// Zero point computation.
  // In float, solve the affine equation for any known pair
  // (real value, corresponding quantized value), of which, two such pairs
  // are known: (rmin, qmin), (rmax, qmax).
  // The arithmetic error on the zero point computed from either pair will be
  // roughly machine_epsilon * (sum of absolute values of terms).
  // Use the variant that adds the smaller error.

  (replaceWithValue $arg)>;

// Remove StopGradient op
def RemoveStopGradient : Pat<
  (TF_StopGradientOp $arg),
  (replaceWithValue $arg)>;

// Converts tf.FusedBatchNormV3 into a sequence of more primitive arithmetic
// operations. Specifically, performs the following calculation:
//
//   (x - mean) * scale / sqrt(variance + epsilon) + offset
//
// Let multiplier = scale / sqrt(variance + epsilon),
// to compute

// license that can be found in the LICENSE file.

//go:build !math_big_pure_go

#include "textflag.h"

// This file provides fast assembly versions for the elementary
// arithmetic operations on vectors implemented in arith.go.

// func addVV(z, x, y []Word) (c Word)
TEXT ·addVV(SB),NOSPLIT,$0
	MOVL z+0(FP), DI
	MOVL x+12(FP), SI
	MOVL y+24(FP), CX
	MOVL z_len+4(FP), BP

// license that can be found in the LICENSE file.

//go:build !math_big_pure_go

#include "textflag.h"

// This file provides fast assembly versions for the elementary
// arithmetic operations on vectors implemented in arith.go.

// func addVV(z, x, y []Word) (c Word)
TEXT ·addVV(SB),NOSPLIT,$0
	ADD.S	$0, R0		// clear carry flag
	MOVW	z+0(FP), R1
	MOVW	z_len+4(FP), R4
	MOVW	x+12(FP), R2

	a[i] = bits.RotateLeft32(a[i], r) & 0xFFF00FFF
	i++
	// ppc64x: "RLWNM\t[$]4, R[0-9]+, [$]20, [$]11, R[0-9]+"
	a[i] = bits.RotateLeft32(a[i], 4) & 0xFFF00FFF
	i++
}

// combined arithmetic and rotate on arm64
func checkArithmeticWithRotate(a *[1000]uint64) {
	// arm64: "AND\tR[0-9]+@>51, R[0-9]+, R[0-9]+"
	a[2] = a[1] & bits.RotateLeft64(a[0], 13)
	// arm64: "ORR\tR[0-9]+@>51, R[0-9]+, R[0-9]+"

import com.google.common.annotations.J2ktIncompatible;
import java.math.BigInteger;
import javax.annotation.CheckForNull;

/**
 * A wrapper class for unsigned {@code int} values, supporting arithmetic operations.
 *
 * <p>In some cases, when speed is more important than code readability, it may be faster simply to
 * treat primitive {@code int} values as unsigned, using the methods from {@link UnsignedInts}.
 *

import com.google.common.annotations.J2ktIncompatible;
import java.math.BigInteger;
import javax.annotation.CheckForNull;

/**
 * A wrapper class for unsigned {@code int} values, supporting arithmetic operations.
 *
 * <p>In some cases, when speed is more important than code readability, it may be faster simply to
 * treat primitive {@code int} values as unsigned, using the methods from {@link UnsignedInts}.
 *

// input and value tensors. It also accepts the copy callback provided by
// pluggable vendor to do the copying of the tensors and the update callback to
// apply the arithmetic operation. The caller takes ownership of the `source`,
// `dest`, `tensor` and `value` tensors and is responsible for freeing them with
// TF_DeleteTensor.
TF_CAPI_EXPORT extern void TF_AssignUpdateVariable(

		// Otherwise use the raw address, so that things like relative
		// jumps show up as JMP 0x123 instead of JMP f+10(SB).
		// It is usually easier to search for 0x123 than to do the mental
		// arithmetic to find f+10.
		addr := pc + uint64(inst.Len) + uint64(a)
		if s, base := symname(addr); s != "" && addr == base {
			return fmt.Sprintf("%s(SB)", s)
		}
		return fmt.Sprintf("%#x", addr)

	case Imm:

	_ = unsafe.Sizeof(func() { type _ [iota /* ERROR "iota outside constant decl" */ ]byte })
	_ = unsafe.Sizeof(func() { func() int { return iota /* ERROR "iota outside constant decl" */ }() })
)

// constant arithmetic precision and rounding must lead to expected (integer) results
var _ = []int64{
	0.0005 * 1e9,
	0.001 * 1e9,
	0.005 * 1e9,
	0.01 * 1e9,
	0.05 * 1e9,
	0.1 * 1e9,
	0.5 * 1e9,
	1 * 1e9,
	5 * 1e9,