Skip to content

[SYCL][Doc] Update E8M0 #21289

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
KornevNikita merged 1 commit into from
Feb 16, 2026
Merged

[SYCL][Doc] Update E8M0 #21289

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
58 changes: 18 additions & 40 deletions sycl/doc/extensions/proposed/sycl_ext_oneapi_fp8.asciidoc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1256,22 +1256,16 @@ class fp8_e8m0 {

// Construct from an array of half, bfloat16, float, double.

explicit fp8_e8m0(half const (&vals)[N], rounding r = rounding::upward,
saturation s = saturation::finite);
explicit fp8_e8m0(bfloat16 const (&vals)[N], rounding r = rounding::upward,
saturation s = saturation::finite);
explicit fp8_e8m0(float const (&vals)[N], rounding r = rounding::upward,
saturation s = saturation::finite);
explicit fp8_e8m0(half const (&vals)[N], rounding r = rounding::upward);
explicit fp8_e8m0(bfloat16 const (&vals)[N], rounding r = rounding::upward);
explicit fp8_e8m0(float const (&vals)[N], rounding r = rounding::upward);
explicit fp8_e8m0(double const (&vals)[N]);

// Construct from an marray of half, bfloat16, float, double.

explicit fp8_e8m0(const marray<half,N>& vals, rounding r = rounding::upward,
saturation s = saturation::finite);
explicit fp8_e8m0(const marray<bfloat16,N>& vals, rounding r = rounding::upward,
saturation s = saturation::finite);
explicit fp8_e8m0(const marray<float,N>& vals, rounding r = rounding::upward,
saturation s = saturation::finite);
explicit fp8_e8m0(const marray<half,N>& vals, rounding r = rounding::upward);
explicit fp8_e8m0(const marray<bfloat16,N>& vals, rounding r = rounding::upward);
explicit fp8_e8m0(const marray<float,N>& vals, rounding r = rounding::upward);
explicit fp8_e8m0(const marray<double,N>& vals);

// Construct from integer types.
Expand Down Expand Up @@ -1399,47 +1393,41 @@ Each value is converted using the `rounding::upward` rounding mode and the

[source,c++]
----
explicit fp8_e8m0(half const (&vals)[N], rounding r = rounding::upward, (1)
saturation s = saturation::finite);
explicit fp8_e8m0(bfloat16 const (&vals)[N], rounding r = rounding::upward, (2)
saturation s = saturation::finite);
explicit fp8_e8m0(float const (&vals)[N], rounding r = rounding::upward, (3)
saturation s = saturation::finite);
explicit fp8_e8m0(double const (&vals)[N]); (4)
explicit fp8_e8m0(half const (&vals)[N], rounding r = rounding::upward); (1)
explicit fp8_e8m0(bfloat16 const (&vals)[N], rounding r = rounding::upward); (2)
explicit fp8_e8m0(float const (&vals)[N], rounding r = rounding::upward); (3)
explicit fp8_e8m0(double const (&vals)[N]); (4)
----

_Effects:_ Initializes each element of this `fp8_e8m0` object from the
corresponding element in the array `vals`.
In overloads (1) - (3), each value is converted using the `r` rounding mode and
the `s` saturation mode.
the `saturation::finite` saturation mode.
In overload (4), each value is converted using the `rounding::upward` rounding
mode and the `saturation::finite` saturation mode.

_Target Support:_ Host and device code support only `rounding::upward` and
`saturation::finite`.
`rounding::toward_zero`.

'''

[source,c++]
----
explicit fp8_e8m0(const marray<half,N>& vals, rounding r = rounding::upward, (1)
saturation s = saturation::finite);
explicit fp8_e8m0(const marray<bfloat16,N>& vals, rounding r = rounding::upward, (2)
saturation s = saturation::finite);
explicit fp8_e8m0(const marray<float,N>& vals, rounding r = rounding::upward, (3)
saturation s = saturation::finite);
explicit fp8_e8m0(const marray<double,N>& vals); (4)
explicit fp8_e8m0(const marray<half,N>& vals, rounding r = rounding::upward); (1)
explicit fp8_e8m0(const marray<bfloat16,N>& vals, rounding r = rounding::upward); (2)
explicit fp8_e8m0(const marray<float,N>& vals, rounding r = rounding::upward); (3)
explicit fp8_e8m0(const marray<double,N>& vals); (4)
----

_Effects:_ Initializes each element of this `fp8_e8m0` object from the
corresponding element in the `marray` object `vals`.
In overloads (1) - (3), each value is converted using the `r` rounding mode and
the `s` saturation mode.
the `saturation::finite` saturation mode.
In overload (4), each value is converted using the `rounding::upward` rounding
mode and the `saturation::finite` saturation mode.

_Target Support:_ Host and device code support only `rounding::upward` and
`saturation::finite`.
`rounding::toward_zero`.

'''

Expand Down Expand Up @@ -1603,16 +1591,6 @@ Conversions to E8M0 using one of the non-stochastic rounding modes and with
* If the resulting value is larger in magnitude than the max normal value, it is
converted to the max normal value while dropping the sign.

==== Non-stochastic rounding modes without saturation

Conversions to E8M0 using one of the non-stochastic rounding modes and with
`saturation::none` work as follows:

* NaN is converted to NaN while dropping the sign.
* Other values are rounded according to the rounding mode.
* If the resulting value is larger in magnitude than the max normal value, it is
converted to NaN while dropping the sign.

==== Integer conversions

Conversions from E8M0 to non-boolean integral types work as follows:
Expand Down