Explore SME matrix operations with interactive visualizations
Introduction
Explore the ZA tile layout
Transpose a 32-bit matrix in place with ZA
Transpose a matrix with an incomplete edge
Multiply matrices with vector outer products
Transpose and pack matrix A
Pad matrix B to the ZA tile width
Trace packed GEMM through ZA
Compare a GEMM with gather loads
Interleave matrix A
Trace GEMM with interleaved matrix A
Next Steps
Explore SME matrix operations with interactive visualizations
Introduction
Explore the ZA tile layout
Transpose a 32-bit matrix in place with ZA
Transpose a matrix with an incomplete edge
Multiply matrices with vector outer products
Transpose and pack matrix A
Pad matrix B to the ZA tile width
Trace packed GEMM through ZA
Compare a GEMM with gather loads
Interleave matrix A
Trace GEMM with interleaved matrix A
Next Steps
See how ZA transposes values
A transpose exchanges rows and columns. ZA tiles make this easy because you can move vectors of values into rows. You can then extract those values back as columns, or vice versa.
The following visualization shows a transpose in-place, where no external memory is used. The transposed result overwrites the original data. To keep this Learning Path open while exploring, open the 32-bit transpose visualization in a separate browser tab. Move the tab to a separate window to view the visualization and Learning Path side by side.
The description of each step appears immediately under the 32-bit Transpose title. It includes the tile addressing. The highlights show the memory and ZA locations involved.
Trace the algorithm
With Scalable Vector Length equal to 256 bits and accessing 32-bit data elements, there are four logical ZA tiles available. The algorithm uses all of them to maximize memory efficiency.
The transpose is achieved by loading data into rows and storing it back from columns. Select the play button or advance manually to follow the algorithm:
| Stage | Number of steps | Operation |
|---|---|---|
| Load first two tiles | 16 | Load data sequentially from memory into rows of the ZA0 and ZA1 tiles. |
| Load next two tiles | 16 | Load data sequentially from memory into rows of the ZA2 and ZA3 tiles. |
| Store two tiles | 16 | Store data sequentially into memory from columns of the ZA0 and ZA2 tiles. |
| Store remaining tiles | 16 | Store data sequentially into memory from columns of the ZA1 and ZA3 tiles. |
The original matrix had data increasing sequentially along each row. The resulting transposed matrix instead has data increasing sequentially down each column. To see the original data, use browser reload to reset the visualization.
Why the data is mirrored left-to-right
Reload the page to reset the visualization, and advance one step. The first eight values of Matrix Data read as follows:
0, 1, 2, 3, 4, 5, 6, 7
The values loaded into ZA0 read as follows:
7, 6, 5, 4, 3, 2, 1, 0
There’s no mirroring. Memory isn’t addressed as a two dimensional grid. Memory is addressed as one-dimensional storage from low to high addresses, similar to the floors of a skyscraper. The visualization arranges Matrix Data as a grid so that you can see how the data values relate to the matrix.
By convention, we draw matrices with the lowest indexed value at the top left. Here, the top-left element of the grid would have the lowest memory address. Elements to the right would have increasing memory addresses.
Moving down from the end of one row to the beginning of the next also increases memory address. Therefore, in the first data read, the 0 is at the lowest address. The 7 is at the highest address.
By convention, ZA tiles are drawn with column 0 on the right side and row 0 at the top. This is purely a convention. The ZA storage is a register file with no left, right, up, or down. Because of this convention, in the first data read, the 0 is stored in the top-right element of ZA0. The 7 is stored in the top-left.
What you’ve learned and what’s next
You’ve learned that ZA tiles provide row and column access to stored values. This access can be used to transpose data in memory.
For a related C example, see Optimize memory layout with transposition . The kernel in the Learning Path prepares a matrix operand for multiplication. It doesn’t reproduce this animation step for step.
Next, you’ll see what changes when the matrix size isn’t an exact multiple of the ZA tile size.