// CHECK: %[[scast:.*]] = "quantfork.scast"(%[[conv]]
// CHECK: %[[fcast:.*]] = "tf.Cast"(%[[scast]]) <{Truncate = false}> : (tensor<*xi8>) -> tensor<*xf32>
// CHECK: %[[avgpool_f32:.*]] = "tf.AvgPool"(%[[fcast]])
// CHECK-SAME: (tensor<*xf32>) -> tensor<*xf32>
// CHECK: %[[round:.*]] = "tf.Round"(%[[avgpool_f32]])
// CHECK: %[[icast:.*]] = "tf.Cast"(%[[round]]) <{Truncate = false}> : (tensor<*xf32>) -> tensor<*xi8>

// A QuantizeCast (qcast) represents a potential type shift from a quantizable
// type to a quantized type.
//
// At runtime, a qcast will apply the transformation expressed by its
// operand and result type. For flexibility during transformation, it is also
// possible to have a qcast that performs no transformation (both its
// operand and result type are quantizable).
//
// A qcast will typically originate from either:

// A QuantizeCast (qcast) represents a potential type shift from a quantizable
// type to a quantized type.
//
// At runtime, a qcast will apply the transformation expressed by its
// operand and result type. For flexibility during transformation, it is also
// possible to have a qcast that performs no transformation (both its
// operand and result type are quantizable).
//
// A qcast will typically originate from either:

    %0 = "quantfork.qcast"(%cst) : (tensor<2x2x3x2xf32>) -> tensor<2x2x3x2x!quant.uniform<i8<-127:127>:f32:3, {4.000000e-03,5.000000e-03}>>
    %1 = "quantfork.dcast"(%0) : (tensor<2x2x3x2x!quant.uniform<i8<-127:127>:f32:3, {4.000000e-03,5.000000e-03}>>) -> tensor<*xf32>
    %2 = "quantfork.qcast"(%arg0) : (tensor<1x2x2x3xf32>) -> tensor<1x2x2x3x!quant.uniform<i8:f32, 8.000000e-03>>

    %5 = "quantfork.dcast"(%4) : (tensor<2x3x1024x3x!quant.uniform<i8<-127:127>:f32, 4.000000e-01>>) -> tensor<2x3x1024x3xf32>
    %6 = "quantfork.qcast"(%arg0) {volatile} : (tensor<2x3x1x1024xf32>) -> tensor<2x3x1x1024x!quant.uniform<i8:f32, 5.000000e-01:2>>

    // CHECK: %[[DQ:.*]] = "quantfork.dcast"(%[[RESHAPE]]) : (tensor<3x1x!quant.uniform<i8:f32, 0.13170163023705575:-1>>) -> tensor<3x1xf32>
    // CHECK: return %[[DQ]]
    %0 = "quantfork.qcast"(%arg0) {volatile} : (tensor<1x2xf32>) -> tensor<1x2x!quant.uniform<i8:f32, 5.000000e-03>>
    %1 = "quantfork.dcast"(%0) : (tensor<1x2x!quant.uniform<i8:f32, 5.000000e-03>>) -> tensor<1x2xf32>

// CHECK: %[[CST:.+]] = "tf.Const"() <{value = dense<3.000000e-01> : tensor<2x3x3x2xf32>}> : () -> tensor<2x3x3x2xf32>
// CHECK: %[[Q:.+]] = "quantfork.qcast"(%[[CST]]) : (tensor<2x3x3x2xf32>) -> tensor<2x3x3x2x!quant.uniform<i8<-127:127>:f32, 0.0023622048182750312>>
// CHECK: %[[DQ:.+]] = "quantfork.dcast"(%[[Q]]) : (tensor<2x3x3x2x!quant.uniform<i8<-127:127>:f32, 0.0023622048182750312>>) -> tensor<2x3x3x2xf32>

    %0 = "quantfork.qcast"(%cst) : (tensor<2x2x3x2xf32>) -> tensor<2x2x3x2x!quant.uniform<i8<-127:127>:f32:3, {4.000000e-03,5.000000e-03}>>
    %1 = "quantfork.dcast"(%0) : (tensor<2x2x3x2x!quant.uniform<i8<-127:127>:f32:3, {4.000000e-03,5.000000e-03}>>) -> tensor<*xf32>
    %2 = "quantfork.qcast"(%arg0) : (tensor<1x2x2x3xf32>) -> tensor<1x2x2x3x!quant.uniform<i8:f32, 8.000000e-03>>

  std::vector<Value> sliced_lhs =
      sliceInput(input_lhs, bcast.x_batch_size(), loc, rewriter);
  std::vector<Value> sliced_rhs =
      sliceInput(input_rhs, bcast.y_batch_size(), loc, rewriter);

  // Compute (single batch) MatMul for each output batch.
  std::vector<Value> matmuls;
  matmuls.reserve(bcast.output_batch_size());
  for (int batch_idx : llvm::seq<int>(0, bcast.output_batch_size())) {
    int lhs_batch_idx, rhs_batch_idx;

  ];
}

def PrepareQuantizePass : Pass<"stablehlo-prepare-quantize", "mlir::ModuleOp"> {
  let summary = "Prepare StableHLO dialect for static range quantization by converting quantfork.stats into quantfork.qcast and dcast ops.";
  let options = [
    Option<"enable_per_channel_quantized_weight_",
        "enable-per-channel-quantized-weight",
        "bool", /*default=*/"true",
        "Whether to enable per-channel quantized weights.">,