// addition/subtraction of arguments represented by Bits values with the
// respective Float addition/subtraction for a variety of precisions
// and rounding modes.
func TestFloatAdd(t *testing.T) {
	for _, xbits := range bitsList {
		for _, ybits := range bitsList {
			// exact values
			x := xbits.Float()
			y := ybits.Float()
			zbits := xbits.add(ybits)
			z := zbits.Float()

	mgf1XOR(db, hash, h)

	// 11. Set the leftmost 8 * emLen - emBits bits of the leftmost octet in
	//     maskedDB to zero.

	db[0] &= 0xff >> (8*emLen - emBits)

	// 12. Let EM = maskedDB || H || 0xbc.
	em[emLen-1] = 0xbc

	// 13. Output EM.
	return em, nil
}

func emsaPSSVerify(mHash, em []byte, emBits, sLen int, hash hash.Hash) error {
	// See RFC 8017, Section 9.1.2.

func bitSetTest(x int) bool {
	// amd64:"ANDL\t[$]9, AX"
	// amd64:"CMPQ\tAX, [$]9"
	return x&9 == 9
}

// mask contiguous one bits
func cont1Mask64U(x uint64) uint64 {
	// s390x:"RISBGZ\t[$]16, [$]47, [$]0,"
	return x & 0x0000ffffffff0000
}

// mask contiguous zero bits
func cont0Mask64U(x uint64) uint64 {
	// s390x:"RISBGZ\t[$]48, [$]15, [$]0,"
	return x & 0xffff00000000ffff
}

  %dq_weight = "quantfork.dcast"(%q_weight) : (tensor<2x3x3x2x!quant.uniform<i8:f32, 0.074855112561992565:-1>>) -> tensor<2x3x3x2xf32>
  %q_bias = "quantfork.qcast"(%bias) : (tensor<2xf32>) -> tensor<2x!quant.uniform<i32:f32, 0.044022349891595126>>
  %dq_bias = "quantfork.dcast"(%q_bias) : (tensor<2x!quant.uniform<i32:f32, 0.044022349891595126>>) -> tensor<2xf32>

  %dq_weight = "quantfork.dcast"(%q_weight) : (tensor<2x3x3x2x!quant.uniform<i8:f32, 0.074855112561992565:-1>>) -> tensor<2x3x3x2xf32>
  %q_bias = "quantfork.qcast"(%bias) : (tensor<2xf32>) -> tensor<2x!quant.uniform<i32:f32, 0.044022349891595126>>
  %dq_bias = "quantfork.dcast"(%q_bias) : (tensor<2x!quant.uniform<i32:f32, 0.044022349891595126>>) -> tensor<2xf32>

}

// writeLSB writes the code c for "Least Significant Bits first" data.
func (w *Writer) writeLSB(c uint32) error {
	w.bits |= c << w.nBits
	w.nBits += w.width
	for w.nBits >= 8 {
		if err := w.w.WriteByte(uint8(w.bits)); err != nil {
			return err
		}
		w.bits >>= 8
		w.nBits -= 8
	}
	return nil
}

// writeMSB writes the code c for "Most Significant Bits first" data.
func (w *Writer) writeMSB(c uint32) error {

// bytes, and throws if it finds an unpinned Go pointer. The gcbits mark each
// pointer value. The src pointer is off bytes into the gcbits.
//
//go:nosplit
//go:nowritebarrier
func cgoCheckBits(src unsafe.Pointer, gcbits *byte, off, size uintptr) {
	skipMask := off / goarch.PtrSize / 8
	skipBytes := skipMask * goarch.PtrSize * 8
	ptrmask := addb(gcbits, skipMask)
	src = add(src, skipBytes)
	off -= skipBytes

		if s.state.get() != mSpanInUse {
			continue
		}

		// Find all allocated objects.
		abits := s.allocBitsForIndex(0)
		for i := uintptr(0); i < uintptr(s.nelems); i++ {
			if abits.index < uintptr(s.freeindex) || abits.isMarked() {
				x := s.base() + i*s.elemsize
				trace.HeapObjectExists(x, s.typePointersOfUnchecked(x).typ)
			}
			abits.advance()
		}
	}

	// Write out all the goroutine stacks.
	forEachGRace(func(gp *g) {

// RUN: tf-opt %s -tfl-prepare-quantize="quantize-signed=true post-training-quantize=true activation-number-of-bits=16" -cse | FileCheck %s

// CHECK-LABEL: QuantizeUnidirectionalLstmFullPerTensor
func.func @QuantizeUnidirectionalLstmFullPerTensor(%arg0: tensor<1x2x3xf32>) -> (tensor<1x2x3xf32>) {
  %input = "quantfork.stats"(%arg0) {layerStats = dense<[0.0, 1.0]> : tensor<2xf32>} : (tensor<1x2x3xf32>) -> tensor<1x2x3xf32>

      bias_shape: Optional[Sequence[int]] = None,
  ) -> module.Module:
    class EinsumModel(module.Module):
      """Einsum class."""

      def __init__(self):
        self._bias = None
        if bias_shape is not None:
          self._bias = array_ops.constant(
              np.random.uniform(size=bias_shape), dtype=dtypes.float32
          )

        self._kernel = np.random.uniform(size=y_shape).astype('f4')