message XlaAutoClusteringSummary {
  // Represents a single element in a histogram of ops ("op" as in "TensorFlow
  // operation").
  //
  // Next ID: 3
  message OpAndCount {
    // The TensorFlow operation (like MatMult, Add etc.)
    string op = 1;

    // The number of times this occurs.
    int32 count = 2;
  }

  // Describes a single XLA cluster.
  //
  // Next ID: 4
  message Cluster {
    string name = 1;

==============================================================================*/

// This transformation pass decomposes dense operations that assume
// support for hybrid quantization. These cases cover when a dense operation
// (e.g. matmul) has both quantized and unquantized inputs by dequantizing
// the quantized inputs, performing the operation in the expressed type, then
// requantizing if a quantized output is required.
//

  %dq_weight = "quantfork.dcast"(%q_weight) : (tensor<144x12x!quant.uniform<i8:f32, 0.074855112561992565:-1>>) -> tensor<144x12xf32>
  %9 = "tf.MatMul"(%7, %dq_weight) {transpose_a = false, transpose_b = false} : (tensor<*xf32>, tensor<144x12xf32>) -> tensor<*xf32>
  %10 = "quantfork.qcast"(%9) {volatile} : (tensor<*xf32>) -> tensor<*x!quant.uniform<i8:f32, 4.000000e-03:-12>>

             AbstractTensorHandle* const y, AbstractTensorHandle** z,
             const char* name = nullptr, const char* raw_device_name = nullptr);

// Multiply the matrix "a" by the matrix "b".
Status MatMul(AbstractContext* ctx, AbstractTensorHandle* const a,
              AbstractTensorHandle* const b, AbstractTensorHandle** product,
              bool transpose_a = false, bool transpose_b = false,

  }

  func.func private @composite_matmul_with_bias_fn_1(%arg0: tensor<1x4xf32>, %arg1: tensor<4x3xf32>, %arg2: tensor<3xf32>) -> tensor<1x3xf32> attributes {tf_quant.composite_function} {
    %0 = "tf.MatMul"(%arg0, %arg1) <{grad_a = false, grad_b = false, transpose_a = false, transpose_b = false}> {attr_map = "0:transpose_a,1:transpose_b", device = ""} : (tensor<1x4xf32>, tensor<4x3xf32>) -> tensor<1x3xf32>

                        std::vector<Output>* grad_outputs) {
  if (is_batch == false) {
    auto dx =
        MatMul(scope, x0, x1, MatMul::TransposeA(adj_x0).TransposeB(adj_x1));
    grad_outputs->push_back(dx);
    auto dy =
        MatMul(scope, y0, y1, MatMul::TransposeA(adj_y0).TransposeB(adj_y1));
    grad_outputs->push_back(dy);
  } else {
    auto dx = BatchMatMulV3(scope, x0, x1, x_data_type,

    } : (tensor<*xf32>, tensor<*xf32>) -> tensor<*xf32>
    %6 = "tf.Cast"(%5) : (tensor<*xf32>) -> tensor<*xi32>
    func.return %6 : tensor<*xi32>
  }

  // Matmul with int32 accumulation.
  func.func private @internal_matmul_fn(
                         %input : tensor<*xi8>, %weight : tensor<*xi8>,
                         %input_scale : tensor<*xf32>, %input_zp : tensor<*xi32>,

  TF_DeleteStatus(status);
  TFE_OpSetAttrType(op, "T", TFE_TensorHandleDataType(a));

  return op;
}

TFE_Op* MatMulOp(TFE_Context* ctx, TFE_TensorHandle* a, TFE_TensorHandle* b) {
  TF_Status* status = TF_NewStatus();

  TFE_Op* op = TFE_NewOp(ctx, "MatMul", status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_OpAddInput(op, a, status);

  // May have been filtered so check for lack of failure instead of success.
  EXPECT_EQ(compilation_status.Delta(kMlirWithFallbackModeFailure), 0);
}

TEST(LegalizeTFTest, MatMul) {
  static constexpr char kMatMulModuleStr[] = R"(
  module attributes {tf.versions = {bad_consumers = [], min_consumer = 0 : i32, producer = 268 : i32}} {
    func.func @main() -> (tensor<5x11xf32>) {

    return success();
  }
};

// Lowers `SparseMatMulOp` to `MatMulOp`, ignoring the sparseness hints,
// since we currently don't have an implementation that can use this
// information. Adds appropriate casts where necessary to align element types
// of operands and result for `MatMulOp`.
class LowerSparseMatMulOp : public RewritePattern {
 public: