Data placement - note
Performing this testing highlighted an important issue about data placement.
The data placement operates on a ëfirst touchí policy ñ the thread which first touches a data item (that is not already placed) will determine how the data is placed, which may or may not be appropriate for subsequent uses of the data.
The data placement for dgemm and the explicit code may be different ñ whichever is used, it is important to ensure that the first placement matches the distribution required for optimal performance in the critical loop.