ippsThreshold_LTAbsVal_64f weird behaviour

[agildr]

Hi,

I'm trying to use ippsThreshold_LTAbsVal_64f to replace small numbers by zero.

According to the documentation a call like ippsThreshold_LTAbsVal_64f(xPtr, resultPtr, length, eps, 0.0) should replace all numbers in x that are in magnitude smaller than eps by zero.

To quote the docs, the logic is:

if( ABS(x[i]) < level ) y[i] = value;
else y[i] = x[i];

I have written the following C# program. I hope it is OK to post C# code, I also had no problems with any other IPP function from C# so far, only this one.

static class IPP
{
	[SuppressUnmanagedCodeSecurity, DllImport("libippcore.so", CallingConvention = CallingConvention.Cdecl, ExactSpelling = true)]
	public static extern int ippInit();

	[SuppressUnmanagedCodeSecurity, DllImport("libipps.so", CallingConvention = CallingConvention.Cdecl, ExactSpelling = true)]
	public static extern unsafe int ippsThreshold_LTAbsVal_64f(double* pSrc, double* pDst, int len, double level, double value);
}

static class Program
{
	static unsafe void IppSanityCheck(int index1, int index2)
	{
		const int length = 10000;
		var array = new double[length];
		var result = new double[length];

		const double eps = 1e-6;
		// setting some indices to values |.| > eps
		array[index1] = 100.0;
		array[index2] = -95.0;

		for (var i = 0; i < length; i++)
		{
			if (array[i] != 0.0)
			{
				Console.WriteLine("Input index {0} contains non-zero value: {1}", i, array[i]);
			}
		}

		fixed (double* arrayPtr = array, resultPtr = result)
		{
			var statusCode = IPP.ippsThreshold_LTAbsVal_64f(arrayPtr, resultPtr, length, eps, 0);
			Console.WriteLine("{0} exited with {1}", nameof(IPP.ippsThreshold_LTAbsVal_64f), statusCode);
		}

		// expect same vector as result
		for (var i = 0; i < length; i++)
		{
			if (Math.Abs(array[i] - result[i]) > 1e-16)
			{
				Console.WriteLine("Result index {0} differs from input: {1} vs. {2}", i, array[i], result[i]);
			}
		}
	}

	static void Main(string[] args)
	{
		if (IPP.ippInit() == 0)
		{
			IppSanityCheck(534, 1096);
			IppSanityCheck(1,2);
		}
		return;
	}
}

This is the output:

Input index 534 contains non-zero value: 100
Input index 1096 contains non-zero value: -95
ippsThreshold_LTAbsVal_64f exited with 0
Result index 534 differs from input: 100 vs. 0
Result index 1096 differs from input: -95 vs. 0
Input index 1 contains non-zero value: 100
Input index 2 contains non-zero value: -95
ippsThreshold_LTAbsVal_64f exited with 0
The program 'dotnet' has exited with code 0 (0x0).

As you can see when calling it with the first input (all zero except for indices 534 and 1096) the non-zero indices get replaced by zero (wrong). When calling it with the second input (now indices 1 and 2) they are not replaced (correct).

Running on Docker/WSL2, x64, Debian 10, .NET 5, IPP is from the Intel APT repo:

# uname -a
Linux 2d1918de08d8 5.10.16.3-microsoft-standard-WSL2 #1 SMP Fri Apr 2 22:23:49 UTC 2021 x86_64 GNU/Linux
# dpkg -l | grep intel
ii  intel-oneapi-runtime-ipp           2021.3.0-333           amd64        Intel® Integrated Performance Primitives runtime
ii  intel-oneapi-runtime-ipp-common    2021.3.0-333           all          Intel® Integrated Performance Primitives runtime common
ii  intel-oneapi-runtime-openmp        2021.3.0-3350          amd64        Intel® OpenMP* Runtime Library runtime
ii  intel-oneapi-runtime-openmp-common 2021.3.0-3350          all          l_openmp.runtime.description>
ii  intel-oneapi-runtime-tbb           2021.3.0-511           amd64        Intel® oneAPI Threading Building Blocks runtime
ii  intel-oneapi-runtime-tbb-common    2021.3.0-511           all          Intel® oneAPI Threading Building Blocks runtime common
#

But I have tried on Windows as well and the behavior is the same.

Best,

agildr

[ShanmukhS_Intel]

Hi,

Thanks for reaching out to us.

We have tried compiling and executing the shared code. 

>>if( ABS(x[i]) < level ) y[i] = value;

else y[i] = x[i];

As per the absolute threshold formula, among the indices shared by you,

Considering 1st case 100 

Absolute of 100 which is 100 and greater than 1e-6. Hence the threshold doesn't get affected.

Considering 2nd case -95

Absolute of -95 which is 95 and greater than 1e-6. Hence the threshold doesn't get affected.

Kindly let us know if any difference in understanding.

Best Regards,

Shanmukh.SS

[agildr]

Hi Shanmukh.SS,

Yes, with the values supplied (0, -95, 100) the result should be the same as the input, since nothing ever is under the threshold (in magnitude), and zero already is equal to zero.

This works with the vector [0,100,-95,0,...,0] (length 10000). However, in the case where I put the 100 at index 534 and -95 at index 1096 I get a different vector (all zeroes) back.

Best,

agildr

[ShanmukhS_Intel]

Hi,

Thank you for sharing the details. The issue is reproducible with indices 534 and 1096 for which the results should have been 100 and -95. Instead, it is returning 0.

We are looking into your issue. We will get back to you soon.

Best Regards,

Shanmukh.SS

[Andrey_B_Intel]

Hi Agildr.

I confirm this issue with ippsThreshold_LTAbsVal_64f.

Fix will be available in one of next IPP releases.

Thanks.

Andrey.

[agildr]

Hi Andrey,

Thanks for the reply.

Can you maybe offer any insights on which constellations trigger this bug?

I'm asking because I have another test case, posting it below.

const int len = 10000;
var x = new double[len];
var y = new double[len];
x[1] = 100;
x[2] = -95;
unsafe
{
  fixed (double* xPtr = x, yPtr = y)
  {
    IPP.ippsThreshold_LTAbsVal_64f(xPtr, yPtr, len, 0.01, 1000);
  }
}

First elements of y:

1000
100	
-95	
1000
1000
0	
0	
1000
1000
0	
0	
1000
1000
0	
0	
1000

Best,

Luca