amd/blis
1
0
Fork 0
mirror of https://github.com/amd/blis.git synced 2026-04-26 18:48:51 +00:00
Code Issues Packages Projects Releases Wiki Activity

Defined setijv, getijv to set/get vector elements.

Browse Source 
Details:
- Defined getijv, setijv operations to get and set elements of a vector,
  in bli_setgetijv.c and .h.
- Renamed bli_setgetij.c and .h to bli_setgetijm.c and .h, respectively.
- Added additional bounds checking to getijm and setijm to prevent
  actions with negative indices.
- Added documentation to BLISObjectAPI.md and BLISTypedAPI.md for getijv
  and setijv.
- Added documentation to BLISTypedAPI.md for getijm and setijm, which
  were inadvertently missing.
- Added a new entry to the FAQ titled "Why does BLIS have vector
  (level-1v) and matrix (level-1m) variations of most level-1
  operations?"
- Comment updates.
This commit is contained in:
Field G. Van Zee
2021-05-01 18:54:48 -05:00
parent 4534daffd1
commit 6a89c7d8f9
8 changed files with 352 additions and 33 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
docs/FAQ.md
View File

@@ -17,6 +17,7 @@ project, as well as those we think a new user or developer might ask. If you do
* [What is a macrokernel?](FAQ.md#what-is-a-macrokernel)
* [What is a context?](FAQ.md#what-is-a-context)
* [I am used to thinking in terms of column-major/row-major storage and leading dimensions. What is a "row stride" / "column stride"?](FAQ.md#im-used-to-thinking-in-terms-of-column-majorrow-major-storage-and-leading-dimensions-what-is-a-row-stride--column-stride)
* [Why does BLIS have vector (level-1v) and matrix (level-1m) variations of most level-1 operations?](FAQ.md#why-does-blis-have-vector-level-1v-and-matrix-level-1m-variations-of-most-level-1-operations)
* [What does it mean when a matrix with general stride is column-tilted or row-tilted?](FAQ.md#what-does-it-mean-when-a-matrix-with-general-stride-is-column-tilted-or-row-tilted)
* [I am not really interested in all of these newfangled features in BLIS. Can I just use BLIS as a BLAS library?](FAQ.md#im-not-really-interested-in-all-of-these-newfangled-features-in-blis-can-i-just-use-blis-as-a-blas-library)
* [What about CBLAS?](FAQ.md#what-about-cblas)
@@ -117,6 +118,16 @@ In generalized storage, we have a row stride and a column stride. The row stride
BLIS also supports situations where both the row stride and column stride are non-unit. We call this situation "general stride".
### Why does BLIS have vector (level-1v) and matrix (level-1m) variations of most level-1 operations?
At first glance, it might appear that an element-wise operation such as `copym` or `axpym` would be sufficiently general purpose to cover the cases where the operands are vectors. After all, an *m x 1* matrix can be viewed as a vector of length m and vice versa. But in BLIS, operations on vectors are treated slightly differently than operations on matrices.
If an application wishes to perform an element-wise operation on two objects, and the application calls a level-1m operation, the dimensions of those objects must be conformal, or "match up" (after any transposition implied by the object properties). This includes situations where one of the dimensions is unit.
However, if an application instead decides to perform an element-wise operation on two objects, and the application calls a level-1v operation, the dimension constraints are slightly relaxed. In this scenario, BLIS only checks that the vector *lengths* are equal. This allows for the vectors to have different orientations (row vs column) while still being considered conformal. So, you could perform a `copyv` operation to copy from an *m x 1* vector to a *1 x m* vector. A `copym` operation on such objects would not be allowed (unless it was executed with the source object containing an implicit transposition).
Another way to think about level-1v operations is that they will work with any two matrix objects in situations where (a) the corresponding level-1m operation *would have* worked if the input had been transposed, and (b) all operands happen to be vectors (i.e., have one unit dimension).
### What does it mean when a matrix with general stride is column-tilted or row-tilted?
When a matrix is stored with general stride, both the row stride and column stride (let's call them `rs` and `cs`) are non-unit. When `rs` < `cs`, we call the general stride matrix "column-tilted" because it is "closer" to being column-stored (than row-stored). Similarly, when `rs` > `cs`, the matrix is "row-tilted" because it is closer to being row-stored.