Change the fb buffer as 16 bytes aligned

[jason-mao]

The 128 bit aligned buffer have good performance for esp_jpeg encoding. The aligned data to improve the memory load time, combined with esp32-S3 128bit SIMD instructions optimization, we got about 10 fps on OV3660 camera with YUV422 and VGA.

[kxbin]

Hello, may I ask if there is test data, how much fps has been improved?

[igrr]

Thank you for the PR @jason-mao!

Have left one note about the usage of heap_caps_aligned_alloc in older releases.

Also would you mind sharing, what is the observed performance improvement?
Perhaps you could add a test case or some other way to benchmark the improvement you have observed?

[dongmenhai]

@jason-mao Hello，How can I test this feature to take effect? I use esp32s3 to run a demo of pic_server, call the frame2jpg function 30 times and calculate the time. By comparison, I found that the method of modifying cam_obj->frames[x].fb.buf to 128 bit alignment did not change the speed of jpeg encoding

[jason-mao]

Hello, may I ask if there is test data, how much fps has been improved?

This modification just to make the 128bit SIMD instructions to load data at theoretical performance, the unaligned data has extra time. This is only the first step. Most of the real improvement time is in code assembly optimization. I think I can add some code on esp-adf to test it.

[jason-mao]

Thank you for the PR @jason-mao!

Have left one note about the usage of heap_caps_aligned_alloc in older releases.

Also would you mind sharing, what is the observed performance improvement? Perhaps you could add a test case or some other way to benchmark the improvement you have observed?

I have update on #419 (comment)

[jason-mao]

@jason-mao Hello，How can I test this feature to take effect? I use esp32s3 to run a demo of pic_server, call the frame2jpg function 30 times and calculate the time. By comparison, I found that the method of modifying cam_obj->frames[x].fb.buf to 128 bit alignment did not change the speed of jpeg encoding

Sorry, maybe I caused some misunderstanding. I explanation here is clearly #419 (comment)

[igrr]

Latest version LGTM.

Given the typical framebuffer sizes, loosing a few bytes of heap due to alignment doesn't seem to be a big issue. So I'm okay with this change even without adding SIMD optimizations to the rest of this project.

@me-no-dev PTAL as well.

[me-no-dev]

Thanks @jason-mao

[weilian1977]

The 128 bit aligned buffer have good performance for esp_jpeg encoding. The aligned data to improve the memory load time, combined with esp32-S3 128bit SIMD instructions optimization, we got about 10 fps on OV3660 camera with YUV422 and VGA.

can provide test code？

[weilian1977]

Hello, may I ask if there is test data, how much fps has been improved?

This modification just to make the 128bit SIMD instructions to load data at theoretical performance, the unaligned data has extra time. This is only the first step. Most of the real improvement time is in code assembly optimization. I think I can add some code on esp-adf to test it.

https://github.com/espressif/esp-adf-libs/blob/master/esp_codec/include/codec/esp_jpeg_version.h？？