a : AAA3 AAA2 AAA1 AAA0 (32 bits)
b : BBB3 BBB2 BBB1 BBB0 (32 bits)
-For all partition settings this copies verbatim. However if A
+For all partition settings this copies verbatim. Also,
+when A is longer, a truncated version of A is always
+copied verbatim, regardless of partition settings.
+However if A
is shorter than B:
partition: p p p (3 bits)
| partition | o3 | o2 | o1 | o0 |
| --------- | -- | -- | -- | -- |
| 000 | [A7A7A7A7] | [A7A7A7A7] | A7A6A5A4 | A3A2A1A0 |
+| 001 | [A7A7A7A7] | [A7A7]A7A6 | A5A4A3A2 | [A1A1]A1A0 |
+| 010 | [A7A7A7A7] | A7A6A5A4 | [A3A3A3A3] | A3A2A1A0 |
+| 011 | [A7A7A7A7] | A7A6A5A4 | [A3A3]A3A2 | [A1A1]A1A0 |
+| 100 | [A7A7]A7A6 | [A5A5A5A5] | [A5A5]A5A4 | A3A2A1A0 |
+| 101 | [A7A7]A7A6 | [A5A5A5A5] | A5A4A3A2 | [A1A1]A1A0 |
+| 110 | [A7A7]A7A6 | [A5A5]A5A4 | [A3A3A3A3] | A3A2A1A0 |
| 111 | [A7A7]A7A6 | [A5A5]A5A4 | [A3A3]A3A2 | [A1A1]A1A0 |
where square brackets are zero if A is unsigned, and contains
-the specified bits if signed.
+the specified bits if signed. Here, each partition copies the
+smaller value (A) into the larger partition (B) then, depending
+on whether A is signed or unsigned, sign-extends or zero-extends
+*on a per-partition basis*.
+
+# Scalar source
+
+When the source A is scalar and is equal or larger than
+the destination it requires copying across multiple
+partitions:
+
+ partition: p p p (3 bits)
+ a : AAAA AAAA AAAA AAAA (16 bits)
+ b : B7B6 B5B4 B3B2 B1B0 (8 bits)
+
+The partition options are:
+
+| partition | o3 | o2 | o1 | o0 |
+| --------- | -- | -- | -- | -- |
+| 000 | A7A6 | A5A4 | A3A2 | A1A0 |
+| 001 | A5A4 | A3A2 | A1A0 | A1A0 |
+| 010 | A3A2 | A1A0 | A3A2 | A1A0 |
+| 011 | A3A2 | A1A0 | A1A0 | A1A0 |
+| 100 | A1A0 | A5A4 | A3A2 | A1A0 |
+| 101 | A1A0 | A3A2 | A1A0 | A1A0 |
+| 110 | A1A0 | A1A0 | A3A2 | A1A0 |
+| 111 | A1A0 | A1A0 | A1A0 | A1A0 |
+
+When the partitions are all open (1x) only the bits that will fit across
+the whole of the target are copied. In this example, B is 8 bits so only
+8 bits of A are copied.
+
+When the partitions are all closed (4x SIMD) each partition of B is
+2 bits wide, therefore only the *first two* bits of A are copied into
+*each* of the four 2-bit partitions in B.
+
+For the case where A is shorter than B output, sign or zero
+extension is required. Here we assume A is 8 bits, B is 16.
+This is similar to the parallel case except A is repeated
+(broadcast) across all of B.
+
+
+| partition | o3 | o2 | o1 | o0 |
+| --------- | -- | -- | -- | -- |
+| 000 | [A7A7A7A7] | [A7A7A7A7] | A7A6A5A4 | A3A2A1A0 |
+| 001 | [A7A7A7A7] | A7A6A5A4 | A3A2A1A0 | A3A2A1A0 |
+| 010 | A7A6A5A4 | A3A2A1A0 | A7A6A5A4 | A3A2A1A0 |
+| 011 | A7A6A5A4 | A3A2A1A0 | A3A2A1A0 | A3A2A1A0 |
+| 100 | A3A2A1A0 | [A7A7A7A7] | A7A6A5A4 | A3A2A1A0 |
+| 101 | A3A2A1A0 | A7A6A5A4 | A3A2A1A0 | A3A2A1A0 |
+| 110 | A3A2A1A0 | A3A2A1A0 | A7A6A5A4 | A3A2A1A0 |
+| 111 | A3A2A1A0 | A3A2A1A0 | A3A2A1A0v | A3A2A1A0 |
+
+Note how when the entire partition set is open (1x 16-bit output)
+that all of A is copied out, and either zero or sign extended
+in the top half of the output. At the other extreme is the
+4x 4-bit output partitions, which have four copies of A, truncated
+from the first 4 bits of A.
+
+Unlike the parallel case, A is not itself partitioned, so is copied
+over as much as is possible. In some cases such as `1x 4-bit, 1x 12-bit`
+(partition mask = `0b100`, above) the 8-bit scalar source will need sign or zero extending.
projects / libreriscv.git / blobdiff