c++ - Performance AVX/SSE assembly vs. intrinsics -

-

i'm trying check optimum approach optimizing basic routines. in case tried example of multiplying 2 float vectors together:

void mul(float *src1, float *src2, float *dst) {     (int i=0; i<cnt; i++) dst[i] = src1[i] * src2[i]; };

plain c implementation slow. did external asm using avx , tried using intrinsics. these test results (time, smaller better):

asm: 0.110 ipp: 0.125 intrinsics: 0.18 plain c++: 4.0

(compiled using msvc 2013, sse2, tried intel compiler, results pretty same)

as can see asm code beaten intel performance primitives (probably because did lots of branches ensure can use avx aligned instructions). i'd utilize intrinsic approach, it's easier manage , thinking compiler should best job optimizing branches , stuff (my asm code sucks in matter imho, yet faster). here's code using intrinsics:

    int i;     (i=0; (minteger)(dst + i) % 32 != 0 && < cnt; i++) dst[i] = src1[i] * src2[i];      if ((minteger)(src1 + i) % 32 == 0)     {         if ((minteger)(src2 + i) % 32 == 0)         {             (; i<cnt-8; i+=8)             {                 __m256 x = _mm256_load_ps( src1 + i);                  __m256 y = _mm256_load_ps( src2 + i);                  __m256 z = _mm256_mul_ps(x, y);                  _mm256_store_ps(dst + i, z);             };         }         else         {             (; i<cnt-8; i+=8)             {                 __m256 x = _mm256_load_ps( src1 + i);                  __m256 y = _mm256_loadu_ps( src2 + i);                  __m256 z = _mm256_mul_ps(x, y);                  _mm256_store_ps(dst + i, z);             };         };     }     else     {         (; i<cnt-8; i+=8)         {             __m256 x = _mm256_loadu_ps( src1 + i);              __m256 y = _mm256_loadu_ps( src2 + i);              __m256 z = _mm256_mul_ps(x, y);              _mm256_store_ps(dst + i, z);         };     };      (; i<cnt; i++) dst[i] = src1[i] * src2[i];

simple: first address dst aligned 32 bytes, branch check sources aligned.

one problem c++ implementations in beginning , @ end not using avx unless enable avx in compiler, not want, because should avx specialization, software should work on platform, avx not available. , sadly there seems no intrinsics instructions such vmovss, there's penalty mixing avx code sse, compiler uses. if enabled avx in compiler, still didn't below 0.14.

any ideas how optimize make instrisics reach speed of asm code?

your implementation intrinsics not same function implementation in straight c: e.g. if function called arguments mul(p, p, p+1)? you'll different results. pure c version slow because compiler ensuring code exactly said.

if want compiler make optimizations based on assumption 3 arrays not overlap, need make explicit:

void mul(float *src1, float *src2, float *__restrict__ dst)

or better

void mul(const float *src1, const float *src2, float *__restrict__ dst)

(i think it's enough have __restrict__ on output pointer, although wouldn't hurt add input pointers too)


Comments

Post a Comment

Popular posts from this blog

node.js - Mongoose: Cast to ObjectId failed for value on newly created object after setting the value -

-
this question has answer here: what's mongoose error cast objectid failed value xxx @ path “_id”? 7 answers even though has been asked many times because entity has no valid objectid field. i came across situation objectid set value can casted objectid still error. the schema: var serviceschema = parammongoose.schema({ servicename:{ type: parammongoose.schema.types.objectid , ref: 'servicenames',required:true } }); pseudo code: servicefromdb=new service({servicename:'some name'}); servicefromdb.servicename='000000000000000000000001'; servicefromdb.save(function(paramerror,paramdata){ if(paramerror){ console.log('but but...',servicefromdb,paramerror) } }); the output of code is: but but... { servicename: 000000000000000000000001, _id: 55079a90286f49280364f78b } { [casterror: cast objectid faile...
Read more

gradle error "Cannot convert the provided notation to a File or URI" -

-
i have defined 2 system properties in gradle.properties: systemprop.builddir=build systemprop.cfgdir=build\\cfg and have following task defined in build.gradle: task clean(group:'clean',description:'clean project') << { ant.sequential { delete(dir: system.properties.builddir) mkdir(dir: system.properties.builddir) delete(dir: system.properties.cfgdir) mkdir(dir: system.properties.cfgdir) } } this generates following error: execution failed task ':clean'. > cannot convert provided notation file or uri: {dir=build}. following types/formats supported: - string or charsequence path, e.g 'src/main/java' or '/usr/include' - string or charsequence uri, e.g 'file:/usr/include' - file instance. - uri or url instance. but equivalent block of code not generate error , works expected: task clean(group:'clean',description:'clean project') << { ...
Read more

python - NameError: name 'subprocess' is not defined -

-
i'm trying make rng website: http://jamesdotcom.com/?p=417 . when run following code: captureimage = subprocess.popen( [ "fswebcam", "-r", "356x292", "-d", "/dev/video0", "static.jpg", "--skip", "10" ], stdout=devnull, stderr=devnull) captureimage.communicate() i following error message: nameerror: name 'subprocess' not defined what should do? thanks have imported first? import subprocess captureimage = subprocess.popen(["fswebcam", ...) subprocess included in python standard library since python 2.4. should not necessary install it.
Read more