Bug: 220192608

Clone this repo:
  1. 3ed0ddb Android.bp: neon has been on by default for years now. by Elliott Hughes · 11 days ago main master
  2. d1f5247 Merge Android 14 QPR2 to AOSP main by Xin Li · 6 weeks ago
  3. eaaba96 Merge Android 24Q1 Release (ab/11220357) by Xin Li · 3 months ago
  4. 33a0ebe Track Bitmap#getConfig() nullability fix by John Reck · 8 months ago
  5. 7924321 Change OWNERS. am: 9f4fcb4b67 am: 4b6ce2a801 am: 1847e71bf7 am: a3b9e4b9ab am: 585c0bdf4f by Xusong Wang · 1 year, 9 months ago android14-dev android14-mainline-healthfitness-release android14-qpr1-release android14-qpr1-s2-release aml_cfg_341510000 aml_hef_341114030 aml_hef_341311010 aml_hef_341415040 aml_hef_341512030 aml_rkp_341012000 aml_rkp_341015010 aml_rkp_341114000 aml_rkp_341311000 aml_rkp_341510000 android-14.0.0_r16 android-14.0.0_r17 android-14.0.0_r18 android-14.0.0_r19 android-14.0.0_r20 android-14.0.0_r21 android-14.0.0_r22 android-14.0.0_r23 android-14.0.0_r24 android-14.0.0_r25 android-14.0.0_r26 android-14.0.0_r27

RenderScript Intrinsics Replacement Toolkit - v0.8 BETA

This Toolkit provides a collection of high-performance image manipulation functions like blur, blend, and resize. It can be used as a stand-alone replacement for most of the deprecated RenderScript Intrinsics functions.

The Toolkit provides ten image manipulation functions:

  • blend,
  • blur,
  • color matrix,
  • convolve,
  • histogram and histogramDot,
  • LUT (lookup table) and LUT 3D,
  • resize, and
  • YUV to RGB.

The Toolkit provides a C++ and a Java/Kotlin interface. It is packaged as an Android library that you can add to your project.

These functions execute multithreaded on the CPU. They take advantage of Neon/AdvSimd on Arm processors and SSE on Intel's.

Compared to the RenderScript Intrinsics, this Toolkit is simpler to use and twice as fast when executing on the CPU. However RenderScript Intrinsics allow more flexibility for the type of allocations supported. This toolkit does not support allocations of floats; most the functions support ByteArrays and Bitmaps.

You should instantiate the Toolkit once and reuse it throughout your application. On instantiation, the Toolkit creates a thread pool that's used for processing all the functions. You can limit the number of poolThreads used by the Toolkit via the constructor. The poolThreads are destroyed once the Toolkit is destroyed, after any pending work is done.

This library is thread safe. You can call methods from different poolThreads. The functions will execute sequentially.

Future improvement ideas:

  • Turn the Java version of the Toolkit into a singleton, to reduce the chance that someone inadventarly create multiple threadpools.

  • Support ByteBuffer. It should be straightforward to use GetDirectBufferAddress in JniEntryPoints.cpp. See https://developer.android.com/training/articles/perf-jni and jni_helper.h.

  • The RenderScript Intrinsics support floats for colorMatrix, convolve, and resize. The Toolkit does not.

  • Allow in place update of buffers, or writing to an existing byte array.

  • For Blur, we could have a version that accepts a mask. This is commonly used for background blurring. We should allow the mask to be smaller than the original, since neural networks models that do segmentation are downscaled.

  • Allow yuvToRgb to have a Restriction.

  • Add support for YUV_420_888, the YUV format favored by Camera2. Allow various strides to be specified.

  • When passing a Restriction, it would be nice to say “Create a smaller output”. The original RenderScript does not allow that. It's not that useful when outputing new buffers as our Java library does.

  • For Resize, Restriction working on input buffer would be more useful but that's not RenderScript.

  • Integrate and test with imageprocessing_jb. Do the same for github/renderscript-samples/

  • Allow Bitmaps with rowSize != width * vectorSize. We could do this also for ByteArray.

  • In TaskProcessor.cpp, the code below is fine and clean, but probably a bit inefficient. When this wakes up another thread, it may have to immediately go back to sleep, since we still hold the lock. It could instead set a need_to_notify flag and test that after releasing the lock (both places). That might avoid some context switches.
if (mTilesInProcess == 0 && mTilesNotYetStarted == 0) {
  • When compiled as part of Android, librenderscript_toolkit.so is 101,456 bytes. When compiled by Android Studio as part of an .aar, it's 387K. Figure out why and slim it down.