Add GPU architecture BLACKWELL

configs/spock_intel_RTX5090.config

@@ -0,0 +1,40 @@
+# NTU-spock
+CUDA_PATH       /work1/koarakawaii/NVIDIA_HPC_SDK/nvidia_hpc_sdk/Linux_x86_64/25.9/compilers
+FFTW2_PATH      /software/fftw/2.1.5-intel-2023.1.0-openmpi-4.1.5-ucx_mt
+FFTW3_PATH      /software/fftw/3.3.10-intel-2023.1.0-openmpi-4.1.5-ucx_mt
+MPI_PATH        /software/openmpi/4.1.5-ucx_mt-intel-2023.1.0
+HDF5_PATH       /software/hdf5/1.10.6-intel-2023.1.0
+GRACKLE_PATH
+GSL_PATH        /software/gsl/2.6-intel-2023.1.0
+LIBYT_PATH
+CUFFTDX_PATH    /software/cuFFTDx/22.11
+
+# compilers
+CXX     icpc
+CXX_MPI mpicxx
+
+# flags
+CXXFLAG -g
+CXXFLAG -O2
+CXXFLAG -fp-model precise
+#CXXFLAG -std=c++11
+#CXXFLAG -gxx-name=YOUR_G++
+CXXFLAG -w1
+CXXFLAG -Wno-unknown-pragmas -diag-disable 3180 -diag-disable 10441
+
+OPENMPFLAG -fopenmp
+
+LIBFLAG -limf -L/work1/koarakawaii/NVIDIA_HPC_SDK/nvidia_hpc_sdk/Linux_x86_64/25.9/cuda/13.0/targets/x86_64-linux/lib
+
+NVCCFLAG_COM -O3
+#NVCCFLAG_COM -use_fast_math
+NVCCFLAG_FLU -Xptxas -dlcm=ca -prec-div=false -ftz=true
+NVCCFLAG_POT -Xptxas -dlcm=ca
+
+# for debugging
+#CXXFLAG -fstack-protector-all
+#CXXFLAG -fstack-protector-strong  # somehow it can capture issues not detected by -fstack-protector-all
+#LIBFLAG -lssp
+
+# gpu
+GPU_COMPUTE_CAPABILITY 1200    # 5090

include/CUFLU.h

@@ -355,6 +355,12 @@
 #     else
 #     define FLU_BLOCK_SIZE_X       512      // not optimized yet
 #     endif
+#  elif ( GPU_ARCH == BLACKWELL )
+#     ifdef FLOAT8
+#     define FLU_BLOCK_SIZE_X       256
+#     else
+#     define FLU_BLOCK_SIZE_X       512      // not optimized yet
+#     endif
 #  else
 #     define FLU_BLOCK_SIZE_X       NULL_INT
 #     ifdef GPU
@@ -420,6 +426,12 @@
 #     else
 #     define FLU_BLOCK_SIZE_X       512      // not optimized yet
 #     endif
+#  elif ( GPU_ARCH == BLACKWELL )
+#     ifdef FLOAT8
+#     define FLU_BLOCK_SIZE_X       256
+#     else
+#     define FLU_BLOCK_SIZE_X       512      // not optimized yet
+#     endif
 #  else
 #     define FLU_BLOCK_SIZE_X       NULL_INT
 #     ifdef GPU
@@ -502,6 +514,13 @@
 #        define FLU_BLOCK_SIZE_Y    32    // not optimized yet
 #     endif
 
+#  elif ( GPU_ARCH == BLACKWELL )
+#     ifdef FLOAT8
+#        define FLU_BLOCK_SIZE_Y    16    // not optimized yet
+#     else
+#        define FLU_BLOCK_SIZE_Y    32    // not optimized yet
+#     endif
+
 #  else
 #        define FLU_BLOCK_SIZE_Y    NULL_INT
 #        ifdef GPU
@@ -577,7 +596,8 @@ using complex_type = typename FFT::value_type;
 
 // use shuffle reduction in the KEPLER and later GPUs
 #  if ( GPU_ARCH == KEPLER  ||  GPU_ARCH == MAXWELL  ||  GPU_ARCH == PASCAL        ||  GPU_ARCH == VOLTA  ||  \
-        GPU_ARCH == TURING  ||  GPU_ARCH == AMPERE   ||  GPU_ARCH == ADA_LOVELACE  ||  GPU_ARCH == HOPPER )
+        GPU_ARCH == TURING  ||  GPU_ARCH == AMPERE   ||  GPU_ARCH == ADA_LOVELACE  ||  GPU_ARCH == HOPPER ||  \
+        GPU_ARCH == BLACKWELL )
 #     define DT_FLU_USE_SHUFFLE
 #  endif
 
@@ -593,7 +613,7 @@ using complex_type = typename FFT::value_type;
 //     for information on warp size
 #ifdef __CUDACC__
 #if ( GPU_ARCH == FERMI   ||  GPU_ARCH == KEPLER  ||  GPU_ARCH == MAXWELL       ||  GPU_ARCH == PASCAL  ||  GPU_ARCH == VOLTA  ||  \
-      GPU_ARCH == TURING  ||  GPU_ARCH == AMPERE  ||  GPU_ARCH == ADA_LOVELACE  ||  GPU_ARCH == HOPPER )
+      GPU_ARCH == TURING  ||  GPU_ARCH == AMPERE  ||  GPU_ARCH == ADA_LOVELACE  ||  GPU_ARCH == HOPPER  ||  GPU_ARCH == BLACKWELL )
 // CUPOT.h will define WARP_SIZE as well
 #  ifndef WARP_SIZE
 #  define WARP_SIZE 32

include/CUPOT.h

@@ -107,6 +107,12 @@
 #     else
 #        define POT_BLOCK_SIZE_Z      4      // not optimized yet
 #     endif
+#  elif ( GPU_ARCH == BLACKWELL )
+#     ifdef FLOAT8
+#        define POT_BLOCK_SIZE_Z      2      // not optimized yet
+#     else
+#        define POT_BLOCK_SIZE_Z      4      // not optimized yet
+#     endif
 #  else
 #        define POT_BLOCK_SIZE_Z      NULL_INT
 #     ifdef GPU
@@ -128,7 +134,8 @@
 // --> although strictly speaking the shuffle functions do NOT work for double precision, but experiments
 //     show that residual_sum += (float)residual, where residual_sum is double, gives acceptable accuracy
 #  if ( GPU_ARCH == KEPLER  ||  GPU_ARCH == MAXWELL  ||  GPU_ARCH == PASCAL        ||  GPU_ARCH == VOLTA  ||  \
-        GPU_ARCH == TURING  ||  GPU_ARCH == AMPERE   ||  GPU_ARCH == ADA_LOVELACE  ||  GPU_ARCH == HOPPER )
+        GPU_ARCH == TURING  ||  GPU_ARCH == AMPERE   ||  GPU_ARCH == ADA_LOVELACE  ||  GPU_ARCH == HOPPER ||  \
+        GPU_ARCH == BLACKWELL )
 #     define SOR_USE_SHUFFLE
 #  endif
 
@@ -185,7 +192,8 @@
 
 // use shuffle reduction in the KEPLER and later GPUs
 #if ( GPU_ARCH == KEPLER  ||  GPU_ARCH == MAXWELL  ||  GPU_ARCH == PASCAL        ||  GPU_ARCH == VOLTA  ||  \
-      GPU_ARCH == TURING  ||  GPU_ARCH == AMPERE   ||  GPU_ARCH == ADA_LOVELACE  ||  GPU_ARCH == HOPPER )
+      GPU_ARCH == TURING  ||  GPU_ARCH == AMPERE   ||  GPU_ARCH == ADA_LOVELACE  ||  GPU_ARCH == HOPPER ||  \
+      GPU_ARCH == BLACKWELL )
 #   define DT_GRA_USE_SHUFFLE
 #endif
 
@@ -195,7 +203,7 @@
 //     for information on warp size
 #ifdef __CUDACC__
 #if ( GPU_ARCH == FERMI   ||  GPU_ARCH == KEPLER  ||  GPU_ARCH == MAXWELL       ||  GPU_ARCH == PASCAL  ||  GPU_ARCH == VOLTA  ||  \
-      GPU_ARCH == TURING  ||  GPU_ARCH == AMPERE  ||  GPU_ARCH == ADA_LOVELACE  ||  GPU_ARCH == HOPPER )
+      GPU_ARCH == TURING  ||  GPU_ARCH == AMPERE  ||  GPU_ARCH == ADA_LOVELACE  ||  GPU_ARCH == HOPPER  ||  GPU_ARCH == BLACKWELL )
 // CUFLU.h will define WARP_SIZE as well
 #  ifndef WARP_SIZE
 #  define WARP_SIZE 32

include/Macro.h

@@ -32,26 +32,29 @@
 #define AMPERE       7
 #define ADA_LOVELACE 8
 #define HOPPER       9
+#define BLACKWELL    10
 
 #ifdef GPU
-#if   ( GPU_COMPUTE_CAPABILITY >= 200  &&  GPU_COMPUTE_CAPABILITY < 300 )
+#if   ( GPU_COMPUTE_CAPABILITY >= 200   &&  GPU_COMPUTE_CAPABILITY < 300 )
 # define GPU_ARCH FERMI
-#elif ( GPU_COMPUTE_CAPABILITY >= 300  &&  GPU_COMPUTE_CAPABILITY < 500 )
+#elif ( GPU_COMPUTE_CAPABILITY >= 300   &&  GPU_COMPUTE_CAPABILITY < 500 )
 # define GPU_ARCH KEPLER
-#elif ( GPU_COMPUTE_CAPABILITY >= 500  &&  GPU_COMPUTE_CAPABILITY < 600 )
+#elif ( GPU_COMPUTE_CAPABILITY >= 500   &&  GPU_COMPUTE_CAPABILITY < 600 )
 # define GPU_ARCH MAXWELL
-#elif ( GPU_COMPUTE_CAPABILITY >= 600  &&  GPU_COMPUTE_CAPABILITY < 700 )
+#elif ( GPU_COMPUTE_CAPABILITY >= 600   &&  GPU_COMPUTE_CAPABILITY < 700 )
 # define GPU_ARCH PASCAL
-#elif ( GPU_COMPUTE_CAPABILITY >= 700  &&  GPU_COMPUTE_CAPABILITY < 750 )
+#elif ( GPU_COMPUTE_CAPABILITY >= 700   &&  GPU_COMPUTE_CAPABILITY < 750 )
 # define GPU_ARCH VOLTA
-#elif ( GPU_COMPUTE_CAPABILITY >= 750  &&  GPU_COMPUTE_CAPABILITY < 800 )
+#elif ( GPU_COMPUTE_CAPABILITY >= 750   &&  GPU_COMPUTE_CAPABILITY < 800 )
 # define GPU_ARCH TURING
-#elif ( GPU_COMPUTE_CAPABILITY >= 800  &&  GPU_COMPUTE_CAPABILITY < 890 )
+#elif ( GPU_COMPUTE_CAPABILITY >= 800   &&  GPU_COMPUTE_CAPABILITY < 890 )
 # define GPU_ARCH AMPERE
-#elif ( GPU_COMPUTE_CAPABILITY >= 890  &&  GPU_COMPUTE_CAPABILITY < 900 )
+#elif ( GPU_COMPUTE_CAPABILITY >= 890   &&  GPU_COMPUTE_CAPABILITY < 900 )
 # define GPU_ARCH ADA_LOVELACE
-#elif ( GPU_COMPUTE_CAPABILITY >= 900  &&  GPU_COMPUTE_CAPABILITY < 1000 )
+#elif ( GPU_COMPUTE_CAPABILITY >= 900   &&  GPU_COMPUTE_CAPABILITY < 1000 )
 # define GPU_ARCH HOPPER
+#elif ( GPU_COMPUTE_CAPABILITY >= 1000  &&  GPU_COMPUTE_CAPABILITY < 1210 )
+# define GPU_ARCH BLACKWELL
 #else
 # error : ERROR : Unknown GPU_COMPUTE_CAPABILITY !!
 #endif // GPU_COMPUTE_CAPABILITY

src/Auxiliary/Aux_TakeNote.cpp

@@ -417,6 +417,8 @@ void Aux_TakeNote()
       fprintf( Note, "GPU_ARCH                        ADA_LOVELACE\n" );
 #     elif ( GPU_ARCH == HOPPER )
       fprintf( Note, "GPU_ARCH                        HOPPER\n" );
+#     elif ( GPU_ARCH == BLACKWELL )
+      fprintf( Note, "GPU_ARCH                        BLACKWELL\n" );
 #     else
       fprintf( Note, "GPU_ARCH                        UNKNOWN\n" );
 #     endif

src/GPU_API/CUAPI_DiagnoseDevice.cu

@@ -53,6 +53,7 @@ void CUAPI_DiagnoseDevice()
    else if ( DeviceProp.major == 8  &&  DeviceProp.minor == 6 )  NCorePerMP = 128;
    else if ( DeviceProp.major == 8  &&  DeviceProp.minor == 9 )  NCorePerMP = 128;
    else if ( DeviceProp.major == 9 )                             NCorePerMP = 128;
+   else if ( DeviceProp.major == 12 &&  DeviceProp.minor == 0 )  NCorePerMP = 128;
    else
       fprintf( stderr, "WARNING : unable to determine the number of cores per multiprocessor for version %d.%d ...\n",
                DeviceProp.major, DeviceProp.minor );
@@ -86,6 +87,9 @@ void CUAPI_DiagnoseDevice()
 
          Aux_GetCPUInfo( FileName );
 
+         int clockRate; // in unit of kHz
+         CUDA_CHECK_ERROR(  cudaDeviceGetAttribute(&clockRate, cudaDevAttrClockRate, GetDeviceID)  );
+
          fprintf( Note, "\n" );
          fprintf( Note, "GPU Info :\n" );
          fprintf( Note, "Number of GPUs                          : %d\n"     , DeviceCount );
@@ -95,7 +99,7 @@ void CUAPI_DiagnoseDevice()
          fprintf( Note, "CUDA Runtime Version                    : %d.%d\n"  , RuntimeVersion/1000, RuntimeVersion%100 );
          fprintf( Note, "CUDA Major Revision Number              : %d\n"     , DeviceProp.major );
          fprintf( Note, "CUDA Minor Revision Number              : %d\n"     , DeviceProp.minor );
-         fprintf( Note, "Clock Rate                              : %f GHz\n" , DeviceProp.clockRate/1.0e6 );
+         fprintf( Note, "Clock Rate                              : %f GHz\n" , clockRate/1.0e6 );
          fprintf( Note, "Global Memory Size                      : %ld MB\n" , (long)DeviceProp.totalGlobalMem/1024/1024 );
          fprintf( Note, "Constant Memory Size                    : %ld KB\n" , (long)DeviceProp.totalConstMem/1024 );
          fprintf( Note, "Shared Memory Size per Block            : %ld KB\n" , (long)DeviceProp.sharedMemPerBlock/1024 );

src/GPU_API/CUAPI_SetDevice.cu

@@ -53,6 +53,7 @@ void CUAPI_SetDevice( const int Mode )
 // set the device ID
    void **d_TempPtr = NULL;
    int SetDeviceID, GetDeviceID = 999;
+   int computeMode;
    cudaDeviceProp DeviceProp;
 
    switch ( Mode )
@@ -77,9 +78,9 @@ void CUAPI_SetDevice( const int Mode )
 
 //       make sure that the "exclusive" compute mode is adopted
          CUDA_CHECK_ERROR(  cudaGetDevice( &GetDeviceID )  );
-         CUDA_CHECK_ERROR(  cudaGetDeviceProperties( &DeviceProp, GetDeviceID )  );
+         CUDA_CHECK_ERROR(  cudaDeviceGetAttribute(&computeMode, cudaDevAttrComputeMode, GetDeviceID)  );
 
-         if ( DeviceProp.computeMode != cudaComputeModeExclusive )
+         if ( computeMode != cudaComputeModeExclusive )
          {
             Aux_Message( stderr, "WARNING : \"exclusive\" compute mode is NOT enabled for \"%s\" at Rank %2d",
                          "OPT__GPUID_SELECT == -2", MPI_Rank );

src/GPU_API/CUAPI_SetMemSize.cu

@@ -36,7 +36,10 @@ void CUAPI_SetMemSize( int &GPU_NStream, int &Flu_GPU_NPGroup, int &Pot_GPU_NPGr
 // (1) GPU_NSTREAM
    if ( GPU_NStream <= 0 )
    {
-      if ( DeviceProp.deviceOverlap )
+      int gpuOverlap;
+      CUDA_CHECK_ERROR(  cudaDeviceGetAttribute(&gpuOverlap, cudaDevAttrGpuOverlap, GetDeviceID)  );
+
+      if ( gpuOverlap )
       {
 #        if   ( MODEL == HYDRO )
 #           if   ( GPU_ARCH == FERMI )
@@ -57,6 +60,8 @@ void CUAPI_SetMemSize( int &GPU_NStream, int &Flu_GPU_NPGroup, int &Pot_GPU_NPGr
             GPU_NStream = 4;
 #           elif ( GPU_ARCH == HOPPER )
             GPU_NStream = 4;
+#           elif ( GPU_ARCH == BLACKWELL )
+            GPU_NStream = 4;
 #           else
 #           error : UNKNOWN GPU_ARCH !!
 #           endif
@@ -80,6 +85,8 @@ void CUAPI_SetMemSize( int &GPU_NStream, int &Flu_GPU_NPGroup, int &Pot_GPU_NPGr
             GPU_NStream = 4;
 #           elif ( GPU_ARCH == HOPPER )
             GPU_NStream = 4;
+#           elif ( GPU_ARCH == BLACKWELL )
+            GPU_NStream = 4;
 #           else
 #           error : ERROR : UNKNOWN GPU_ARCH !!
 #           endif
@@ -118,6 +125,8 @@ void CUAPI_SetMemSize( int &GPU_NStream, int &Flu_GPU_NPGroup, int &Pot_GPU_NPGr
          Flu_GPU_NPGroup = 1*GPU_NStream*DeviceProp.multiProcessorCount;
 #        elif ( GPU_ARCH == HOPPER )
          Flu_GPU_NPGroup = 1*GPU_NStream*DeviceProp.multiProcessorCount;
+#        elif ( GPU_ARCH == BLACKWELL )
+         Flu_GPU_NPGroup = 1*GPU_NStream*DeviceProp.multiProcessorCount;
 #        else
 #        error : UNKNOWN GPU_ARCH !!
 #        endif
@@ -141,6 +150,8 @@ void CUAPI_SetMemSize( int &GPU_NStream, int &Flu_GPU_NPGroup, int &Pot_GPU_NPGr
          Flu_GPU_NPGroup = 1*GPU_NStream*DeviceProp.multiProcessorCount;
 #        elif ( GPU_ARCH == HOPPER )
          Flu_GPU_NPGroup = 1*GPU_NStream*DeviceProp.multiProcessorCount;
+#        elif ( GPU_ARCH == BLACKWELL )
+         Flu_GPU_NPGroup = 1*GPU_NStream*DeviceProp.multiProcessorCount;
 #        else
 #        error : UNKNOWN GPU_ARCH !!
 #        endif
@@ -173,6 +184,8 @@ void CUAPI_SetMemSize( int &GPU_NStream, int &Flu_GPU_NPGroup, int &Pot_GPU_NPGr
       Pot_GPU_NPGroup = 1*GPU_NStream*DeviceProp.multiProcessorCount;
 #     elif ( GPU_ARCH == HOPPER )
       Pot_GPU_NPGroup = 1*GPU_NStream*DeviceProp.multiProcessorCount;
+#     elif ( GPU_ARCH == BLACKWELL )
+      Pot_GPU_NPGroup = 1*GPU_NStream*DeviceProp.multiProcessorCount;
 #     else
 #     error : UNKNOWN GPU_ARCH !!
 #     endif
@@ -203,6 +216,8 @@ void CUAPI_SetMemSize( int &GPU_NStream, int &Flu_GPU_NPGroup, int &Pot_GPU_NPGr
       Che_GPU_NPGroup = 1*GPU_NStream*DeviceProp.multiProcessorCount;
 #     elif ( GPU_ARCH == HOPPER )
       Che_GPU_NPGroup = 1*GPU_NStream*DeviceProp.multiProcessorCount;
+#     elif ( GPU_ARCH == BLACKWELL )
+      Che_GPU_NPGroup = 1*GPU_NStream*DeviceProp.multiProcessorCount;
 #     else
 #     error : UNKNOWN GPU_ARCH !!
 #     endif
@@ -232,6 +247,8 @@ void CUAPI_SetMemSize( int &GPU_NStream, int &Flu_GPU_NPGroup, int &Pot_GPU_NPGr
       Src_GPU_NPGroup = 1*GPU_NStream*DeviceProp.multiProcessorCount;
 #     elif ( GPU_ARCH == HOPPER )
       Src_GPU_NPGroup = 1*GPU_NStream*DeviceProp.multiProcessorCount;
+#     elif ( GPU_ARCH == BLACKWELL )
+      Src_GPU_NPGroup = 1*GPU_NStream*DeviceProp.multiProcessorCount;
 #     else
 #     error : UNKNOWN GPU_ARCH !!
 #     endif

src/configure.py

@@ -957,7 +957,7 @@ def set_gpu( gpus, flags, args ):
                 gpu_opts["MAXRREGCOUNT_FLU"] = "--maxrregcount=128"
             else:
                 gpu_opts["MAXRREGCOUNT_FLU"] = "--maxrregcount=70"
-        elif 500 <= compute_capability and compute_capability <= 900:
+        elif 500 <= compute_capability and compute_capability <= 1210:
             if args["double"]:
                 gpu_opts["MAXRREGCOUNT_FLU"] = "--maxrregcount=192"
             else: