// In the layout of fragment A and fragment C during WGMMA, data each thread holds resides in two particular rows. This function converts the local_row_idx (0~2) to the actual row_idx
// You may refer to this link for the detailed layout: https://docs.nvidia.com/cuda/parallel-thread-execution/#wgmma-64n16-a
// If is_no_split is True, then this request is exclusively assigned to this sm_part, so we shall write the result directly into params.o_ptr and params.softmax_lse_ptr. Otherwise, write to oaccum_ptr and softmax_lseaccum_ptr, with the corresponding split idx being (n_split_idx + num_splits_ptr[batch_idx])
// For the complete schedule of the kernel, please read our deep-dive write-up (link can be found in the README.md file).
// We don't use __ldg here, otherwise NVCC (ptxas, in particular) will do instruction reorder and place __ldg (LDG.E.128.CONSTANT in SASS) in front of cudaGridDependencySynchronize() (ACQBULK in SASS), leading to data race.
// Besides, a token may have some extra masks other than the common mask. We use rRightBorderForQSeq to denote it, which means the right border of the k-sequence for the particular q token. In this way, (seqlen_k-common_mask_len) - rRightBorderForQSeq < 64 holds, which means that we only need to apply the causal mask to the last two KV blocks
// NOTE This may lead to start_block_idx >= end_block_idx which needs some special handling
// In the layout of fragment A and fragment C during WGMMA, data each thread holds resides in two particular rows. This function converts the local_row_idx (0~1) to the actual row_idx
// You may refer to this link for the detailed layout: https://docs.nvidia.com/cuda/parallel-thread-execution/#wgmma-64n16-a
