// RUN: flatbuffer_translate -mlir-to-tflite-flatbuffer %s -o - | flatbuffer_to_string - | FileCheck %s

func.func @main() -> tensor<3x2xi32> {
  %0 = "tfl.pseudo_const" () {value = dense<0> : tensor<3x2xi32>, tfl.is_variable} : () -> tensor<3x2xi32> loc("variable")
  func.return %0 : tensor<3x2xi32>
}

// CHECK:      {
// CHECK-NEXT:     version: 3,
// CHECK-NEXT:     operator_codes: [ ],
// CHECK-NEXT:     subgraphs: [ {
// CHECK-NEXT:       tensors: [ {

require (
	github.com/google/uuid v1.6.0
	github.com/jinzhu/now v1.1.5
	github.com/lib/pq v1.10.9
	github.com/stretchr/testify v1.9.0
	gorm.io/driver/mysql v1.5.7
	gorm.io/driver/postgres v1.5.9
	gorm.io/driver/sqlite v1.5.6
	gorm.io/driver/sqlserver v1.5.3
	gorm.io/gorm v1.25.10
)

require (
	filippo.io/edwards25519 v1.1.0 // indirect
	github.com/davecgh/go-spew v1.1.1 // indirect

	Log2E  = 1 / Ln2
	Ln10   = 2.30258509299404568401799145468436420760110148862877297603332790 // https://oeis.org/A002392
	Log10E = 1 / Ln10
)

// Floating-point limit values.
// Max is the largest finite value representable by the type.
// SmallestNonzero is the smallest positive, non-zero value representable by the type.
const (
	MaxFloat32             = 0x1p127 * (1 + (1 - 0x1p-23)) // 3.40282346638528859811704183484516925440e+38

// RUN: flatbuffer_translate -mlir-to-tflite-flatbuffer %s -o - | flatbuffer_translate --tflite-flatbuffer-to-mlir - -o - | FileCheck %s

// CHECK-LABEL: main
func.func @main(%arg0: tensor<1x224x224x3xf32>) -> tensor<1x401408xf32> {
// CHECK:   %{{.*}} = "tfl.quantize"(%{{.*}}) <{qtype = tensor<1x224x224x3x!quant.uniform<u8:f32, 7.812500e-03:128>>}> : (tensor<1x224x224x3xf32>) -> tensor<1x224x224x3x!quant.uniform<u8:f32, 7.812500e-03:128>>

namespace mlir {
namespace TFL {
namespace {
#define GEN_PASS_DEF_LEGALIZEHASHTABLESPASS
#include "tensorflow/compiler/mlir/lite/transforms/passes.h.inc"

// This file has Legalize hash tables pass which is responsible for:
// - Converting static hash table ops to the TFLite equivalent ops.
//
// There are needs to fall back to Flex for the following cases:
// - Mutable hash table cases

It is compatible with:

* PostgreSQL
* MySQL
* SQLite

In this example, we'll use **SQLite**, because it uses a single file and Python has integrated support. So, you can copy this example and run it as is.

        "//tensorflow/compiler/mlir/lite/experimental/tac:device_transform_patterns",
        "//tensorflow/compiler/mlir/lite/experimental/tac/hardwares:simple_hardware",
    ],
    alwayslink = 1,
)

tf_cc_binary(
    name = "example-hardware-translate",
    deps = [
        ":example_hardware",
        "//tensorflow/compiler/mlir/lite/experimental/tac:tac-translate-lib",

// RUN: flatbuffer_translate -mlir-to-tflite-flatbuffer %s -emit-custom-ops -o - | flatbuffer_to_string - | FileCheck %s

func.func @main(tensor<4x5xbf16>) -> tensor<4x5xbf16> {
^bb0(%arg0: tensor<4x5xbf16>):

// CHECK:  {
// CHECK-NEXT:      version: 3,
// CHECK-NEXT:      operator_codes: [ {
// CHECK-NEXT:        deprecated_builtin_code: 53,
// CHECK-NEXT:        version: 7,
// CHECK-NEXT:        builtin_code: CAST
// CHECK-NEXT:      } ],

	switch {
	case x == 0:
		return MinInt32
	case IsNaN(x):
		return MaxInt32
	case IsInf(x, 0):
		return MaxInt32
	}
	return ilogb(x)
}

// ilogb returns the binary exponent of x. It assumes x is finite and
// non-zero.
func ilogb(x float64) int {
	x, exp := normalize(x)
	return int((Float64bits(x)>>shift)&mask) - bias + exp

	// Inf or NaN or zero involved. At most one rounding will occur.
	if x == 0.0 || y == 0.0 || z == 0.0 || bx&uvinf == uvinf || by&uvinf == uvinf {
		return x*y + z
	}
	// Handle non-finite z separately. Evaluating x*y+z where
	// x and y are finite, but z is infinite, should always result in z.
	if bz&uvinf == uvinf {
		return z
	}

	// Inputs are (sub)normal.
	// Split x, y, z into sign, exponent, mantissa.