More fixes.

JuliaGPU · Feb 6, 2025 · 252cfe6 · 252cfe6
1 parent 6bb7f86
commit 252cfe6
Show file tree

Hide file tree

Showing 4 changed files with 22 additions and 15 deletions.
diff --git a/lib/cublas/CUBLAS.jl b/lib/cublas/CUBLAS.jl
@@ -104,11 +104,9 @@ function handle()
         end
 
         cublasSetStream_v2(new_handle, cuda.stream)
+        cublasSetPointerMode_v2(new_handle, CUBLAS_POINTER_MODE_DEVICE)
         math_mode!(new_handle, cuda.math_mode)
 
-        # default to device pointers everywhere
-        cublasSetPointerMode_v2(state.handle, CUBLAS_POINTER_MODE_DEVICE)
-
         (; handle=new_handle, cuda.stream, cuda.math_mode)
     end
     state = get!(states, cuda.context) do

diff --git a/lib/cublas/wrappers.jl b/lib/cublas/wrappers.jl
@@ -1445,13 +1445,9 @@ for (fname, fname_64, elty) in ((:cublasSgemmGroupedBatched, :cublasSgemmGrouped
             Cptrs = unsafe_batch(reduce(vcat, C))
 
             try
-                ## XXX: does not seem to support device pointers
+                ## XXX: cublasXgemmGroupedBatched does not seem to support device pointers
                 cublasSetPointerMode_v2(handle(), CUBLAS_POINTER_MODE_HOST)
 
-                mode = Ref{cublasPointerMode_t}()
-                cublasGetPointerMode_v2(handle(), mode)
-                @show mode[]
-
                 if CUBLAS.version() >= v"12.0"
                     $fname_64(handle(), transa, transb, m, n, k, alpha, Aptrs, lda,
                               Bptrs, ldb, beta, Cptrs, ldc, group_count, group_size)
@@ -1507,12 +1503,19 @@ for (fname, fname_64, elty) in ((:cublasSgemmGroupedBatched, :cublasSgemmGrouped
             Bptrs = unsafe_batch(B)
             Cptrs = unsafe_batch(C)
 
-            if CUBLAS.version() >= v"12.0"
-                $fname_64(handle(), transa, transb, m, n, k, alpha, Aptrs, lda,
-                          Bptrs, ldb, beta, Cptrs, ldc, group_count, group_size)
-            else
-                $fname(handle(), transa, transb, m, n, k, alpha, Aptrs, lda,
-                          Bptrs, ldb, beta, Cptrs, ldc, group_count, group_size)
+            try
+                ## XXX: cublasXgemmGroupedBatched does not seem to support device pointers
+                cublasSetPointerMode_v2(handle(), CUBLAS_POINTER_MODE_HOST)
+
+                if CUBLAS.version() >= v"12.0"
+                    $fname_64(handle(), transa, transb, m, n, k, alpha, Aptrs, lda,
+                              Bptrs, ldb, beta, Cptrs, ldc, group_count, group_size)
+                else
+                    $fname(handle(), transa, transb, m, n, k, alpha, Aptrs, lda,
+                              Bptrs, ldb, beta, Cptrs, ldc, group_count, group_size)
+                end
+            finally
+                cublasSetPointerMode_v2(handle(), CUBLAS_POINTER_MODE_DEVICE)
             end
             unsafe_free!(Cptrs)
             unsafe_free!(Bptrs)

diff --git a/test/libraries/cublas/level1.jl b/test/libraries/cublas/level1.jl
@@ -26,8 +26,9 @@ k = 13
         @test testf(*, rand(T, m)', rand(T, m))
         @test testf(norm, rand(T, m))
         @test testf(BLAS.asum, rand(T, m))
+
         @test testf(axpy!, rand(), rand(T, m), rand(T, m))
-        #@test testf(LinearAlgebra.axpby!, rand(), rand(T, m), rand(), rand(T, m))
+        @test testf(LinearAlgebra.axpby!, rand(), rand(T, m), rand(), rand(T, m))
         if T <: Complex
             @test testf(dot, rand(T, m), rand(T, m))
             x = rand(T, m)

diff --git a/test/libraries/cublas/level3/gemm.jl b/test/libraries/cublas/level3/gemm.jl
@@ -225,6 +225,7 @@ k = 13
             end
             @test_throws DimensionMismatch CUBLAS.gemm_batched('N','N',alpha,bd_A,bd_bad)
         end
+
         @testset "gemmBatchedEx!" begin
             # C = (alpha*A)*B + beta*C
             CUBLAS.gemmBatchedEx!('N','N',alpha,bd_A,bd_B,beta,bd_C)
@@ -236,6 +237,7 @@ k = 13
             end
             @test_throws DimensionMismatch CUBLAS.gemmBatchedEx!('N','N',alpha,bd_A,bd_bad,beta,bd_C)
         end
+
         nbatch = 10
         bA = rand(elty, m, k, nbatch)
         bB = rand(elty, k, n, nbatch)
@@ -256,6 +258,7 @@ k = 13
             @test bC ≈ h_C
             @test_throws DimensionMismatch CUBLAS.gemm_strided_batched!('N', 'N', alpha, bd_A, bd_B, beta, bd_bad)
         end
+
         @testset "gemmStridedBatchedEx!" begin
             CUBLAS.gemmStridedBatchedEx!('N', 'N', alpha, bd_A, bd_B, beta, bd_C)
             for i in 1:nbatch
@@ -265,6 +268,7 @@ k = 13
             @test bC ≈ h_C
             @test_throws DimensionMismatch CUBLAS.gemmStridedBatchedEx!('N', 'N', alpha, bd_A, bd_B, beta, bd_bad)
         end
+
         @testset "gemm_strided_batched" begin
             bd_C = CUBLAS.gemm_strided_batched('N', 'N', bd_A, bd_B)
 
@@ -402,6 +406,7 @@ k = 13
                 @test C ≈ Array(dC) rtol=rtol
             end
         end
+
         # also test an unsupported combination (falling back to GPUArrays)
         if VERSION < v"1.11-"   # JuliaGPU/CUDA.jl#2441
             AT=BFloat16