atomr-accel-cuda 0.10.0

GPU acceleration via the actor model. Wraps NVIDIA CUDA libraries (cuBLAS, cuDNN, cuFFT, cuRAND, cuSOLVER, cuSPARSE, cuTENSOR, cuBLASLt, NVRTC, NCCL) as supervised atomr actors with generation-validated buffers and a uniform async surface.
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

//! Kernel-actor wrappers around CUDA library handles (§3.2).
//!
//! Each library actor follows a uniform shape:
//!
//! * a `Real { handle, stream, completion, state, … }` variant holding
//!   the cudarc handle plus the per-actor caches it needs;
//! * a `Mock` variant for GPU-free tests;
//! * a `props(stream, allocator, completion, state)` constructor that
//!   panics with `"ContextPoisoned: <Lib>::new failed: …"` if the
//!   handle can't be created, so the supervisor restarts;
//! * a `mock_props()` constructor that replies `Unrecoverable("…not
//!   supported in mock mode")` to every variant.
//!
//! The shared kernel-enqueue body lives in
//! [`envelope::run_kernel`] — every library actor calls it instead of
//! reimplementing the validate / enqueue / spawn-completion-await /
//! reply / drop-keep-alive sequence.
//!
//! F2 ships: `BlasActor`, `CudnnActor`, `FftActor`, `RngActor`.
//! F3 adds: `SolverActor`, `BlasLtActor`, `NvrtcActor`.
//! F4 adds: `CollectiveActor` (NCCL).

pub mod dispatch;
pub mod envelope;
pub mod record;

#[cfg(feature = "cublaslt")]
pub use dispatch::{BlasLtDispatch, BlasLtDispatchCtx};
#[cfg(feature = "nccl")]
pub use dispatch::{CollectiveDispatch, CollectiveDispatchCtx};
#[cfg(feature = "cudnn")]
pub use dispatch::{CudnnDispatch, CudnnDispatchCtx};
pub use dispatch::{
    DevSliceArg, GemmDispatchCtx, NvrtcDispatchCtx, NvrtcLaunchDispatch, RngDispatch, ScalarArg,
};
#[cfg(feature = "cufft")]
pub use dispatch::{FftDispatch, FftDispatchCtx};
#[cfg(feature = "cusparse")]
pub use dispatch::{SendSparseHandle, SparseDispatch, SparseDispatchCtx, SparseOp};
#[cfg(feature = "cutensor")]
pub use dispatch::{TensorDispatch, TensorDispatchCtx, WorkspacePool};

pub mod blas;

pub use blas::{
    AsumRequest, AxpyRequest, BlasActor, BlasMsg, CopyRequest, DotRequest, GeamRequest,
    GemmRequest, GemmStridedBatchedRequest, GemvRequest, GerRequest, IamaxRequest, IaminRequest,
    Nrm2Request, RotRequest, ScalRequest, SwapRequest, SyrkRequest, TrsmRequest,
};

#[cfg(feature = "cudnn")]
pub mod cudnn;
#[cfg(feature = "cudnn")]
pub use cudnn::{
    ActivationFwdRequest, ActivationKind, ActivationRequest, AttentionMask, AttentionParams,
    BatchNormRequest, ConvBwdDataRequest, ConvBwdFilterRequest, ConvDescParams, ConvForwardRequest,
    ConvFwdRequest, ConvParams, CudnnActor, CudnnMsg, DropoutFwdRequest, EpilogueKind,
    GroupNormRequest, InstanceNormRequest, LayerNormRequest, LrnFwdRequest, LrnParams,
    MultiHeadAttnBwdRequest, MultiHeadAttnFwdRequest, NormBwdRequest, NormMode, NormPhase,
    PoolBwdRequest, PoolFwdRequest, PoolMode, PoolParams, RnnBwdRequest, RnnDirection,
    RnnFwdRequest, RnnMode, RnnParams, SoftmaxFwdRequest, SoftmaxMode, SoftmaxRequest,
    TensorLayout,
};

#[cfg(feature = "cufft")]
pub mod fft;
#[cfg(feature = "cufft")]
pub use fft::{
    FftActor, FftCallbackKind, FftDirection, FftKind, FftMsg, FftPlan, FftPlanMany, FftRequest,
    PlanKey,
};

#[cfg(feature = "curand")]
pub mod rng;
#[cfg(feature = "curand")]
pub use rng::{Distribution, FillRequest, RngActor, RngGeneratorKind, RngMsg};

#[cfg(feature = "cusolver")]
pub mod solver;
#[cfg(feature = "cusolver")]
pub use solver::{
    CholeskyRequest, GesvdjBatchedRequest, GetrfBatchedRequest, HegvdRequest, LuRequest,
    LuSolveRequest, PotrfBatchedRequest, QrRequest, SolverActor, SolverDispatch, SolverMsg,
    SvdRequest, SyevdRequest, SygvdRequest, Uplo,
};
#[cfg(all(feature = "cusolver", feature = "cusolver-sp"))]
pub use solver::{SparseCholeskyRequest, SparseLuRequest, SparseQrRequest};

#[cfg(feature = "cublaslt")]
pub mod blas_lt;
#[cfg(feature = "cublaslt")]
pub use blas_lt::{
    Activation, BlasLtActor, BlasLtMsg, Epilogue, HeuristicCacheRef, MatmulRequest, ScaleSet,
    WorkspacePool as BlasLtWorkspacePool,
};

#[cfg(feature = "nvrtc")]
pub mod nvrtc;
#[cfg(feature = "nvrtc")]
pub use nvrtc::{KernelArg, KernelHandle, NvrtcActor, NvrtcMsg, NvrtcOpts};

#[cfg(feature = "nccl")]
pub mod collective;
#[cfg(feature = "nccl")]
pub use collective::{
    AllGatherRequest, AllReduceRequest, AllToAllRequest, AllToAllvRequest, BroadcastRequest,
    CollectiveActor, CollectiveMsg, GroupGuard, NcclCapabilities, NcclReduceSupported, PreMulSumOp,
    RecvRequest, ReduceOp, ReduceRequest, ReduceScatterRequest, SendRequest,
};

#[cfg(feature = "cusparse")]
mod sparse;
#[cfg(feature = "cusparse")]
pub use sparse::{CsrMatrix, SparseActor, SparseMsg};

#[cfg(feature = "cutensor")]
pub mod tensor;
#[cfg(feature = "cutensor")]
pub use tensor::{
    ComputeDesc, ContractRequest, ElementwiseBinaryRequest, ElementwiseTrinaryRequest, OperandSpec,
    PermutationRequest, ReductionRequest, TensorActor, TensorMsg, TensorSpec,
};