document issue

[foricee]

Thank you for your excellent work and detailed document！I think there's a small typo in the document. According to NVIDIA's doc, below should be above.

https://github.com/wangsiping97/FastGEMV/blob/main/method_and_result.md

The __shfl_down_sync(mask, var, delta) function is used to perform the sum reduction. It takes a mask (0xffffffff, representing all threads in a warp) which indicates the threads participating in the operation, var which is the value to be shifted, and delta which is the number of positions to shift. This function returns the value of var held by the thread delta positions below the calling thread in the warp. If the target thread is out of bounds, the calling thread's own value is returned.

https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#id36

__shfl_down_sync() calculates a source lane ID by adding delta to the caller’s lane ID.