Skip to content

Rework of filter boundary behaviour#774

Open
Quickmarble wants to merge 7 commits intoimage-rs:mainfrom
Quickmarble:filter-refactor
Open

Rework of filter boundary behaviour#774
Quickmarble wants to merge 7 commits intoimage-rs:mainfrom
Quickmarble:filter-refactor

Conversation

@Quickmarble
Copy link
Copy Markdown
Contributor

@Quickmarble Quickmarble commented Apr 7, 2026

This extends the work started in #769 (specifying boundary behaviour with Border<P>) to filter*() functions and friends. Refactoring all of them while keeping track of performance might take a while.

Closes #2.

@Quickmarble
Copy link
Copy Markdown
Contributor Author

There might be more places where boundary behaviour should be reviewed, but I guess it's enough changes for the scope of this PR. I need to update docs and then it'll be ready.

The following functions have been modified to account for Border<P>:

  • filter(), filter_parallel(), gaussian_blue_f32(), separable_filter(), separable_filter_equal(), filter_clamped(), {horizontal,vertical}_filter(), laplacian_filter(), laplacian_filter_parallel();
  • sharpen3x3(), sharpen3x3_parallel(), sharpen_gaussian();
  • bilateral_filter();
  • box_filter();
  • median_filter();
  • gradients_grayscale(), gradients(), {horizontal,vertical}_{sobel,scharr,prewitt}(), {sobel,prewitt}_gradients();
  • cell_histograms();
  • {row,column}_running_sum().

One- and two-dimensional filters now employ BoundaryAccess::convolve_at() and BoundaryAccess::convolve_{horizontal,vertical}_at().

@Quickmarble
Copy link
Copy Markdown
Contributor Author

Benchmarks before:

test edges::benches::bench_canny                                                      ... bench:   1,546,706.70 ns/iter (+/- 22,128.71)
test filter::benches::bench_filter_clamped_gray_3x3                                   ... bench:     584,080.35 ns/iter (+/- 1,641.65)
test filter::benches::bench_filter_clamped_gray_3x3_ref                               ... bench:     156,551.45 ns/iter (+/- 161.06)
test filter::benches::bench_filter_clamped_gray_5x5                                   ... bench:   1,266,857.00 ns/iter (+/- 8,958.98)
test filter::benches::bench_filter_clamped_gray_7x7                                   ... bench:   2,244,828.10 ns/iter (+/- 23,081.02)
test filter::benches::bench_filter_clamped_parallel_gray_3x3                          ... bench:     253,829.89 ns/iter (+/- 353,977.36)
test filter::benches::bench_filter_clamped_parallel_gray_5x5                          ... bench:     342,850.21 ns/iter (+/- 83,774.43)
test filter::benches::bench_filter_clamped_parallel_gray_7x7                          ... bench:     513,707.80 ns/iter (+/- 94,758.57)
test filter::benches::bench_filter_clamped_parallel_rgb_3x3                           ... bench:     305,939.88 ns/iter (+/- 75,138.22)
test filter::benches::bench_filter_clamped_parallel_rgb_5x5                           ... bench:     486,178.99 ns/iter (+/- 74,637.19)
test filter::benches::bench_filter_clamped_parallel_rgb_7x7                           ... bench:     738,947.28 ns/iter (+/- 113,300.36)
test filter::benches::bench_filter_clamped_rgb_3x3                                    ... bench:     983,044.05 ns/iter (+/- 10,010.20)
test filter::benches::bench_filter_clamped_rgb_5x5                                    ... bench:   2,349,108.60 ns/iter (+/- 30,522.30)
test filter::benches::bench_filter_clamped_rgb_7x7                                    ... bench:   4,226,682.10 ns/iter (+/- 30,645.08)
test filter::benches::bench_gaussian_f32_stdev_1                                      ... bench:     120,661.23 ns/iter (+/- 1,003.70)
test filter::benches::bench_gaussian_f32_stdev_10                                     ... bench:     750,532.90 ns/iter (+/- 15,513.70)
test filter::benches::bench_gaussian_f32_stdev_3                                      ... bench:     262,145.30 ns/iter (+/- 1,175.09)
test filter::benches::bench_horizontal_filter                                         ... bench:     619,759.00 ns/iter (+/- 1,090.36)
test filter::benches::bench_separable_filter                                          ... bench:     447,399.10 ns/iter (+/- 986.46)
test filter::benches::bench_vertical_filter                                           ... bench:     619,702.70 ns/iter (+/- 1,900.86)
test filter::bilateral::benches::bench_bilateral_filter_greyscale                     ... bench:  23,385,243.60 ns/iter (+/- 20,811.49)
test filter::bilateral::benches::bench_bilateral_filter_rgb                           ... bench:  25,062,125.20 ns/iter (+/- 163,893.23)
test filter::box_filter::benches::bench_box_filter                                    ... bench:     656,247.15 ns/iter (+/- 2,190.00)
test filter::median::benches::bench_median_filter_s100_r1                             ... bench:      38,876.62 ns/iter (+/- 604.49)
test filter::median::benches::bench_median_filter_s100_r4                             ... bench:      72,981.55 ns/iter (+/- 652.40)
test filter::median::benches::bench_median_filter_s100_r8                             ... bench:     121,941.80 ns/iter (+/- 773.93)
test filter::median::benches::bench_median_filter_s100_rx1_ry4                        ... bench:      40,694.45 ns/iter (+/- 881.24)
test filter::median::benches::bench_median_filter_s100_rx1_ry8                        ... bench:      40,419.63 ns/iter (+/- 328.81)
test filter::median::benches::bench_median_filter_s100_rx4_ry8                        ... bench:      70,971.05 ns/iter (+/- 905.82)
test filter::median::benches::bench_median_filter_s100_rx8_ry1                        ... bench:     117,630.32 ns/iter (+/- 1,359.44)
test gradients::benches::bench_sobel_gradients                                        ... bench:   3,758,538.50 ns/iter (+/- 38,712.83)
test hog::benches::bench_hog                                                          ... bench:     303,231.67 ns/iter (+/- 864.03)
test integral_image::benches::bench_column_running_sum                                ... bench:         311.97 ns/iter (+/- 3.76)
test integral_image::benches::bench_row_running_sum                                   ... bench:         191.78 ns/iter (+/- 2.25)
test seam_carving::benches::bench_shrink_width_s100_r1                                ... bench:     147,886.17 ns/iter (+/- 344.99)
test seam_carving::benches::bench_shrink_width_s100_r4                                ... bench:     583,191.40 ns/iter (+/- 2,128.95)
test seam_carving::benches::bench_shrink_width_s100_r8                                ... bench:   1,145,089.20 ns/iter (+/- 19,530.64)

Benchmarks after:

test edges::benches::bench_canny                                                      ... bench:   1,606,450.30 ns/iter (+/- 13,825.03)
test filter::benches::bench_filter_clamped_gray_3x3                                   ... bench:     618,154.45 ns/iter (+/- 3,379.63)
test filter::benches::bench_filter_clamped_gray_3x3_ref                               ... bench:     156,640.23 ns/iter (+/- 684.63)
test filter::benches::bench_filter_clamped_gray_5x5                                   ... bench:   1,323,662.10 ns/iter (+/- 6,155.14)
test filter::benches::bench_filter_clamped_gray_7x7                                   ... bench:   2,320,877.20 ns/iter (+/- 17,540.28)
test filter::benches::bench_filter_clamped_parallel_gray_3x3                          ... bench:     250,470.22 ns/iter (+/- 375,711.28)
test filter::benches::bench_filter_clamped_parallel_gray_5x5                          ... bench:     336,580.32 ns/iter (+/- 68,551.22)
test filter::benches::bench_filter_clamped_parallel_gray_7x7                          ... bench:     496,986.41 ns/iter (+/- 72,033.12)
test filter::benches::bench_filter_clamped_parallel_rgb_3x3                           ... bench:     301,910.75 ns/iter (+/- 61,504.24)
test filter::benches::bench_filter_clamped_parallel_rgb_5x5                           ... bench:     467,288.24 ns/iter (+/- 63,345.47)
test filter::benches::bench_filter_clamped_parallel_rgb_7x7                           ... bench:     710,636.18 ns/iter (+/- 107,407.17)
test filter::benches::bench_filter_clamped_rgb_3x3                                    ... bench:   1,014,798.80 ns/iter (+/- 15,952.69)
test filter::benches::bench_filter_clamped_rgb_5x5                                    ... bench:   2,472,048.90 ns/iter (+/- 23,803.58)
test filter::benches::bench_filter_clamped_rgb_7x7                                    ... bench:   4,440,285.70 ns/iter (+/- 89,498.21)
test filter::benches::bench_gaussian_f32_stdev_1                                      ... bench:     125,662.53 ns/iter (+/- 644.16)
test filter::benches::bench_gaussian_f32_stdev_10                                     ... bench:     794,682.10 ns/iter (+/- 29,472.12)
test filter::benches::bench_gaussian_f32_stdev_3                                      ... bench:     266,603.67 ns/iter (+/- 586.30)
test filter::benches::bench_horizontal_filter                                         ... bench:     621,516.65 ns/iter (+/- 1,472.73)
test filter::benches::bench_separable_filter                                          ... bench:     449,685.30 ns/iter (+/- 2,281.07)
test filter::benches::bench_vertical_filter                                           ... bench:     621,339.20 ns/iter (+/- 6,215.51)
test filter::bilateral::benches::bench_bilateral_filter_greyscale                     ... bench:  17,407,408.50 ns/iter (+/- 94,707.10)
test filter::bilateral::benches::bench_bilateral_filter_rgb                           ... bench:  25,102,466.00 ns/iter (+/- 30,795.66)
test filter::box_filter::benches::bench_box_filter                                    ... bench:     656,730.50 ns/iter (+/- 6,125.16)
test filter::median::benches::bench_median_filter_s100_r1                             ... bench:      36,239.47 ns/iter (+/- 166.48)
test filter::median::benches::bench_median_filter_s100_r4                             ... bench:      72,511.96 ns/iter (+/- 1,528.24)
test filter::median::benches::bench_median_filter_s100_r8                             ... bench:     120,008.00 ns/iter (+/- 1,108.74)
test filter::median::benches::bench_median_filter_s100_rx1_ry4                        ... bench:      37,737.98 ns/iter (+/- 212.96)
test filter::median::benches::bench_median_filter_s100_rx1_ry8                        ... bench:      39,023.13 ns/iter (+/- 260.45)
test filter::median::benches::bench_median_filter_s100_rx4_ry8                        ... bench:      71,151.80 ns/iter (+/- 1,219.23)
test filter::median::benches::bench_median_filter_s100_rx8_ry1                        ... bench:     118,220.14 ns/iter (+/- 1,505.95)
test gradients::benches::bench_sobel_gradients                                        ... bench:   3,942,084.35 ns/iter (+/- 15,059.41)
test hog::benches::bench_hog                                                          ... bench:     304,469.85 ns/iter (+/- 1,022.83)
test integral_image::benches::bench_column_running_sum                                ... bench:         312.00 ns/iter (+/- 3.12)
test integral_image::benches::bench_row_running_sum                                   ... bench:         190.43 ns/iter (+/- 0.97)
test seam_carving::benches::bench_shrink_width_s100_r1                                ... bench:     155,330.17 ns/iter (+/- 389.11)
test seam_carving::benches::bench_shrink_width_s100_r4                                ... bench:     612,569.00 ns/iter (+/- 3,089.73)
test seam_carving::benches::bench_shrink_width_s100_r8                                ... bench:   1,202,706.30 ns/iter (+/- 4,013.73)

@Quickmarble Quickmarble marked this pull request as ready for review April 7, 2026 22:23
@Quickmarble
Copy link
Copy Markdown
Contributor Author

Should be ready now!

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Boundary conditions

1 participant