vec Common Interface

The SIMD toolkit common interface is a work in progress. We’ve implemented many types and functions already, but many more remain to be done. Care to contribute?

Specification

For each function, we compare the return values of the vector functions against the return values of their scalar equivalents. Any runtime denormal handling should be on. You can run these tests yourself in the vectest executable.

For integral functions, both sets of values shall match over the domain of the function.

For floating point functions, either: both values are not-a-number, or both values have the same sign and are either infinite or differ by at most 5 ulp. The domain shall be all numbers, except if the compiler finite math option is on, then it shall be all finite numbers.

Stages

Given the specification above, there are four stages of implementation:

Unplanned grey. This function has not been planned, usually because it doesn’t make sense for the domain type.
Planned red. This function has been planned but not yet implemented.
Implemented yellow. This function has been implemented but is not yet at specification. It might still be useful for testing and in limited situations.
At Specification green. This function has been implemented and empirically verified to conform to specifications. A clear run of the exhaustive test is sufficient.

These are the SIMD types implemented:

vec <unsigned char, 16> — u8x16.
vec <signed char, 16> — i8x16.
vec <boolean <char>, 16> — b8x16.
vec <unsigned short, 8> — u16x8.
vec <short, 8> — i16x8.
vec <boolean <short>, 8> — b16x8.
vec <unsigned int, 4> — u32x4.
vec <int, 4> — i32x4.
vec <boolean <int>, 4> — b32x4.
vec <float, 4> — f32x4.
vec <boolean <float>, 4> (SSE/SSE2/SSE3 only) — f32bx4.
vec <unsigned long long, 2> (SSE/SSE2/SSE3 only) — u64x2.
vec <long long, 2> (SSE/SSE2/SSE3 only) — i64x2.
vec <signed char, 16> — b64x2.
vec <double, 2> (SSE/SSE2/SSE3 only) — f64x2.
vec <boolean <double>, 2> (SSE/SSE2/SSE3 only) — f64bx2.
vec <complex <float>, 2> (Altivec only) — c64x2.

These are the common interface functions implemented:

Altivec Implementation
types	U8x16	I8x16	B8x16	U16x8	I16x8	B16x8	U32x4	I32x4	F32x4	B32x4	C64x2	B64x2
summarizers
sum
max
min
unary operators
operator-
operator~
operator!
binary operators
operator*
operator/
operator%
operator+
operator-
operator&
operator\|
operator^
operator<<
operator>>
operator==
operator!=
operator<
operator>
operator<=
operator>=
operator&&
operator\|\|
transcendental functions
abs
acos
asin
atan
atan2
cos
cosh
exp
log
max
min
pow
sin
sinh
sqrt
tan
tanh

As you can see, Altivec has the most complete implementation. Only some float and complex transcendentals need to be filled in.

SSE/SSE2/SSE3 Implementation
types	U8x16	I8x16	B8x16	U16x8	I16x8	B16x8	U32x4	I32x4	B32x4	F32x4	F32Bx4	U64x2	I64x2	B64x2	F64x2	F64Bx2
summarizers
sum
max
min
unary operators
operator-
operator~
operator!
binary operators
operator*
operator/
operator%
operator+
operator-
operator&
operator\|
operator^
operator<<
operator>>
operator==
operator!=
operator<
operator>
operator<=
operator>=
operator&&
operator\|\|
transcendental functions
abs
acos
asin
atan
atan2
cos
cosh
exp
log
max
min
pow
sin
sinh
sqrt
tan
tanh

SSE/SSE2/SSE3 has a workable but incomplete implementation. The summarizers, multiplicative and shift functions are largely unimplemented, as are long long, float and double functions.

» mach