Merge pull request #1412 from JuliaGPU/tb/sparse_coo

Fix sparse COO to CSR conversion.

Author: maleadt

.buildkite/pipeline.yml

@@ -20,31 +20,6 @@ steps:
     if: build.message !~ /\[skip tests\]/
     timeout_in_minutes: 120
 
-  - label: "Julia 1.6 (debug)"
-    plugins:
-      - JuliaCI/julia#v1:
-          version: 1.6
-      - JuliaCI/julia-test#v1:
-          julia_args: "-g2"
-          test_args: "--sanitize --quickfail --jobs=1"
-      - JuliaCI/julia-coverage#v1:
-          codecov: true
-          dirs:
-            - src
-            - lib
-            - examples
-    agents:
-      queue: "juliagpu"
-      cuda: "11.0"
-      cap: "sm_80"
-    env:
-      JULIA_CUDA_VERSION: '11.6'
-      JULIA_CUDA_USE_COMPAT: 'false'  # NVIDIA bug #3418723: injection tools prevent probing libcuda
-      JULIA_CUDA_USE_BINARYBUILDER: 'true'
-    if: build.message !~ /\[skip tests\]/ &&
-        !build.pull_request.draft
-    timeout_in_minutes: 180
-
   - label: "Julia 1.7"
     plugins:
       - JuliaCI/julia#v1:
@@ -301,6 +276,31 @@ steps:
     if: build.message !~ /\[skip tests\]/ && !build.pull_request.draft
     timeout_in_minutes: 60
 
+  - label: "Compute Sanitizer"
+    plugins:
+      - JuliaCI/julia#v1:
+          version: 1.6
+      - JuliaCI/julia-test#v1:
+          julia_args: "-g2"
+          test_args: "--sanitize --quickfail --jobs=1"
+      - JuliaCI/julia-coverage#v1:
+          codecov: true
+          dirs:
+            - src
+            - lib
+            - examples
+    agents:
+      queue: "juliagpu"
+      cuda: "11.0"
+      cap: "sm_80"
+    env:
+      JULIA_CUDA_VERSION: '11.6'
+      JULIA_CUDA_USE_COMPAT: 'false'  # NVIDIA bug #3418723: injection tools prevent probing libcuda
+      JULIA_CUDA_USE_BINARYBUILDER: 'true'
+    if: build.message !~ /\[skip benchmarks\]/ &&
+        build.branch =~ /^master$$/
+    timeout_in_minutes: 240
+
 
   # other tasks
 
@@ -337,8 +337,7 @@ steps:
       queue: "benchmark"
       cuda: "*"
     if: build.message !~ /\[skip benchmarks\]/ &&
-        build.branch =~ /^master$$/ &&
-        !build.pull_request.draft
+        build.branch =~ /^master$$/
     timeout_in_minutes: 30
 
   - label: "Benchmarks on 1.7"
@@ -368,8 +367,7 @@ steps:
       queue: "benchmark"
       cuda: "*"
     if: build.message !~ /\[skip benchmarks\]/ &&
-        build.branch =~ /^master$$/ &&
-        !build.pull_request.draft
+        build.branch =~ /^master$$/
     timeout_in_minutes: 30
 
   - wait

lib/cusparse/conversions.jl

@@ -380,23 +380,23 @@ end
 
 ## CSR to COO and vice-versa
 
+# TODO: we can do similar for CSC conversions, but that requires the columns to be sorted
+
 function CuSparseMatrixCSR(coo::CuSparseMatrixCOO{Tv}, ind::SparseChar='O') where {Tv}
     m,n = size(coo)
-    nnz = coo.nnz
-    csrRowPtr = CUDA.zeros(Cint, nnz)
-    cusparseXcoo2csr(handle(), coo.rowInd, nnz, m, csrRowPtr, ind)
-    CuSparseMatrixCSR{Tv}(csrRowPtr, coo.colInd, coo.nzVal, size(coo))
+    csrRowPtr = CuVector{Cint}(undef, m+1)
+    cusparseXcoo2csr(handle(), coo.rowInd, nnz(coo), m, csrRowPtr, ind)
+    CuSparseMatrixCSR{Tv}(csrRowPtr, coo.colInd, nonzeros(coo), size(coo))
 end
 
 function CuSparseMatrixCOO(csr::CuSparseMatrixCSR{Tv}, ind::SparseChar='O') where {Tv}
     m,n = size(csr)
-    nnz = csr.nnz
-    cooRowInd = CUDA.zeros(Cint, nnz)
-    cusparseXcsr2coo(handle(), csr.rowPtr, nnz, m, cooRowInd, ind)
-    CuSparseMatrixCOO{Tv}(cooRowInd, csr.colVal, csr.nzVal, size(csr), nnz)
+    cooRowInd = CuVector{Cint}(undef, nnz(csr))
+    cusparseXcsr2coo(handle(), csr.rowPtr, nnz(csr), m, cooRowInd, ind)
+    CuSparseMatrixCOO{Tv}(cooRowInd, csr.colVal, nonzeros(csr), size(csr), nnz(csr))
 end
 
-### CSC/BST to COO and viceversa
+### CSC/BSR to COO and viceversa
 
 CuSparseMatrixCSC(coo::CuSparseMatrixCOO) = CuSparseMatrixCSC(CuSparseMatrixCSR(coo)) # no direct conversion
 CuSparseMatrixCOO(csc::CuSparseMatrixCSC) = CuSparseMatrixCOO(CuSparseMatrixCSR(csc)) # no direct conversion

src/array.jl

@@ -128,9 +128,9 @@ is_unified(a::CuArray) = isa(a.storage.buffer, Mem.UnifiedBuffer)
 
 # type and dimensionality specified, accepting dims as series of Ints
 CuArray{T,N,B}(::UndefInitializer, dims::Integer...) where {T,N,B} =
-  CuArray{T,N,B}(undef, dims)
+  CuArray{T,N,B}(undef, convert(Tuple{Vararg{Int}}, dims))
 CuArray{T,N}(::UndefInitializer, dims::Integer...) where {T,N} =
-  CuArray{T,N}(undef, dims)
+  CuArray{T,N}(undef, convert(Tuple{Vararg{Int}}, dims))
 
 # type but not dimensionality specified
 CuArray{T}(::UndefInitializer, dims::Dims{N}) where {T,N} =

test/runtests.jl

@@ -179,7 +179,7 @@ if !has_cutensor() || first(picks).cap < v"6.0"
 end
 if do_sanitize
     # XXX: some library tests fail under compute-sanitizer
-    append!(skip_tests, ["cutensor", "cusparse"])
+    append!(skip_tests, ["cutensor"])
     # XXX: others take absurdly long
     push!(skip_tests, "cusolver")
     # XXX: these hang for some reason