JuliaGPU/CUDA.jl

Weird bug with matrix multiplication on Array view

Open

#2,821 opened on Jul 24, 2025

View on GitHub
 (0 comments) (0 reactions) (0 assignees)Julia (1,408 stars) (274 forks)batch import
cuda librarieshelp wanted

Description

Describe the bug

I observe that the matrix multiplication with Array view returns wrong result, for majority of type combinations, with no clear pattern, I can think of. Following are the examples:

To reproduce

The Minimal Working Example (MWE) for this bug:


julia> z = CUDA.rand(2,2,1)
2×2×1 CuArray{Float32, 3, CUDA.DeviceMemory}:
[:, :, 1] =
 0.615746  0.619146
 0.585899  0.0667595

julia> zf = Float64.(z)
2×2×1 CuArray{Float64, 3, CUDA.DeviceMemory}:
[:, :, 1] =
 0.615746  0.619146
 0.585899  0.0667595

julia> rot = CuArray(([1.0f0 0.0f0; 0.0f0 1.0f0]))
2×2 CuArray{Float32, 2, CUDA.DeviceMemory}:
 1.0  0.0
 0.0  1.0

julia> rotf = Float64.(rot)
2×2 CuArray{Float64, 2, CUDA.DeviceMemory}:
 1.0  0.0
 0.0  1.0

julia> view(z,1,:,:)
2×1 view(::CuArray{Float32, 3, CUDA.DeviceMemory}, 1, :, :) with eltype Float32:
 0.61574584
 0.61914593

julia> view(zf,1,:,:)
2×1 view(::CuArray{Float64, 3, CUDA.DeviceMemory}, 1, :, :) with eltype Float64:
 0.6157458424568176
 0.619145929813385

# wrong result
julia> rot * view(z,1,:,:)
2×1 CuArray{Float32, 2, CUDA.DeviceMemory}:
 0.61574584
 0.5858987

# right result
julia> rotf * view(z,1,:,:)
2×1 CuArray{Float64, 2, CUDA.DeviceMemory}:
 0.6157458424568176
 0.619145929813385

# wrong result
julia> rot * view(zf,1,:,:)
2×1 CuArray{Float64, 2, CUDA.DeviceMemory}:
 0.6157458424568176
 0.5858986973762512

# wrong result
julia> rotf * view(zf,1,:,:)
2×1 CuArray{Float64, 2, CUDA.DeviceMemory}:
 0.6157458424568176
 0.5858986973762512

I am on [052768ef] CUDA v5.8.2

Version info

Details on Julia:

julia> versioninfo()

Julia Version 1.11.6
Commit 9615af0f269 (2025-07-09 12:58 UTC)
Build Info:
  Official https://julialang.org/ release
Platform Info:
  OS: Linux (x86_64-linux-gnu)
  CPU: 96 × Intel(R) Xeon(R) Platinum 8268 CPU @ 2.90GHz
  WORD_SIZE: 64
  LLVM: libLLVM-16.0.6 (ORCJIT, cascadelake)
Threads: 1 default, 0 interactive, 1 GC (on 96 virtual cores)
Environment:
  JULIA_PKG_USE_CLI_GIT = true
  JULIA_DEPOT_PATH = /scratch/bt62/sn8885/.julia

Details on CUDA:

julia> CUDA.versioninfo()
CUDA runtime 12.9, artifact installation
CUDA driver 12.9
NVIDIA driver 570.124.6

CUDA libraries: 
- CUBLAS: 12.9.1
- CURAND: 10.3.10
- CUFFT: 11.4.1
- CUSOLVER: 11.7.5
- CUSPARSE: 12.5.10
- CUPTI: 2025.2.1 (API 28.0.0)
- NVML: 12.0.0+570.124.6

Julia packages: 
- CUDA: 5.8.2
- CUDA_Driver_jll: 0.13.1+0
- CUDA_Runtime_jll: 0.17.1+0

Toolchain:
- Julia: 1.11.6
- LLVM: 16.0.6

1 device:
  0: Tesla V100-SXM2-32GB (sm_70, 31.352 GiB / 32.000 GiB available)

Contributor guide

Tech stack
None
Domain
backend
Issue type
bug
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
4
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
3-5 days
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
fresh
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
needs investigation
Prerequisites
Julia programmingGPU computing basics
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
20
Research direction
Investigate the matrix multiplication implementation in CUDA.jl, focusing on how it handles SubArray views and mixed element types. Look at the relevant source files under src/linalg/ to understand type promotion and memory layout. Check if the issue is related to how views are passed to underlying CUBLAS routines. Since there are no comments yet, the first step is to reproduce the bug and identify the specific code path that fails.