Use cusparseSpSM_updateMatrix in CUSOLVERRF.jl (#20)

exanauts · Apr 15, 2024 · 428b3d0 · 428b3d0
1 parent e30d7ea
commit 428b3d0
Show file tree

Hide file tree

Showing 4 changed files with 8 additions and 29 deletions.
diff --git a/.github/workflows/action.yml b/.github/workflows/action.yml
@@ -19,7 +19,7 @@ jobs:
     strategy:
       matrix:
         julia-version: ['1.9']
-        cuda-version: ['11.8', '12.3']
+        cuda-version: ['11.8', '12.4']
         julia-arch: [x64]
         os: [ubuntu-22.04]
 

diff --git a/Project.toml b/Project.toml
@@ -4,7 +4,7 @@ uuid = "a8cc9031-bad2-4722-94f5-40deabb4245c"
 version = "0.2.4"
 
 [compat]
-CUDA = "4, 5.1.1"
+CUDA = "4, 5.3.0"
 KLU = "0.3, 0.4"
 julia = "1.6"
 

diff --git a/src/backsolve.jl b/src/backsolve.jl
@@ -118,8 +118,8 @@ function CuSparseSM(
 ) where T
     chktrans(transa)
 
-    # CUDA 12.3 is required for the routine cusparseSpSV_updateMatrix
-    @assert CUDA.runtime_version() ≥ v"12.3"
+    # CUDA 12.4 is required for the routine cusparseSpSM_updateMatrix
+    @assert CUDA.runtime_version() ≥ v"12.4"
 
     m, n = size(A)
     @assert m == n
@@ -173,8 +173,10 @@ function backsolve!(s::CuSparseSM, A::CUSPARSE.CuSparseMatrixCSR{T}, X::CuMatrix
     alpha = one(T)
 
     descX = CUSPARSE.CuDenseMatrixDescriptor(X)
-    transx = 'N'
+    CUSPARSE.cusparseSpSM_updateMatrix(CUSPARSE.handle(), s.infoL, A.nzVal, CUSPARSE.CUSPARSE_SPSM_UPDATE_GENERAL)
+    CUSPARSE.cusparseSpSM_updateMatrix(CUSPARSE.handle(), s.infoU, A.nzVal, CUSPARSE.CUSPARSE_SPSM_UPDATE_GENERAL)
 
+    transx = 'N'
     if s.transa == 'N'
         CUSPARSE.cusparseSpSM_solve(
             CUSPARSE.handle(), s.transa, transx, Ref{T}(alpha), s.descL, descX, descX, T, s.algo, s.infoL,

diff --git a/src/interface.jl b/src/interface.jl
@@ -62,7 +62,7 @@ function RFLU(
     r = CuVector{Tv}(undef, n)       ; fill!(r, zero(Tv))
     T = CuMatrix{Tv}(undef, n, nrhs) ; fill!(T, zero(Tv))
 
-    if CUDA.runtime_version() ≥ v"12.3"
+    if CUDA.runtime_version() ≥ v"12.4"
         tsv = CuSparseSV(M, 'T')
         dsm = CuSparseSM(M, 'N', T)
         tsm = CuSparseSM(M, 'T', T)
@@ -85,29 +85,6 @@ end
 # Refactoring
 function LinearAlgebra.lu!(rf::RFLU, J::CuSparseMatrixCSR)
     rf_refactor!(rf.rf, J)
-    # Perform the analysis again of SpSM
-    if CUDA.runtime_version() ≥ v"12.3"
-        T = eltype(J)
-        alpha = one(T)
-        # Update rf.dsm
-        dsm = rf.dsm
-        descX = CUSPARSE.CuDenseMatrixDescriptor(T, dsm.n, dsm.nrhs)
-        CUSPARSE.cusparseSpSM_analysis(
-            CUSPARSE.handle(), dsm.transa, 'N', Ref{T}(alpha), dsm.descL, descX, descX, T, dsm.algo, dsm.infoL, dsm.bufferL,
-        )
-        CUSPARSE.cusparseSpSM_analysis(
-            CUSPARSE.handle(), dsm.transa, 'N', Ref{T}(alpha), dsm.descU, descX, descX, T, dsm.algo, dsm.infoU, dsm.bufferU,
-        )
-        # Update rf.tsm
-        tsm = rf.tsm
-        descX = CUSPARSE.CuDenseMatrixDescriptor(T, tsm.n, tsm.nrhs)
-        CUSPARSE.cusparseSpSM_analysis(
-            CUSPARSE.handle(), tsm.transa, 'N', Ref{T}(alpha), tsm.descL, descX, descX, T, tsm.algo, tsm.infoL, tsm.bufferL,
-        )
-        CUSPARSE.cusparseSpSM_analysis(
-            CUSPARSE.handle(), tsm.transa, 'N', Ref{T}(alpha), tsm.descU, descX, descX, T, tsm.algo, tsm.infoU, tsm.bufferU,
-        )
-    end
 end
 
 # Direct solve