Fixes for compilation with CUDA 13.0

cuda/Makefile

@@ -51,14 +51,15 @@ lib: objs ilib cudalib spgpulib
 	$(AR) $(LIBNAME) $(OBJS)
 	/bin/cp -p $(LIBNAME) $(LIBDIR)
 
-$(COBJS): spgpuinc
 
 objs: spgpuinc $(OBJS) iobjs cudaobjs spgpuobjs
 	/bin/cp -p *$(.mod) $(MODDIR)
 	/bin/cp -p *.h $(INCDIR)
 
+#$(COBJS): spgpuinc
+
 spgpuinc:
-	$(MAKE) -C spgpu includes	
+	$(MAKE) -C spgpu incl	
 spgpuobjs:
 	$(MAKE) -C spgpu objs
 spgpulib:

cuda/cuda_util.c

@@ -37,6 +37,9 @@ static int hasUVA=-1;
 static struct cudaDeviceProp *prop=NULL;
 static spgpuHandle_t psb_cuda_handle = NULL;
 static cublasHandle_t psb_cublas_handle = NULL;
+#if PSB_CUDA_VERSION >= 13000
+static int memoryClockRate=-1;
+#endif
 #if defined(TRACK_CUDA_MALLOC)
 static int64_t total_cuda_mem = 0;
 #endif
@@ -310,17 +313,28 @@ int getGPUMultiProcessors()
 
 int getGPUMemoryBusWidth()
 { int count=0;
-#if CUDART_VERSION >= 5000
+#if PSB_CUDA_VERSION >= 5000
   if (prop!=NULL) 
     count = prop->memoryBusWidth;
 #endif
   return(count);
 }
+
+#if PSB_CUDA_VERSION >= 13000
+int getGPUMemoryClockRate(int dev)
+#else
 int getGPUMemoryClockRate()
+#endif
 { int count=0;
-#if CUDART_VERSION >= 5000
+#if PSB_CUDA_VERSION >= 5000
+#if PSB_CUDA_VERSION >= 13000
+  cudaDeviceGetAttribute(&memoryClockRate,
+			 cudaDevAttrMemoryClockRate, dev);
+  count = memoryClockRate;
+#else
   if (prop!=NULL) 
     count = prop->memoryClockRate;
+#endif
 #endif
   return(count);
 }

cuda/cuda_util.h

@@ -59,7 +59,11 @@ int getDeviceHasUVA();
 int setDevice(int dev);
 int getGPUMultiProcessors();
 int getGPUMemoryBusWidth();
+#if PSB_CUDA_VERSION >= 13000
+int getGPUMemoryClockRate(int dev);
+#else
 int getGPUMemoryClockRate();
+#endif
 int getGPUWarpSize();
 int getGPUMaxThreadsPerBlock();
 int getGPUMaxThreadsPerMP();

cuda/psb_cuda_env_mod.F90

@@ -145,11 +145,20 @@ module psb_cuda_env_mod
       use iso_c_binding
       integer(c_int) :: res
     end function psb_C_get_MemoryBusWidth
+#if (PSB_CUDA_VERSION >= 13000)    
+    function psb_C_get_MemoryClockRate(dev) &
+         & result(res) bind(c,name='getGPUMemoryClockRate')
+      use iso_c_binding
+      integer(c_int), value :: dev
+      integer(c_int) :: res
+    end function psb_C_get_MemoryClockRate
+#else
     function psb_C_get_MemoryClockRate() &
          & result(res) bind(c,name='getGPUMemoryClockRate')
       use iso_c_binding
       integer(c_int) :: res
     end function psb_C_get_MemoryClockRate
+#endif
     function psb_C_get_WarpSize() &
          & result(res) bind(c,name='getGPUWarpSize')
       use iso_c_binding
@@ -174,8 +183,9 @@ module psb_cuda_env_mod
     end subroutine psb_C_cpy_NameString
   end interface
 
-  logical, private :: gpu_do_maybe_free_buffer = .false.
-
+  logical, private, save :: gpu_do_maybe_free_buffer = .false.
+  integer(c_int), private, save :: myDevice
+  
 Contains
   
   function psb_cuda_get_maybe_free_buffer() result(res)
@@ -217,6 +227,7 @@ Contains
 
     if (present(dev)) then 
       info = psb_C_gpu_init(dev)
+      myDevice = dev
     else
       if (count >0) then 
         dev_ = mod(iam,count)
@@ -224,6 +235,7 @@ Contains
         dev_ = 0
       end if
       info = psb_C_gpu_init(dev_)
+      myDevice = dev_
     end if
     if (info == 0) info = initFcusparse()
     if (info /= 0) then 
@@ -283,7 +295,11 @@ Contains
 
   function psb_cuda_MemoryClockRate() result(res)     
     integer(psb_ipk_) :: res
+#if (PSB_CUDA_VERSION >= 13000)    
+    res =  psb_C_get_MemoryClockRate(myDevice)
+#else
     res =  psb_C_get_MemoryClockRate()
+#endif
   end function psb_cuda_MemoryClockRate
 
   function psb_cuda_MemoryBusWidth() result(res)     
@@ -296,7 +312,11 @@ Contains
     ! Formula here: 2*ClockRate(KHz)*BusWidth(bit)
     ! normalization: bit/byte, KHz/MHz
     ! output: MBytes/s
+#if PSB_CUDA_VERSION >= 13000
+    res =  2.d0*0.125d0*1.d-3*psb_C_get_MemoryBusWidth()*psb_C_get_MemoryClockRate(myDevice)
+#else
     res =  2.d0*0.125d0*1.d-3*psb_C_get_MemoryBusWidth()*psb_C_get_MemoryClockRate()
+#endif
   end function psb_cuda_MemoryPeakBandwidth
 
   function psb_cuda_DeviceName() result(res)     

cuda/spgpu/Makefile

@@ -11,7 +11,7 @@ LIBNAME=libspgpu.a
 OBJS=coo.o core.o dia.o ell.o hdia.o hell.o 
 CU_INCLUDES=-I$(INCDIR) $(CUDA_INCLUDES)
 
-all: includes objs
+all: incl objs
 
 objs: $(OBJS) iobjs
 
@@ -24,7 +24,7 @@ iobjs:
 ilib:
 	$(MAKE) -C kernels lib LIBNAME=$(LIBNAME)
 
-includes:
+incl:
 	/bin/cp -p *.h $(INCDIR)
 
 clean: iclean

test/cudakern/Makefile

@@ -27,7 +27,7 @@ dir:
 	(if test ! -d $(EXEDIR); then mkdir $(EXEDIR); fi)
 dpdegenmv.o spdegenmv.o: data_input.o
 dpdegenmv: $(DPGOBJS) 
-	$(FLINK) $(LOPT) $(DPGOBJS) -o dpdegenmv $(FINCLUDES) $(PSBLAS_LIB) $(LDLIBS) 
+	$(FLINK) $(LOPT) $(DPGOBJS) -o dpdegenmv $(FINCLUDES) $(PSBLAS_LIB) $(LDLIBS) -lstdc++
 	/bin/mv dpdegenmv $(EXEDIR) 
 spdegenmv: $(SPGOBJS)
 	$(FLINK) $(LOPT) $(SPGOBJS) -o spdegenmv $(PSBLAS_LIB) $(LDLIBS)