// CHECK: %[[LHS_SPLIT:.*]]:6 = "tf.Split"(%[[SPLITTING_AXIS]], %[[LHS_RESHAPED]]) : (tensor<i32>, tensor<6x5x4xf32>) -> (tensor<1x5x4xf32>, tensor<1x5x4xf32>, tensor<1x5x4xf32>, tensor<1x5x4xf32>, tensor<1x5x4xf32>, tensor<1x5x4xf32>)
  // CHECK: %[[LHS_1:.*]] = "tf.Reshape"(%[[LHS_SPLIT]]#0, %[[MATMUL_LHS_SHAPE]]) : (tensor<1x5x4xf32>, tensor<2xi64>) -> tensor<5x4xf32>

      return %0 : tensor<1x1x4xf32>
    }
  )mlir";

  OwningOpRef<ModuleOp> module_op =
      ParseModuleOpString(kModuleXlaCallModuleOpWithDefaultQuantizationMethod);
  ASSERT_TRUE(module_op);

// CHECK-LABEL: QuantizeUnidirectionalLstmFullPerTensor
func.func @QuantizeUnidirectionalLstmFullPerTensor(%arg0: tensor<1x2x3xf32>) -> (tensor<1x2x3xf32>) {
  %input = "quantfork.stats"(%arg0) {layerStats = dense<[0.0, 1.0]> : tensor<2xf32>} : (tensor<1x2x3xf32>) -> tensor<1x2x3xf32>
  %1 = "tfl.pseudo_const"() {value = dense<[[0.1]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>

      return %0 : tensor<1x1x4xf32>
    }
  )mlir";

  const OwningOpRef<ModuleOp> module_op =
      ParseModuleOpString(kXlaCallModuleOpWithQuantizationMethodAttr);
  ASSERT_TRUE(module_op);

}

func.func @einsum_matrixdotprod(%arg0: tensor<2x5x7x3xf32>, %arg1: tensor<7x3x4xf32>) -> tensor<2x5x4xf32> {
  %0 = "tf.Einsum"(%arg0, %arg1) {T = "tfdtype$DT_FLOAT", equation = "bfnd,ndh->bfh"}: (tensor<2x5x7x3xf32>, tensor<7x3x4xf32>) -> tensor<2x5x4xf32>
  func.return %0 : tensor<2x5x4xf32>
  // CHECK-LABEL: einsum_matrixdotprod
  // CHECK-DAG: %[[cst:.*]] = arith.constant dense<[2, 5, 21]> : tensor<3xi64>

  func.return %0, %1, %2 : tensor<1x4x4xf32>, tensor<1x4x4xf32>, tensor<14x4x4xf32>
}

// -----

func.func @testSplitVOpWithBadSizeSplitsConstantSum(%arg0: tensor<16x4x4xf32>) -> (tensor<0x4x4xf32>, tensor<16x4x4xf32>) {

    %14 = stablehlo.multiply %12, %13 : tensor<1x4x3xf32>  // s1 * s2
    %15 = call @uniform_quantize_1(%14, %5, %6) : (tensor<1x4x3xf32>, tensor<1x1x1xf32>, tensor<1x1x1xi8>) -> tensor<1x4x3xi8>
    %16 = call @uniform_dequantize_0(%15, %5, %6) : (tensor<1x4x3xi8>, tensor<1x1x1xf32>, tensor<1x1x1xi8>) -> tensor<1x4x3xf32>
    return %16 : tensor<1x4x3xf32>
  }

    func.func @main() -> (tensor<1x4x4xf32>) {
      %%arg0 = "tf.Const"() {value = dense<-3.0> : tensor<1x4x2xf32>} : () -> tensor<1x4x2xf32>
      %%arg1 = "tf.Const"() {value = dense<-3.0> : tensor<1x2x4xf32>} : () -> tensor<1x2x4xf32>

      %%1 = "tf.%s"(%%arg0, %%arg1) {T = f32, adj_x = false, adj_y = false, grad_x = false, grad_y = false, device = ""} : (tensor<1x4x2xf32>, tensor<1x2x4xf32>) -> tensor<1x4x4xf32>

  // CHECK-LABEL: test_batch_matmul_adj_to_einsum
  // CHECK: %[[RES_EINSUM:[0-9]*]] = "tf.Einsum"(%arg0, %arg1) <{equation = "...mk,...nk->...mn"}> : (tensor<1x2x3xf32>, tensor<4x3xf32>) -> tensor<1x2x4xf32>
  // CHECK: return %[[RES_EINSUM]] : tensor<1x2x4xf32>
  %0 = "tf.BatchMatMul"(%arg0, %arg1) {adj_x = false, adj_y = true} : (tensor<1x2x3xf32>, tensor<4x3xf32>) -> tensor<1x2x4xf32>

//CHECK:func.func private @slice(%arg0: tensor<160x20x1xf32>) -> tensor<1x1x1xf32> {
//CHECK-NEXT: %0 = "vhlo.slice_v1"(%arg0) <{limit_indices = #vhlo.tensor_v1<dense<0> : tensor<3xi64>>, start_indices = #vhlo.tensor_v1<dense<1> : tensor<3xi64>>, strides = #vhlo.tensor_v1<dense<1> : tensor<3xi64>>}> : (tensor<160x20x1xf32>) -> tensor<1x1x1xf32> 
//CHECK-NEXT: return %0 : tensor<1x1x1xf32>
//CHECK-NEXT:}