Halide: How to process image in (overlapping) blocks? - c++

Question

I'm discovering Halide and got some success with a pipeline doing various
transformations. Most of these are based on the examples within the sources (color-transformations, various filters, hist-eq).
My next step needs to process the image in blocks. In a more general form,
partially-overlapping blocks.
Examples
Input:
[ 1, 2, 3, 4, 5, 6, 7, 8,
9, 10, 11, 12, 13, 14, 15, 16,
17, 18, 19, 20, 21, 22, 23, 24,
25, 26, 27, 28, 29, 30, 31, 32]
Non-overlapping blocks:
Size: 2x4
[ 1, 2, 3, 4,
9, 10, 11, 12]
[ 5, 6, 7, 8,
13, 14, 15, 16]
[ 17, 18, 19, 20,
25, 26, 27, 28]
[ 21, 22, 23, 24,
29, 30, 31, 32]
Overlapping blocks:
Size: 2x4 with 50% overlap (both axes)
[ 1, 2, 3, 4,
9, 10, 11, 12]
[ 3, 4, 5, 6,
11, 12, 13, 14]
[ 5, 6, 7, 8,
13, 14, 15, 16]
-
[ 9, 10, 11, 12,
17, 18, 19, 20]
[11, 12, 13, 14,
19, 20, 21, 22]
...
I suspect there should be a nice way to express these, as those are also quite common
in many algorithms (e.g. macroblocks).
What i checked out
I tried to gather ideas from the tutorial and example apps and found the following,
which seem somewhat connected to what i want to implement:
Halide tutorial lesson 6: Realizing Funcs over arbitrary domains
// We start by creating an image that represents that rectangle
Image<int> shifted(5, 7); // In the constructor we tell it the size
shifted.set_min(100, 50); // Then we tell it the top-left corner
The problem i have: how to generalize this to multiple shifted domains without looping?
Halide tutorial lesson 9: Multi-pass Funcs, update definitions, and reductions
Here RDom is introduced which looks nice to create a block-view
Most examples using RDom seem to be sliding-window like approaches where there are no jumps
Target
So in general i'm asking how to implement a block-based view which can then be processed by
other steps.
It would be nice if the approach will be general enough to realize both, overlapping & no overlapping
Somehow generating the top-left indices first?
In my case, the image-dimension is known at compile-time which simplifies this
But i still would like some compact form which is nice to work with from Halide's perspective (no handcoded stuff like those examples with small filter-boxes)
The approach used might be depending on the output per block, which is a scalar in my case
Maybe someone can give me some ideas and/or some examples (which would be very helpful).
I'm sorry for not providing code, as i don't think i could produce anything helpful.
Edit: Solution
After dsharlet's answer and some tiny debugging/discussion here, the following very simplified self-containing code works (assuming an 1-channel 64x128 input like this one i created).
#include "Halide.h"
#include "Halide/tools/halide_image_io.h"
#include <iostream>
int main(int argc, char **argv) {
Halide::Buffer<uint8_t> input = Halide::Tools::load_image("TestImages/block_example.png");
// This is a simple example assuming an input of 64x128
std::cout << "dim 0: " << input.width() << std::endl;
std::cout << "dim 1: " << input.height() << std::endl;
// The "outer" (block) and "inner" (pixel) indices that describe a pixel in a tile.
Halide::Var xo, yo, xi, yi, x, y;
// The distance between the start of each tile in the input.
int tile_stride_x = 32;
int tile_stride_y = 64;
int tile_size_x = 32;
int tile_size_y = 64;
Halide::Func tiled_f;
tiled_f(xi, yi, xo, yo) = input(xo * tile_stride_x + xi, yo * tile_stride_y + yi);
Halide::RDom tile_dom(0, tile_size_x, 0, tile_size_y);
Halide::Func tile_means;
tile_means(xo, yo) = sum(Halide::cast<uint32_t>(tiled_f(tile_dom.x, tile_dom.y, xo, yo))) / (tile_size_x * tile_size_y);
Halide::Func output;
output(xo, yo) = Halide::cast<uint8_t>(tile_means(xo, yo));
Halide::Buffer<uint8_t> output_(2, 2);
output.realize(output_);
Halide::Tools::save_image(output_, "block_based_stuff.png");
}

Answer 1

Here's an example that breaks a Func into blocks of abitrary stride and size:
Func f = ... // The thing being blocked
// The "outer" (block) and "inner" (pixel) indices that describe a pixel in a tile.
Var xo, yo, xi, yi;
// The distance between the start of each tile in the input.
int tile_stride_x, tile_stride_y;
Func tiled_f;
tiled_f(xi, yi, xo, yo) = f(xo * tile_stride_x + xi, yo * tile_stride_y + yi);
Func tiled_output;
tiled_output(xi, yi, xo, yo) = ... // Your tiled processing here
To compute some reduction (like statistics) on each block, you can do the following:
RDom tile_dom(0, tile_size_x, 0, tile_size_y);
Func tile_means;
tile_means(xo, yo) = sum(tiled_output(tile_dom.x, tile_dom.y, xo, yo)) / (tile_size_x * tile_size_y);
To flatten the tiles back into a result is a bit tricky. It probably depends on your method of combining the results in overlapped areas. If you want to add up the overlapping tiles, the simplest way is probably to use an RDom:
RDom tiles_dom(
0, tile_size_x,
0, tile_size_y,
min_tile_xo, extent_tile_xo,
min_tile_yo, extent_tile_yo);
Func output;
Expr output_x = tiles_dom[2] * tile_stride_x + tiles_dom[0];
Expr output_y = tiles_dom[3] * tile_stride_y + tiles_dom[1];
output(x, y) = 0;
output(output_x, output_y) += tiled_output(tiles_dom[0], tiles_dom[1], tiles_dom[2], tiles_dom[3]);
Note that in the above two blocks of code, tile_stride_x and tile_size_x are independent parameters, allowing for any tile size and overlap.
In both of your examples, tile_size_x = 4, and tile_size_y = 2. To get non-overlapping tiles, set the tile strides equal to the tile size. To get 50% overlapping tiles, set tile_stride_x = 2, and tile_stride_y = 1.
A useful schedule for an algorithm like this is:
// Compute tiles as needed by the output.
tiled_output.compute_at(output, tile_dom[2]);
// or
tiled_output.compute_at(tile_means, xo);
There are other options, like using a pure func (no update/RDom) that uses the mod operator to figure out tile inner and outer indices. However, this approach can be difficult to schedule efficiently with overlapping tiles (depending on the processing you do at each tile). I use the RDom approach when this problem comes up.
Note that with the RDom approach, you have to supply the bounds of the tile indices you want computed (min_tile_xo, extent_tile_xo, ...), which can be tricky for overlapped tiles.