first cuda kernel try

build.sh

+#!/bin/sh
+
+TORCH=$(python -c "import os; import torch; print(os.path.dirname(torch.__file__))")
+SRC_DIR=torch_scatter/kernel
+BUILD_DIR=torch_scatter/build
+
+mkdir -p $BUILD_DIR
+nvcc -c -o $BUILD_DIR/kernel.so $SRC_DIR/kernel.cu -Xcompiler -fPIC -shared -I$TORCH/lib/include/TH -I$TORCH/lib/include/THC -I$TORCH/lib/include -I$SRC_DIR

torch_scatter/kernel/generic/kernel.cu

+#ifndef THC_GENERIC_FILE
+#define THC_GENERIC_FILE "generic/kernel.cu"
+#else
+
+void scatter_(mul)(THCState *state, int dim, THCTensor *output, THCudaLongTensor *index, THCTensor *input) {
+  printf("mul");
+}
+
+void scatter_(div)(THCState *state, int dim, THCTensor *output, THCudaLongTensor *index, THCTensor *input) {
+  printf("div");
+}
+
+void scatter_(mean)(THCState *state, int dim, THCTensor *output, THCudaLongTensor *index, THCTensor *input, THCTensor *num_output) {
+  printf("mean");
+}
+
+void scatter_(max)(THCState *state, int dim, THCTensor *output, THCudaLongTensor *index, THCTensor *input, THCudaLongTensor *arg_output) {
+  printf("max");
+}
+
+void scatter_(min)(THCState *state, int dim, THCTensor *output, THCudaLongTensor *index, THCTensor *input, THCudaLongTensor *arg_output) {
+  printf("min");
+}
+
+void index_backward(THCState *state, int dim, THCTensor *output, THCudaLongTensor *index, THCTensor *grad, THCudaLongTensor *arg_grad) {
+  printf("index_backward");
+}
+
+#endif

torch_scatter/kernel/kernel.cu

+#include <THC/THC.h>
+
+#define scatter_(NAME) TH_CONCAT_4(scatter_, NAME, _kernel_, Real)
+#define index_backward TH_CONCAT_2(index_backward_kernel_, Real)
+
+#include "generic/kernel.cu"
+#include "THCGenerateAllTypes.h"