new saturated kernels

[Ka-zam]

New saturated sum kernels

[marcusmueller]

should be named saturated_add (or add_saturated) to keep in style with the existing add-named kernels (we use sum only for "complicated" polynomial sums, not element-wise stuff)

[Ka-zam]

#807 handles this log2 failure

[jdemel]

LGTM

I'm impressed by how many kernels you implement. Thanks a lot.

[jdemel]

Can you rebase this PR on latest main? The error should be fixed but I'd like to make sure before I merge anything. Thanks!

[Ka-zam]

LGTM

I'm impressed by how many kernels you implement. Thanks a lot.

We should be in a better place now! 😄

[jdemel]

There's still @marcusmueller comment on naming.
To stick with our naming convention and thus go the way of least surprise, the kernels should be named add_saturated. We use the term sum like e.g. NumPy does when all elements in a vector get summed up.

[Ka-zam]

There's still @marcusmueller comment on naming. To stick with our naming convention and thus go the way of least surprise, the kernels should be named add_saturated. We use the term sum like e.g. NumPy does when all elements in a vector get summed up.

Oops, forgot to remove the unused kernels. Should be good now.

$ volk_profile -R satu

RUN_VOLK_TESTS: volk_8i_x2_add_saturated_8i(vlen=131071, iter=1987, tol=1e-06)
arch                       |         time |     throughput |    max_abs |
---------------------------+--------------+----------------+------------+
generic                    |    7.6207 ms |  102529.3 MB/s |          - |
u_sse2                     |    8.3234 ms |   93873.1 MB/s |    0.0e+00 |
u_avx2                     |    6.2392 ms |  125231.5 MB/s |    0.0e+00 | *
a_sse2                     |    7.7166 ms |  101255.1 MB/s |    0.0e+00 |
a_avx2                     |    6.2236 ms |  125545.4 MB/s |    0.0e+00 | *
Best aligned arch          | a_avx2 (1.22x)
Best unaligned arch        | u_avx2 (1.22x)
--------------------------------------------------------------------------------

RUN_VOLK_TESTS: volk_8u_x2_add_saturated_8u(vlen=131071, iter=1987, tol=1e-06)
arch                       |         time |     throughput |    max_abs |
---------------------------+--------------+----------------+------------+
generic                    |    6.2801 ms |  124415.0 MB/s |          - |
u_sse2                     |    6.0746 ms |  128624.6 MB/s |    0.0e+00 | *
u_avx2                     |    6.1945 ms |  126134.9 MB/s |    0.0e+00 |
a_sse2                     |    5.9526 ms |  131260.8 MB/s |    0.0e+00 |
a_avx2                     |    5.9446 ms |  131437.5 MB/s |    0.0e+00 | *
Best aligned arch          | a_avx2 (1.06x)
Best unaligned arch        | u_sse2 (1.03x)
--------------------------------------------------------------------------------

RUN_VOLK_TESTS: volk_16i_x2_add_saturated_16i(vlen=131071, iter=1987, tol=1e-06)
arch                       |         time |     throughput |    max_abs |
---------------------------+--------------+----------------+------------+
generic                    |   13.6151 ms |  114775.7 MB/s |          - | *
u_sse2                     |   16.0701 ms |   97241.9 MB/s |    0.0e+00 |
u_avx2                     |   13.9325 ms |  112161.2 MB/s |    0.0e+00 |
a_sse2                     |   16.0001 ms |   97667.6 MB/s |    0.0e+00 |
a_avx2                     |   14.4234 ms |  108344.3 MB/s |    0.0e+00 |
Best aligned arch          | generic
Best unaligned arch        | generic
--------------------------------------------------------------------------------

RUN_VOLK_TESTS: volk_16u_x2_add_saturated_16u(vlen=131071, iter=1987, tol=1e-06)
arch                       |         time |     throughput |    max_abs |
---------------------------+--------------+----------------+------------+
generic                    |   14.0463 ms |  111253.0 MB/s |          - |
u_sse2                     |   16.9538 ms |   92173.2 MB/s |    0.0e+00 |
u_avx2                     |   13.7991 ms |  113246.0 MB/s |    0.0e+00 | *
a_sse2                     |   13.6123 ms |  114800.1 MB/s |    0.0e+00 | *
a_avx2                     |   13.9583 ms |  111954.2 MB/s |    0.0e+00 |
Best aligned arch          | a_sse2 (1.03x)
Best unaligned arch        | u_avx2 (1.02x)
--------------------------------------------------------------------------------

Session summary (4 kernels):
  Average speedup vs generic: 1.08x
  Max speedup vs generic: 1.22x (volk_8i_x2_add_saturated_8i)