Open Radar

 
Summary:
Accelerate/LAPACK 32-bit float matrix inversion returns incorrect values (Haswell+10.11).

Summary:
Accelerate/LAPACK 32-bit float matrix inversion returns incorrect values.

Steps to Reproduce:
I could not isolate the problem outside of our app, but I can provide steps to reproduce the problem inside of our app.

Here's the code leading to erroneous results:
template<>
	bool matrix<float>::invert()  {
		//assert(cols == rows);
		__CLPK_integer n = rows;
		__CLPK_integer *ipiv = new __CLPK_integer[n];
		__CLPK_integer nwork = n*n;
		float *work = new float[nwork];
		__CLPK_integer info;
		bool res = false;
		
		// There is a pb with non-4/8 multiple on AVX2 (Haswell) + 10.11 (float, double are OK). 
		//			print("mat");
		sgetrf_(&n,&n,values(),&n,ipiv,&info);
		if(info == 0) {
			sgetri_(&n,values(),&n,ipiv,work,&nwork,&info);
			res = (info == 0);
		}
		//			print("imat");
		
		delete[] ipiv;
		delete[] work;
		
		return res;
	}

To reproduce the problem in Prizmo:
- Unzip the provided app on an Haswell or more recent Mac (supposedly AVX2)
- Open Prizmo.app
- Menu > Help > Sample Documents > Open Text Document
- Notice the document preview appears all black instead of actual contents.


Expected Results:
- Actual document contents.

Actual Results:
- All black image. The inverted matrix mentioned above has a number of 'nan' lines instead of correct values, as compared to pre-Haswell CPUs (same code)

Regression:
- Problem occurs with Haswell or more recent CPU, running OS X 10.11.x
- No problem with pre-10.11
- No problem with pre-Haswell
- No problem with double version (dgetrf_ / dgetri_)