2.3.6 代码 batch_size 调整后 结果

[GeLee-Q]

浅浅的上传一下实验结果，可能粒度不够细致

[zzq96]

为什么64的时候最快那?难道L1 cache只有0.25KB(64 x 4 = 256B=0.25KB)?

[GeLee-Q]

https://tvm.apache.org/docs/how_to/optimize_operators/opt_gemm.html#blocking
首先查看官方文档，知道block是如何计算大小的

A important trick to enhance the cache hit rate is blocking — data chunk will be computed block by block. The memory access inside the block is a small neighbourhood which is with high memory locality. In this tutorial, I picked up 32 as the blocking factor. So the block will fill 32 * 32 * sizeof(float) which is 4KB in the cache whose total size is 32KB (L1 data cache)

设置为32 就是 32 * 32 * 4 = 4K
设置为64 就是 64 * 64 * 4 = 16K
设置为 128 就 就是64K 超出了L1 Cache 32K的大小，所以64是最佳

[SilkSong]

我在本地跑了一下，发现每次结果都不太一样，best block size 有时是256，有时是1024，并不是64，是否是因为电脑不同导致的？