1+ #include < librecuda.h>
2+
3+ #include < iostream>
4+ #include < iomanip>
5+ #include < vector>
6+ #include < fstream>
7+ #include < cstring>
8+
9+ inline void cudaCheck (libreCudaStatus_t error, const char *file, int line) {
10+ if (error != LIBRECUDA_SUCCESS) {
11+ const char *error_string;
12+ libreCuGetErrorString (error, &error_string);
13+ printf (" [CUDA ERROR] at file %s:%d: %s\n "
14+ exit (EXIT_FAILURE);
15+ }
16+ };
17+ #define CUDA_CHECK (err ) (cudaCheck(err, __FILE__, __LINE__))
18+
19+ struct struct_t {
20+ int x, y, z;
21+ int w, h, d;
22+ char str[32 ];
23+ char me_ugly;
24+ };
25+ static_assert (sizeof (struct_t ) == 60);
26+
27+ int main () {
28+ CUDA_CHECK (libreCuInit (0 ));
29+
30+ int device_count{};
31+ CUDA_CHECK (libreCuDeviceGetCount (&device_count));
32+ std::cout << " Device count: " std::to_string (device_count) << std::endl;
33+
34+ LibreCUdevice device{};
35+ CUDA_CHECK (libreCuDeviceGet (&device, 0 ));
36+
37+ LibreCUcontext ctx{};
38+ CUDA_CHECK (libreCuCtxCreate_v2 (&ctx, CU_CTX_SCHED_YIELD, device));
39+
40+ char name_buffer[256 ] = {};
41+ libreCuDeviceGetName (name_buffer, 256 , device);
42+ std::cout << " Device Name: " std::string (name_buffer) << std::endl;
43+ LibreCUmodule module{};
44+
45+ // read cubin file
46+ uint8_t *image;
47+ size_t n_bytes;
48+ {
49+ std::ifstream input (" read_from_struct.cubin"
50+ std::vector<uint8_t > bytes (
51+ (std::istreambuf_iterator<char >(input)),
52+ (std::istreambuf_iterator<char >()));
53+ input.close ();
54+ image = new uint8_t [bytes.size ()];
55+ memcpy (image, bytes.data (), bytes.size ());
56+ n_bytes = bytes.size ();
57+ }
58+ CUDA_CHECK (libreCuModuleLoadData (&module, image, n_bytes));
59+
60+ // read functions
61+ uint32_t num_funcs{};
62+ CUDA_CHECK (libreCuModuleGetFunctionCount (&num_funcs, module));
63+ std::cout << " Num functions: "
64+
65+ auto *functions = new LibreCUFunction[num_funcs];
66+ CUDA_CHECK (libreCuModuleEnumerateFunctions (functions, num_funcs, module));
67+
68+ for (size_t i = 0 ; i < num_funcs; i++) {
69+ LibreCUFunction func = functions[i];
70+ const char *func_name{};
71+ CUDA_CHECK (libreCuFuncGetName (&func_name, func));
72+ std::cout << " function \" " " \" "
73+ }
74+
75+ delete[] functions;
76+
77+ // find function
78+ LibreCUFunction func{};
79+ CUDA_CHECK (libreCuModuleGetFunction (&func, module, " read_from_struct"
80+ // create stream
81+ LibreCUstream stream{};
82+ CUDA_CHECK (libreCuStreamCreate (&stream, 0 ));
83+
84+ void *w_dst_va{};
85+ CUDA_CHECK (libreCuMemAlloc (&w_dst_va, sizeof (int ), true ));
86+
87+ struct_t s = {
88+ .w =64 ,
89+ };
90+
91+ void *params[] = {
92+ &s, // struct
93+ &w_dst_va, // dst
94+ };
95+
96+ CUDA_CHECK (
97+ libreCuLaunchKernel (func,
98+ 1 , 1 , 1 ,
99+ 1 , 1 , 1 ,
100+ 8192 ,
101+ stream,
102+ params, sizeof (params) / sizeof (void *),
103+ nullptr
104+ )
105+ );
106+
107+ // dispatch built up command buffer to GPU
108+ CUDA_CHECK (libreCuStreamCommence (stream));
109+
110+ // wait for work to complete
111+ CUDA_CHECK (libreCuStreamAwait (stream));
112+ std::cout << " Dst value (post exec): " int *) (w_dst_va) << std::endl;
113+
114+ // free memory
115+ CUDA_CHECK (libreCuMemFree (w_dst_va));
116+
117+ // destroy stream
118+ CUDA_CHECK (libreCuStreamDestroy (stream));
119+
120+ // unload module
121+ CUDA_CHECK (libreCuModuleUnload (module));
122+
123+ // destroy ctx
124+ CUDA_CHECK (libreCuCtxDestroy (ctx));
125+ return 0 ;
126+ }
0 commit comments