Repository navigation
Conversation
Summary: `WorkWave` now lists its codec runs in plan order, and `WorkGroup` is gone. `Device::chooseDecodeBatches` groups the runs of a wave by kernel, and each kernel serves one transform, so the device decides every grouping. The schedule decides only which runs can start together, and it no longer reads transforms. Reviewed By: Cyan4973 Differential Revision: D128407138
Summary: Adds `bindDecode`, which gives every planned stream and codec run its device address, so the executor can hand each batch to `Stream::decodeBatch`. - A `Stored` stream uses its address in the frame, read from the chunk's prepared DCtx. - A `Destination` or `StreamArena` stream uses its arena's base address plus its offset in the layout. Binding gets those addresses from `ArenaAddresses`, so it does not read the placement format. - A `StreamRef` uses its backing stream's address plus its offset. The planner gives a backing stream a lower index than its aliases, so one pass in stream order resolves a chain of aliases. - Each final `Stored` or `StreamRef` output gets a terminal copy into the caller's output, in chunk order and then output order. - String streams return `temporaryLibraryLimitation` for now. `buildDecodeInfo` rejects them, so a decode fails before any allocation. `bindDecode` binds the runs of each batch from `planDecodeBatches`, in the batch's order, and gives each batch a `BoundBatch` with its kernel and its runs. A run that no batch holds, such as an aliasing run, gets no `BoundCodecRun`. `BoundDecode` holds the spans and the arrays they point into, so it is move-only. `viewOf(PreparedGPUChunk&)` builds the planner's view of a chunk. `decompressChunks` and the new `testkit::StagedHostFrame` both use it, so the tests stage frames exactly as production does, including each chunk's `frameIndex`. Reviewed By: Cyan4973 Differential Revision: D128343911
|
@mmandina has exported this pull request. If you are a Meta employee, you can view the originating Diff in D128343911. |
|
This pull request has been merged in 516c546. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary:
Adds
bindDecode, which gives every planned stream and codec run its device address, so the executor can hand each batch toStream::decodeBatch.Storedstream uses its address in the frame, read from the chunk's prepared DCtx.DestinationorStreamArenastream uses its arena's base address plus its offset in the layout. Binding gets those addresses fromArenaAddresses, so it does not read the placement format.StreamRefuses its backing stream's address plus its offset. The planner gives a backing stream a lower index than its aliases, so one pass in stream order resolves a chain of aliases.StoredorStreamRefoutput gets a terminal copy into the caller's output, in chunk order and then output order.temporaryLibraryLimitationfor now.buildDecodeInforejects them, so a decode fails before any allocation.bindDecodebinds the runs of each batch fromplanDecodeBatches, in the batch's order, and gives each batch aBoundBatchwith its kernel and its runs. A run that no batch holds, such as an aliasing run, gets noBoundCodecRun.BoundDecodeholds the spans and the arrays they point into, so it is move-only.viewOf(PreparedGPUChunk&)builds the planner's view of a chunk.decompressChunksand the newtestkit::StagedHostFrameboth use it, so the tests stage frames exactly as production does, including each chunk'sframeIndex.Reviewed By: Cyan4973
Differential Revision: D128343911