[Impeller] Create reusable prefix sum. #42167
Conversation
| reinterpret_cast<CS::OutputData<kCount>*>(view.contents); | ||
| EXPECT_TRUE(output); | ||
|
|
||
| constexpr uint32_t expected[kCount + 1] = {0, 1, 3, 6, 10, 15}; |
There was a problem hiding this comment.
this should really start at 1, right? or should I be expected to shift the sequence in the shader
| // Parallel prefix sum computes the prefix in place in storage. | ||
| // BLOCK_SIZE is the overall storage size while ident must be the global | ||
| // x identifier. | ||
| #define PrefixSum(ident, storage, BLOCK_SIZE) \ |
There was a problem hiding this comment.
This is a simple and not particularly fast implementation, but its good enough to start with and could be swapped for a more efficient implementation later.
There was a problem hiding this comment.
This won't be as efficient as the hardware implementations (e.g. subgroupInclusive/ExclusiveAdd).
We also should define whether this is exclusive or inclusive.
There was a problem hiding this comment.
I was thinking we could if/def or function constant our way into more efficient implementations. I have more reading to do though. Where should I look into the subgroupInclusive/ExclusiveAdd?
|
Ahh the issue was that this was an exclusive sum and I need to convert that to an inclusive sum. Fixed |
| // Parallel exclusive prefix sum computes the prefix in place in storage. | ||
| // BLOCK_SIZE is the overall storage size while ident must be the global | ||
| // x identifier. | ||
| #define ExclusivePrefixSum(ident, storage, BLOCK_SIZE) \ |
There was a problem hiding this comment.
Why a macro and not a function?
There was a problem hiding this comment.
Needs to work over threadgroup shared memory which can't be passed as a ptr in glsl as far as I can tell
There was a problem hiding this comment.
Right, GLSL doesn't have pointers, you can't pass shared memory to a function defined in another file (it will try to do a copy)
chinmaygarde
changed the title
jonahwilliams
added
the
autosubmit
flutter/engine@1ed9fc0...a342a91 2023-05-22 skia-flutter-autoroll@skia.org Roll Skia from 2612bb159848 to d448fe07ea46 (5 revisions) (flutter/engine#42213) 2023-05-22 aam@google.com Skip and ignore boringssl/src/rust when looking for the licenses. (flutter/engine#42210) 2023-05-22 jonahwilliams@google.com [Impeller] Create reusable prefix sum. (flutter/engine#42167) If this roll has caused a breakage, revert this CL and stop the roller using the controls here: https://autoroll.skia.org/r/flutter-engine-flutter-autoroll Please CC rmistry@google.com,zra@google.com on the revert to ensure that a human is aware of the problem. To file a bug in Flutter: https://github.com/flutter/flutter/issues/new/choose To report a problem with the AutoRoller itself, please file a bug: https://bugs.chromium.org/p/skia/issues/entry?template=Autoroller+Bug Documentation for the AutoRoller is here: https://skia.googlesource.com/buildbot/+doc/main/autoroll/README.md
…7352) flutter/engine@1ed9fc0...a342a91 2023-05-22 skia-flutter-autoroll@skia.org Roll Skia from 2612bb159848 to d448fe07ea46 (5 revisions) (flutter/engine#42213) 2023-05-22 aam@google.com Skip and ignore boringssl/src/rust when looking for the licenses. (flutter/engine#42210) 2023-05-22 jonahwilliams@google.com [Impeller] Create reusable prefix sum. (flutter/engine#42167) If this roll has caused a breakage, revert this CL and stop the roller using the controls here: https://autoroll.skia.org/r/flutter-engine-flutter-autoroll Please CC rmistry@google.com,zra@google.com on the revert to ensure that a human is aware of the problem. To file a bug in Flutter: https://github.com/flutter/flutter/issues/new/choose To report a problem with the AutoRoller itself, please file a bug: https://bugs.chromium.org/p/skia/issues/entry?template=Autoroller+Bug Documentation for the AutoRoller is here: https://skia.googlesource.com/buildbot/+doc/main/autoroll/README.md
3 participants
Creates a reusable function macro for performing a prefix sum. Eventually, we'll need this for polyline decomposition
Reasonably fast with max (1024) input elements.