN_ATTRS __inline__ void __gpu_exit (void)
static _DEFAULT_FN_ATTRS __inline__ void __gpu_thread_suspend (void)
 _Pragma ("omp end declare variant")
 _Pragma ("omp end declare target")

Macro Definition Documentation

◆ __CUDA_ARCH__

#define __CUDA_ARCH__   0

Definition at line 21 of file nvptxintrin.h.

◆ __gpu_constant

#define __gpu_constant   __attribute__((address_space(4)))

Definition at line 29 of file nvptxintrin.h.

◆ __gpu_generic

#define __gpu_generic   __attribute__((address_space(0)))

Definition at line 32 of file nvptxintrin.h.

◆ __gpu_global

#define __gpu_global   __attribute__((address_space(1)))

Definition at line 31 of file nvptxintrin.h.

◆ __gpu_kernel

#define __gpu_kernel   __attribute__((nvptx_kernel, visibility("protected")))

Definition at line 35 of file nvptxintrin.h.

◆ __gpu_local

#define __gpu_local   __attribute__((address_space(3)))

Definition at line 30 of file nvptxintrin.h.

◆ __gpu_private

#define __gpu_private   __attribute__((address_space(5)))

Definition at line 28 of file nvptxintrin.h.

Function Documentation

◆ __gpu_ballot()

_DEFAULT_FN_ATTRS __inline__ uint64_t __gpu_ballot ( uint64_t __lane_mask,
bool __x )
static

Definition at line 121 of file nvptxintrin.h.

References _DEFAULT_FN_ATTRS.

◆ __gpu_block_id_x()

_DEFAULT_FN_ATTRS __inline__ uint32_t __gpu_block_id_x ( void )
static

Definition at line 53 of file nvptxintrin.h.

References _DEFAULT_FN_ATTRS.

◆ __gpu_block_id_y()

_DEFAULT_FN_ATTRS __inline__ uint32_t __gpu_block_id_y ( void )
static

Definition at line 58 of file nvptxintrin.h.

References _DEFAULT_FN_ATTRS.

◆ __gpu_block_id_z()

_DEFAULT_FN_ATTRS __inline__ uint32_t __gpu_block_id_z ( void )
static

Definition at line 63 of file nvptxintrin.h.

References _DEFAULT_FN_ATTRS.

◆ __gpu_exit()

_DEFAULT_FN_ATTRS __inline__ void __gpu_exit ( void )
static

Definition at line 206 of file nvptxintrin.h.

References _DEFAULT_FN_ATTRS, and noreturn.

◆ __gpu_is_ptr_local()

_DEFAULT_FN_ATTRS __inline__ bool __gpu_is_ptr_local ( void * ptr)
static

Definition at line 196 of file nvptxintrin.h.

References _DEFAULT_FN_ATTRS.

◆ __gpu_is_ptr_private()

_DEFAULT_FN_ATTRS __inline__ bool __gpu_is_ptr_private ( void * ptr)
static

Definition at line 201 of file nvptxintrin.h.

References _DEFAULT_FN_ATTRS.

◆ __gpu_lane_id()

_DEFAULT_FN_ATTRS __inline__ uint32_t __gpu_lane_id ( void )
static

Definition at line 103 of file nvptxintrin.h.

References _DEFAULT_FN_ATTRS.

◆ __gpu_lane_mask()

_DEFAULT_FN_ATTRS __inline__ uint64_t __gpu_lane_mask ( void )
static

Definition at line 108 of file nvptxintrin.h.

References _DEFAULT_FN_ATTRS.

◆ __gpu_match_all_u32()

_DEFAULT_FN_ATTRS __inline__ uint64_t __gpu_match_all_u32 ( uint64_t __lane_mask,
uint32_t __x )
static

Definition at line 173 of file nvptxintrin.h.

References __gpu_match_all_u32_impl().

◆ __gpu_match_all_u64()

_DEFAULT_FN_ATTRS __inline__ uint64_t __gpu_match_all_u64 ( uint64_t __lane_mask,
uint64_t __x )
static

Definition at line 185 of file nvptxintrin.h.

References __gpu_match_all_u64_impl().

◆ __gpu_match_any_u32()

_DEFAULT_FN_ATTRS __inline__ uint64_t __gpu_match_any_u32 ( uint64_t __lane_mask,
uint32_t __x )
static

Definition at line 151 of file nvptxintrin.h.

References __gpu_match_any_u32_impl().

◆ __gpu_match_any_u64()

_DEFAULT_FN_ATTRS __inline__ uint64_t __gpu_match_any_u64 ( uint64_t __lane_mask,
uint64_t __x )
static

Definition at line 162 of file nvptxintrin.h.

References __gpu_match_any_u64_impl().

◆ __gpu_num_blocks_x()

_DEFAULT_FN_ATTRS __inline__ uint32_t __gpu_num_blocks_x ( void )
static

Definition at line 38 of file nvptxintrin.h.

References _DEFAULT_FN_ATTRS.

◆ __gpu_num_blocks_y()

_DEFAULT_FN_ATTRS __inline__ uint32_t __gpu_num_blocks_y ( void )
static

Definition at line 43 of file nvptxintrin.h.

References _DEFAULT_FN_ATTRS.

◆ __gpu_num_blocks_z()

_DEFAULT_FN_ATTRS __inline__ uint32_t __gpu_num_blocks_z ( void )
static

Definition at line 48 of file nvptxintrin.h.

References _DEFAULT_FN_ATTRS.

◆ __gpu_num_lanes()

_DEFAULT_FN_ATTRS __inline__ uint32_t __gpu_num_lanes ( void )
static

Definition at line 98 of file nvptxintrin.h.

References _DEFAULT_FN_ATTRS.

Referenced by __gpu_read_first_lane_u32(), and __gpu_shuffle_idx_u32().

◆ __gpu_num_threads_x()

_DEFAULT_FN_ATTRS __inline__ uint32_t __gpu_num_threads_x ( void )
static

Definition at line 68 of file nvptxintrin.h.

References _DEFAULT_FN_ATTRS.

◆ __gpu_num_threads_y()

_DEFAULT_FN_ATTRS __inline__ uint32_t __gpu_num_threads_y ( void )
static

Definition at line 73 of file nvptxintrin.h.

References _DEFAULT_FN_ATTRS.

◆ __gpu_num_threads_z()

_DEFAULT_FN_ATTRS __inline__ uint32_t __gpu_num_threads_z ( void )
static

Definition at line 78 of file nvptxintrin.h.

References _DEFAULT_FN_ATTRS.

◆ __gpu_read_first_lane_u32()

_DEFAULT_FN_ATTRS __inline__ uint32_t __gpu_read_first_lane_u32 ( uint64_t __lane_mask,
uint32_t __x )
static

Definition at line 114 of file nvptxintrin.h.

References __gpu_num_lanes().

◆ __gpu_shuffle_idx_u32()

_DEFAULT_FN_ATTRS __inline__ uint32_t __gpu_shuffle_idx_u32 ( uint64_t __lane_mask,
uint32_t __idx,
uint32_t __x,
uint32_t __width )
static

Definition at line 139 of file nvptxintrin.h.

References __gpu_num_lanes().

◆ __gpu_sync_lane()

_DEFAULT_FN_ATTRS __inline__ void __gpu_sync_lane ( uint64_t __lane_mask)
static

Definition at line 133 of file nvptxintrin.h.

References _DEFAULT_FN_ATTRS.

◆ __gpu_sync_threads()

_DEFAULT_FN_ATTRS __inline__ void __gpu_sync_threads ( void )
static

Definition at line 128 of file nvptxintrin.h.

References _DEFAULT_FN_ATTRS.

◆ __gpu_thread_id_x()

_DEFAULT_FN_ATTRS __inline__ uint32_t __gpu_thread_id_x ( void )
static

Definition at line 83 of file nvptxintrin.h.

References _DEFAULT_FN_ATTRS.

◆ __gpu_thread_id_y()

_DEFAULT_FN_ATTRS __inline__ uint32_t __gpu_thread_id_y ( void )
static

Definition at line 88 of file nvptxintrin.h.

References _DEFAULT_FN_ATTRS.

◆ __gpu_thread_id_z()

_DEFAULT_FN_ATTRS __inline__ uint32_t __gpu_thread_id_z ( void )
static

Definition at line 93 of file nvptxintrin.h.

References _DEFAULT_FN_ATTRS.

◆ __gpu_thread_suspend()

_DEFAULT_FN_ATTRS __inline__ void __gpu_thread_suspend ( void )
static

Definition at line 211 of file nvptxintrin.h.

References _DEFAULT_FN_ATTRS.

◆ _Pragma() [1/4]

_Pragma ( "omp begin declare target device_type(nohost)" )

◆ _Pragma() [2/4]

_Pragma ( "omp begin declare variant match(device = {arch(nvptx64)})" )

◆ _Pragma() [3/4]

_Pragma ( "omp end declare target" )

◆ _Pragma() [4/4]

_Pragma ( "omp end declare variant" )
Generated on for clang by doxygen 1.14.0