sse C ++内存命令

Question

SSE asm has SQRTPS command.

SQRTPS command have 2 versions:

SQRTPS xmm1, xmm2
SQRTPS xmm1, m128

gcc/clang/vs (all) compilers have helper function _mm_sqrt_ps .

But _mm_sqrt_ps can work only with preloaded xmm (with _mm_set_ps / _mm_load_ps).

From Visual Studio, for example: http://msdn.microsoft.com/en-us/library/vstudio/8z67bwwk%28v=vs.100%29.aspx

What I expect:

__attribute__((aligned(16))) float data[4];
__attribute__((aligned(16))) float result[4];
asm{
    sqrtps xmm0, data             // DIRECTLY FROM MEMORY
    movaps result, xmm0
}

What I have (in C):

__attribute__((aligned(16))) float data[4];
__attribute__((aligned(16))) float result[4];
auto xmm  = _mm_load_ps(&data)   // or _mm_set_ps 
xmm = _mm_sqrt_ps(xmm);
_mm_store_ps(&result[0], xmm);

(in asm):

movaps xmm1, data
sqrtps xmm0, xmm1               // FROM REGISTER
movaps result, xmm0

---

In other words, I would like to see something like this:

__attribute__((aligned(16))) float data[4];
__attribute__((aligned(16))) float result[4];
auto xmm = _mm_sqrt_ps(data);                  // DIRECTLY FROM MEMORY, no need to load (because there is such instruction)
_mm_store_ps(&result[0], xmm);

Answer 1

Quick research: I made the following file, called mysqrt.cpp :

#include <pmmintrin.h>

extern "C" __m128 MySqrt(__m128* a) {
    return _mm_sqrt_ps(a[1]);
}

Trying gcc, namely g++4.8 -msse3 -O3 -S mysqrt.cpp && cat mysqrt.s :

_MySqrt:
LFB526:
    sqrtps  16(%rdi), %xmm0
    ret

Clang ( clang++3.6 -msse3 -O3 -S mysqrt.cpp && cat mysqrt.s ):

_MySqrt:                                ## @MySqrt
    .cfi_startproc
## BB#0:                                ## %entry
    pushq   %rbp
Ltmp0:
    .cfi_def_cfa_offset 16
Ltmp1:
    .cfi_offset %rbp, -16
    movq    %rsp, %rbp
Ltmp2:
    .cfi_def_cfa_register %rbp
    sqrtps  16(%rdi), %xmm0
    popq    %rbp
    retq

Don't know about VS, but at least both gcc and clang seem to produce memory version of sqrtps if needed.

UPDATE Example of function usage:

#include <iostream>
#include <pmmintrin.h>

extern "C" __m128 MySqrt(__m128* a);

int main() {
    __m128 x[2];
    x[1] = _mm_set_ps1(4);
    __m128 y = MySqrt(x);
    std::cout << y[0] << std::endl;
}

// output:
2

UPDATE 2: Regarding your code, you should just do:

auto xmm = _mm_sqrt_ps(*reinterpret_cast<__m128*>(data));

And of course it will be at your own risk, you should guarantee that data contains valid __m128 and is properly aligned.

Answer 2

I think you misunderstood the interface provided by the primitive _mm_sqrt_ps(__m128) . The argument type here can be a variable hold in memory or in register. The extension type __m128 acts like any normal builtin type, eg double , and is not bound to an xmm register but can also be stored in memory.

EDIT Unless you use asm , the compiler determines if and when a variable is loaded into register or left in memory. So, in the following code snippet

__m128 foo(const __m128 x, const __m128*y, std::size_t n)
{
  __m128 result = _mm_set_ps(1.0);
  while(n--)
    result = _mm_mul_ps(result,_mm_add_ps(x,_mm_sqrt_ps(*y++)));
  return result;
}

it's up to the compiler which variables are stored in register. I would think that the compiler puts x and result into xmm registers, but gets *y directly from memory.

Answer 3

The answer to your question is that you can't control this , at least for aligned loads, with intrinsics. It's up to the compiler to decide if it uses SQRTPS xmm1, xmm2 or SQRTPS xmm1, m128. If you want to be 100% certain then you have to write it in assembly. This is one of the deficiencies of intrinsics (at least as they are currently implemented) in my opinion.

Some code can help explain this.

We can get GCC (64-bit with -O3) to generate both version using aligned and unaligned loads

float x[4], y[4]
__m128 x4 = _mm_loadu_ps(x);
__m128 y4 = _mm_sqrt_ps(x4);
_mm_storeu_ps(y,y4);

This gives (with Intel syntax)

movups  xmm0, XMMWORD PTR [rdx]
sqrtps  xmm0, xmm0

However, if we do an aligned load we get the other form

float x[4], y[4]
__m128 x4 = _mm_load_ps(x);
__m128 y4 = _mm_sqrt_ps(x4);
_mm_storeu_ps(y,y4);

This combines the load and square root into one instruction

sqrtps  xmm0, XMMWORD PTR [rax]

Most people would say "trust the compiler." I disagree. If you're using intrinsics then it should be assumed that YOU know what you're doing and NOT the compiler. Here is an example difference-in-performance-between-msvc-and-gcc-for-highly-optimized-matrix-multp where GCC chose one form and MSVC chose the other form (for multiplication instead of the sqrt) and it made a difference in performance.

So once again, if you're using aligned loads, you can only pray that the compiler does what you want. And then maybe on the next version of the compiler it does something different...