[Pass] Simplify consecutive casts in Relay

src/relay/transforms/simplify_expr.cc

@@ -78,11 +78,11 @@ class SimplifyReshape : public DFPatternRewrite {
 };
 
 /*!
- * \brief SimplifyCast matches the pattern of cast data to the same dtype.
+ * \brief SimplifySameCast matches the pattern of cast data to the same dtype.
  */
-class SimplifyCast : public DFPatternRewrite {
+class SimplifySameCast : public DFPatternRewrite {
  public:
- SimplifyCast() {
+ SimplifySameCast() {
  data_pat_ = IsWildcard();
  like_pat_ = IsWildcard();
  pattern_ = IsOp("cast_like")({data_pat_, like_pat_}) || IsOp("cast")({data_pat_});
@@ -104,6 +104,60 @@ class SimplifyCast : public DFPatternRewrite {
  DFPattern like_pat_;
 };
 
+/*!
+ * \brief SimplifyConsecutiveCast matches the pattern of consecutive cast/cast_like ops
+ */
+class SimplifyConsecutiveCast : public DFPatternRewrite {
+ public:
+ SimplifyConsecutiveCast() {
+ data_ = IsWildcard();
+ cast1_ = IsOp("cast_like")({data_, IsWildcard()}) || IsOp("cast")({data_});
+ pattern_ = IsOp("cast_like")({cast1_, IsWildcard()}) || IsOp("cast")({cast1_});
+ }
+
+ Expr Callback(const Expr& pre, const Expr& post,
+ const Map<DFPattern, Array<Expr>>& node_map) const override {
+ auto data = node_map[data_][0];
+ auto cast1 = Downcast<Call>(node_map[cast1_][0]);
+ auto data_type = Downcast<TensorType>(data->checked_type());
+ DataType cast1_dtype = Downcast<TensorType>(cast1->checked_type())->dtype;
+
+ if (!IsWidenCast(data_type->dtype, cast1_dtype)) {
+ // Cannot remove the narrow cast
+ return post;
+ }
+
+ const CallNode* cast2 = post.as<CallNode>();
+ DataType cast2_dtype = Downcast<TensorType>(cast2->checked_type())->dtype;
+ auto expr = MakeCast(data, cast2_dtype);
+
+ // We need to set the checked type as it may be needed in the next callback
+ expr->checked_type_ = TensorType(data_type->shape, cast2_dtype);
+ return expr;
+ }
+
+ bool IsWidenCast(DataType origin, DataType cast) const {
+ /* Return whether casting from origin to cast results in more or the same precision.*/
+ if (origin.code() == cast.code() && origin.bits() <= cast.bits()) {
+ return true;
+ }
+ if (origin.code() == DataType::kBFloat || cast.code() == DataType::kBFloat) {
+ // BFloat cast cannot be omitted
+ return false;
+ }
+ if (origin.code() < cast.code()) {
+ // Loosely have a hiearchy to datatypes
+ // e.g. int --> uint --> float has increasing range of numbers they can represent
+ return true;
+ }
+ return false;
+ }
+
+ protected:
+ DFPattern data_;
+ DFPattern cast1_;
+};
+
 /*!
  * \brief SimplifyTranspose matches the pattern of consecutive transpose op,
  * and merges or cancels them.
@@ -640,7 +694,8 @@ Expr SimplifyExpr(const Expr& expr, const IRModule& mod) {
  composer.AddRewrite<EliminateIdentityRewrite>();
  composer.AddRewrite<SimplifyReshape>();
  composer.AddRewrite<SimplifyTranspose>();
- composer.AddRewrite<SimplifyCast>();
+ composer.AddRewrite<SimplifySameCast>();
+ composer.AddRewrite<SimplifyConsecutiveCast>();
  composer.AddRewrite<FullElementwise>();
  composer.AddRewrite<SimplifyConsecutiveAdd>();
  return RewritePatterns(composer.MakeCallbacks(), expr, mod);

tests/python/relay/aot/test_crt_aot_usmp.py

@@ -211,7 +211,7 @@ def test_byoc_microtvm(merge_compiler_regions):
  "model_url, usmp_algo, workspace_size,",
  [
  (MOBILENET_V1_URL, "greedy_by_size", 4845696),
- (MOBILENET_V1_URL, "greedy_by_conflicts", 4845696),
+ (MOBILENET_V1_URL, "greedy_by_conflicts", 4444288),
  ],
 )
 def test_tflite_model(model_url, usmp_algo, workspace_size):

tests/python/relay/test_pass_simplify_expr.py

@@ -402,7 +402,7 @@ def check(x, y=None, do_nothing=False):
  check(id_op(const, x), id_op(op_like(x), x))
 
 
-def test_simplify_cast():
+def test_simplify_same_cast():
  dtype = "int32"
  data = relay.var("data", shape=(3, 4, 5), dtype=dtype)
  expr1 = relay.cast(data, dtype)
@@ -416,6 +416,36 @@ def test_simplify_cast():
  assert tvm.ir.structural_equal(actual2, expected)
 
 
+def test_simplify_consecutive_cast():
+ x = relay.var("x", shape=(3, 4, 5), dtype="int8")
+ y = relay.var("y", shape=(3, 4), dtype="int64")
+ z = relay.var("z", shape=(3,), dtype="float32")
+
+ expr1 = relay.cast(x, "int16")
+ expr2 = relay.cast(expr1, "int32")
+ expr3 = relay.cast_like(expr2, y)
+ expr4 = relay.cast_like(expr3, z)
+
+ actual1 = run_opt_pass(expr2, relay.transform.SimplifyExpr())
+ expected = run_infer_type(relay.cast(x, "int32"))
+ assert tvm.ir.structural_equal(actual1, expected)
+ actual2 = run_opt_pass(expr3, relay.transform.SimplifyExpr())
+ expected = run_infer_type(relay.cast(x, "int64"))
+ assert tvm.ir.structural_equal(actual2, expected)
+ actual3 = run_opt_pass(expr4, relay.transform.SimplifyExpr())
+ expected = run_infer_type(relay.cast(x, "float32"))
+ assert tvm.ir.structural_equal(actual3, expected)
+
+ # cannot simplify the narrow cast
+ x = relay.var("x", shape=(3, 4, 5), dtype="float32")
+ y = relay.var("y", shape=(3, 4), dtype="float32")
+ expr1 = relay.cast(x, "int32")
+ expr2 = relay.cast_like(expr1, y)
+ actual = run_opt_pass(expr2, relay.transform.SimplifyExpr())
+ expected = run_infer_type(expr2)
+ assert tvm.ir.structural_equal(actual, expected)
+
+
 def test_concretize_reshape_like():
  data = relay.var("data", shape=(2, 3, 4), dtype="float32")
  shape_like = relay.var("shape_like", shape=(6, 2, 2), dtype="float32")