func.return %mm : tensor<1x3x3x12xf32>

// CHECK: %[[w:.*]] = "tfl.pseudo_qconst"() <{qtype = tensor<512x12x!quant.uniform<i8<-127:127>:f32, 1.000000e+00>>,
// CHECK: %[[mm:.*]] = "tfl.batch_matmul"(%arg0, %[[w]]) <{adj_x = false, adj_y = false
// CHECK-SAME: , asymmetric_quantize_inputs = true
// CHECK: return %[[mm:.*]]

// PerTensor: %[[w:.*]] = "tfl.pseudo_qconst"() <{qtype = tensor<512x12x!quant.uniform<i8<-127:127>:f32, 1.000000e+00>>,

  %1 = "tfl.pseudo_qconst"() {qtype = tensor<32x3x3x3x!quant.uniform<i8<-127:127>:f32, 0.021826678373682216:151>>, value = dense<-76> : tensor<32x3x3x3xi8>} : () -> tensor<32x3x3x3x!quant.uniform<i8<-127:127>:f32, 0.021826678373682216>>

// CHECK-NEXT: tfl.dequantize
}

// CHECK-LABEL: dequantize_pushdown_gather_with_reduction
func.func @dequantize_pushdown_gather_with_reduction(%arg0: tensor<2xi32>) -> tensor<2x2xf32> {
  %w = "tfl.pseudo_qconst"() {qtype = tensor<12x2x!quant.uniform<i8<-127:127>:f32, 1.000000e+00>>, value = dense<127> : tensor<12x2xi8>} : () -> tensor<12x2x!quant.uniform<i8<-127:127>:f32, 1.000000e+00>>

// CHECK-NEXT:   %{{.*}} = "tfl.pseudo_qconst"() <{qtype = tensor<32x3x3x3x!quant.uniform<u8<1:255>:f32, 0.021826678{{[0-9]*}}:151>>, value = dense<-76> : tensor<32x3x3x3xi8>}> : () -> tensor<32x3x3x3x!quant.uniform<u8<1:255>:f32, 0.021826678{{[0-9]*}}:151>>

  // CHECK-NEXT:       has_rank: true
  // CHECK-NEXT:    }, {
  // CHECK-NEXT:      shape: [ 3 ],
  // CHECK-NEXT:      type: INT8,
  // CHECK-NEXT:      buffer: 2,
  // CHECK-NEXT:      name: "tfl.pseudo_qconst",
  // CHECK-NEXT:      quantization: {
  // CHECK-NEXT:        scale: [ 0.1 ],
  // CHECK-NEXT:        zero_point: [ 0 ]
  // CHECK-NEXT:       },
  // CHECK-NEXT:       has_rank: true
  // CHECK-NEXT:    }, {

  // CHECK-NEXT:       has_rank: true
  // CHECK-NEXT:    }, {
  // CHECK-NEXT:      shape: [ 3 ],
  // CHECK-NEXT:      type: INT8,
  // CHECK-NEXT:      buffer: 2,
  // CHECK-NEXT:      name: "tfl.pseudo_qconst",
  // CHECK-NEXT:      quantization: {
  // CHECK-NEXT:        scale: [ 1.0 ],
  // CHECK-NEXT:        zero_point: [ 0 ]
  // CHECK-NEXT:       },
  // CHECK-NEXT:       has_rank: true
  // CHECK-NEXT:    }, {

      // uint8 doesn't require same operands and results scales.
      bool is_uint8 = !sign && (bit_width == 8);
      return !is_uint8;
    }
  }];
}

def TFL_ConstOp : Op<TFL_Dialect, "pseudo_const", [ConstantLike, Pure,
    FirstAttrDerivedResultType,
    QuantizableResult,
    DeclareOpInterfaceMethods<TFL_RuntimeVerification>]> {
  let summary = "Constant pseudo op.";

  let description = [{

                                           Type type, Location loc) {
  // If this is a constant bytes attribute or the result type doesn't match the
  // attribute type, then generate a tfl.pseudo_const.
  if (value.isa<ConstBytesAttr>() ||
      (value.isa<ElementsAttr>() &&
       value.cast<ElementsAttr>().getType() != type))
    return builder.create<ConstOp>(loc, type, value.cast<ElementsAttr>());

    For example, if the kernel does not support dynamic range quantization the
    graph will be converted into the following IR:

    %q_w = "tfl.pseudo_qconst"() {
         qtype = tensor<64x3x3x3x!quant.uniform<i8<-127:127>:f32, 1.000000e+00>>
    %w = "tfl.dequantize"(%q_w) :
         (tensor<64x3x3x3x!quant.uniform<i8<-127:127>:f32, 1.000000e+00>>) ->

# MLIR-SAME:  outputs = "output"
# MLIR:         %[[shape:.*]] = arith.constant dense<[1, -1, 31]> : tensor<3xi32>
# MLIR:         %[[bias:.*]] = "tfl.pseudo_qconst"() <{qtype = tensor<186x!quant.uniform<i32:f32:0
# MLIR:         %[[weight:.*]] = "tfl.pseudo_qconst"() <{qtype = tensor<186x1x1x256x!quant.uniform<i8<-127:127>:f32:0, {0.12581039038230116,