Question about data format document Fig17 about doc HOT 5 CLOSED

NVDLA supports Winograd not only 3x3 kernels but also other cases.
The condition of kernel supported by Winograd:
int((original_kernel_width + conv_x_stride - 1) / conv_x_stride) == 3 &&
int((original_kernel_height + conv_y_stride - 1) / conv_y_stride) == 3
Once the condition is reached, that kernel can be supported for Winograd with some extra steps.

简单来说，NVDLA的Winograd模式可以支持3x3以外的kernel，只要保证kernel的水平大小除以水平stride等于3（上取整）以及垂直大小除以垂直stride也等于3（上取整），那么这个kernel就可以被Winograd加速，当然需要一些额外的kernel调整步骤

from doc.

非常感谢您的回复，对我来说非常重要的信息，之前我在研究winograd相关算法的时候一直没有考虑到相关问题，而且我看到winograd的output activation 的反变换是基于固定连接关系的加法器实现，所以一直以为只能支持一种变换； 关于kernel的调整有相关的paper可以参考么，再次感谢！
Thanks so much! It's very important for me since I never considerate that case and I think the inverse transform of output activation is fixed by the adders . Could you please share me some related papers about the adjustment of kernel.
By the way , I think "int" is a bit of confusing , ceiling is better.

from doc.

Actually it's still 3x3 case. A non-3x3 kernel which meetis the formulas above can be transformed to a 3x3 kernel whose stride is 1x1. It is showed in first two diagram of http://nvdla.org/_images/format_channel_extension_and_conversion_for_wingorad.svg
The input activition should be remapped accordingly. CDMA in NVDLA will handle the conversion of input activation automatically.
PS: I'm used to C code :-)

实际上这依然是3x3的kernel。对于非3x3的kernel而言，只要满足上述两个表达式，就可以被转化为3x3并且stride是1x1的kernel
相应的，输入的activation也需要重新排列来适配转化后的kernel。这一点是由NVDLA中的CDMA来完成的。

from doc.

Thanks! happy new year

from doc.

@pengwubj can you please get in touch with me at [email protected]? It concerns a systems programming course I teach.

from doc.