numba/numba

CUDA doesn't warn the error if their launch configuration exceeds the capability of their device

Open

#1,055 建立於 2015年3月26日

在 GitHub 查看
 (0 留言) (0 反應) (0 負責人)Python (9,177 star) (1,132 fork)batch import
CUDATaskgood first issue

描述

The launch configuration is not checked and the CUDA target proceeds with attempting to call cuLaunchKernel. This results in the user being presented with the generic error:

numba.cuda.cudadrv.driver.CudaAPIError: Call to cuLaunchKernel results in CUDA_ERROR_INVALID_VALUE

It would be a bit more user- and beginner-friendly to check the launch configuration and point out the limits if they are exceeded.

E.g. mailing list thread: https://groups.google.com/a/continuum.io/forum/?utm_medium=email&utm_source=footer#!msg/numba-users/TkW3Bdjq_Yw/Qv5xq3e6r_sJ

貢獻者指南

技術棧
python
領域
backenddeveloper experience
議題類型
feature
難度面向新貢獻者的預計實作難度，1 表示很小改動，5 表示專家級工作。
3
預計時間有經驗貢獻者完成調查、實作、測試並準備 pull request 的粗略時間範圍。
1-3 hours
活動狀態議題目前的可參與程度：新鮮、活躍、陳舊、阻塞或等待維護者輸入。
stale
清晰度議題是否清楚說明預期改動、驗收標準和下一步。
clear
前置要求
Basic knowledge of CUDAFamiliarity with numba's CUDA targetUnderstanding of device properties
新手友善度1-100 的估計分數，表示該議題對首次貢獻者的友善程度。
40
研究方向
The issue requests a warning when CUDA launch configuration exceeds device capabilities. To implement, query device properties using `cudaGetDeviceProperties()` and compare grid/block sizes and shared memory with device limits. If exceeded, raise a descriptive error before calling `cuLaunchKernel`. Relevant files are in `numba/cuda/` directory, specifically the kernel launch code. See the linked mailing list thread for user pain point.