# intel-intrinsics 1.2.2

The most practical D SIMD solution! Using SIMD intrinsics with Intel syntax with D.

To use this package, run the following command in your project's root directory:

## intel-intrinsics

Use Intel intrinsics in D code with a wide range of compilers.

### Usage

``````
import inteli.mmx;       // allow MMX intrinsics
import inteli.xmmintrin; // allow SSE1 intrinsics
import inteli.emmintrin; // allow SSE2 intrinsics
import inteli.pmmintrin; // allow SSE3 intrinsics

// distance between two points in 4D
float distance(float[4] a, float[4] b) nothrow @nogc
{

// core.simd is publicly imported, or emulated if need be.
// One can use arithmetic operators / indexing on SIMD types.
__m128 diffSquared = va - vb;
diffSquared = _mm_mul_ps(diffSquared, diffSquared);
sum += _mm_srli_ps!4(sum);

return _mm_cvtss_f32(_mm_sqrt_ss(sum));
}
assert(distance([0, 2, 0, 0], [0, 0, 0, 0]) == 2);

``````

### Why?

#### Capabilities

Some instructions aren't accessible using `core.simd` and `ldc.simd` capabilities. For example: `pmaddwd` which is so important in digital video. In this case one need to generate the right IR, or use the right LLVM intrinsic call.

#### Familiar syntax

Intel intrinsic syntax is more familiar to C++ programmers and there is a convenient online guide provided by Intel: https://software.intel.com/sites/landingpage/IntrinsicsGuide/

Without this critical Intel documentation, it's much more difficult to write sizeable SIMD code.

In `intel-intrinsics` it is extended with indexing and arithmetic operators, for convenience.

#### Future-proof

`intel-intrinsics` is a set of stable SIMD intrinsic that compiler teams don't have the manpower to maintain. It is mimicked on the set of similar intrinsics in GCC, clang, ICC...

LDC SIMD builtins are a moving target (https://github.com/ldc-developers/ldc/issues/2019), and you need a layer over it if you want to be sure your code won't break. (The reason is that as things become expressible in IR only in LLVM, x86 builtins get removed).

#### Portability

Because D code or LLVM IR is portable, one goal of `intel-intrinsics` is to be one day platform-independent. One could target ARM one day and still get comparable speed-up.

The long-term goal is:

Write the same SIMD code for LDC, GDC, and DMD. Support x86 well, and ARM eventually. Get top speed..

#### Supported instructions set

DMDLDCGDC
MMXYes but slow (#16)YesYes (slow in 32-bit)
SSEYes but slow (#16)YesYes (slow in 32-bit)
SSE2Yes but slow (#16)YesYes (slow in 32-bit)
SSE3Yes but slow (#16)Yes (use -mattr=+sse3)Yes but slow (#39)
SSSE3NoNoNo
...NoNoNo

#### Important difference

Every implicit conversion of similarly-sized vectors should be done with a `cast` instead.

``````__m128i b = _mm_set1_epi32(42);
__m128 a = b;             // NO, only works in LDC
__m128 a = cast(__m128)b; // YES, works in all D compilers

``````

This is because D does not allow user-defined implicit conversions, and `core.simd` might be emulated (DMD).

### Who is using it?

• Auburn Sounds demonstrated 3.5x speed-up for some loops in the talk: intel-intrinsics: Not intrinsically about intrinsics.
• Pixel Perfect Engine is using `intel-intrinsics` for blitting images: https://github.com/ZILtoid1991/CPUblit/blob/master/src/CPUblit/composing.d
• Please get in touch to get on that list!
Dependencies:
none
Versions:
 1.2.2 2019-Dec-04 1.2.1 2019-Dec-04 1.2.0 2019-Oct-06 1.1.3 2019-Sep-14 1.1.2 2019-Sep-09