你好，有一个问题需要请教在8.3中，通过__shared__ real S[TILE_DIM][TILE_DIM + 1]来解决bank冲突的问题。原先S[TIL

再明确回答一下，如果定义了 <div class="snippet-clipboard-content notranslate position-relative

bank冲突避免 about cuda-programming HOT 3 CLOSED

brucefan1983 commented on June 30, 2024

bank冲突避免

from cuda-programming.

fever-Wong commented on June 30, 2024

谢谢，您发给我的邮件已经收到，我会尽快处理。Thank you,the email you sent me has been received and I will handle it as soon as possible.王景博fever wong

from cuda-programming.

brucefan1983 commented on June 30, 2024

感谢您的提问。书中没有说共享内存的层数一定是32层，只不过图8.1显示的是32层（因为是以数组转置为例进行讲解）。共享内存只规定每一层有32个bank，至于有多少层，那是由共享内存数组的数据量决定的。如果您仔细看，会发现书中第87页的第一段所举的例子就只有4层。

from cuda-programming.

brucefan1983 commented on June 30, 2024

再明确回答一下，如果定义了

#define TILE_DIM 32

// in the kernel:
__shared__ float S[TILE_DIM][TILE_DIM+1];

那么该共享内存数组 S 的数据需要使用33层，每层有32个4字节的bank。

from cuda-programming.