decuda 0.1.1

CUDA to HIP, SYCL, OpenCL, and Rust GPU migration tool — automatic source-code translator for porting CUDA C++ kernels to AMD ROCm HIP, Intel oneAPI SYCL, Khronos OpenCL, and Rust GPU (cust / rust-gpu)
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289

//! Database of CUDA runtime / driver APIs and how each maps to the supported
//! targets. The database is loaded once at startup; entries can be queried by
//! API name and target backend.

#![allow(dead_code)]

use std::collections::BTreeMap;

use once_cell::sync::Lazy;

use crate::cli::Target;

/// Per-API metadata.
#[derive(Debug, Clone)]
pub struct ApiInfo {
    pub header: &'static str,
    pub hip: Option<&'static str>,
    pub sycl: Option<&'static str>,
    pub rust: Option<&'static str>,
    pub opencl: Option<&'static str>,
    pub note: &'static str,
}

impl ApiInfo {
    pub fn supports(&self, t: Target) -> bool {
        self.mapping(t).is_some()
    }

    pub fn mapping(&self, t: Target) -> Option<&'static str> {
        match t {
            Target::Hip => self.hip,
            Target::Sycl => self.sycl,
            Target::Rust => self.rust,
            Target::Opencl => self.opencl,
            Target::All => None,
        }
    }

    pub fn summarize(&self, t: Target) -> String {
        if let Some(m) = self.mapping(t) {
            format!("-> {t:?} = {m}")
        } else {
            format!("(no mapping for {t:?})")
        }
    }
}

pub type ApiMap = BTreeMap<&'static str, ApiInfo>;

pub fn database() -> &'static ApiMap {
    &DATABASE
}

/// All known CUDA runtime API entries.
///
/// Entries marked `None` for a backend mean: no automatic translation; the
/// rewriter preserves the original token and emits a warning.
static DATABASE: Lazy<ApiMap> = Lazy::new(|| {
    let mut m: ApiMap = BTreeMap::new();

    // ---- Memory management -------------------------------------------------
    insert(&mut m, "cudaMalloc", api("cuda_runtime.h",
        Some("hipMalloc"), None, None, None,
        "opencl: use clCreateBuffer + clEnqueueWriteBuffer"));
    insert(&mut m, "cudaFree", api("cuda_runtime.h",
        Some("hipFree"), None, None, None,
        "opencl: use clReleaseMemObject"));
    insert(&mut m, "cudaMemcpy", api("cuda_runtime.h",
        Some("hipMemcpy"), None, None, None,
        "opencl: clEnqueueCopyBuffer / clEnqueueReadBuffer"));
    insert(&mut m, "cudaMemcpyAsync", api("cuda_runtime.h",
        Some("hipMemcpyAsync"), None, None, None,
        "async memcpy needs a queue on every backend"));
    insert(&mut m, "cudaMemset", api("cuda_runtime.h",
        Some("hipMemset"), None, None, None, ""));
    insert(&mut m, "cudaMemcpyDeviceToHost", header_only("cuda_runtime.h"));
    insert(&mut m, "cudaMemcpyHostToDevice", header_only("cuda_runtime.h"));
    insert(&mut m, "cudaMemcpyDeviceToDevice", header_only("cuda_runtime.h"));
    insert(&mut m, "cudaHostAlloc", api("cuda_runtime.h",
        Some("hipHostMalloc"), None, None, None,
        "opencl: use clCreateBuffer with CL_MEM_USE_HOST_PTR"));
    insert(&mut m, "cudaMallocHost", api("cuda_runtime.h",
        Some("hipHostMalloc"), None, None, None,
        "opencl: clCreateBuffer with CL_MEM_ALLOC_HOST_PTR"));
    insert(&mut m, "cudaMallocManaged", api("cuda_runtime.h",
        Some("hipMallocManaged"), None, None, None,
        "sycl: sycl::malloc_shared; opencl: clSVMAlloc (different signatures)"));
    insert(&mut m, "cudaFreeHost", api("cuda_runtime.h",
        Some("hipFreeHost"), None, None, None,
        "sycl: sycl::free; opencl: clReleaseMemObject (different semantics)"));
    insert(&mut m, "cudaMemcpyToSymbol", api("cuda_runtime.h",
        Some("hipMemcpyToSymbol"), None, None, None,
        "sycl/opencl: no symbol-table equivalent; use constant memory"));
    insert(&mut m, "cudaMemcpyFromSymbol", api("cuda_runtime.h",
        Some("hipMemcpyFromSymbol"), None, None, None,
        "sycl/opencl: no symbol-table equivalent; use constant memory"));

    // ---- Streams & events ---------------------------------------------------
    insert(&mut m, "cudaStreamCreate", api("cuda_runtime.h",
        Some("hipStreamCreate"), None, None, None,
        "sycl/opencl: this maps to creating a queue / command queue"));
    insert(&mut m, "cudaStreamDestroy", api("cuda_runtime.h",
        Some("hipStreamDestroy"), None, None, None, ""));
    insert(&mut m, "cudaStreamSynchronize", api("cuda_runtime.h",
        Some("hipStreamSynchronize"), None, None, None,
        "sycl: queue.wait(); opencl: clFinish / clWaitForEvents"));
    insert(&mut m, "cudaEventCreate", api("cuda_runtime.h",
        Some("hipEventCreate"), None, None, None, ""));
    insert(&mut m, "cudaEventRecord", api("cuda_runtime.h",
        Some("hipEventRecord"), None, None, None, ""));
    insert(&mut m, "cudaEventSynchronize", api("cuda_runtime.h",
        Some("hipEventSynchronize"), None, None, None, ""));
    insert(&mut m, "cudaEventElapsedTime", api("cuda_runtime.h",
        Some("hipEventElapsedTime"), None, None, None, ""));
    insert(&mut m, "cudaEventDestroy", api("cuda_runtime.h",
        Some("hipEventDestroy"), None, None, None, ""));

    // ---- Device & error handling -------------------------------------------
    insert(&mut m, "cudaGetDeviceCount", api("cuda_runtime.h",
        Some("hipGetDeviceCount"), None, None, None, ""));
    insert(&mut m, "cudaGetDevice", api("cuda_runtime.h",
        Some("hipGetDevice"), None, None, None, ""));
    insert(&mut m, "cudaSetDevice", api("cuda_runtime.h",
        Some("hipSetDevice"), None, None, None, ""));
    insert(&mut m, "cudaGetLastError", api("cuda_runtime.h",
        Some("hipGetLastError"), None, None, None, ""));
    insert(&mut m, "cudaPeekAtLastError", api("cuda_runtime.h",
        Some("hipPeekAtLastError"), None, None, None, ""));
    insert(&mut m, "cudaGetErrorName", api("cuda_runtime.h",
        Some("hipGetErrorName"), None, None, None, ""));
    insert(&mut m, "cudaGetErrorString", api("cuda_runtime.h",
        Some("hipGetErrorString"), None, None, None, ""));
    insert(&mut m, "cudaDeviceSynchronize", api("cuda_runtime.h",
        Some("hipDeviceSynchronize"), None, None, None,
        "sycl: queue.wait_and_throw(); opencl: clFinish (require queue object)"));
    insert(&mut m, "cudaDeviceReset", api("cuda_runtime.h",
        Some("hipDeviceReset"), None, None, None,
        "sycl/opencl: no direct equivalent; device reset is implicit"));
    insert(&mut m, "cudaDeviceGetAttribute", api("cuda_runtime.h",
        Some("hipDeviceGetAttribute"), None, None, None,
        "sycl: device.get_info; opencl: clGetDeviceInfo (different signatures)"));

    // ---- Types (lookup only) -----------------------------------------------
    insert(&mut m, "cudaError_t", type_alias("cuda_runtime.h",
        Some("hipError_t"), Some("sycl::errc"), Some("cust::CUresult"), Some("cl_int")));
    insert(&mut m, "cudaStream_t", type_alias("cuda_runtime.h",
        Some("hipStream_t"), Some("sycl::queue"), Some("cust::Stream"), Some("cl_command_queue")));
    insert(&mut m, "cudaEvent_t", type_alias("cuda_runtime.h",
        Some("hipEvent_t"), Some("sycl::event"), None, Some("cl_event")));
    insert(&mut m, "dim3", type_alias("",
        Some("dim3"), Some("sycl::range<3>"), Some("(u32, u32, u32)"), Some("[usize; 3]")));

    // ---- Math intrinsics (mostly 1:1) --------------------------------------
    for name in [
        "__sinf", "__cosf", "__tanf", "__expf", "__logf", "__log2f", "__log10f",
        "__powf", "__sqrtf", "__fdividef", "__fmaf_rn", "__fmul_rn",
        "__fmadd_rd", "__float2int_rn", "__float2uint_rn",
    ] {
        insert(&mut m, name, math_intrinsic(name));
    }
    m
});

fn insert(m: &mut ApiMap, name: &'static str, info: ApiInfo) {
    m.insert(name, info);
}

const fn api(
    header: &'static str,
    hip: Option<&'static str>,
    sycl: Option<&'static str>,
    rust: Option<&'static str>,
    opencl: Option<&'static str>,
    note: &'static str,
) -> ApiInfo {
    ApiInfo {
        header,
        hip,
        sycl,
        rust,
        opencl,
        note,
    }
}

fn header_only(header: &'static str) -> ApiInfo {
    ApiInfo {
        header,
        hip: Some("hipMemcpyKind"),
        sycl: None,
        rust: None,
        opencl: None,
        note: "enum constant; backend value is structural",
    }
}

fn type_alias(
    header: &'static str,
    hip: Option<&'static str>,
    sycl: Option<&'static str>,
    rust: Option<&'static str>,
    opencl: Option<&'static str>,
) -> ApiInfo {
    ApiInfo {
        header,
        hip,
        sycl,
        rust,
        opencl,
        note: "type alias; no automatic code rewrite",
    }
}

fn math_intrinsic(name: &'static str) -> ApiInfo {
    ApiInfo {
        header: "device_functions.h",
        hip: Some(name),
        sycl: Some("sycl::native::"),
        rust: Some(name),
        opencl: Some(name.trim_start_matches('_')),
        note: "math intrinsic; mostly identical across HIP/OpenCL/Rust",
    }
}

/// Look up an API entry, returning None if the name is not in the database.
pub fn lookup(name: &str) -> Option<&'static ApiInfo> {
    DATABASE.get(name)
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn cuda_malloc_maps_to_hip() {
        let e = lookup("cudaMalloc").unwrap();
        assert_eq!(e.hip, Some("hipMalloc"));
    }

    #[test]
    fn cuda_event_t_maps_everywhere() {
        let e = lookup("cudaEvent_t").unwrap();
        assert!(e.hip.is_some());
        assert!(e.sycl.is_some());
        assert!(e.rust.is_none());
        assert!(e.opencl.is_some());
    }

    #[test]
    fn unknown_apis_return_none() {
        assert!(lookup("cudaDefinitelyNotARealApi").is_none());
    }

    #[test]
    fn cuda_device_synchronize_maps_to_hip() {
        let e = lookup("cudaDeviceSynchronize").unwrap();
        assert_eq!(e.hip, Some("hipDeviceSynchronize"));
        assert!(e.sycl.is_none(), "SYCL requires queue object — semantic");
        assert!(e.opencl.is_none(), "OpenCL requires command queue — semantic");
    }

    #[test]
    fn cuda_malloc_managed_maps_to_hip() {
        let e = lookup("cudaMallocManaged").unwrap();
        assert_eq!(e.hip, Some("hipMallocManaged"));
    }

    #[test]
    fn cuda_free_host_maps_to_hip() {
        let e = lookup("cudaFreeHost").unwrap();
        assert_eq!(e.hip, Some("hipFreeHost"));
    }

    #[test]
    fn cuda_memcpy_to_from_symbol_maps_to_hip() {
        let to = lookup("cudaMemcpyToSymbol").unwrap();
        assert_eq!(to.hip, Some("hipMemcpyToSymbol"));
        let from = lookup("cudaMemcpyFromSymbol").unwrap();
        assert_eq!(from.hip, Some("hipMemcpyFromSymbol"));
    }

    #[test]
    fn cuda_device_reset_and_attribute_map_to_hip() {
        let reset = lookup("cudaDeviceReset").unwrap();
        assert_eq!(reset.hip, Some("hipDeviceReset"));
        let attr = lookup("cudaDeviceGetAttribute").unwrap();
        assert_eq!(attr.hip, Some("hipDeviceGetAttribute"));
    }
}