cuIdx is a single header CUDA/C++ library that aims to assist with indexing in CUDA kernels. In particular it provides functions to retrieve:
- The 1D local index of a thread (within its block)
- The 1D global index of a thread
- The 1D local warp index of a thread (within its block)
- The 1D global warp index of a thread
- The the index of a thread w.r.t to it's warp (lane)
- The 1D block size of a thread block
for any number of grid and block dimensions.
All library functions are meant to be called from within CUDA kernels. (Almost) all functions accept template arguments that denote the number of grid/block dimensions as specified on kernel launch. Non-template versions exist and can cover any number of grid/block dimensions at the expense of a slightly more costly computation (3 dimensions are assumed to cover all cases.)
cuIdx.cuh should be compiled with NVCC. Minimum supported CUDA Version: 2.1
Todo
- Retrieve global thread index
#include "cuidx.cuh"
using namespace cuidx;
__global__ void vecadd(int *A, int *B, int *C, unsigned len) {
// auto pos = blockIdx.x * blockDim.x + threadIdx.x; // vanilla cuda, 1D grid/blocks
// auto pos = blockIdx.x * blockDim.x * blockDim.y
// + threadIdx.y * blockDim.x + threadIdx.x; // vanilla cuda, 1D grid, 2D blocks
// auto pos = blockIdx.x * blockDim.x * blockDim.y * blockDim.z
// + threadIdx.z * blockDim.y * blockDim.x
// + threadIdx.y * blockDim.x
// + threadIdx.x // vanilla cuda, 1D grid, 3D blocks
auto pos = gtid(); // <- cuidx
if (pos < len)
C[pos] = A[pos] + B[pos];
}
- Retrieve global thread index (faster)
#include "cuidx.cuh"
using namespace cuidx;
__global__ void vecadd(int *A, int *B, int *C, unsigned len) {
// auto pos = gtid(); // <- cuidx, assumes 3D grid/blocks
auto pos = gtid<1,1>(); // <- cuidx, for 1D grid/blocks (faster)
if (pos < len)
C[pos] = A[pos] + B[pos];
}
- Warp level
#include "cuidx.cuh"
using namespace cuidx;
__global__ void kernel(int *A, int *B, int *C, unsigned len) {
__shared__ int shmem[BLOCK_SIZE / WARPSIZE] = {0};
int warp_sum = 0;
...
__syncwarp();
// vanilla cuda
auto tid = threadIdx.x * blockDim.x + threadIdx.x; // get tid in 2D block
if ( tid % 32 == 0) // check laneid == 0 (leader)
shmem[tid / 32] = warp_sum; // write at warp index
// cuidx
if ( wleader())
shmem[ wid()] = warp_sum;
}