Documentation of TMA: part 2 (#2041)

This PR does the following:
- Added a few more texts to `doc/dev/tma.md` describing how to create
box.
- Largely extents `doc/reading/divisibility-of-split.md` discussing
indivisible splits.
- Added `doc/reading/tma-modeling-in-depth.md` discussing deeper and
more formally about predication and correctness of TMA.

Suggested order of reading:
1.
https://github.com/NVIDIA/Fuser/blob/tma-step-2/doc/reading/divisibility-of-split.md
2. https://github.com/NVIDIA/Fuser/blob/tma-step-2/doc/dev/tma.md
3.
https://github.com/NVIDIA/Fuser/blob/tma-step-2/doc/reading/tma-modeling-in-depth.md

.gitignore

@@ -49,4 +49,5 @@ foo.bin
 
 *_generated.*
 
+# Mac OS internal file
 .DS_Store

csrc/non_divisible_split.h

@@ -15,6 +15,7 @@
 
 namespace nvfuser {
 
+//! See doc/reading/divisibility-of-split.md#predication
 //! If an IterDomain is split and its inner output domain is
 //! eventually split too, the second split must be divisible or the
 //! inner domain must be predicated. This class finds Split

doc/dev/tma.md

@@ -4,7 +4,7 @@
  * SPDX-License-Identifier: BSD-3-Clause
 -->
 
-# TMA support in nvFuser
+# TMA Support in NVFuser
 
 ## Introduction
 
@@ -72,6 +72,26 @@ Instead, it is a virtual domain that only exists in the user's mind.
 Also note that the IterDomain expressions between the global tensor's allocation domain and the TMA domain must be a view,
 for example, we can not merge discontiguous IterDomains ([why?](../reading/divisibility-of-split.md#merging-discontiguous-iterdomains)), and we can not have indivisible splits either.
 
-### Step 2: create box
+### Step 2: Define box
+
+After having scheduled a TMA domain, the next step is to define box.
+There are two ways of defining box: partitioning and compositing.
+
+#### Define box by partitioning
+
+Defining box by partitioning is as simple as: select an IterDomain in the TMA domain, then
+inner split that IterDomain by the box size of that dimension.
+
+We call this split expression a "*boxing split*", the input of this split a "*partitioned IterDomain*",
+the inner output of this split a "*box IterDomain*", and the outer output of this split a "*coordinate IterDomain*".
+
+For the case of Figure 1, if both box dimensions are defined by partitioning,
+the schedule should look like the Figure 3 below:
+
+![Figure 3: Boxing by partitioning](tma/box-by-partitioning.svg)
+
+Please note that, although in the above example, the split is divisible, this does not have to be the case in general.
+
+#### Define box by compositing
 
 TODO: this documentation is under construction