[SYCL] Add external arbitrary floating point functions #2114
Conversation
Add support of functions that operate on arbitrary precision floating point numbers. This datatype and its corresponding operations can be useful on targets that can take advantage of narrower representation such as FPGAs. Arbitrary precision floating point numbers are represented using arbitrary precision integers (_ExtInt(N)). Additional parameters are: * _E_: represents the number of exponent bits * _M_: represents the number of mantissa bits The total width of the type is `E+M+1` where the last bit is used to represent the sign. The data layout is shown below: `[ S (sign bit) ][ E (Exponent) ][ M (Mantissa) ]` The width of the data `(E+M+1)` is encoded with the width of the arbitrary precision integer variable. Mantissa (namely `M`) is encoded within each of the signature of the operations. Exponent values (namely `E`) are inferred from the width of corresponding variable. Usage example: _ExtInt(Wa) a; _ExtInt(Wb) b; _ExtInt(Wout) res = __spirv_ArbitraryFloatAddINTEL<Wa, Wb, Wout>(a, Ma, b, Mb, Mout); Signed-off-by: Mikhail Lychkov <[email protected]>
mlychkov
force-pushed
the
private/mlychkov/add_ihs_float_built-ins
branch
from
e03ae23 to
57f58b5
Compare
I think that @GarveyJoe better knows about that? |
It is intentional, variety of functions follows https://hlslibs.org/ declarations. |
Interesting. I hope the SYCL implementation will be in modern C++ instead... |
I have a feeling that yes :) |
Add support of functions that operate on arbitrary precision floating
point numbers. This datatype and its corresponding operations can be
useful on targets that can take advantage of narrower representation
such as FPGAs.
Arbitrary precision floating point numbers are represented using
arbitrary precision integers (_ExtInt(N)). Additional parameters are:
The total width of the type is
E+M+1where the last bit is used torepresent the sign. The data layout is shown below:
[ S (sign bit) ][ E (Exponent) ][ M (Mantissa) ]The width of the data
(E+M+1)is encoded with the width of thearbitrary precision integer variable. Mantissa (namely
M) is encodedwithin each of the signature of the operations. Exponent values
(namely
E) are inferred from the width of corresponding variable.Usage example:
_ExtInt(Wa) a;
_ExtInt(Wb) b;
_ExtInt(Wout) res = __spirv_ArbitraryFloatAddINTEL<Wa, Wb, Wout>(a, Ma, b, Mb, Mout);
Signed-off-by: Mikhail Lychkov [email protected]