diff --git a/ext/AMDGPU/operators.jl b/ext/AMDGPU/operators.jl
index a88bb6a..a2c2632 100644
--- a/ext/AMDGPU/operators.jl
+++ b/ext/AMDGPU/operators.jl
@@ -102,8 +102,8 @@ for (SparseMatrixType, BlasType) in ((:(ROCSparseMatrixCSR{T}), :BlasFloat),
             m,n = size(A)
             alpha = Ref{T}(one(T))
             descA = rocSPARSE.ROCSparseMatrixDescriptor(A, 'O')
-            rocsparse_uplo = Ref{rocSPARSE.rocsparse_diag_type}(uplo)
-            rocsparse_diag = Ref{rocSPARSE.rocsparse_matrix_type}(diag)
+            rocsparse_uplo = Ref{rocSPARSE.rocsparse_fill_mode}(uplo)
+            rocsparse_diag = Ref{rocSPARSE.rocsparse_diag_type}(diag)
             rocSPARSE.rocsparse_spmat_set_attribute(descA, rocSPARSE.rocsparse_spmat_fill_mode, rocsparse_uplo, Csize_t(sizeof(rocsparse_uplo)))
             rocSPARSE.rocsparse_spmat_set_attribute(descA, rocSPARSE.rocsparse_spmat_diag_type, rocsparse_diag, Csize_t(sizeof(rocsparse_diag)))
             if nrhs == 1
@@ -122,10 +122,10 @@ for (SparseMatrixType, BlasType) in ((:(ROCSparseMatrixCSR{T}), :BlasFloat),
                 descY = rocSPARSE.ROCDenseMatrixDescriptor(T, m, nrhs)
                 algo = rocSPARSE.rocsparse_spsm_alg_default
                 buffer_size = Ref{Csize_t}()
-                rocSPARSE.rocsparse_spsm(rocSPARSE.handle(), transa, alpha, descA, descX, descY, T, algo,
+                rocSPARSE.rocsparse_spsm(rocSPARSE.handle(), transa, 'N', alpha, descA, descX, descY, T, algo,
                                          rocSPARSE.rocsparse_spsm_stage_buffer_size, buffer_size, C_NULL)
                 buffer = ROCVector{UInt8}(undef, buffer_size[])
-                rocSPARSE.rocsparse_spsm(rocSPARSE.handle(), transa, alpha, descA, descX, descY, T, algo,
+                rocSPARSE.rocsparse_spsm(rocSPARSE.handle(), transa, 'N', alpha, descA, descX, descY, T, algo,
                                          rocSPARSE.rocsparse_spsm_stage_preprocess, buffer_size, buffer)
                 return AMD_TriangularOperator{T}(T, m, n, nrhs, transa, descA, buffer_size, buffer)
             end
@@ -161,6 +161,6 @@ function LinearAlgebra.ldiv!(Y::ROCMatrix{T}, A::AMD_TriangularOperator{T}, X::R
     descX = rocSPARSE.ROCDenseMatrixDescriptor(X)
     algo = rocSPARSE.rocsparse_spsm_alg_default
     alpha = Ref{T}(one(T))
-    rocSPARSE.rocsparse_spsm(rocSPARSE.handle(), A.transa, alpha, A.descA, descX, descY, T,
+    rocSPARSE.rocsparse_spsm(rocSPARSE.handle(), A.transa, 'N', alpha, A.descA, descX, descY, T,
                              algo, rocSPARSE.rocsparse_spsm_stage_compute, A.buffer_size, A.buffer)
 end
diff --git a/test/gpu/gpu.jl b/test/gpu/gpu.jl
index 20ddcb8..498f913 100644
--- a/test/gpu/gpu.jl
+++ b/test/gpu/gpu.jl
@@ -152,19 +152,17 @@ function test_triangular(FC, V, DM, SM)
       ldiv!(y_gpu, opA_gpu, x_gpu)
       @test collect(y_gpu) ≈ y_cpu
     end
-    if diag == 'N'
-      for j = 1:5
-        y_cpu = rand(FC, n)
-        x_cpu = rand(FC, n)
-        A_cpu2 = A_cpu + j*I
-        ldiv!(y_cpu, triangle(A_cpu2), x_cpu)
-        y_gpu = V(y_cpu)
-        x_gpu = V(x_cpu)
-        A_gpu2 = SM(A_cpu2)
-        update!(opA_gpu, A_gpu2)
-        ldiv!(y_gpu, opA_gpu, x_gpu)
-        @test collect(y_gpu) ≈ y_cpu
-      end
+    for j = 1:5
+      y_cpu = rand(FC, n)
+      x_cpu = rand(FC, n)
+      A_cpu2 = A_cpu + j*tril(A_cpu,-1) + j*triu(A_cpu,1)
+      ldiv!(y_cpu, triangle(A_cpu2), x_cpu)
+      y_gpu = V(y_cpu)
+      x_gpu = V(x_cpu)
+      A_gpu2 = SM(A_cpu2)
+      update!(opA_gpu, A_gpu2)
+      ldiv!(y_gpu, opA_gpu, x_gpu)
+      @test collect(y_gpu) ≈ y_cpu
     end
 
     nrhs = 3
@@ -178,11 +176,11 @@ function test_triangular(FC, V, DM, SM)
       ldiv!(Y_gpu, opA_gpu, X_gpu)
       @test collect(Y_gpu) ≈ Y_cpu
     end
-    if diag == 'N' && V.body.name.name != :CuArray
+    if V.body.name.name != :CuArray
       for j = 1:5
         Y_cpu = rand(FC, n, nrhs)
         X_cpu = rand(FC, n, nrhs)
-        A_cpu2 = A_cpu + j*I
+        A_cpu2 = A_cpu + j*tril(A_cpu,-1) + j*triu(A_cpu,1)
         ldiv!(Y_cpu, triangle(A_cpu2), X_cpu)
         Y_gpu = DM(Y_cpu)
         X_gpu = DM(X_cpu)