TFLiteConverter: Support QuantizeAndDequantizeV4

Author: lgeiger

tensorflow/compiler/mlir/lite/tests/legalize-tf.mlir

@@ -2068,3 +2068,15 @@ func @all_i64axes(%arg0: tensor<8x16x16xi1>, %arg1: tensor<2xi64>) -> tensor<?xi
   // CHECK: %[[V0:.*]] = "tfl.cast"(%arg1) : (tensor<2xi64>) -> tensor<2xi32>
   // CHECK: "tfl.reduce_all"(%arg0, %[[V0]]) {keep_dims = false} : (tensor<8x16x16xi1>, tensor<2xi32>) -> tensor<?xi1>
 }
+
+func @quantize_dequantize_v4(%arg0 : tensor<?x?xf32>) -> tensor<?x?xf32> {
+  %cst = constant dense<0.0> : tensor<f32>
+  %cst_0 = constant dense<255.0> : tensor<f32>
+  %0 = "tf.QuantizeAndDequantizeV4"(%arg0, %cst, %cst_0) : (tensor<?x?xf32>, tensor<f32>, tensor<f32>) -> tensor<?x?xf32>
+  return %0 : tensor<?x?xf32>
+
+// CHECK-LABEL: quantize_dequantize_v4
+// CHECK:  %[[QUANT:.*]] = "tfl.quantize"(%arg0) {qtype = tensor<?x?x!quant.uniform<u8:f32, 1.000000e+00>>} : (tensor<?x?xf32>) -> tensor<?x?x!quant.uniform<u8:f32, 1.000000e+00>>
+// CHECK:  %[[DEQUANT:.*]] = "tfl.dequantize"(%[[QUANT]]) : (tensor<?x?x!quant.uniform<u8:f32, 1.000000e+00>>) -> tensor<?x?xf32>
+// CHECK:  return %[[DEQUANT]]
+}

tensorflow/compiler/mlir/lite/transforms/legalize_patterns.td

@@ -297,8 +297,8 @@ def LegalizeFakeQuantWithMinMaxVars: Pat<
 
 // TODO(rocky): Not all of the attributes are handled correctly.  Make this
 // more general if there is a need.
-def LegalizeQuantizeAndDequantizeV2 : Pat<
-  (TF_QuantizeAndDequantizeV2Op $inputs, (ConstantOp F32ElementsAttr:$min),
+def LegalizeQuantizeAndDequantizeV4 : Pat<
+  (TF_QuantizeAndDequantizeV4Op $inputs, (ConstantOp F32ElementsAttr:$min),
     (ConstantOp F32ElementsAttr:$max),
     $signed_input, $num_bits, $range_given, $round_mode, $narrow_range, $axis),
   (TFL_DequantizeOp

tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td

@@ -9862,6 +9862,8 @@ tensor.}]>:$input_max,
   );
 
   TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<0>;
+
+  let hasCanonicalizer = 1;
 }
 
 def TF_QuantizeAndDequantizeV3Op : TF_Op<"QuantizeAndDequantizeV3", [NoSideEffect]> {

tensorflow/compiler/mlir/tensorflow/ir/tf_ops_n_z.cc

@@ -491,6 +491,15 @@ OpFoldResult PowOp::fold(ArrayRef<Attribute> operands) {
   return {};
 }
 
+//===----------------------------------------------------------------------===//
+// QuantizeAndDequantizeV2Op
+//===----------------------------------------------------------------------===//
+
+void QuantizeAndDequantizeV2Op::getCanonicalizationPatterns(
+    OwningRewritePatternList& results, MLIRContext* context) {
+  results.insert<QuantizeAndDequantizeV2ToQuantizeAndDequantizeV4>(context);
+}
+
 //===----------------------------------------------------------------------===//
 // QrOp
 //===----------------------------------------------------------------------===//

tensorflow/compiler/mlir/tensorflow/tests/canonicalize.mlir

@@ -1695,3 +1695,11 @@ func @while_with_id_passthrough(%arg0: tensor<7xf32> {tf._user_specified_name =
   %7 = "tf.Identity"(%6#2) {device = ""} : (tensor<?xf32>) -> tensor<?xf32>
   return %7 : tensor<?xf32>
 }
+
+// CHECK-LABEL: testConvertQuantizeAndDequantizeV2ToQuantizeAndDequantizeV4
+func @testConvertQuantizeAndDequantizeV2ToQuantizeAndDequantizeV4(%arg0 : tensor<?x?xf32>, %arg1 : tensor<f32>, %arg2 : tensor<f32>) -> tensor<?x?xf32> {
+  %0 = "tf.QuantizeAndDequantizeV2"(%arg0, %arg1, %arg2) : (tensor<?x?xf32>, tensor<f32>, tensor<f32>) -> tensor<?x?xf32>
+  return %0 : tensor<?x?xf32>
+  // CHECK: %[[QUANT:.*]] = "tf.QuantizeAndDequantizeV4"(%arg0, %arg1, %arg2) {axis = -1 : i64, narrow_range = false, num_bits = 8 : i64, range_given = false, round_mode = "HALF_TO_EVEN", signed_input = true} : (tensor<?x?xf32>, tensor<f32>, tensor<f32>) -> tensor<?x?xf32>
+  // CHECK: return %[[QUANT]] : tensor<?x?xf32>
+}

tensorflow/compiler/mlir/tensorflow/transforms/canonicalize.td

@@ -188,6 +188,16 @@ def MatrixSetDiagV2ToV3 : Pat<(TF_MatrixSetDiagV2Op $input, $diag, $k),
                               (TF_MatrixSetDiagV3Op $input, $diag, $k,
                                 (GetStrAttr<"LEFT_LEFT">))>;
 
+//===----------------------------------------------------------------------===//
+// QuantizeAndDequantizeV2 op patterns.
+//===----------------------------------------------------------------------===//
+
+def QuantizeAndDequantizeV2ToQuantizeAndDequantizeV4 : Pat<
+  (TF_QuantizeAndDequantizeV2Op $inputs, $min, $max, $signed_input, $num_bits,
+    $range_given, $round_mode, $narrow_range, $axis),
+  (TF_QuantizeAndDequantizeV4Op $inputs, $min, $max, $signed_input, $num_bits,
+    $range_given, $round_mode, $narrow_range, $axis)>;
+
 //===----------------------------------------------------------------------===//
 // RealDiv op patterns.
 //===----------------------------------------------------------------------===//