Details:

Due to uninitialized internal state, the Matrix Math Accelerator (MMA) attached to the C71x may report errors in the FirstErrorCode and LastErrorCode fields of the HWA_STATUS register after power-on. Because these fields are sticky, any subsequent HWARCV instruction may throw a C71x exception.

Workaround(s):

After power-on, a short instruction sequence running on the C71x can initialize the internal MMA state before the first execution of normal MMA operation. Only one execution of the sequence is required.

The sequence generates a valid HWA_CONFIG and HWA_OFFSET value, loads it into the MMA, then clears the sticky error codes.

The sequence, in C71x assembly code is:

PROT 
    MVK32 .M2 0x0,B0 ; clear low word of VB0 
    VDUPW .C2 B0,VB0 ; duplicate word across VB0 
    HWAOPEN .L2 VB0,VB0,0 ; clear HWA_CONFIG and HWA_OFFSET 
    HWACLOSE .S1 0 ; clear any error conditions