[MPS] Add scatter_reduce.two #141948
Closed
[MPS] Add scatter_reduce.two #141948
Conversation
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
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/141948
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 New FailureAs of commit 1ff4994 with merge base 78543e6 ( NEW FAILURE - The following job has failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
pytorch-bot
bot
added
ciflow/mps
Attention! native_functions.yaml was changedIf you are adding a new function or defaulted argument to native_functions.yaml, you cannot use it from pre-existing Python frontend code until our FC window passes (two weeks). Split your PR into two PRs, one which adds the new C++ functionality, and one that makes use of it from Python, and land them two weeks apart. See https://github.com/pytorch/pytorch/wiki/PyTorch's-Python-Frontend-Backward-and-Forward-Compatibility-Policy#forwards-compatibility-fc for more info. Caused by: |
Which is just a flavor of out-of-box scatter-reduce
malfet
force-pushed
the
malfet/mps-add-scatter-reduce-two
branch
from
b8c9b2d to
c63b817
Compare
manuelcandales
approved these changes
Dec 3, 2024
|
@pytorchbot merge -f "MPS tests + Lint are green" |
Merge startedYour change will be merged immediately since you used the force (-f) flag, bypassing any CI checks (ETA: 1-5 minutes). Please use Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
pobin6
pushed a commit
to pobin6/pytorch
that referenced
this pull request
Dec 5, 2024
Which has been request 20+ times on pytorch#77764 is just a flavor of out-of-box scatter-reduce, so all this op does is redispatches existing implementation. Unsupported dtype/reduction type combinations: - min/max for int64 - min/max for int32 on MacOS-14 or older Following swift code demonstrates problem with scatterAlongAxis MPS call ```swift import Metal import MetalPerformanceShadersGraph func scatterMPS(device: MTLDevice, inp_buf: MTLBuffer, upd_buf: MTLBuffer, idx_buf: MTLBuffer, out_buf: MTLBuffer, inp_elem: Int, upd_elem: Int) { let graph = MPSGraph() let inputPlaceholder = graph.placeholder(shape: [inp_elem as NSNumber], dataType: .int64, name: nil) let updatesPlaceholder = graph.placeholder(shape: [upd_elem as NSNumber], dataType: .int64, name: nil) let indicesPlaceholder = graph.placeholder(shape: [upd_elem as NSNumber], dataType: .int64, name: nil) let outNode = graph.scatterAlongAxis(0, data: inputPlaceholder, updates: updatesPlaceholder, indices: indicesPlaceholder, mode: .min, name: nil) let mpsInputBuffer = MPSGraphTensorData(inp_buf, shape: [inp_elem as NSNumber], dataType: .int64) let mpsUpdatesBuffer = MPSGraphTensorData(upd_buf, shape: [upd_elem as NSNumber], dataType: .int64) let mpsIndicesBuffer = MPSGraphTensorData(idx_buf, shape: [upd_elem as NSNumber], dataType: .int64) let mpsOutputBuffer = MPSGraphTensorData(out_buf, shape: [inp_elem as NSNumber], dataType: .int64) guard let queue = device.makeCommandQueue() else { fatalError("Can't make queue") } graph.run(with: queue, feeds: [inputPlaceholder: mpsInputBuffer, updatesPlaceholder: mpsUpdatesBuffer, indicesPlaceholder: mpsIndicesBuffer ], targetOperations: nil, resultsDictionary: [outNode: mpsOutputBuffer]) } func makeBufferWithValues(device: MTLDevice, values: [Int64]) -> MTLBuffer { guard let buf = device.makeBuffer(length: values.count * MemoryLayout<Int64>.size, options: [.storageModeShared]) else { fatalError("Can't alloc") } let buf_data = buf.contents().assumingMemoryBound(to: Int64.self) for i in 0..<values.count { buf_data[i] = values[i] } return buf } guard let device = MTLCopyAllDevices().first else { fatalError("Not Metal device found") } print("Using device \(device.name)") let inp_elem = 4 let upd_elem = 4 let inp_buf = makeBufferWithValues(device: device, values: [1, 2, 3, 4]) let upd_buf = makeBufferWithValues(device: device, values: [Int64.max - 1, Int64.max - 2 , Int64.max >> 16 , 11]) let idx_buf = makeBufferWithValues(device: device, values: [0, 1, 2, 3]) guard let out_buf = device.makeBuffer(length:inp_elem * MemoryLayout<Int64>.size, options: [.storageModeShared]) else { fatalError("Can't alloc") } scatterMPS(device: device, inp_buf: inp_buf, upd_buf: upd_buf, idx_buf: idx_buf, out_buf: out_buf, inp_elem: inp_elem, upd_elem: upd_elem) let obuf_data = out_buf.contents().assumingMemoryBound(to: Int64.self) for i in 0..<inp_elem { print("out_buf[\(i)] = \(obuf_data[i])") } ``` that prints `4294967294, 4294967293, 4294967295, 4` instead of expected `1, 2, 3, 4` Where `torch.tensor([[1, 9223372036854775806], [2, 9223372036854775805], [3, 140737488355327], [4, 11]], dtype=torch.int64, device='mps').max(1)` yields an expected results Pull Request resolved: pytorch#141948 Approved by: https://github.com/manuelcandales
AmdSampsa
pushed a commit
to AmdSampsa/pytorch
that referenced
this pull request
Dec 9, 2024
Which has been request 20+ times on pytorch#77764 is just a flavor of out-of-box scatter-reduce, so all this op does is redispatches existing implementation. Unsupported dtype/reduction type combinations: - min/max for int64 - min/max for int32 on MacOS-14 or older Following swift code demonstrates problem with scatterAlongAxis MPS call ```swift import Metal import MetalPerformanceShadersGraph func scatterMPS(device: MTLDevice, inp_buf: MTLBuffer, upd_buf: MTLBuffer, idx_buf: MTLBuffer, out_buf: MTLBuffer, inp_elem: Int, upd_elem: Int) { let graph = MPSGraph() let inputPlaceholder = graph.placeholder(shape: [inp_elem as NSNumber], dataType: .int64, name: nil) let updatesPlaceholder = graph.placeholder(shape: [upd_elem as NSNumber], dataType: .int64, name: nil) let indicesPlaceholder = graph.placeholder(shape: [upd_elem as NSNumber], dataType: .int64, name: nil) let outNode = graph.scatterAlongAxis(0, data: inputPlaceholder, updates: updatesPlaceholder, indices: indicesPlaceholder, mode: .min, name: nil) let mpsInputBuffer = MPSGraphTensorData(inp_buf, shape: [inp_elem as NSNumber], dataType: .int64) let mpsUpdatesBuffer = MPSGraphTensorData(upd_buf, shape: [upd_elem as NSNumber], dataType: .int64) let mpsIndicesBuffer = MPSGraphTensorData(idx_buf, shape: [upd_elem as NSNumber], dataType: .int64) let mpsOutputBuffer = MPSGraphTensorData(out_buf, shape: [inp_elem as NSNumber], dataType: .int64) guard let queue = device.makeCommandQueue() else { fatalError("Can't make queue") } graph.run(with: queue, feeds: [inputPlaceholder: mpsInputBuffer, updatesPlaceholder: mpsUpdatesBuffer, indicesPlaceholder: mpsIndicesBuffer ], targetOperations: nil, resultsDictionary: [outNode: mpsOutputBuffer]) } func makeBufferWithValues(device: MTLDevice, values: [Int64]) -> MTLBuffer { guard let buf = device.makeBuffer(length: values.count * MemoryLayout<Int64>.size, options: [.storageModeShared]) else { fatalError("Can't alloc") } let buf_data = buf.contents().assumingMemoryBound(to: Int64.self) for i in 0..<values.count { buf_data[i] = values[i] } return buf } guard let device = MTLCopyAllDevices().first else { fatalError("Not Metal device found") } print("Using device \(device.name)") let inp_elem = 4 let upd_elem = 4 let inp_buf = makeBufferWithValues(device: device, values: [1, 2, 3, 4]) let upd_buf = makeBufferWithValues(device: device, values: [Int64.max - 1, Int64.max - 2 , Int64.max >> 16 , 11]) let idx_buf = makeBufferWithValues(device: device, values: [0, 1, 2, 3]) guard let out_buf = device.makeBuffer(length:inp_elem * MemoryLayout<Int64>.size, options: [.storageModeShared]) else { fatalError("Can't alloc") } scatterMPS(device: device, inp_buf: inp_buf, upd_buf: upd_buf, idx_buf: idx_buf, out_buf: out_buf, inp_elem: inp_elem, upd_elem: upd_elem) let obuf_data = out_buf.contents().assumingMemoryBound(to: Int64.self) for i in 0..<inp_elem { print("out_buf[\(i)] = \(obuf_data[i])") } ``` that prints `4294967294, 4294967293, 4294967295, 4` instead of expected `1, 2, 3, 4` Where `torch.tensor([[1, 9223372036854775806], [2, 9223372036854775805], [3, 140737488355327], [4, 11]], dtype=torch.int64, device='mps').max(1)` yields an expected results Pull Request resolved: pytorch#141948 Approved by: https://github.com/manuelcandales
73 tasks
|
Results running final rc for 2.6 on MacOS 15.1.1: Final rc running on MacOs 14.4: |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
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.
Which has been request 20+ times on #77764 is just a flavor of out-of-box scatter-reduce, so all this op does is redispatches existing implementation.
Unsupported dtype/reduction type combinations:
Following swift code demonstrates problem with scatterAlongAxis MPS call
that prints
4294967294, 4294967293, 4294967295, 4instead of expected1, 2, 3, 4Where
torch.tensor([[1, 9223372036854775806], [2, 9223372036854775805], [3, 140737488355327], [4, 11]], dtype=torch.int64, device='mps').max(1)yields an expected results