Modern CPUs often support specialized vector types and vector operations, sometimes called "media instructions". Vector types are a fixed array of floating or integer types, and vector operations operate simultaneously on them, thus achieving great speedups.
When the compiler takes advantage of these instructions with standard D code to speed up loops over arithmetic data, this is called auto-vectorization. Auto-vectorization, however, has had only limited success and has not been able to really take advantage of the richness (and often quirkiness) of the native vector instructions.
D has the array operation notation, such as:
int a,b; ... a += b;
which can be vectorized by the compiler, but again success is limited for the same reason auto-vectorization is.
The difficulties with trying to use vector instructions on regular arrays are:
- The vector types have stringent alignment requirements that are not and cannot be met by conventional arrays.
- C ABI's often have vector extensions and have special name mangling for them, call/return conventions, and symbolic debug support.
- The only way to get at the full vector instruction set would be to use inline assembler - but the compiler cannot do register allocation across inline assembler blocks (or other optimizations), leading to poor code performance.
- Interleaving conventional array code with vector operations on the same data can unwittingly lead to extremely poor runtime performance.
These issues are cleared up by using special vector types.
Vector types and operations are introduced to D code by importing core.simd:
These types and operations will be the ones defined for the architecture the compiler is targetting. If a particular CPU family has varying support for vector types, an additional runtime check may be necessary. The compiler does not emit runtime checks; those must be done by the programmer.
The types defined will all follow the naming convention:
where type is the vector element type and NN is the number of those elements in the vector type. The type names will not be keywords.
Vector types have the property:
|.array||Returns static array representation|
All the properties of the static array representation also work.
Vector types of the same size can be implicitly converted among each other. Vector types can be cast to the static array representation.
Integers and floating point values can be implicitly converted to their vector equivalents:
int4 v = 7; v = 3 * v; // multiply each element in v by 3
Accessing Individual Vector Elements
They cannot be accessed directly, but can be when converted to an array type:
int4 v; (cast(int*)&v) = 2; // set 3rd element of the 4 int vector cast(int)v = 2; // set 3rd element of the 4 int vector v.array = 2; // set 3rd element of the 4 int vector v.ptr = 2; // set 3rd element of the 4 int vector
If vector extensions are implemented, the version identifier D_SIMD is set.
Whether a type exists or not can be tested at compile time with an IsExpression:
static if (is(typeNN)) ... yes, it is supported ... else ... nope, use workaround ...
Whether a particular operation on a type is supported can be tested at compile time with:
float4 a,b; static if (__traits(compiles, a+b)) ... yes, it is supported ... else ... nope, use workaround ...
For runtime testing to see if certain vector instructions are available, see the functions in core.cpuid.
A typical workaround would be to use array vector operations instead:
float4 a,b; static if (__traits(compiles, a/b)) c = a / b; else c = a / b;
X86 And X86_64 Vector Extension Implementation
The rest of this document describes the specific implementation of the vector types for the X86 and X86_64 architectures.
The vector extensions are currently implemented for the OS X 32 bit target, and all 64 bit targets.
core.simd defines the following types:
|Type Name||Description||gcc Equivalent|
|void16||16 bytes of untyped data||no equivalent|
|byte16||16 byte's||signed char __attribute((vector_size(16)))|
|ubyte16||16 ubyte's||unsigned char __attribute((vector_size(16)))|
|short8||8 short's||short __attribute((vector_size(16)))|
|ushort8||8 ushort's||ushort __attribute((vector_size(16)))|
|int4||4 int's||int __attribute((vector_size(16)))|
|uint4||4 uint's||unsigned __attribute((vector_size(16)))|
|long2||2 long's||long __attribute((vector_size(16)))|
|ulong2||2 ulong's||unsigned long __attribute((vector_size(16)))|
|float4||4 float's||float __attribute((vector_size(16)))|
|double2||2 double's||double __attribute((vector_size(16)))|
Note: for 32 bit gcc, it's long long instead of long.
Operators not listed are not supported at all.
Vector Operation Intrinsics
See core.simd for the supported intrinsics.