nvidia__compute-eval / DOMAIN_MAP.md
HYUKI5's picture
nvidia/compute-eval@60301f090273 を複製し、カードを日本語で書く
9f38d8f verified
|
Raw History Blame Contribute Delete
3.51 kB

ComputeEval Domain Map

591 problems across 9 groups.


What We Cover

CUDA C++ (528 problems)

Group Problems Covers
cuda-runtime 29 Kernel launch, memory management, streams, events, CUDA Graphs, cluster launch, occupancy
cuda-kernels 114 Shared memory, warp intrinsics, reductions, scans, stencils, tensor cores, cooperative groups
cccl 80 Thrust, CUB, libcu++
cublas 81 BLAS levels 1–3, extensions, applications
mathlibs 101 cuSPARSE, cuSOLVER, cuFFT, cuRAND
cudnn 113 Convolutions, attention, matmul, normalization via cuDNN Graph API
cutlass 10 CUTLASS templates, collectives, schedulers, epilogues, and integration patterns

CUDA Python (63 problems)

Group Problems Covers
cutile 48 Tile-based kernels: matmul, attention, normalization, element-wise ops (SM 10.0+)
cute-dsl 15 CuTe DSL layouts, tensors, copy/MMA atoms, pipelines, and custom kernels

Most problems are generation — write a function or kernel from a specification. The cutlass and cute-dsl groups add 14 repair problems: fix a defective or under-performing implementation.


What We Don't Cover

Domains

NVIDIA technologies not yet in ComputeEval.

Libraries

Domain One-liner
cuBLASLt Lightweight matmul API with epilogue fusion and algorithm selection
cuTENSOR Tensor contractions, reductions, element-wise ops
cuBLASDx / cuFFTDx Device-side BLAS and FFT (inside your own kernels)
cuSPARSELt 2:4 structured sparsity for Ampere+
cuDSS Direct sparse solver
cuCollections GPU hash maps and concurrent data structures
TensorRT Network definition, plugins, ONNX import, quantization
Transformer Engine FP8/FP4 GEMMs, delayed scaling, fused attention
Video Codec SDK Hardware-accelerated video encode and decode
nvJPEG / NPP Image and signal processing
nvCOMP GPU compression (LZ4, Snappy, zstd)
OptiX Ray tracing with RT cores
cuQuantum / cuPQC Quantum circuit simulation, post-quantum cryptography
RAPIDS cuDF, cuML, cuGraph, cuVS — GPU-accelerated data science

CUDA programming model

Domain One-liner
Tensor Memory Accelerator Hopper+ async bulk copy engine
Virtual memory management cuMemCreate / cuMemMap — fine-grained memory control
CUDA Math API Device-side intrinsics and special functions

Multi-GPU & communication

Domain One-liner
NCCL Collective and point-to-point communication
NVSHMEM GPU-initiated communication (partitioned global address space)
GPUDirect Direct GPU-to-NIC and GPU-to-storage transfers
CUDA-Aware MPI GPU buffers in MPI calls

Python GPU programming

Domain One-liner
Triton GPU kernel language with automatic optimization
CuPy GPU arrays, custom kernels, array protocols
PyTorch CUDA extensions Custom ops, cpp_extension, custom autograd
Numba CUDA JIT-compiled CUDA kernels in Python
Warp Differentiable GPU programming
JAX custom kernels Pallas, custom_call, XLA integration

Problem shapes

Shape Description
Optimization Given correct code, make it faster
Translation Given working code, port it to a different API, language, or architecture