GLuint vbo;
struct cudaGraphicsResource *cudaVboResource;

glGenBuffers(1, &vbo);
glBindBuffer(GL_ARRAY_BUFFER, vbo);
glBufferData(GL_ARRAY_BUFFER, n * sizeof(float4), nullptr, GL_DYNAMIC_DRAW);
cudaGraphicsGLRegisterBuffer(&cudaVboResource, vbo, cudaGraphicsMapFlagsWriteDiscard);

cudaGraphicsMapResources(1, &cudaVboResource, 0);
float4 *d_positions;
size_t numBytes;
cudaGraphicsResourceGetMappedPointer((void**)&d_positions, &numBytes, cudaVboResource);

updatePositionsKernel<<<blocks, threads>>>(d_positions, n);

cudaGraphicsUnmapResources(1, &cudaVboResource, 0);
// OpenGL can now render the buffer CUDA just wrote directly, with no CPU copy
