l2_38 := l2 * 38 l3_38 := l3 * 38 l3_19 := l3 * 19 l4_19 := l4 * 19 // r0 = l0×l0 + 19×(l1×l4 + l2×l3 + l3×l2 + l4×l1) = l0×l0 + 19×2×(l1×l4 + l2×l3) r0 := mul64(l0, l0) r0 = addMul64(r0, l1_38, l4) r0 = addMul64(r0, l2_38, l3) // r1 = l0×l1 + l1×l0 + 19×(l2×l4 + l3×l3 + l4×l2) = 2×l0×l1 + 19×2×l2×l4 + 19×l3×l3 r1 := mul64(l0_2, l1) r1 = addMul64(r1, l2_38, l4) r1 = addMul64(r1, l3_19, l3) // r2 = l0×l2 + l1×l1 + l2×l0 + 19×(l3×l4 + l4×l3) = 2×l0×l2 + l1×l1 + 19×2×l3×l4 r2 := mul64(l0_2, l2) r2 =...

### Bug Fixes and Other Changes

* GPU
    * Support for NVIDIA GPUs with compute capability 8.9 (e.g. L4 & L40) has been added to TF binary distributions (Python wheels).
* Replace `DebuggerOptions` of TensorFlow Quantizer, and migrate to `DebuggerConfig` of StableHLO Quantizer.