Cleanup

Author: kshyatt

lib/cusparse/broadcast.jl

@@ -375,11 +375,13 @@ function sparse_to_sparse_broadcast_kernel(f::F, output::CuSparseDeviceVector{Tv
     row_ix > output.nnz && return
     row_and_ptrs = @inbounds offsets[row_ix]
     row          = @inbounds row_and_ptrs[1]
-    args_are_nnz = @inbounds row_and_ptrs[2]
+    arg_ptrs     = @inbounds row_and_ptrs[2]
     vals = ntuple(Val(N)) do i
         arg = @inbounds args[i]
-        arg_is_nnz = @inbounds args_are_nnz[i]
-        _getindex(arg, row, arg_is_nnz)::Tv
+        # ptr is 0 if the sparse vector doesn't have an element at this row
+        # ptr is 0 if the arg is a scalar AND f preserves zeros
+        ptr = @inbounds arg_ptrs[i]
+        _getindex(arg, row, ptr)::Tv
     end
     output_val = f(vals...)
     @inbounds output.iPtr[row_ix]  = row 
@@ -455,14 +457,16 @@ function sparse_to_dense_broadcast_kernel(::Type{<:CuSparseVector}, f::F,
     row_ix > length(output) && return
     row_and_ptrs = @inbounds offsets[row_ix]
     row          = @inbounds row_and_ptrs[1]
-    args_are_nnz = @inbounds row_and_ptrs[2]
+    arg_ptrs     = @inbounds row_and_ptrs[2]
     vals = ntuple(Val(length(args))) do i
         arg = @inbounds args[i]
-        arg_is_nnz = @inbounds args_are_nnz[i]
-        _getindex(arg, row, arg_is_nnz)::Tv
+        # ptr is 0 if the sparse vector doesn't have an element at this row
+        # ptr is row if the arg is dense OR a scalar with non-zero-preserving f
+        # ptr is 0 if the arg is a scalar AND f preserves zeros
+        ptr = @inbounds arg_ptrs[i]
+        _getindex(arg, row, ptr)::Tv
     end
-    out_val = f(vals...)
-    @inbounds output[row] = out_val 
+    @inbounds output[row] = f(vals...)
     return
 end
 ## COV_EXCL_STOP

test/Project.toml

@@ -5,7 +5,6 @@ BFloat16s = "ab4f0b2a-ad5b-11e8-123f-65d77653426b"
 CUDA_Driver_jll = "4ee394cb-3365-5eb0-8335-949819d2adfc"
 CUDA_Runtime_jll = "76a88914-d11a-5bdc-97e0-2f5a05c973a2"
 ChainRulesCore = "d360d2e6-b24c-11e9-a2a3-2a2ae2dbcce4"
-Cthulhu = "f68482b8-f384-11e8-15f7-abe071a5a75f"
 DataStructures = "864edb3b-99cc-5e75-8d2d-829cb0a9cfe8"
 Dates = "ade2ca70-3891-5945-98fb-dc099432e06a"
 Distributed = "8ba89e20-285c-5b6f-9357-94700520ee1b"