grid: Add GPU backend

INSTALL.md

@@ -189,6 +189,7 @@ version of CP2K (ssmp or psmp), the FFTW3 threading library libfftw3_threads
 - Use `-D__CUDA_PROFILING` to turn on Nvidia Tools Extensions.
   It requires to link `-lnvToolsExt`.
 - Link to a blas/scalapack library that accelerates large DGEMMs (e.g. libsci_acc)
+- Use the `-D__GRID_CUDA` to compile the GPU and HYBRID backends for the grid library.
 
 ### 2k. libxc (optional, wider choice of xc functionals)
 

src/cp2k_info.F

@@ -274,6 +274,10 @@ FUNCTION cp2k_flags() RESULT(flags)
       flags = TRIM(flags)//" check_diag"
 #endif
 
+#if defined __GRID_CUDA
+      flags = TRIM(flags)//" grid_cuda"
+#endif
+
    END FUNCTION cp2k_flags
 
 ! **************************************************************************************************

src/grid/Makefile

@@ -5,7 +5,10 @@ all: grid_collocate_miniapp.x grid_collocate_unittest.x
 clean:
 	rm -fv grid_*.o */grid_*.o grid_*.x
 
-CFLAGS := -fopenmp -g -O3 -std=c99 -march=native -Wall -Wextra -Wpedantic
+CFLAGS := -fopenmp -g -O3 -std=c99 -march=native -Wall -Wextra -Wpedantic -I${CUDA_PATH}/include -D__GRID_CUDA
+NVFLAGS := -g -O3 -arch sm_50 -Wno-deprecated-gpu-targets -Xcompiler "-fopenmp -Wall -Wextra" -D__GRID_CUDA
+
+LIBS := -lm -lblas -lcudart -lcuda -L${CUDA_PATH}/lib64
 
 ALL_HEADERS := $(shell find . -name "*.h")
 
@@ -18,15 +21,20 @@ ALL_OBJECTS := grid_collocate_replay.o \
                ref/grid_ref_task_list.o \
                ref/grid_ref_collocate.o \
                ref/grid_ref_integrate.o \
-               ref/grid_ref_prepare_pab.o
+               ref/grid_ref_prepare_pab.o \
+               gpu/grid_gpu_task_list.o \
+               gpu/grid_gpu_collocate.o
 
 %.o: %.c $(ALL_HEADERS)
 	cd $(dir $<); $(CC) -c $(CFLAGS) $(notdir $<)
 
+%.o: %.cu $(ALL_HEADERS)
+	cd $(dir $<); nvcc -c $(NVFLAGS) $(notdir $<)
+
 grid_collocate_miniapp.x: grid_collocate_miniapp.o $(ALL_OBJECTS)
-	$(CC) $(CFLAGS) -o $@ $^ -lm -lblas
+	nvcc $(NVFLAGS) -o $@ $^ $(LIBS)
 
 grid_collocate_unittest.x: grid_collocate_unittest.o $(ALL_OBJECTS)
-	$(CC) $(CFLAGS) -o $@ $^ -lm -lblas
+	nvcc $(NVFLAGS) -o $@ $^ $(LIBS)
 
 #EOF

src/grid/PACKAGE

@@ -10,6 +10,7 @@
         "./common",
         "./ref",
         "./cpu",
+        "./gpu",
     ],
     "public": ["grid_api.F"],
 }

src/grid/common/grid_basis_set.h

@@ -25,6 +25,8 @@ typedef struct {
   double *zet;
 } grid_basis_set;
 
+#ifndef __cplusplus
+
 /*******************************************************************************
  * \brief Allocates a basis set which can be passed to grid_create_task_list.
  *
@@ -63,4 +65,6 @@ void grid_free_basis_set(grid_basis_set *basis_set);
 
 #endif
 
+#endif
+
 // EOF

src/grid/common/grid_constants.h

@@ -35,5 +35,6 @@
 #define GRID_BACKEND_AUTO 10
 #define GRID_BACKEND_REF 11
 #define GRID_BACKEND_CPU 12
+#define GRID_BACKEND_GPU 13
 
 // EOF

src/grid/common/grid_prepare_pab.h

@@ -251,7 +251,13 @@ GRID_DEVICE static void prepare_pab_Di2(const int ider, const orbital a,
 }
 
 /*******************************************************************************
- * \brief TODO
+ * \brief Transforms a given element of the density matrix according to func.
+ * \param func          Transformation function to apply, one of GRID_FUNC_*.
+ * \param {a,b}         Orbital angular momenta.
+ * \param zet_{a,b}     Gaussian exponents.
+ * \param pab_val       Input matrix element of pab.
+ * \param n             Leading dimensions of output matrix cab.
+ * \param cab           Output matrix.
  * \author Ole Schuett
  ******************************************************************************/
 GRID_DEVICE static void prepare_pab(const int func, const orbital a,

src/grid/gpu/PACKAGE

@@ -0,0 +1,5 @@
+{
+    "description": "Optimized GPU backend",
+    "requires": ["../common",],
+    "archive": "libcp2kgridgpu",
+}