}
				if ft.Size() == 8 {
					return s.float64ToUint32(n, v, ft, tt)
				}
			}
		}

		if !ok {
			s.Fatalf("weird float conversion %v -> %v", ft, tt)
		}
		op1, op2, it := conv.op1, conv.op2, conv.intermediateType

		if op1 != ssa.OpInvalid && op2 != ssa.OpInvalid {
			// normal case, not tripping over unsigned 64
			if op1 == ssa.OpCopy {
				if op2 == ssa.OpCopy {
					return v
				}

};

using ConvertConv2DDynamic =
    ConvertConvDynamic<TF::Conv2DOp, /*num_spatial_dims=*/2>;

// Converts the TensorFlow conv op in template to the generic HLO conv op by
// converting TensorFlow op attributes to HLO op attributes.
//
// Sample result for Conv2D:
//
//   %conv = "mhlo.convolution"(%input, %filter) {
//     strides = [1, 2],
//     paddings = [[1, 0], [1, 1]],
//     ...
//   }
//

  // CHECK-DAG: %[[ONE:.*]] = mhlo.constant dense<1.000000e+00> : tensor<f32>
  // CHECK: %[[CONV:.*]] = mhlo.convert %arg0 : (tensor<3xi32>) -> tensor<3xi64>
  // CHECK: %[[F32:.*]] = "mhlo.rng"(%[[ZERO]], %[[ONE]], %[[CONV]]) {{.*UNIFORM.*}} -> tensor<12x?x64xf32>
  %0 = "tf.RandomUniform"(%arg0) : (tensor<3xi32>) -> tensor<12x?x64xf32>
  // CHECK: return %[[F32]]

  // CHECK: %[[dq:.*]] = "tfl.dequantize"(%[[q]])
  // CHECK: %[[conv:.*]] = "tfl.conv_2d"(%arg0, %[[dq]], %[[cst]])
  // CHECK: return %[[conv]] : tensor<256x8x7x3xf32>
}

// CHECK-LABEL: @fuseMulIntoFullyConnectedWithOptionalAttribute