*   Added colocation constraints to `StagingArea`.
*   `sparse_matmul_op` reenabled for Android builds.
*   Restrict weights rank to be the same as the broadcast target, to avoid
    ambiguity on broadcast rules.
*   Upgraded libxsmm to 1.7.1 and applied other changes for performance and
    memory usage.
*   Fixed bfloat16 integration of LIBXSMM sparse mat-mul.

XC2 // 0xe1 VSLDOI $15, XC2, ZERO, XC2 // 0xe1... VSLDOI $1, ZERO, T0, T1 // ...1 VADDUBM XC2, XC2, XC2 // 0xc2... VSPLTISB $7, T2 VOR XC2, T1, XC2 // 0xc2....01 VSPLTB $0, H, T1 // most significant byte VSL H, T0, H // H<<=1 VSRAB T1, T2, T1 // broadcast carry bit VAND T1, XC2, T1 VXOR H, T1, IN // twisted H VSLDOI $8, IN, IN, H // twist even more ... VSLDOI $8, ZERO, XC2, XC2 // 0xc2.0 VSLDOI $8, ZERO, H, HL // ... and split VSLDOI $8, H, ZERO, HH STXVD2X VXC2, (XIP+R0) // save pre-computed table...

XC2 // 0xe1 VSLDOI $15, XC2, ZERO, XC2 // 0xe1... VSLDOI $1, ZERO, T0, T1 // ...1 VADDUBM XC2, XC2, XC2 // 0xc2... VSPLTISB $7, T2 VOR XC2, T1, XC2 // 0xc2....01 VSPLTB $0, H, T1 // most significant byte VSL H, T0, H // H<<=1 VSRAB T1, T2, T1 // broadcast carry bit VAND T1, XC2, T1 VXOR H, T1, IN // twisted H VSLDOI $8, IN, IN, H // twist even more ... VSLDOI $8, ZERO, XC2, XC2 // 0xc2.0 VSLDOI $8, ZERO, H, HL // ... and split VSLDOI $8, H, ZERO, HH STXVD2X VXC2, (XIP+R0) // save pre-computed table...