ryankert01 commented on code in PR #1399:
URL: https://github.com/apache/mahout/pull/1399#discussion_r3448788190
##########
qdp/qdp-core/src/gpu/cuda_ffi.rs:
##########
@@ -37,65 +47,247 @@ pub(crate) struct CudaPointerAttributes {
pub allocation_flags: u32,
}
-// CUDA error codes
+// CUDA/HIP error codes (numerically identical for the codes used).
pub(crate) const CUDA_SUCCESS: i32 = 0;
-// Note: CUDA_ERROR_NOT_READY may be used in future optimizations for
non-blocking event checks
-// Reference:
https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__TYPES.html#group__CUDART__TYPES_1g3f51e3575c2178246db0a94a430e0028
#[allow(dead_code)]
pub(crate) const CUDA_ERROR_NOT_READY: i32 = 34;
-unsafe extern "C" {
- pub(crate) fn cudaHostAlloc(pHost: *mut *mut c_void, size: usize, flags:
u32) -> i32;
- pub(crate) fn cudaFreeHost(ptr: *mut c_void) -> i32;
+// ---- CUDA backend: bind libcudart directly ----
+#[cfg(all(feature = "cuda", not(feature = "hip")))]
+pub(crate) use cuda_rt::*;
+
+#[cfg(all(feature = "cuda", not(feature = "hip")))]
+mod cuda_rt {
+ use super::CudaPointerAttributes;
+ use std::ffi::c_void;
+
+ unsafe extern "C" {
+ pub(crate) fn cudaHostAlloc(pHost: *mut *mut c_void, size: usize,
flags: u32) -> i32;
+ pub(crate) fn cudaFreeHost(ptr: *mut c_void) -> i32;
+
+ #[allow(dead_code)]
+ pub(crate) fn cudaPointerGetAttributes(
+ attributes: *mut CudaPointerAttributes,
+ ptr: *const c_void,
+ ) -> i32;
+
+ pub(crate) fn cudaMemGetInfo(free: *mut usize, total: *mut usize) ->
i32;
+
+ pub(crate) fn cudaMemcpyAsync(
+ dst: *mut c_void,
+ src: *const c_void,
+ count: usize,
+ kind: u32,
+ stream: *mut c_void,
+ ) -> i32;
+
+ #[allow(dead_code)]
+ pub(crate) fn cudaMemcpy(
+ dst: *mut c_void,
+ src: *const c_void,
+ count: usize,
+ kind: u32,
+ ) -> i32;
+
+ pub(crate) fn cudaEventCreateWithFlags(event: *mut *mut c_void, flags:
u32) -> i32;
+ pub(crate) fn cudaEventRecord(event: *mut c_void, stream: *mut c_void)
-> i32;
+ pub(crate) fn cudaEventDestroy(event: *mut c_void) -> i32;
+ pub(crate) fn cudaStreamWaitEvent(
+ stream: *mut c_void,
+ event: *mut c_void,
+ flags: u32,
+ ) -> i32;
+ pub(crate) fn cudaStreamSynchronize(stream: *mut c_void) -> i32;
+
+ pub(crate) fn cudaMemsetAsync(
+ devPtr: *mut c_void,
+ value: i32,
+ count: usize,
+ stream: *mut c_void,
+ ) -> i32;
+
+ #[allow(dead_code)]
+ pub(crate) fn cudaEventQuery(event: *mut c_void) -> i32;
+ pub(crate) fn cudaEventSynchronize(event: *mut c_void) -> i32;
+ pub(crate) fn cudaEventElapsedTime(
+ ms: *mut f32,
+ start: *mut c_void,
+ end: *mut c_void,
+ ) -> i32;
+ }
+}
+
+// ---- HIP backend: bind libamdhip64, expose the same cuda* names ----
+#[cfg(feature = "hip")]
+pub(crate) use hip_rt::*;
+
+// The wrapper functions deliberately keep the cuda* spelling so call sites are
+// vendor-agnostic; suppress the snake_case lint for that intentional naming.
+#[cfg(feature = "hip")]
+#[allow(non_snake_case)]
+mod hip_rt {
+ use super::{CUDA_MEMORY_TYPE_DEVICE, CUDA_MEMORY_TYPE_MANAGED,
CudaPointerAttributes};
+ use std::ffi::c_void;
+
+ // hipMemoryType enum values are NOT guaranteed equal to CUDA's across ROCm
+ // releases (older HIP used Host=0/Device=1; the hip_runtime_api.h note
flags
+ // this explicitly). So we read the real hipPointerAttribute_t and compare
its
+ // `type` against the named hipMemoryType* constants rather than a magic
+ // number, then translate to the CUDA convention the caller expects.
+ const HIP_MEMORY_TYPE_DEVICE: i32 = 2; // hipMemoryTypeDevice
+ const HIP_MEMORY_TYPE_MANAGED: i32 = 3; // hipMemoryTypeManaged
+
+ // Mirror of hipPointerAttribute_t (ROCm hip_runtime_api.h): the leading
+ // `type` field is the hipMemoryType enum read by cudaPointerGetAttributes.
+ #[repr(C)]
+ struct HipPointerAttributes {
Review Comment:
+1
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
