[Texture support][Part 2] Add opencl adreno target, topi schedules, and relay op strategies #7687
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
|
Need to add FP32 support. Currently if we do not convert convolutions to FP16 and try to compile it for Adreno, we get the follow error: |
schedule utilizing NCHW4c layout and texture memory.
topi impl. for conv2d.
…provided texture buffers. Need to propagate this to allocated textures when cache_read(texture) is used for weights.
graph runtime behavior of providing external texture buffers.
jroesch
added
needs-triage
|
Closing as duplicated by #11161. |
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Outdated suggestions cannot be applied.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.
This PR introduces the
opencl --device=adrenotarget and corresponding relay strategies. The conv2d schedules introduced here utilize spatial and channel-wise packing for the weights (OIHW4o) and activations (NCHW4c), respectively, both with vector length 4 to support lowering to RGBA texture memory.AutoTVM support
AutoTVM doesn't currently capture the runtime context of extracted tasks. Without capturing information about the runtime buffer scopes, the codegeneration during tuning will only occur on flat buffers (not texture). For now, we utilize a cache_read("texture") stage when tuning to explore the performance benefit of utilizing texture memory. We believe that with a sufficient number of iteration's per trial, the copy time to texture prior to running the main compute kernel (which results from a cache_read) should be constant over the search and therefore not greatly impact the tuning results.
Note that the cache_read is not needed when using the graph_runtime which supports passing in external texture buffers (see: [Texture support][Part 3] Support storage scope tag in graph runtime codegen, planning, runtime and compile engine #7688). Therefore, in the schedules one will observe
if autotvm.GLOBAL_SCOPE.in_tuning:which around scheduling related to adding a cache_read to texture stage.The schedules can be simplified once either 1) AutoTVM tuning supports capturing this runtime information during task extraction (thereby removing the need for a cache_read to texture) or 2) once texture lowering in tir.TextureFlatten fully supports cache_read cancellation (external buffer forwarding through the cache_read). Cancellation currently is supported except for the cases of padding wherein an extra texture to texture copy results.
See RFC here: https://discuss.tvm.apache.org/t/rfc-texture-memory-support/9467