From 82c7c780a6b0b8d728fdf9f2df2d820a78617e5b Mon Sep 17 00:00:00 2001
From: Yong He <yonghe@outlook.com>
Date: Sat, 11 Feb 2023 13:19:15 -0800
Subject: Update 07-autodiff.md

---
 docs/user-guide/07-autodiff.md | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

(limited to 'docs/user-guide')

diff --git a/docs/user-guide/07-autodiff.md b/docs/user-guide/07-autodiff.md
index 488800a98..ddd4fe342 100644
--- a/docs/user-guide/07-autodiff.md
+++ b/docs/user-guide/07-autodiff.md
@@ -387,7 +387,7 @@ DifferentialPair<R> derivative(DifferentialPair<T0> p0, inout DifferentialPair<T
 
 A backward derivative propagation function propagates the derivative of the function output to all the input parameters simultaneously.
 
-Given an orignal function `f`, the general rule for determining the signature of its backward propagation function is that a differentiable output `o` becomes an input parameter holding the partial derivative of a downstream output with regard to the this differentiable output, i.e. $$\partial y/\partial o\$$); an input differentiable parameter `i` in the original function will become an output in the backward propagation function, holding the propagated partial derivative $$partial y/\partial i$$; and any non-differentiable outputs are dropped from the backward propagation function. This means that the backward propagation function never returns any values computed in the original function.
+Given an orignal function `f`, the general rule for determining the signature of its backward propagation function is that a differentiable output `o` becomes an input parameter holding the partial derivative of a downstream output with regard to the this differentiable output, i.e. $$\partial y/\partial o\$$); an input differentiable parameter `i` in the original function will become an output in the backward propagation function, holding the propagated partial derivative $$\partial y/\partial i$$; and any non-differentiable outputs are dropped from the backward propagation function. This means that the backward propagation function never returns any values computed in the original function.
 
 More specifically, the signature of its backward propagation function is determined using the following rules:
 - A backward propagation function always returns `void`.
@@ -567,6 +567,6 @@ The compiler can generate forward derivative and backward propagation implementa
 - No access to global variables or shader parameters within a differentiable function.
 - All operations to global resources, including texture reads or atomic writes, are treating as a non-differentiable operation.
 - If a differentiable function contains calls that cause side-effects such as updates to global memory, there will not be a guarantee on how many times the side-effect will occur during the resulting derivative function or back-propagation function.
-- `for` loops: In a backward differentiable function, loops currently cannot have `continue` statements although `break` statements are supported. Loops must use the attribute `[MaxIters(<count>)]` to specify a maximum number of iterations. This will be used by compiler to allocate space to store intermediate data. If the actual number of iterations exceeds the provided maximum, the behavior is undefined.
+- Loops: In a backward differentiable function, loops currently cannot have `continue` statements although `break` statements are supported. Loops must use the attribute `[MaxIters(<count>)]` to specify a maximum number of iterations. This will be used by compiler to allocate space to store intermediate data. If the actual number of iterations exceeds the provided maximum, the behavior is undefined. You can always mark loops with the `[ForceUnroll]` attribute to instruct the Slang compiler to unroll the loop before generating derivative propagation functions. Unrolled loops will be treated the same way as ordinary code and is not subject to any additional restrictions.
 
 The above restrictions do not apply if a user-defined derivative or backward propagation function is provided.
-- 
cgit v1.2.3