Coalesce stores in Slice for smaller output types #3568
Merged
Conversation
This file contains 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
This change coalesces stores to global memory in SliceGPU when OutputType is smaller than 4 bytes in order to improve performance. Signed-off-by: Szymon Karpiński <[email protected]>
mzient
reviewed
Dec 14, 2021
dali/kernels/slice/slice_gpu.cuh
Outdated
| @@ -72,6 +72,9 @@ struct SliceBlockDesc { | |||
| uint64_t size; | |||
| }; | |||
|
|
|||
| template<typename OutputType> | |||
| constexpr int coalesced_pixels = sizeof(OutputType) >= 4 ? 1 : 4 / sizeof(OutputType); | |||
There was a problem hiding this comment.
Nitpick: perhaps it should be "values", not "pixels". An RGB pixel consists of 3 values and I don't think this is recognized anyhow.
There was a problem hiding this comment.
You're right, this is a bit confusing. Renamed to "values"
mzient
approved these changes
Dec 14, 2021
jantonguirao
approved these changes
Dec 14, 2021
|
!build |
|
CI MESSAGE: [3588408]: BUILD FAILED |
banasraf
approved these changes
Dec 14, 2021
|
CI MESSAGE: [3588408]: BUILD STARTED |
Signed-off-by: Szymon Karpiński <[email protected]>
|
Clang fails to unroll a loop (the outer loop, I guess). There are two options:
|
Signed-off-by: Szymon Karpiński <[email protected]>
szkarpinski
force-pushed
the
slice-coalesce
branch
from
8fb1a8e to
58ea73d
Compare
|
CI MESSAGE: [3589111]: BUILD STARTED |
|
CI MESSAGE: [3589111]: BUILD PASSED |
23 tasks
cyyever
pushed a commit
to cyyever/DALI
that referenced
this pull request
Jan 23, 2022
* Coalesce stores in Slice for smaller output types This change coalesces stores to global memory in SliceGPU when OutputType is smaller than 4 bytes in order to improve performance. Signed-off-by: Szymon Karpiński <[email protected]>
cyyever
pushed a commit
to cyyever/DALI
that referenced
this pull request
Jan 23, 2022
* Coalesce stores in Slice for smaller output types This change coalesces stores to global memory in SliceGPU when OutputType is smaller than 4 bytes in order to improve performance. Signed-off-by: Szymon Karpiński <[email protected]>
cyyever
pushed a commit
to cyyever/DALI
that referenced
this pull request
Jan 23, 2022
* Coalesce stores in Slice for smaller output types This change coalesces stores to global memory in SliceGPU when OutputType is smaller than 4 bytes in order to improve performance. Signed-off-by: Szymon Karpiński <[email protected]>
cyyever
pushed a commit
to cyyever/DALI
that referenced
this pull request
Jan 23, 2022
* Coalesce stores in Slice for smaller output types This change coalesces stores to global memory in SliceGPU when OutputType is smaller than 4 bytes in order to improve performance. Signed-off-by: Szymon Karpiński <[email protected]>
cyyever
pushed a commit
to cyyever/DALI
that referenced
this pull request
Jan 23, 2022
* Coalesce stores in Slice for smaller output types This change coalesces stores to global memory in SliceGPU when OutputType is smaller than 4 bytes in order to improve performance. Signed-off-by: Szymon Karpiński <[email protected]>
cyyever
pushed a commit
to cyyever/DALI
that referenced
this pull request
Jan 23, 2022
* Coalesce stores in Slice for smaller output types This change coalesces stores to global memory in SliceGPU when OutputType is smaller than 4 bytes in order to improve performance. Signed-off-by: Szymon Karpiński <[email protected]>
cyyever
pushed a commit
to cyyever/DALI
that referenced
this pull request
Jan 23, 2022
* Coalesce stores in Slice for smaller output types This change coalesces stores to global memory in SliceGPU when OutputType is smaller than 4 bytes in order to improve performance. Signed-off-by: Szymon Karpiński <[email protected]>
cyyever
pushed a commit
to cyyever/DALI
that referenced
this pull request
Jan 23, 2022
* Coalesce stores in Slice for smaller output types This change coalesces stores to global memory in SliceGPU when OutputType is smaller than 4 bytes in order to improve performance. Signed-off-by: Szymon Karpiński <[email protected]>
cyyever
pushed a commit
to cyyever/DALI
that referenced
this pull request
Jan 23, 2022
* Coalesce stores in Slice for smaller output types This change coalesces stores to global memory in SliceGPU when OutputType is smaller than 4 bytes in order to improve performance. Signed-off-by: Szymon Karpiński <[email protected]>
cyyever
pushed a commit
to cyyever/DALI
that referenced
this pull request
Jan 23, 2022
* Coalesce stores in Slice for smaller output types This change coalesces stores to global memory in SliceGPU when OutputType is smaller than 4 bytes in order to improve performance. Signed-off-by: Szymon Karpiński <[email protected]>
cyyever
pushed a commit
to cyyever/DALI
that referenced
this pull request
Jan 23, 2022
* Coalesce stores in Slice for smaller output types This change coalesces stores to global memory in SliceGPU when OutputType is smaller than 4 bytes in order to improve performance. Signed-off-by: Szymon Karpiński <[email protected]>
cyyever
pushed a commit
to cyyever/DALI
that referenced
this pull request
Jan 23, 2022
* Coalesce stores in Slice for smaller output types This change coalesces stores to global memory in SliceGPU when OutputType is smaller than 4 bytes in order to improve performance. Signed-off-by: Szymon Karpiński <[email protected]>
cyyever
pushed a commit
to cyyever/DALI
that referenced
this pull request
Jan 23, 2022
* Coalesce stores in Slice for smaller output types This change coalesces stores to global memory in SliceGPU when OutputType is smaller than 4 bytes in order to improve performance. Signed-off-by: Szymon Karpiński <[email protected]>
cyyever
pushed a commit
to cyyever/DALI
that referenced
this pull request
Feb 21, 2022
* Coalesce stores in Slice for smaller output types This change coalesces stores to global memory in SliceGPU when OutputType is smaller than 4 bytes in order to improve performance. Signed-off-by: Szymon Karpiński <[email protected]>
Closed
cyyever
pushed a commit
to cyyever/DALI
that referenced
this pull request
May 13, 2022
* Coalesce stores in Slice for smaller output types This change coalesces stores to global memory in SliceGPU when OutputType is smaller than 4 bytes in order to improve performance. Signed-off-by: Szymon Karpiński <[email protected]>
cyyever
pushed a commit
to cyyever/DALI
that referenced
this pull request
Jun 7, 2022
* Coalesce stores in Slice for smaller output types This change coalesces stores to global memory in SliceGPU when OutputType is smaller than 4 bytes in order to improve performance. Signed-off-by: Szymon Karpiński <[email protected]>
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.
Description
What happened in this PR
Note: The diff presented by GitHub looks terrible and huge. Use
Hide whitespaceoption to see the real changes, and not the indent change ;)Currently, each thread in a block of
SliceGPUkernel is responsible for processing and storing pixels at indices spaced byBlockDim.x. WhenOutputTypeis small, for example isuint8, this results in single-byte stores from each thread.This proposal makes each thread process few subsequent pixels, so that the stores can be easily coalesced to a full 32-bit global memory access.
This works as follows. A number of
OutputTypes fitting in a 32-bit word is computed (in compile time) ascoalesced_pixels. Then, on each step, a thread processesscoalesced_pixelspixels instead of one, and then moves right byBlockDim.x * coalesced_pixelsindices instead ofBlockDim.x. Those stores to subsequent locations (should) get coealesced to a single request by the GPU, thus reducing total number of global memory requests issued.This significantly improves Slice's performance for arrays of
uint8s. The plots below present the throughput improvement in cropping 2000x2000 RGBuint8images to 1000x1000 on Titan V. As you can see, the throughput improvement is up to 40%.Additional information
OutputTypeis smaller than 4 bytes.Checklist
Tests
Documentation
DALI team only
Requirements
REQ IDs: N/A
JIRA TASK: N/A