Hello,
The ippsFilterMedian_32f does not yield to a correct result when any AVX cpu feature is enable/available.
Tested with IPP 11.0, tested on both Linux and Windows, on multiple computers with AVX, AVX2 and AVX512.
Minimal reproduction code :
#include <iostream>
#include <ipp.h>
#define PX_FM ( ippCPUID_MMX | ippCPUID_SSE | ippCPUID_SSE2 )
#define M7_FM ( PX_FM | ippCPUID_SSE3 )
#define U8_FM ( M7_FM | ippCPUID_SSSE3 )
#define N8_FM ( U8_FM | ippCPUID_MOVBE )
#define Y8_FM ( U8_FM | ippCPUID_SSE41 | ippCPUID_SSE42 )
int main()
{
ippInit();
// ippSetCpuFeatures(Y8_FM);
int len = 10;
int maskSize = 7;
float pSrc[10] = { 1, 2, 3, 4, 5, 6, 7, 8, 9, 10 };
float* pDst = ippsMalloc_32f(17);
// Have pDst be misaligned
pDst = pDst+7;
int bufferSize;
ippsFilterMedianGetBufferSize(maskSize, ipp32f, &bufferSize);
unsigned char* pBuffer = ippsMalloc_8u(bufferSize);
ippsFilterMedian_32f(pSrc, pDst, len, maskSize, nullptr, nullptr, pBuffer);
std::cout << "dst = [ ";
for(int i = 0; i<10; ++i)
std::cout << pDst[i] << " ";
std::cout << "]" << std::endl;
}
The shared code computes the effect of a median filter on the input :
src=[ 1 2 3 4 5 6 7 8 9 10 ]
The expected result is :
dst = [ 1 1 1 1 2 3 4 5 6 7 ]
The actual result (on a computer with AVX) is :
dst = [ 1 1 2 2 3 3 4 5 6 7 ]
The actual result is erroneous.
Either
- uncommenting the line 13 "ippSetCpuFeatures(Y8_FM);" to disable AVX
- commenting the line 21 "pDst = pDst+7;" to have the destination aligned to 64 bytes yield the expected result.
