warpAffineLinear performance

giacomo_c_ — Tue, 17 Jan 2017 08:24:05 GMT

I'm trying to convert an application built with Ipp 8.0.1 to Ipp 2017.1.
Problem arises with function ippiResizeSqrPixel_16s_C1R removed from current library.
I need precise scale factor and x and y shift control with subpixel accuracy so the substitute function found is warpAffineLinear (are there other options?) with transformation coeffs used for scaling and offset.
I wrote a simple test program (source code attached ) to verify the performance differences between the old and the new api. This program compares warpAffineLinear and ippiResizeSqrPixel_16s_C1R ( IPP 9.0 legacy libraries version ipp90legacy_win_9.0.0.008) used with LINEAR_INTERPOLATION.

Parameters explored are :
- image edge size (square images)
- zoom scale factor
- single thread vs multithread (only for warpAffine)

Below there is the output obtained. Last 2 columns show execution time in ms.

1) Observing only single thread cases (lines 0,1,2, 6,7,8, 12,13,14) warpAffineLinear is a little slower than ippiResizeSqrPixel_16s_C1R for zoom factor greater or equal 1 but is VERY SLOWER for zoom factor less than 1 (lines 0,6,12)

2) Using a multi thread implementation with image tiling (8 thread for 8 tiles) warpAffineLinear perform better for scale factors greater or equal than 1 (lines 4,5, 10,11, 16,17) but still there are performance problems with scale factor less than 1 (line 3, 9, 15). For example for a 1024x1024 image with a scale factor of 0.99 (line 3) warpAffine takes 1.28 ms while for a scale factor of 1.01 (line 5) it takes 0.61 ms.

My real application executes thousand of image resize operation with images of size about 1500x1500.

-----------------------------------------------------------------------------------------------------------------------
line image scale threads warpAffine(ms) ippiResizeSqrPixel_16s_C1R(ms)
0 1024 0.9900 1 3.4699 1.2519
1 1024 1.0000 1 1.3585 1.2664
2 1024 1.0100 1 1.3355 1.2734
3 1024 0.9900 8 1.2801 1.4048
4 1024 1.0000 8 0.5745 1.5947
5 1024 1.0100 8 0.6185 1.2758

6 2048 0.9900 1 17.7146 5.2773
7 2048 1.0000 1 6.2307 5.3990
8 2048 1.0100 1 6.3813 5.3994
9 2048 0.9900 8 5.3796 5.5666
10 2048 1.0000 8 2.7698 5.6583
11 2048 1.0100 8 2.6872 5.3551

12 4096 0.9900 1 57.0254 19.7586
13 4096 1.0000 1 23.9561 20.2563
14 4096 1.0100 1 23.2165 20.0609
15 4096 0.9900 8 16.9511 20.3904
16 4096 1.0000 8 10.6648 20.2441
17 4096 1.0100 8 10.6935 20.3870

Thank's in advance

Hi Giacomo,

Valentin_K_Intel — Thu, 19 Jan 2017 10:17:51 GMT

Hi Giacomo,

Thank you for the information. We will investigate it.

Best regards,
Valentin

topic Hi Giacomo, in Intel® Integrated Performance Primitives

warpAffineLinear performance

Hi Giacomo,