Issue with __m256 type of intel intrinsics

Question

I\'m trying to test some of the Intel Intrinsics to see how they work. So, i created a function to do that for me and this is the code:

void test_intel_256()         


        

Answer 1

MMX and SSE2 are baseline for x86-64, but AVX is not. You do need to specifically enable AVX, where you didn't for SSE2.

Build with -march=haswell or whatever CPU you actually have. Or just use -mavx.

Beware that gcc -mavx with the default tune=generic will split 256b loadu/storeu intrinsics into vmovups xmm / vinsertf128, which is bad if your data is actually aligned most of the time, and especially bad on Haswell with limited shuffle-port throughput.

It's good for Sandybridge and Bulldozer-family if your data really is unaligned, though. See https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80568: it even affects AVX2 vector-integer code, even though all AVX2 CPUs (except maybe Excavator and Ryzen) are harmed by this tuning. tune=generic doesn't take into account what instruction-set extension are enabled, and there's no tune=generic-avx2.

You could use -mavx2 -mno-avx256-split-unaligned-load -mno-avx256-split-unaligned-store. That still doesn't enable other tuning options (like optimizing for macro-fusion of compare and branch) that all modern x86 CPUs have (except low-power ones), but that isn't enabled by gcc's tune=generic. (https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78855).

---

Also:

I'm including these libraries mmintrin.h, emmintrin.h, xmmintrin.h

Don't do that. Always just include immintrin.h in SIMD code. It pulls in all Intel SSE/AVX extensions. This is why you get error: unknown type name ‘__m256’

---

Keep in mind that subscripting vector types lie __m256 is non-standard and non-portable. They're not arrays, and there's no reason you should expect [] to work like an array. Extracting the 3rd element or something from a SIMD vector in a register requires a shuffle instruction, not a load.

---

If you want handy wrappers for vector types that let you do stuff like use operator[] to extract scalars from elements of vector variables, have a look at Agner Fog's Vector Class Library. It's GPLed, so you'll have to look at other wrapper libraries if that's a problem.

It lets you do stuff like

// example from the manual for operator[]
Vec4i a(10,11,12,13);
int b = a[2];   // b = 12

You can use normal intrinsics on VCL types. Vec8f is a transparent wrapper on __m256, so you can use it with _mm256_mul_ps.

Answer 2

try this out

res=_MM_ADD_PS(vec1,vec2); because the prototype of the __M256_MM_ADD_PS is

__m256 _MM_ADD_PS(__m256,__m256);

it takes two __m256 data types as the parameters and returns their sum as __m256 data, just like

int add(int , int);

for initializing

vec=_MM_setr_PS(7.0,7.0,7.0,7.0,7.0,7.0,7.0,7.0) or

vec =_MM_LOAD_PS(&arr) or

vec =_MM_LOAD_PS(ptr)