US 12,481,878 B2
Applying a two dimensional (2D) kernel on an input feature map
Orly Weisel, Efrat (IL); Yaniv Fais, Tel Aviv (IL); and Shira Hirsch, Jerusalem (IL)
Assigned to Mobileye Vision Technologies Ltd., Jerusalem (IL)
Filed by Mobileye Vision Technologies Ltd., Jerusalem (IL)
Filed on Aug. 10, 2022, as Appl. No. 17/884,948.
Claims priority of provisional application 63/231,529, filed on Aug. 10, 2021.
Prior Publication US 2023/0046558 A1, Feb. 16, 2023
Int. Cl. G06N 3/08 (2023.01); G06N 3/063 (2023.01)
CPC G06N 3/08 (2013.01) [G06N 3/063 (2013.01)] 19 Claims
OG exemplary drawing
 
1. A method for applying a two dimensional (2D) kernel on an input feature map (IFM) that comprises multiple IFM tiles, the method comprising:
selecting a current IFM tile of the multiple IFM tiles;
writing the current IFM tile to a buffer of a neural network (NN) processor, wherein the current IFM tile comprises overlap current IFM tile rows that are utilized during (a) a calculation of at least one row of a current output feature map (OFM) tile, and (b) a calculation of at least one row of a next OFM tile;
calculating the current OFM tile, by the NN processor, wherein the calculating comprises scanning at least the current IFM tile with the 2D kernel;
selecting a next IFM tile of the multiple IFM tiles;
writing the next IFM tile to the buffer, wherein the writing of the next IFM tile to the buffer comprises:
overwriting rows of the current IFM tile that differ from the overlap current IFM tiles, wherein the overwriting occurs following a completion of the calculating of the current OFM tile; and
maintaining the overlap current IFM tile rows in the buffer until, at least, completion of a calculation of the at least one row of the next OFM tile; and
calculating the next OFM tile, by the NN processor, wherein the calculating comprises scanning the next IFM tile and the overlap current IFM tile rows with the 2D kernel.