@@ -30,6 +30,7 @@ To evaluate metric for a given _grid_ (**n_rows**, **n_cols**),
all macros are packed into the _gridcells_,
and several terms (**empty_ratio**, **ver_waste** and **hor_waste**)
that reflect wasted space are evaluated.
## Packing
Macro packing is performed as follows \[Algorithm 1 Lines 11-22\]:
- Macros are placed in order of non-increasing macro area.
- All macros are placed, one by one, into the (**n_rows**, **n_cols**) _gridcells_.
...
...
@@ -38,9 +39,7 @@ candidate _grid_ is considered.
- A macro is placed at the \textbf{first} (according to row-major order) _gridcell_ where it can be legally placed.
- Placement of a macro means placing that macro's center at the center of some _gridcell_.
- The placement of a macro's center at the center of some _gridcell_ is legal if (1) no part of the macro is outside of the canvas, and (2) no overlap of the macro with any previously-placed macro is induced.
## Metric
After macro packing, we can calculate the **empty_ratio** of current _grid_, i.e.,
the number of empty _gridcells_ over the total number of _gridcells_ (**n_rows**\times **n_cols**).
A _gridcell_ is claimed as an empty _gridcell_ if the intersection area of placed macros with it is less than 0.00001 times its area.
...
...
@@ -61,7 +60,7 @@ After calculating **empty_ratio**, **ver_waste** and **ver_waste**, the **metric
The _grid_ with best **metric** is noted as **n_rows_opt** and **n_cols_opt**.
## Grid Simplification
Once we have found **n_rows_opt** and **n_cols_opt** as described above,
we seek a smaller _grid_ that has similar metric properties. \[Algorithm 1 Lines 33-39\]
Specifically, we find values of **n_rows_actual** and **n_cols_actual** such that
...
...
@@ -71,55 +70,6 @@ This is the grid that is used in Circuit Training.
To our understanding, the foregoing procedure results in grids that are of similar sizes, e.g., with ~25 <= **n_rows_actual** , **n_cols_actual** <= ~40.
The gridding process starts with the dimensions **H_canvas** and **W_canvas** of the layout canvas, as well as a list of hard **macro blocks**, where each block has a width and a height. Blocks are not rotatable. The area of a block is the product of its width and height.
Then, the gridding searches over combinations **(n_rows, n_cols)**, with constraints
- 10 <= **n_rows** , **n_cols** <= 100
-**n_rows*** **n_cols** <= 3000
Note that **n_rows*** **n_cols** is the number of _gridcells_ in the canvas. Each _gridcell_ has width **W_canvas / n_cols** and height **H_canvas / n_rows**.
The main idea is to search for a particular **(n_rows, n_cols)** combination that minimizes a metric called _loss_.
To evaluate _loss_ for a given _grid_, all blocks are packed into the _grid_, and several terms that reflect wasted space are evaluated.
## Packing
Packing is performed as follows.
- All blocks are placed, one by one, into the **(n_rows, n_cols)** _grid_. If the current block cannot be placed, then the _grid_ is infeasible and the next candidate _grid_ is considered.
- Blocks are placed in order of decreasing (i.e., non-increasing) block **area**.
- A block is placed at the **first** (according to row-major order) gridcell where it can be legally placed.
- Placement of a block means placing that block's **center** at the **center** of some gridcell.
- The placement of a block's center at the center of some gridcell is _legal_ if (1) no part of the block is outside of the canvas, and (2) no overlap of the block with any previously-placed block is induced.
- If two placed blocks intersect gridcells of the same row in the _grid_, and the x-spans of the blocks intersect, then the two blocks have _horizontal overlap_. (Note that this definition achieves simplicity at the cost of pessimism - but, the goal is to have a quick-and-dirty packing that feeds evaluation of _loss_.) Then, _vertical overlap_ is similarly defined.
## Loss
_Loss_ is well-defined once all blocks have been legally packed (i.e., placed) into the grid, according to the above process, with no overlap.
We search for the pair **(n_rows_opt, n_cols_opt)** that satisfies constraints above and minimizes total loss, which is defined as **loss_tot = loss_h + loss_v + loss_2d**. Currently, we are unclear on the definition of the **loss_2d** term and for now do not implement it. (However, we believe this term must somehow be based on the ratio between total "used gridcells" (i.e., having nonzero area of placed macro blocks) and "total gridcells" (i.e., **n_rows*** **n_cols**).)
To evaluate horizontal loss, **loss_h**, we calculate
-**width_tot_blocks** = the sum of widths of all blocks in the design
-**width_tot__used_gridcells** = the sum of widths of all gridcells in the grid that have nonempty intersection with some placed block
Vertical loss, **loss_v**, is similarly evaluated.
## Grid Simplification
Once we have found **n_rows_opt** and **n_cols_opt** as described above, we seek a smaller grid that has similar _loss_ properties.
Specifically, we find values of **n_rows_actual** and **n_cols_actual** such that the above constraints are satisfied, **loss_h_actual + loss_v_actual** is within some _tolerance_ (e.g., 5% or 10%) of the optimal **loss_tot**, and **n_rows_actual * n_cols_actual** is minimized. This is the grid that is used in Circuit Training.
To our understanding, the foregoing procedure results in grids that are of similar sizes, e.g., with ~25 <= **n_rows_actual** , **n_cols_actual** <= ~40.
## Thanks
We thank Google engineers for May 19, 2022 discussions that explained the gridding method used in Circuit Training.
All errors of understanding and implementation are the authors'. We will rectify such errors as soon as possible after being made aware of them.
