From d502d7399f87ee0087acfa21a976e936bab9bff1 Mon Sep 17 00:00:00 2001 From: Andrew Adams Date: Wed, 29 Jun 2022 14:47:25 -0700 Subject: [PATCH] Update README.md Suggested typo fix. (Hi Zach!) --- README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/README.md b/README.md index 750c884..6871a5c 100644 --- a/README.md +++ b/README.md @@ -752,7 +752,7 @@ In this way, first-class dims are a way of adapting the nicer syntax of these ar Performance Expectations ======================== -First-class dimensions are not a compiler. They provide syntax for existing PyTorch operations such as advanced indexing that is easier to read and write. For large sized tensors, the performance of any statements including them will be the same as using the already existing operations. An important exception is the pattern matching of products and summation, where performance will be improved by issuing to a matrix-multiply kernel. The C++ implementation of dimensions adds a small overhead of around 2us on top of PyTorch's normal overhead of 8us to each function that uses them. In the future, the implementation can encorporate more fusion optimization to further improve performance of this style of code. +First-class dimensions are not a compiler. They provide syntax for existing PyTorch operations such as advanced indexing that is easier to read and write. For large sized tensors, the performance of any statements including them will be the same as using the already existing operations. An important exception is the pattern matching of products and summation, where performance will be improved by issuing to a matrix-multiply kernel. The C++ implementation of dimensions adds a small overhead of around 2us on top of PyTorch's normal overhead of 8us to each function that uses them. In the future, the implementation can incorporate more fusion optimization to further improve performance of this style of code. ## License