[ET-VK][Test] aten.max_pool2d_with_indices #2548
Closed
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
Due to the below issues, we only check equality of the output tensor and not the index tensor. 1. We can't verify index tensors since VK-float16 vs CPU-float32 deltas can change which index in a pool is the maximum. That can yield completely different integers in the index tensor. Hence, we only verify the output tensor not the index tensor. 2. To actually visualize the index tensor, we need to re-construct the int32 values from the int64 values. Since the `torch.int64` index tensor is serialized as `int32` in Vulkan, Python expects int64 but C++ writes to the buffer as though it is for int32. Hence, we must apply some computation to re-construct the tensor. See below for details. A helper function was included in an earlier version of this change, but was removed for conciseness since we aren't checking that index tensor anyway. For example, if the first and second elements return 16 and 17, we get this value as the first element: ``` 73014444048 = 1000100000000000000000000000000010000 ``` We must split this int64 into two int32 values, and construct a new tensor accordingly. ``` 10001 | 00000000000000000000000000010000 10001 | 10000 17 | 16 ``` Differential Revision: [D54962492](https://our.internmc.facebook.com/intern/diff/D54962492/) [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/2548
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit bb26fea with merge base 80e3989 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
CLA Signed
This was referenced Mar 20, 2024
Closed
|
This pull request was exported from Phabricator. Differential Revision: D54962492 |
jorgep31415
added a commit
that referenced
this pull request
Mar 20, 2024
Due to the below issues, we only check equality of the output tensor and not the index tensor. 1. We can't verify index tensors since VK-float16 vs CPU-float32 deltas can change which index in a pool is the maximum. That can yield completely different integers in the index tensor. Hence, we only verify the output tensor not the index tensor. 2. To actually visualize the index tensor, we need to re-construct the int32 values from the int64 values. Since the `torch.int64` index tensor is serialized as `int32` in Vulkan, Python expects int64 but C++ writes to the buffer as though it is for int32. Hence, we must apply some computation to re-construct the tensor. See below for details. A helper function was included in an earlier version of this change, but was removed for conciseness since we aren't checking that index tensor anyway. For example, if the first and second elements return 16 and 17, we get this value as the first element: ``` 73014444048 = 1000100000000000000000000000000010000 ``` We must split this int64 into two int32 values, and construct a new tensor accordingly. ``` 10001 | 00000000000000000000000000010000 10001 | 10000 17 | 16 ``` Differential Revision: [D54962492](https://our.internmc.facebook.com/intern/diff/D54962492/) ghstack-source-id: 219490380 Pull Request resolved: #2548
Due to the below issues, we only check equality of the output tensor and not the index tensor. 1. We can't verify index tensors since VK-float16 vs CPU-float32 deltas can change which index in a pool is the maximum. That can yield completely different integers in the index tensor. Hence, we only verify the output tensor not the index tensor. 2. To actually visualize the index tensor, we need to re-construct the int32 values from the int64 values. Since the `torch.int64` index tensor is serialized as `int32` in Vulkan, Python expects int64 but C++ writes to the buffer as though it is for int32. Hence, we must apply some computation to re-construct the tensor. See below for details. A helper function was included in an earlier version of this change, but was removed for conciseness since we aren't checking that index tensor anyway. For example, if the first and second elements return 16 and 17, we get this value as the first element: ``` 73014444048 = 1000100000000000000000000000000010000 ``` We must split this int64 into two int32 values, and construct a new tensor accordingly. ``` 10001 | 00000000000000000000000000010000 10001 | 10000 17 | 16 ``` Differential Revision: [D54962492](https://our.internmc.facebook.com/intern/diff/D54962492/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D54962492 |
jorgep31415
added a commit
that referenced
this pull request
Mar 20, 2024
Pull Request resolved: #2548 Due to the below issues, we only check equality of the output tensor and not the index tensor. 1. We can't verify index tensors since VK-float16 vs CPU-float32 deltas can change which index in a pool is the maximum. That can yield completely different integers in the index tensor. Hence, we only verify the output tensor not the index tensor. 2. To actually visualize the index tensor, we need to re-construct the int32 values from the int64 values. Since the `torch.int64` index tensor is serialized as `int32` in Vulkan, Python expects int64 but C++ writes to the buffer as though it is for int32. Hence, we must apply some computation to re-construct the tensor. See below for details. A helper function was included in an earlier version of this change, but was removed for conciseness since we aren't checking that index tensor anyway. For example, if the first and second elements return 16 and 17, we get this value as the first element: ``` 73014444048 = 1000100000000000000000000000000010000 ``` We must split this int64 into two int32 values, and construct a new tensor accordingly. ``` 10001 | 00000000000000000000000000010000 10001 | 10000 17 | 16 ``` ghstack-source-id: 219491401 @exported-using-ghexport Differential Revision: [D54962492](https://our.internmc.facebook.com/intern/diff/D54962492/)
SS-JIA
approved these changes
Mar 21, 2024
Due to the below issues, we only check equality of the output tensor and not the index tensor. 1. We can't verify index tensors since VK-float16 vs CPU-float32 deltas can change which index in a pool is the maximum. That can yield completely different integers in the index tensor. Hence, we only verify the output tensor not the index tensor. 2. To actually visualize the index tensor, we need to re-construct the int32 values from the int64 values. Since the `torch.int64` index tensor is serialized as `int32` in Vulkan, Python expects int64 but C++ writes to the buffer as though it is for int32. Hence, we must apply some computation to re-construct the tensor. See below for details. A helper function was included in an earlier version of this change, but was removed for conciseness since we aren't checking that index tensor anyway. For example, if the first and second elements return 16 and 17, we get this value as the first element: ``` 73014444048 = 1000100000000000000000000000000010000 ``` We must split this int64 into two int32 values, and construct a new tensor accordingly. ``` 10001 | 00000000000000000000000000010000 10001 | 10000 17 | 16 ``` Differential Revision: [D54962492](https://our.internmc.facebook.com/intern/diff/D54962492/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D54962492 |
jorgep31415
added a commit
that referenced
this pull request
Mar 21, 2024
Pull Request resolved: #2548 Due to the below issues, we only check equality of the output tensor and not the index tensor. 1. We can't verify index tensors since VK-float16 vs CPU-float32 deltas can change which index in a pool is the maximum. That can yield completely different integers in the index tensor. Hence, we only verify the output tensor not the index tensor. 2. To actually visualize the index tensor, we need to re-construct the int32 values from the int64 values. Since the `torch.int64` index tensor is serialized as `int32` in Vulkan, Python expects int64 but C++ writes to the buffer as though it is for int32. Hence, we must apply some computation to re-construct the tensor. See below for details. A helper function was included in an earlier version of this change, but was removed for conciseness since we aren't checking that index tensor anyway. For example, if the first and second elements return 16 and 17, we get this value as the first element: ``` 73014444048 = 1000100000000000000000000000000010000 ``` We must split this int64 into two int32 values, and construct a new tensor accordingly. ``` 10001 | 00000000000000000000000000010000 10001 | 10000 17 | 16 ``` ghstack-source-id: 219506266 @exported-using-ghexport Differential Revision: [D54962492](https://our.internmc.facebook.com/intern/diff/D54962492/)
|
This pull request has been merged in a41ac1c. |
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.
Stack from ghstack (oldest at bottom):
Due to the below issues, we only check equality of the output tensor and not the index tensor.
torch.int64index tensor is serialized asint32in Vulkan, Python expects int64 but C++ writes to the buffer as though it is for int32. Hence, we must apply some computation to re-construct the tensor. See below for details. A helper function was included in an earlier version of this change, but was removed for conciseness since we aren't checking that index tensor anyway.For example, if the first and second elements return 16 and 17, we get this value as the first element:
We must split this int64 into two int32 values, and construct a new tensor accordingly.
Differential Revision: D54962492