mwe for CUDA problems

euro-hpc-pl · Mar 19, 2024 · f7cf9fd · f7cf9fd
1 parent 937c494
commit f7cf9fd
Showing 1 changed file with 22 additions and 0 deletions.
diff --git a/benchmarks/MWE.jl b/benchmarks/MWE.jl
@@ -0,0 +1,22 @@
+using CUDA
+using SparseArrays
+
+a = CUDA.rand(1, 1)
+b = cu(sprand(Float64, 1, 1, 1.))
+
+# works
+b * a
+a * b
+
+# explodes for CUDA driver v12 but works for CUDA driver v11.4, tested on julia 1.9 and 1.10
+c = CUSPARSE.CuSparseMatrixCSC{Float64, Int32}(b.colPtr, b.rowVal, b.nzVal, (1, 1))
+d = CUDA.rand(Float64, 1, 1)
+
+d * c
+c * d
+
+e = CUSPARSE.CuSparseMatrixCSR(c)
+e * d
+
+# also explodes for CUDA driver v11.4, on both versions of julia
+d * e