Debian Codegraph @

rts/opencl/prelude.cl

raw
1 // Start of prelude.cl
2
3 #define SCALAR_FUN_ATTR static inline
4 #define FUTHARK_FUN_ATTR static
5
6 typedef char int8_t;
7 typedef short int16_t;
8 typedef int int32_t;
9 typedef long int64_t;
10
11 typedef uchar uint8_t;
12 typedef ushort uint16_t;
13 typedef uint uint32_t;
14 typedef ulong uint64_t;
15
16 #define get_tblock_id(d) get_group_id(d)
17 #define get_num_tblocks(d) get_num_groups(d)
18
19 // Clang-based OpenCL implementations need this for 'static' to work.
20 #ifdef cl_clang_storage_class_specifiers
21 #pragma OPENCL EXTENSION cl_clang_storage_class_specifiers : enable
22 #endif
23 #pragma OPENCL EXTENSION cl_khr_byte_addressable_store : enable
24
25 #ifdef FUTHARK_F64_ENABLED
26 #pragma OPENCL EXTENSION cl_khr_fp64 : enable
27 #endif
28
29 #pragma OPENCL EXTENSION cl_khr_int64_base_atomics : enable
30 #pragma OPENCL EXTENSION cl_khr_int64_extended_atomics : enable
31
32 // NVIDIAs OpenCL does not create device-wide memory fences (see #734), so we
33 // use inline assembly if we detect we are on an NVIDIA GPU.
34 #ifdef cl_nv_pragma_unroll
35 static inline void mem_fence_global() {
36 asm("membar.gl;");
37 }
38 #else
39 static inline void mem_fence_global() {
40 mem_fence(CLK_LOCAL_MEM_FENCE | CLK_GLOBAL_MEM_FENCE);
41 }
42 #endif
43 static inline void mem_fence_local() {
44 mem_fence(CLK_LOCAL_MEM_FENCE);
45 }
46
47 static inline void barrier_local() {
48 barrier(CLK_LOCAL_MEM_FENCE);
49 }
50
51 // Important for this to be int64_t so it has proper alignment for any type.
52 #define SHARED_MEM_PARAM __local uint64_t* shared_mem,
53 #define FUTHARK_KERNEL __kernel
54 #define FUTHARK_KERNEL_SIZED(a,b,c) __attribute__((reqd_work_group_size(a, b, c))) __kernel
55
56 // End of prelude.cl