%11 = "tf.Conv2D"(%8, %arg4)
       {
         data_format = "NCHW",
         dilations = [1, 1, 1, 1],
         explicit_paddings = [],
         padding = "VALID",
         strides = [1, 1, 1, 1]
       } : (tensor<?x64x56x56xf32>, tensor<1x1x64x256xf32>) -> tensor<?x256x56x56xf32>

  // CHECK: %[[CONV2:[0-9]*]] = "tf.Conv2D"(%[[MAX_POOL]], %arg4)
  // CHECK-SAME: data_format = "NHWC"

fused with `automatic double transpose` to reduce extra overhead on the host.

### Extend from Conv2D to Conv3D

SpaceToDepth not only helps with 2D image models but also 3D image models such
as I3D. The plan is to apply automatic space to depth for Conv2D as the first

// CHECK-NOT: func private @internal_conv2d_fn
// CHECK-NOT: func private @internal_matmul_fn
// CHECK: func private @quantized_conv2d_with_bias_fn
// CHECK-SAME: tf_quant.quantized_ops = ["Conv2D", "BiasAdd"]
// CHECK: func private @quantized_conv2d_with_bias_and_relu_fn
// CHECK: func private @quantized_conv2d_with_bias_and_relu6_fn
// CHECK: func private @quantized_conv2d_fn

                                PatternRewriter& rewriter) const override;
};

// Ensure bias for conv2d op.
struct EnsureBiasForConv2d : public OpRewritePattern<TFL::Conv2DOp> {
  using OpRewritePattern<TFL::Conv2DOp>::OpRewritePattern;

  LogicalResult matchAndRewrite(TFL::Conv2DOp conv_op,
                                PatternRewriter& rewriter) const override;
};

// Pad slice to 4d.

Nigel Tao <******@****.***> 1341270545 +1000

  parameters[
    {"quantized_ops": ["MatMul"], "internal_func_name": "internal_matmul_fn"},
    {"quantized_ops": ["Conv2D"], "internal_func_name": "internal_conv2d_fn"},
    {"quantized_ops": ["DepthwiseConv2D"], "internal_func_name": "internal_depthwise_conv2d_fn"},
    {"quantized_ops": ["Conv3D"], "internal_func_name": "internal_conv3d_fn"},
    {"quantized_ops": ["BatchMatMul"], "internal_func_name": "internal_batch_matmul_fn"}
  ]

  %2 = "tf.Cast"(%1) {Truncate = false} : (tensor<1x3x2x2xbf16>) -> tensor<1x3x2x2xf32>
  %3 = "tf.IdentityN"(%2) {device = ""} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>
  return %3 : tensor<1x3x2x2xf32>
// CHECK: tf.Conv2D
// CHECK-SAME: loc(callsite("Model/conv2d@conv2d_with_valid_loc"("Conv2D") at "QuantizationUnit({{.*}})"))
}

    %1 = "tf.Conv2D"(%0, %cst) {data_format = "NHWC", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true}
        : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<1x3x2x2xf32> loc(fused["Conv2D:", "Model/conv2d"])
    %2 = "tf.IdentityN"(%1) {device = ""} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>

  %4 = "tf.Transpose"(%2, %3) : (tensor<1x32x32x8xf32>, tensor<4xi32>) -> tensor<1x8x32x32xf32>

  // Check that Conv2D computed in NCHW format, and all redundant transpose
  // operations removed from the function.

  // CHECK: %[[CONV:[0-9]*]] = "tf.Conv2D"(%arg0, %arg1)
  // CHECK-SAME: data_format = "NCHW"
  // CHECK-SAME: -> tensor<1x8x32x32xf32>

  // CHECK: return %[[CONV]]

      if (function_name.contains("with_bias")) {
        spec->biases_params[2] = {{0, 1},
                                  quant::GetUniformQuantizedTypeForBias};
      }
    } else if (function_name.contains("conv2d")) {
      spec->coeff_op_quant_dim[1] = 3;
      if (function_name.contains("with_bias")) {
        spec->biases_params[2] = {{0, 1},
                                  quant::GetUniformQuantizedTypeForBias};
      }