solver: use bsrsm2 by default for cuda 12

include/gt-solver/solver.h

@@ -8,10 +8,19 @@
 
 #ifdef GTENSOR_DEVICE_CUDA
 #if CUDART_VERSION >= 12000
-// new generic MPI; available since 11.3.1, but performance
+#ifdef GTENSOR_DEVICE_CUDA_CUSPARSE_GENERIC
+// New generic API; available since 11.3.1, but performance
 // is worse than old csrsm2 API in many cases pre-12.
 #include "gt-solver/backend/cuda-generic.h"
 #else
+// bsrsm2 API, which can work with csr format by setting
+// block size 1. In CUDA 12, appears to use less memory and
+// often be faster than generic API. Exists even in 8.0,
+// but not clear it has advantage over csrsm2 API for older
+// cuda versions where csrsm2 is still available.
+#include "gt-solver/backend/cuda-bsrsm2.h"
+#endif // GTENSOR_DEVICE_CUDA_CUSPARSE_GENERIC
+#else
 // legacy API, deprecated since 11.3.1 but still supported until 12
 #include "gt-solver/backend/cuda-csrsm2.h"
 #endif // CUDA_VERSION