In this PR, we increase the spatial dimension of the C32 and D32 ports in GeMMX to 2 (previously it was 1).
With this, we add an extra stride in the spatial address generation for C32 and D32 (now can jump at 2 elements and row granularity), thus enabling more data layout opportunities for reducing the bank conflict.
@jorendumoulin @JosseVanDelm , please have a look at it, now we have two extra spatial strides CSRs. Please let me know if there is anything unclear.
JosseVanDelm
commented
1 month ago
In this PR, we increase the spatial dimension of the C32 and D32 ports in GeMMX to 2 (previously it was 1). With this, we add an extra stride in the spatial address generation for C32 and D32 (now can jump at 2 elements and row granularity), thus enabling more data layout opportunities for reducing the bank conflict.
@jorendumoulin @JosseVanDelm , please have a look at it, now we have two extra spatial strides CSRs. Please let me know if there is anything unclear.