Methods and apparatus for performing a machine learning operation using storage element pointers
Methods, apparatus, systems, and articles of manufacture are disclosed for performing a machine learning operation using storage element pointers. An example computer readable medium comprises instructions that when executed, cause at least one processor to select, in response to a determination that a machine learning operation is to be performed, create first and second storage element pointers based on a type of machine learning operation to be performed, remap input tensor data of the input tensor based on the first storage element pointer without movement of the input tensor data in memory, cause execution of the machine learning operation with the remapped input tensor data to create intermediate tensor data, remap the intermediate tensor data based on the second storage element pointer without movement of the intermediate tensor data in memory, and provide the remapped intermediate tensor data as an output tensor.
1 . An apparatus, comprising:
a computer processor; and a non-transitory computer-readable memory storing computer program instructions executable by the computer processor to perform operations, the operations comprising:
generating a first pointer table for a machine learning operation of a first type, the first pointer table comprising storage element pointers pointing to storage elements of a memory, wherein a storage element is to store a particular portion of an input tensor of the machine learning operation of the first type;
generating a second pointer table by rearranging one or more storage element pointers in the first pointer table;
forming a first tensor based on the second pointer table, the first tensor comprising a data element of the input tensor, wherein a position of the data element in the first tensor is different form a position of the data element in the input tensor;
causing execution of a machine learning operation of a second type with the first tensor to generate a second tensor;
generating a third pointer group, the third pointer group comprising storage element pointers pointing to storage elements storing data elements of the second tensor; and
forming an output tensor of the machine learning operation of the first type from the second tensor based on the third pointer group.
2 . The apparatus of claim 1 , wherein the machine learning operation of the first type is a convolution operation.
3 . The apparatus of claim 2 , wherein the convolution operation of the second type is a group convolution operation.
4 . The apparatus of claim 2 , wherein the convolution operation of the second type is a dilated convolution operation.
5 . The apparatus of claim 1 , wherein the first tensor further comprises an additional data element of the input tensor, wherein the data element and the additional data element are adjacent in the first tensor, wherein the data element and the additional data element are separated by one or more other data elements in the input tensor.
6 . The apparatus of claim 1 , wherein the second pointer table comprises a subset of the storage element pointers of the first pointer table.
7 . The apparatus of claim 1 , wherein the particular portion of the input tensor comprises activations in a plurality of input channels of the machine learning operation of the first type.
8 . At least one non-transitory computer readable medium storing instructions that, when executed, cause at least one processor to perform operations, the operations comprising:
generating a first pointer table for a machine learning operation of a first type, the first pointer table comprising storage element pointers pointing to storage elements of a memory, wherein a storage element is to store a particular portion of an input tensor of the machine learning operation of the first type;
generating a second pointer table by rearranging one or more storage element pointers in the first pointer table;
forming a first tensor based on the second pointer table, the first tensor comprising a data element of the input tensor, wherein a position of the data element in the first tensor is different form a position of the data element in the input tensor;
causing execution of a machine learning operation of a second type with the first tensor to generate a second tensor;
generating a third pointer group, the third pointer group comprising storage element pointers pointing to storage elements storing data elements of the second tensor; and
forming an output tensor of the machine learning operation of the first type from the second tensor based on the third pointer group.
9 . The at least one non-transitory computer readable medium of claim 8 , wherein the machine learning operation of the first type is a convolution operation.
10 . The at least one non-transitory computer readable medium of claim 9 , wherein the convolution operation of the second type is a group convolution operation.
11 . The at least one non-transitory computer readable medium of claim 9 , wherein the convolution operation of the second type is a dilated convolution operation.
12 . The at least one non-transitory computer readable medium of claim 8 , wherein the generating of the first pointer table, the second pointer table, or the third pointer table is performed before the execution of the machine learning operation of the second type.
13 . The at least one non-transitory computer readable medium of claim 8 , wherein the first tensor further comprises an additional data element of the input tensor, wherein the data element and the additional data element are adjacent in the first tensor, wherein the data element and the additional data element are separated by one or more other data elements in the input tensor.
14 . The at least one non-transitory computer readable medium of claim 8 , wherein the second pointer table comprises a subset of the storage element pointers of the first pointer table.
15 . The at least one non-transitory computer readable medium of claim 8 , wherein the particular portion of the input tensor comprises activations in a plurality of input channels of the machine learning operation of the first type.
16 . A method, comprising:
generating a first pointer table for a machine learning operation of a first type, the first pointer table comprising storage element pointers pointing to storage elements of a memory, wherein a storage element is to store a particular portion of an input tensor of the machine learning operation of the first type;
generating a second pointer table by rearranging one or more storage element pointers in the first pointer table;
forming a first tensor based on the second pointer table, the first tensor comprising a data element of the input tensor, wherein a position of the data element in the first tensor is different form a position of the data element in the input tensor;
causing execution of a machine learning operation of a second type on the first tensor to generate a second tensor;
generating a third pointer group, the third pointer group comprising storage element pointers pointing to storage elements storing data elements of the second tensor; and
forming an output tensor of the machine learning operation of the first type from the second tensor based on the third pointer group.
17 . The method of claim 16 , wherein the machine learning operation of the first type is a convolution operation.
18 . The method of claim 17 , wherein the convolution operation of the second type is a group convolution operation.
19 . The method of claim 17 , wherein the convolution operation of the second type is a dilated convolution operation.
20 . The method of claim 16 , wherein the generating of the first pointer table, the second pointer table, or the third pointer table is performed before the execution of the machine learning operation of the second type.
21 . The method of claim 16 , wherein the first tensor further comprises an additional data element of the input tensor, wherein the data element and the additional data element are adjacent in the first tensor, wherein the data element and the additional data element are separated by one or more other data elements in the input tensor.
22 . The method of claim 16 , wherein the second pointer table comprises a subset of the storage element pointers of the first pointer table.
23 . The method of claim 16 , wherein the particular portion of the input tensor comprises activations in a plurality of input channels of the machine learning operation of the first type.