%1 = "tfl.quantize"(%arg1) {qtype = tensor<2x1x!quant.uniform<u8:f32, 1.0:128>>}: (tensor<2x1xf32>) -> tensor<2x1x!quant.uniform<u8:f32, 1.0:128>>
  %2 = "tfl.dequantize"(%0) : (tensor<2x2x!quant.uniform<u8:f32, 1.0:128>>) -> tensor<2x2xf32>
  %3 = "tfl.dequantize"(%1) : (tensor<2x1x!quant.uniform<u8:f32, 1.0:128>>) -> tensor<2x1xf32>

^bb0(%arg0: tensor<1x56x56x24x!quant.uniform<u8:f32, 0.27583434161017922:119>>, %arg1: tensor<1x56x56x24x!quant.uniform<u8:f32, 0.40149296779258581:136>>):

  %0 = "tfl.quantize"(%arg0) {qtype = tensor<4x!quant.uniform<u8:f32, 1.0>>} : (tensor<4xf32>) -> tensor<4x!quant.uniform<u8:f32, 1.0>>
  %1:4 = "tfl.split"(%arg1, %0) {num_splits = 4 : i32} : (tensor<i32>, tensor<4x!quant.uniform<u8:f32, 1.0>>)
  -> (tensor<2x!quant.uniform<u8:f32, 1.0>>, tensor<2x!quant.uniform<u8:f32, 1.0>>,tensor<2x!quant.uniform<u8:f32, 1.0>>, tensor<2x!quant.uniform<u8:f32, 1.0>>)

// no signed byte load.

func set16(x8 int8, u8 *uint8, y8 int8, z8 uint8) {
	// Truncate not needed, load does sign/zero extend

	// ppc64x:-"MOVBZ\tR\\d+,\\sR\\d+"
	val16[0] = uint16(*u8)

	// AND not needed due to size
	// ppc64x:-"ANDCC"
	sval16[1] = 255 & int16(x8+y8)

	// ppc64x:-"ANDCC"
	val16[1] = 255 & uint16(*u8+z8)

}
func shiftidx(u8 *uint8, x16 *int16, u16 *uint16) {

func.func @QuantizeMaximum(tensor<1x6x6x16x!quant.uniform<u8:f32, 0.1>>, tensor<1x6x6x16x!quant.uniform<u8:f32, 0.1>>) -> tensor<1x6x6x16xf32> {
^bb0(%arg0: tensor<1x6x6x16x!quant.uniform<u8:f32, 0.1>>, %arg1: tensor<1x6x6x16x!quant.uniform<u8:f32, 0.1>>):
  %0 = "tfl.dequantize"(%arg0) : (tensor<1x6x6x16x!quant.uniform<u8:f32, 0.1>>) -> tensor<1x6x6x16xf32>
  %1 = "tfl.dequantize"(%arg1) : (tensor<1x6x6x16x!quant.uniform<u8:f32, 0.1>>) -> tensor<1x6x6x16xf32>

  %cst = "tfl.pseudo_qconst"() {qtype = tensor<1x2x!quant.uniform<u8:f32, 1.0>>, value = dense<-76> : tensor<1x2xi8>} : () -> tensor<1x2x!quant.uniform<u8:f32, 1.0>>
  %2 = "tfl.concatenation"(%1, %cst) {axis = 0 : i32, fused_activation_function = "NONE"} : (tensor<1x2x!quant.uniform<u8:f32, 1.0>>, tensor<1x2x!quant.uniform<u8:f32, 1.0>>) -> tensor<2x2x!quant.uniform<u8:f32, 1.0>>

  func.return %6 : tensor<1x112x112x32x!quant.uniform<u8:f32, 0.023528476789885875>>
}

// DEBUG-LABEL: QuantizeSplit
func.func @QuantizeSplit(%arg: tensor<4x!quant.uniform<u8:f32, 1.0>>, %cst: tensor<i32>) -> (tensor<2x!quant.uniform<u8:f32, 1.0>>,tensor<2x!quant.uniform<u8:f32, 1.0>>) {
  %0 = "tfl.dequantize"(%arg) : (tensor<4x!quant.uniform<u8:f32, 1.0>>) -> tensor<4xf32>

  %5 = "tfl.reshape"(%4, %0) : (tensor<1x112x112x32x!quant.uniform<u8:f32, 0.023528476789885875>>, tensor<2xi32>) -> tensor<1x401408x!quant.uniform<u8:f32, 0.023528476789885875>>

}

func.func @qu8() -> tensor<3x!quant.uniform<u8<1:255>:f32, 1.0>> {
  // CHECK-LABEL: @qu8
  // CHECK: <{qtype = tensor<3x!quant.uniform<u8<1:255>:f32, 1.000000e+00>>, value = dense<1> : tensor<3xi8>}> : () -> tensor<3x!quant.uniform<u8<1:255>:f32, 1.000000e+00>>
  %0 = "tfl.pseudo_qconst"() { qtype = tensor<3x!quant.uniform<u8<1:255>:f32, 1.0>>, value = dense<1> : tensor<3xi8>} : () -> tensor<3x!quant.uniform<u8<1:255>:f32, 1.0>>

# MLIR-NEXT:    return %[[add]] : tensor<4x!quant.uniform<u8:f32, 0.0078431372549019607:128>>
# MLIR-NEXT:  }

# CHECK-LABEL: {
# CHECK-NEXT:  version: 3,
# CHECK-NEXT:  operator_codes: [ {
# CHECK-NEXT:    version: 1