Add batch_index_select_dim0 (w/ TBE backend) #1897
Closed
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
✅ Deploy Preview for pytorch-fbgemm-docs canceled.
|
|
This pull request was exported from Phabricator. Differential Revision: D46084590 |
sryap
added a commit
to sryap/FBGEMM
that referenced
this pull request
Jul 27, 2023
Summary: Pull Request resolved: pytorch#1897 Usage: ``` # This target might change in the future torch.ops.load_library("//deeplearning/fbgemm/fbgemm_gpu/codegen:index_select_ops") ... output = torch.ops.fbgemm.batch_index_select_dim0( inputs, # Tensor - 1D tensor (concatenated flatten inputs) indices, # Tensor - 1D tensor (concatenated indices) input_num_indices, # List[int] input_rows, # List[int] input_columns, # List[int] ) ``` Differential Revision: D46084590 fbshipit-source-id: a5ec8c2d45ae39d5eb79b61a8263e112276de50f
|
This pull request was exported from Phabricator. Differential Revision: D46084590 |
|
This pull request was exported from Phabricator. Differential Revision: D46084590 |
sryap
added a commit
to sryap/FBGEMM
that referenced
this pull request
Jul 28, 2023
Summary: Pull Request resolved: pytorch#1897 This diff introduces `batch_index_select_dim0` using the `SplitTBE` implementation (it shares the same code generator as TBE). The new operator is designed to address limitations of `group_index_select_dim0`. Both operators are designed to operate multiple inputs. However, `batch_index_select_dim0` requires all inputs to be contiguous in memory, while `batch_index_select_dim0` can operate on inputs with a discrete memory layout. Implementation-wise, they are different. We plan to merge their backends in the future. Since `batch_index_select_dim0` is backed by TBE, it inherits TBE limitations including: - The column sizes must be a multiple of 4 and not exceed 1024. Moreover, the underlying buffer of the inputs tensor must be 16-byte aligned. This is because the TBE kernel uses a vector load/store which requires the buffer to be 16-byte aligned. The kernel will raise an error if this assumption is violated. - Due to the 16-byte aligned enforcement, during the backward pass, if the output gradient is not 16-byte aligned, the operator will copy the output gradient into a new 16-byte aligned buffer. This can be expensive if the output gradient size is large. Usage: ``` # This target might change in the future torch.ops.load_library("//deeplearning/fbgemm/fbgemm_gpu/codegen:index_select_ops") ... output = torch.ops.fbgemm.batch_index_select_dim0( inputs, # Tensor - 1D tensor (concatenated flatten inputs) indices, # Tensor - 1D tensor (concatenated indices) input_num_indices, # List[int] input_rows, # List[int] input_columns, # List[int] ) ``` Differential Revision: D46084590 fbshipit-source-id: 96fb152d0270e2d09127fbaab349b5ac02068bcb
|
This pull request was exported from Phabricator. Differential Revision: D46084590 |
sryap
added a commit
to sryap/FBGEMM
that referenced
this pull request
Jul 28, 2023
Summary: Pull Request resolved: pytorch#1897 This diff introduces `batch_index_select_dim0` using the `SplitTBE` implementation (it shares the same code generator as TBE). The new operator is designed to address limitations of `group_index_select_dim0`. Both operators are designed to operate multiple inputs. However, `batch_index_select_dim0` requires all inputs to be contiguous in memory, while `batch_index_select_dim0` can operate on inputs with a discrete memory layout. Implementation-wise, they are different. We plan to merge their backends in the future. Since `batch_index_select_dim0` is backed by TBE, it inherits TBE limitations including: - The column sizes must be a multiple of 4 and not exceed 1024. Moreover, the underlying buffer of the inputs tensor must be 16-byte aligned. This is because the TBE kernel uses a vector load/store which requires the buffer to be 16-byte aligned. The kernel will raise an error if this assumption is violated. - Due to the 16-byte aligned enforcement, during the backward pass, if the output gradient is not 16-byte aligned, the operator will copy the output gradient into a new 16-byte aligned buffer. This can be expensive if the output gradient size is large. Usage: ``` # This target might change in the future torch.ops.load_library("//deeplearning/fbgemm/fbgemm_gpu/codegen:index_select_ops") ... output = torch.ops.fbgemm.batch_index_select_dim0( inputs, # Tensor - 1D tensor (concatenated flatten inputs) indices, # Tensor - 1D tensor (concatenated indices) input_num_indices, # List[int] input_rows, # List[int] input_columns, # List[int] ) ``` Differential Revision: D46084590 fbshipit-source-id: 69df7e36784e77bc4d06cec2e9aba1fb59587e42
|
This pull request was exported from Phabricator. Differential Revision: D46084590 |
sryap
added a commit
to sryap/FBGEMM
that referenced
this pull request
Jul 28, 2023
Summary: Pull Request resolved: pytorch#1897 This diff introduces `batch_index_select_dim0` using the `SplitTBE` implementation (it shares the same code generator as TBE). The new operator is designed to address limitations of `group_index_select_dim0`. Both operators are designed to operate multiple inputs. However, `batch_index_select_dim0` requires all inputs to be contiguous in memory, while `batch_index_select_dim0` can operate on inputs with a discrete memory layout. Implementation-wise, they are different. We plan to merge their backends in the future. Since `batch_index_select_dim0` is backed by TBE, it inherits TBE limitations including: - The column sizes must be a multiple of 4 and not exceed 1024. Moreover, the underlying buffer of the inputs tensor must be 16-byte aligned. This is because the TBE kernel uses a vector load/store which requires the buffer to be 16-byte aligned. The kernel will raise an error if this assumption is violated. - Due to the 16-byte aligned enforcement, during the backward pass, if the output gradient is not 16-byte aligned, the operator will copy the output gradient into a new 16-byte aligned buffer. This can be expensive if the output gradient size is large. Usage: ``` # This target might change in the future torch.ops.load_library("//deeplearning/fbgemm/fbgemm_gpu/codegen:index_select_ops") ... output = torch.ops.fbgemm.batch_index_select_dim0( inputs, # Tensor - 1D tensor (concatenated flatten inputs) indices, # Tensor - 1D tensor (concatenated indices) input_num_indices, # List[int] input_rows, # List[int] input_columns, # List[int] ) ``` Differential Revision: D46084590 fbshipit-source-id: 1b8a94d7eb886337c8764751e262df7016cbf7dc
|
This pull request was exported from Phabricator. Differential Revision: D46084590 |
sryap
added a commit
to sryap/FBGEMM
that referenced
this pull request
Jul 28, 2023
Summary: Pull Request resolved: pytorch#1897 This diff introduces `batch_index_select_dim0` using the `SplitTBE` implementation (it shares the same code generator as TBE). The new operator is designed to address limitations of `group_index_select_dim0`. Both operators are designed to operate multiple inputs. However, `batch_index_select_dim0` requires all inputs to be contiguous in memory, while `batch_index_select_dim0` can operate on inputs with a discrete memory layout. Implementation-wise, they are different. We plan to merge their backends in the future. Since `batch_index_select_dim0` is backed by TBE, it inherits TBE limitations including: - The column sizes must be a multiple of 4 and not exceed 1024. Moreover, the underlying buffer of the inputs tensor must be 16-byte aligned. This is because the TBE kernel uses a vector load/store which requires the buffer to be 16-byte aligned. The kernel will raise an error if this assumption is violated. - Due to the 16-byte aligned enforcement, during the backward pass, if the output gradient is not 16-byte aligned, the operator will copy the output gradient into a new 16-byte aligned buffer. This can be expensive if the output gradient size is large. Usage: ``` # This target might change in the future torch.ops.load_library("//deeplearning/fbgemm/fbgemm_gpu/codegen:index_select_ops") ... output = torch.ops.fbgemm.batch_index_select_dim0( inputs, # Tensor - 1D tensor (concatenated flatten inputs) indices, # Tensor - 1D tensor (concatenated indices) input_num_indices, # List[int] input_rows, # List[int] input_columns, # List[int] ) ``` Differential Revision: D46084590 fbshipit-source-id: 190a7bc19a205837da91f268b078a34f145c3273
|
This pull request was exported from Phabricator. Differential Revision: D46084590 |
sryap
added a commit
to sryap/FBGEMM
that referenced
this pull request
Jul 28, 2023
Summary: Pull Request resolved: pytorch#1897 This diff introduces `batch_index_select_dim0` using the `SplitTBE` implementation (it shares the same code generator as TBE). The new operator is designed to address limitations of `group_index_select_dim0`. Both operators are designed to operate multiple inputs. However, `batch_index_select_dim0` requires all inputs to be contiguous in memory, while `batch_index_select_dim0` can operate on inputs with a discrete memory layout. Implementation-wise, they are different. We plan to merge their backends in the future. Since `batch_index_select_dim0` is backed by TBE, it inherits TBE limitations including: - The column sizes must be a multiple of 4 and not exceed 1024. Moreover, the underlying buffer of the inputs tensor must be 16-byte aligned. This is because the TBE kernel uses a vector load/store which requires the buffer to be 16-byte aligned. The kernel will raise an error if this assumption is violated. - Due to the 16-byte aligned enforcement, during the backward pass, if the output gradient is not 16-byte aligned, the operator will copy the output gradient into a new 16-byte aligned buffer. This can be expensive if the output gradient size is large. Usage: ``` # This target might change in the future torch.ops.load_library("//deeplearning/fbgemm/fbgemm_gpu/codegen:index_select_ops") ... output = torch.ops.fbgemm.batch_index_select_dim0( inputs, # Tensor - 1D tensor (concatenated flatten inputs) indices, # Tensor - 1D tensor (concatenated indices) input_num_indices, # List[int] input_rows, # List[int] input_columns, # List[int] ) ``` Differential Revision: D46084590 fbshipit-source-id: ef02e43cff10311b29bff3d351839ac9fde13ddf
|
This pull request was exported from Phabricator. Differential Revision: D46084590 |
sryap
added a commit
to sryap/FBGEMM
that referenced
this pull request
Jul 28, 2023
Summary: Pull Request resolved: pytorch#1897 This diff introduces `batch_index_select_dim0` using the `SplitTBE` implementation (it shares the same code generator as TBE). The new operator is designed to address limitations of `group_index_select_dim0`. Both operators are designed to operate multiple inputs. However, `batch_index_select_dim0` requires all inputs to be contiguous in memory, while `batch_index_select_dim0` can operate on inputs with a discrete memory layout. Implementation-wise, they are different. We plan to merge their backends in the future. Since `batch_index_select_dim0` is backed by TBE, it inherits TBE limitations including: - The column sizes must be a multiple of 4 and not exceed 1024. Moreover, the underlying buffer of the inputs tensor must be 16-byte aligned. This is because the TBE kernel uses a vector load/store which requires the buffer to be 16-byte aligned. The kernel will raise an error if this assumption is violated. - Due to the 16-byte aligned enforcement, during the backward pass, if the output gradient is not 16-byte aligned, the operator will copy the output gradient into a new 16-byte aligned buffer. This can be expensive if the output gradient size is large. Usage: ``` # This target might change in the future torch.ops.load_library("//deeplearning/fbgemm/fbgemm_gpu/codegen:index_select_ops") ... output = torch.ops.fbgemm.batch_index_select_dim0( inputs, # Tensor - 1D tensor (concatenated flatten inputs) indices, # Tensor - 1D tensor (concatenated indices) input_num_indices, # List[int] input_rows, # List[int] input_columns, # List[int] ) ``` Differential Revision: D46084590 fbshipit-source-id: 3c36b52615176b8941f24682e993f63c564942d9
|
This pull request was exported from Phabricator. Differential Revision: D46084590 |
sryap
added a commit
to sryap/FBGEMM
that referenced
this pull request
Jul 28, 2023
Summary: Pull Request resolved: pytorch#1897 This diff introduces `batch_index_select_dim0` using the `SplitTBE` implementation (it shares the same code generator as TBE). The new operator is designed to address limitations of `group_index_select_dim0`. Both operators are designed to operate multiple inputs. However, `batch_index_select_dim0` requires all inputs to be contiguous in memory, while `batch_index_select_dim0` can operate on inputs with a discrete memory layout. Implementation-wise, they are different. We plan to merge their backends in the future. Since `batch_index_select_dim0` is backed by TBE, it inherits TBE limitations including: - The column sizes must be a multiple of 4 and not exceed 1024. Moreover, the underlying buffer of the inputs tensor must be 16-byte aligned. This is because the TBE kernel uses a vector load/store which requires the buffer to be 16-byte aligned. The kernel will raise an error if this assumption is violated. - Due to the 16-byte aligned enforcement, during the backward pass, if the output gradient is not 16-byte aligned, the operator will copy the output gradient into a new 16-byte aligned buffer. This can be expensive if the output gradient size is large. Usage: ``` # This target might change in the future torch.ops.load_library("//deeplearning/fbgemm/fbgemm_gpu/codegen:index_select_ops") ... output = torch.ops.fbgemm.batch_index_select_dim0( inputs, # Tensor - 1D tensor (concatenated flatten inputs) indices, # Tensor - 1D tensor (concatenated indices) input_num_indices, # List[int] input_rows, # List[int] input_columns, # List[int] ) ``` Differential Revision: D46084590 fbshipit-source-id: 59f99f5c2bc5c5424205bd668a6c7777ecf53f7b
|
This pull request was exported from Phabricator. Differential Revision: D46084590 |
sryap
added a commit
to sryap/FBGEMM
that referenced
this pull request
Jul 28, 2023
Summary: Pull Request resolved: pytorch#1897 This diff introduces `batch_index_select_dim0` using the `SplitTBE` implementation (it shares the same code generator as TBE). The new operator is designed to address limitations of `group_index_select_dim0`. Both operators are designed to operate multiple inputs. However, `batch_index_select_dim0` requires all inputs to be contiguous in memory, while `batch_index_select_dim0` can operate on inputs with a discrete memory layout. Implementation-wise, they are different. We plan to merge their backends in the future. Since `batch_index_select_dim0` is backed by TBE, it inherits TBE limitations including: - The column sizes must be a multiple of 4 and not exceed 1024. Moreover, the underlying buffer of the inputs tensor must be 16-byte aligned. This is because the TBE kernel uses a vector load/store which requires the buffer to be 16-byte aligned. The kernel will raise an error if this assumption is violated. - Due to the 16-byte aligned enforcement, during the backward pass, if the output gradient is not 16-byte aligned, the operator will copy the output gradient into a new 16-byte aligned buffer. This can be expensive if the output gradient size is large. Usage: ``` # This target might change in the future torch.ops.load_library("//deeplearning/fbgemm/fbgemm_gpu/codegen:index_select_ops") ... output = torch.ops.fbgemm.batch_index_select_dim0( inputs, # Tensor - 1D tensor (concatenated flatten inputs) indices, # Tensor - 1D tensor (concatenated indices) input_num_indices, # List[int] input_rows, # List[int] input_columns, # List[int] ) ``` Differential Revision: D46084590 fbshipit-source-id: 6225ae90362ca478b6c3120febb1dcf93291a3d6
|
This pull request was exported from Phabricator. Differential Revision: D46084590 |
sryap
added a commit
to sryap/FBGEMM
that referenced
this pull request
Jul 28, 2023
Summary: Pull Request resolved: pytorch#1897 This diff introduces `batch_index_select_dim0` using the `SplitTBE` implementation (it shares the same code generator as TBE). The new operator is designed to address limitations of `group_index_select_dim0`. Both operators are designed to operate multiple inputs. However, `batch_index_select_dim0` requires all inputs to be contiguous in memory, while `batch_index_select_dim0` can operate on inputs with a discrete memory layout. Implementation-wise, they are different. We plan to merge their backends in the future. Since `batch_index_select_dim0` is backed by TBE, it inherits TBE limitations including: - The column sizes must be a multiple of 4 and not exceed 1024. Moreover, the underlying buffer of the inputs tensor must be 16-byte aligned. This is because the TBE kernel uses a vector load/store which requires the buffer to be 16-byte aligned. The kernel will raise an error if this assumption is violated. - Due to the 16-byte aligned enforcement, during the backward pass, if the output gradient is not 16-byte aligned, the operator will copy the output gradient into a new 16-byte aligned buffer. This can be expensive if the output gradient size is large. Usage: ``` # This target might change in the future torch.ops.load_library("//deeplearning/fbgemm/fbgemm_gpu/codegen:index_select_ops") ... output = torch.ops.fbgemm.batch_index_select_dim0( inputs, # Tensor - 1D tensor (concatenated flatten inputs) indices, # Tensor - 1D tensor (concatenated indices) input_num_indices, # List[int] input_rows, # List[int] input_columns, # List[int] ) ``` Differential Revision: D46084590 fbshipit-source-id: 0d2e55c7150cf678a5feb1569beb4ab448565916
|
This pull request was exported from Phabricator. Differential Revision: D46084590 |
sryap
added a commit
to sryap/FBGEMM
that referenced
this pull request
Jul 28, 2023
Summary: Pull Request resolved: pytorch#1897 This diff introduces `batch_index_select_dim0` using the `SplitTBE` implementation (it shares the same code generator as TBE). The new operator is designed to address limitations of `group_index_select_dim0`. Both operators are designed to operate multiple inputs. However, `batch_index_select_dim0` requires all inputs to be contiguous in memory, while `batch_index_select_dim0` can operate on inputs with a discrete memory layout. Implementation-wise, they are different. We plan to merge their backends in the future. Since `batch_index_select_dim0` is backed by TBE, it inherits TBE limitations including: - The column sizes must be a multiple of 4 and not exceed 1024. Moreover, the underlying buffer of the inputs tensor must be 16-byte aligned. This is because the TBE kernel uses a vector load/store which requires the buffer to be 16-byte aligned. The kernel will raise an error if this assumption is violated. - Due to the 16-byte aligned enforcement, during the backward pass, if the output gradient is not 16-byte aligned, the operator will copy the output gradient into a new 16-byte aligned buffer. This can be expensive if the output gradient size is large. Usage: ``` # This target might change in the future torch.ops.load_library("//deeplearning/fbgemm/fbgemm_gpu/codegen:index_select_ops") ... output = torch.ops.fbgemm.batch_index_select_dim0( inputs, # Tensor - 1D tensor (concatenated flatten inputs) indices, # Tensor - 1D tensor (concatenated indices) input_num_indices, # List[int] input_rows, # List[int] input_columns, # List[int] ) ``` Differential Revision: D46084590 fbshipit-source-id: 6dd9d2bc6cf86832b301a855337a28fed76d748a
|
This pull request was exported from Phabricator. Differential Revision: D46084590 |
sryap
added a commit
to sryap/FBGEMM
that referenced
this pull request
Jul 28, 2023
Summary: Pull Request resolved: pytorch#1897 This diff introduces `batch_index_select_dim0` using the `SplitTBE` implementation (it shares the same code generator as TBE). The new operator is designed to address limitations of `group_index_select_dim0`. Both operators are designed to operate multiple inputs. However, `batch_index_select_dim0` requires all inputs to be contiguous in memory, while `batch_index_select_dim0` can operate on inputs with a discrete memory layout. Implementation-wise, they are different. We plan to merge their backends in the future. Since `batch_index_select_dim0` is backed by TBE, it inherits TBE limitations including: - The column sizes must be a multiple of 4 and not exceed 1024. Moreover, the underlying buffer of the inputs tensor must be 16-byte aligned. This is because the TBE kernel uses a vector load/store which requires the buffer to be 16-byte aligned. The kernel will raise an error if this assumption is violated. - Due to the 16-byte aligned enforcement, during the backward pass, if the output gradient is not 16-byte aligned, the operator will copy the output gradient into a new 16-byte aligned buffer. This can be expensive if the output gradient size is large. Usage: ``` # This target might change in the future torch.ops.load_library("//deeplearning/fbgemm/fbgemm_gpu/codegen:index_select_ops") ... output = torch.ops.fbgemm.batch_index_select_dim0( inputs, # Tensor - 1D tensor (concatenated flatten inputs) indices, # Tensor - 1D tensor (concatenated indices) input_num_indices, # List[int] input_rows, # List[int] input_columns, # List[int] ) ``` Differential Revision: D46084590 fbshipit-source-id: 92a7b39c9e0ae6fe9dced906376e294efc6a6bf7
|
This pull request was exported from Phabricator. Differential Revision: D46084590 |
sryap
added a commit
to sryap/FBGEMM
that referenced
this pull request
Aug 1, 2023
Summary: Pull Request resolved: pytorch#1897 This diff introduces `batch_index_select_dim0` using the `SplitTBE` implementation (it shares the same code generator as TBE). The new operator is designed to address limitations of `group_index_select_dim0`. Both operators are designed to operate multiple inputs. However, `batch_index_select_dim0` requires all inputs to be contiguous in memory, while `batch_index_select_dim0` can operate on inputs with a discrete memory layout. Implementation-wise, they are different. We plan to merge their backends in the future. Since `batch_index_select_dim0` is backed by TBE, it inherits TBE limitations including: - The column sizes must be a multiple of 4 and not exceed 1024. Moreover, the underlying buffer of the inputs tensor must be 16-byte aligned. This is because the TBE kernel uses a vector load/store which requires the buffer to be 16-byte aligned. The kernel will raise an error if this assumption is violated. - Due to the 16-byte aligned enforcement, during the backward pass, if the output gradient is not 16-byte aligned, the operator will copy the output gradient into a new 16-byte aligned buffer. This can be expensive if the output gradient size is large. Usage: ``` # This target might change in the future torch.ops.load_library("//deeplearning/fbgemm/fbgemm_gpu/codegen:index_select_ops") ... output = torch.ops.fbgemm.batch_index_select_dim0( inputs, # Tensor - 1D tensor (concatenated flatten inputs) indices, # Tensor - 1D tensor (concatenated indices) input_num_indices, # List[int] input_rows, # List[int] input_columns, # List[int] ) ``` Reviewed By: jianyuh Differential Revision: D46084590 fbshipit-source-id: 04d4487c277e1d164e669af9bb71b4f6d19c1460
|
This pull request was exported from Phabricator. Differential Revision: D46084590 |
sryap
added a commit
to sryap/FBGEMM
that referenced
this pull request
Aug 1, 2023
Summary: Pull Request resolved: pytorch#1897 This diff introduces `batch_index_select_dim0` using the `SplitTBE` implementation (it shares the same code generator as TBE). The new operator is designed to address limitations of `group_index_select_dim0`. Both operators are designed to operate multiple inputs. However, `batch_index_select_dim0` requires all inputs to be contiguous in memory, while `batch_index_select_dim0` can operate on inputs with a discrete memory layout. Implementation-wise, they are different. We plan to merge their backends in the future. Since `batch_index_select_dim0` is backed by TBE, it inherits TBE limitations including: - The column sizes must be a multiple of 4 and not exceed 1024. Moreover, the underlying buffer of the inputs tensor must be 16-byte aligned. This is because the TBE kernel uses a vector load/store which requires the buffer to be 16-byte aligned. The kernel will raise an error if this assumption is violated. - Due to the 16-byte aligned enforcement, during the backward pass, if the output gradient is not 16-byte aligned, the operator will copy the output gradient into a new 16-byte aligned buffer. This can be expensive if the output gradient size is large. Usage: ``` # This target might change in the future torch.ops.load_library("//deeplearning/fbgemm/fbgemm_gpu/codegen:index_select_ops") ... output = torch.ops.fbgemm.batch_index_select_dim0( inputs, # Tensor - 1D tensor (concatenated flatten inputs) indices, # Tensor - 1D tensor (concatenated indices) input_num_indices, # List[int] input_rows, # List[int] input_columns, # List[int] ) ``` Reviewed By: jianyuh Differential Revision: D46084590 fbshipit-source-id: 00a912f39e0b9f92c8c22b3a4b5ca26e0981d858
|
This pull request was exported from Phabricator. Differential Revision: D46084590 |
sryap
added a commit
to sryap/FBGEMM
that referenced
this pull request
Aug 1, 2023
Summary: Pull Request resolved: pytorch#1897 This diff introduces `batch_index_select_dim0` using the `SplitTBE` implementation (it shares the same code generator as TBE). The new operator is designed to address limitations of `group_index_select_dim0`. Both operators are designed to operate multiple inputs. However, `batch_index_select_dim0` requires all inputs to be contiguous in memory, while `batch_index_select_dim0` can operate on inputs with a discrete memory layout. Implementation-wise, they are different. We plan to merge their backends in the future. Since `batch_index_select_dim0` is backed by TBE, it inherits TBE limitations including: - The column sizes must be a multiple of 4 and not exceed 1024. Moreover, the underlying buffer of the inputs tensor must be 16-byte aligned. This is because the TBE kernel uses a vector load/store which requires the buffer to be 16-byte aligned. The kernel will raise an error if this assumption is violated. - Due to the 16-byte aligned enforcement, during the backward pass, if the output gradient is not 16-byte aligned, the operator will copy the output gradient into a new 16-byte aligned buffer. This can be expensive if the output gradient size is large. Usage: ``` # This target might change in the future torch.ops.load_library("//deeplearning/fbgemm/fbgemm_gpu/codegen:index_select_ops") ... output = torch.ops.fbgemm.batch_index_select_dim0( inputs, # Tensor - 1D tensor (concatenated flatten inputs) indices, # Tensor - 1D tensor (concatenated indices) input_num_indices, # List[int] input_rows, # List[int] input_columns, # List[int] ) ``` Reviewed By: jianyuh Differential Revision: D46084590 fbshipit-source-id: eeb58b27c77ebcacba6ad27465c67feb673b4bbe
|
This pull request was exported from Phabricator. Differential Revision: D46084590 |
Summary: Pull Request resolved: pytorch#1897 This diff introduces `batch_index_select_dim0` using the `SplitTBE` implementation (it shares the same code generator as TBE). The new operator is designed to address limitations of `group_index_select_dim0`. Both operators are designed to operate multiple inputs. However, `batch_index_select_dim0` requires all inputs to be contiguous in memory, while `batch_index_select_dim0` can operate on inputs with a discrete memory layout. Implementation-wise, they are different. We plan to merge their backends in the future. Since `batch_index_select_dim0` is backed by TBE, it inherits TBE limitations including: - The column sizes must be a multiple of 4 and not exceed 1024. Moreover, the underlying buffer of the inputs tensor must be 16-byte aligned. This is because the TBE kernel uses a vector load/store which requires the buffer to be 16-byte aligned. The kernel will raise an error if this assumption is violated. - Due to the 16-byte aligned enforcement, during the backward pass, if the output gradient is not 16-byte aligned, the operator will copy the output gradient into a new 16-byte aligned buffer. This can be expensive if the output gradient size is large. Usage: ``` # This target might change in the future torch.ops.load_library("//deeplearning/fbgemm/fbgemm_gpu/codegen:index_select_ops") ... output = torch.ops.fbgemm.batch_index_select_dim0( inputs, # Tensor - 1D tensor (concatenated flatten inputs) indices, # Tensor - 1D tensor (concatenated indices) input_num_indices, # List[int] input_rows, # List[int] input_columns, # List[int] ) ``` Reviewed By: jianyuh Differential Revision: D46084590 fbshipit-source-id: 160ea0810abf3be3fc5f087f8a7a56437e481874
|
This pull request was exported from Phabricator. Differential Revision: D46084590 |
|
This pull request has been merged in 410d264. |
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:
Usage:
Differential Revision: D46084590