Add operator `isnan` (#3979)

* add expr `isnan`

* move to intrinsic

* doc & add to topi

* fix error from ci

docs/api/python/intrin.rst

@@ -36,6 +36,7 @@ tvm.intrin
    tvm.trunc
    tvm.round
    tvm.abs
+   tvm.isnan
 
 .. autofunction:: tvm.call_packed
 .. autofunction:: tvm.call_pure_intrin
@@ -52,3 +53,4 @@ tvm.intrin
 .. autofunction:: tvm.trunc
 .. autofunction:: tvm.round
 .. autofunction:: tvm.abs
+.. autofunction:: tvm.isnan

docs/api/python/topi.rst

@@ -32,6 +32,7 @@ List of operators
    topi.trunc
    topi.round
    topi.abs
+   topi.isnan
    topi.exp
    topi.tanh
    topi.log
@@ -127,6 +128,7 @@ topi
 .. autofunction:: topi.trunc
 .. autofunction:: topi.round
 .. autofunction:: topi.abs
+.. autofunction:: topi.isnan
 .. autofunction:: topi.exp
 .. autofunction:: topi.tanh
 .. autofunction:: topi.log

include/tvm/expr_operator.h

@@ -465,6 +465,12 @@ TVM_DLL Expr pow(Expr x, Expr y);
  * \return The aboslute value of input data x
  */
 TVM_DLL Expr abs(Expr x);
+/*!
+ * \brief Check if x is NaN.
+ * \param x The input data
+ * \return The result expression.
+ */
+TVM_DLL Expr isnan(Expr x);
 
 /*!
  * \brief sum of of source expression over axis

include/tvm/ir.h

@@ -574,6 +574,7 @@ class Call : public ExprNode {
   static constexpr const char* likely = "likely";
   static constexpr const char* glsl_texture_store = "glsl_texture_store";
   static constexpr const char* prefetch = "prefetch";
+  static constexpr const char* isnan = "isnan";
 
   /*! \brief Vectorizable intrinsic list. */
   static const char* vectorizable_intrinsics[];

python/tvm/intrin.py

@@ -434,6 +434,22 @@ def round(x):
     return _make.round(x)
 
 
+def isnan(x):
+    """Check if input value is Nan.
+
+    Parameters
+    ----------
+    x : Expr
+        Input argument.
+
+    Returns
+    -------
+    y : Expr
+        The result.
+    """
+    return _make.isnan(x)
+
+
 def power(x, y):
     """x power y
 

src/api/api_ir.cc

@@ -38,6 +38,9 @@ TVM_REGISTER_API("_Var")
 TVM_REGISTER_API("make.abs")
 .set_body_typed(tvm::abs);
 
+TVM_REGISTER_API("make.isnan")
+.set_body_typed(tvm::isnan);
+
 TVM_REGISTER_API("make.floor")
 .set_body_typed(tvm::floor);
 

src/codegen/codegen_c.cc

@@ -576,6 +576,12 @@ void CodeGenC::VisitExpr_(const Call *op, std::ostream& os) {  // NOLINT(*)
     os << " *)(&(";
     this->PrintExpr(op->args[0], os);
     os << ")))";
+  } else if (op->is_intrinsic(Call::isnan)) {
+    os << "(";
+    this->PrintExpr(op->args[0], os);
+    os << " != ";
+    this->PrintExpr(op->args[0], os);
+    os << ")";
   } else {
     if (op->call_type == Call::Intrinsic ||
         op->call_type == Call::PureIntrinsic) {

src/codegen/llvm/codegen_llvm.cc

@@ -746,6 +746,10 @@ llvm::Value* CodeGenLLVM::CreateIntrinsic(const Call* op) {
   } else if (op->is_intrinsic(Call::reinterpret)) {
     llvm::Type * target = LLVMType(op->type);
     return builder_->CreateBitCast(MakeValue(op->args[0]), target);
+  } else if (op->is_intrinsic(Call::isnan)) {
+    // TODO(hgt312): set fast math flag
+    llvm::Value* a = MakeValue(op->args[0]);
+    return builder_->CreateFCmpUNO(a, a);
   } else if (op->is_intrinsic("vectorlow")) {
     llvm::Value *v = MakeValue(op->args[0]);
     int l = v->getType()->getVectorNumElements();

src/lang/expr_operator.cc

@@ -424,6 +424,30 @@ Expr abs(Expr x) {
   }
 }
 
+Expr isnan(Expr x) {
+  Type t = Bool(x.type().lanes());
+  if (x.type().is_int() || x.type().is_uint()) {
+    return make_const(t, false);
+  } else if (x.type().is_float()) {
+    using ir::FloatImm;
+    const FloatImm* fx = x.as<FloatImm>();
+    if (fx) {
+      return make_const(t, std::isnan(fx->value));
+    }
+    if (x.type().bits() == 16) {
+      return ir::Call::make(t, ir::Call::isnan,
+                               {cast(Float(32, t.lanes()), std::move(x))},
+                               ir::Call::PureIntrinsic);
+    } else {
+      return ir::Call::make(t, ir::Call::isnan, {x}, ir::Call::PureIntrinsic);
+    }
+  } else {
+    LOG(FATAL) << "Data type " << x.type()
+               <<" not supported for isnan op. Skipping isnan op...";
+    return x;
+  }
+}
+
 Expr sum(Expr source, Array<IterVar> rdom) {
   Var x("x", source.type()), y("y", source.type());
   Expr result = ir::Add::make(x, y);

tests/python/unittest/test_lang_basic.py

@@ -17,6 +17,7 @@
 import tvm
 import numpy as np
 
+
 def test_const():
     x = tvm.const(1, "int32")
     print(x.dtype)
@@ -39,6 +40,7 @@ def test_scalar_dtype_inference():
     assert tvm.convert(1).dtype == 'int32'
     assert tvm.convert(1.0).dtype == 'float32'
 
+
 def test_make():
     x = tvm.const(1, "int32")
     y = tvm.var("x")
@@ -46,6 +48,7 @@ def test_make():
     assert isinstance(tvm.max(x, y), tvm.expr.Max)
     assert isinstance(tvm.min(x, y), tvm.expr.Min)
 
+
 def test_ir():
     x = tvm.const(1, "int32")
     y = tvm.make.IntImm('int32', 1)
@@ -53,6 +56,7 @@ def test_ir():
     stmt = tvm.make.Evaluate(z)
     assert isinstance(stmt, tvm.stmt.Evaluate)
 
+
 def test_ir2():
     x = tvm.var("n")
     a = tvm.var("array", tvm.handle)
@@ -60,12 +64,14 @@ def test_ir2():
     assert isinstance(st, tvm.stmt.Store)
     assert(st.buffer_var == a)
 
+
 def test_let():
     x = tvm.var('x')
     y = tvm.var('y')
     stmt = tvm.make.LetStmt(
         x, 10, tvm.make.Evaluate(x + 1));
 
+
 def test_cast():
     x = tvm.var('x', dtype="float32")
     y = x.astype("int32")
@@ -104,10 +110,12 @@ def test_stmt():
                  tvm.stmt.For.Serial, 0,
                  x)
 
+
 def test_dir():
     x = tvm.var('x')
     dir(x)
 
+
 def test_dtype():
     x = tvm.var('x')
     assert x.dtype == 'int32'
@@ -158,6 +166,7 @@ def test_all():
         '(((%s < %s) && (%s > (%s + 1))) && (%s < (%s*2)))' % (
             x.name, y.name, y.name, z.name, x.name, z.name)
 
+
 def test_bitwise():
     x = tvm.var('x')
     y = tvm.var('y')
@@ -172,6 +181,18 @@ def test_bitwise():
     assert(tvm.var("z", "int8x2") << tvm.const(1, "int8x2")).dtype == "int8x2"
 
 
+def test_isnan():
+    x = tvm.var('x', 'float32')
+    assert str(tvm.isnan(x)) == 'isnan(x)'
+    assert str(tvm.isnan(x).dtype) == 'bool'
+    y = tvm.var('y', 'float16')
+    assert str(tvm.isnan(y)) == 'isnan(float32(y))'
+    z = tvm.var('z', 'int32')
+    assert str(tvm.isnan(z)) == '(bool)0'
+    k = tvm.var('k', 'int8x2')
+    assert str(tvm.isnan(k).dtype) == 'uint1x2'
+
+
 def test_equality():
     a = tvm.var('a')
     b = tvm.var('b')
@@ -203,5 +224,6 @@ if __name__ == "__main__":
     test_any()
     test_all()
     test_bitwise()
+    test_isnan()
     test_equality()
     test_equality_string_imm()

topi/include/topi/elemwise.h

@@ -58,6 +58,7 @@ TOPI_DECLARE_UNARY_OP(abs);
 TOPI_DECLARE_UNARY_OP(cos);
 TOPI_DECLARE_UNARY_OP(sin);
 TOPI_DECLARE_UNARY_OP(atan);
+TOPI_DECLARE_UNARY_OP(isnan);
 
 /*
  * \brief Fast_tanh_float implementation from Eigen

topi/python/topi/math.py

@@ -244,6 +244,23 @@ def abs(x):
 
 
 @tvm.tag_scope(tag=tag.ELEMWISE)
+def isnan(x):
+    """Check if value of x is NaN, element-wise.
+
+    Parameters
+    ----------
+    x : tvm.Tensor
+        Input argument.
+
+    Returns
+    -------
+    y : tvm.Tensor
+        The result.
+    """
+    return tvm.compute(x.shape, lambda *i: tvm.isnan(x(*i)))
+
+
+@tvm.tag_scope(tag=tag.ELEMWISE)
 def round(x):
     """Round elements of x to nearest integer.
 

topi/tests/python/test_topi_math.py

@@ -72,6 +72,46 @@ def test_ewise():
         for device in get_all_backend():
             check_device(device)
 
+    def test_isnan(
+        low,
+        high,
+        shape=(20, 3),
+        dtype=tvm.float32,
+        check_round=False,
+        skip_name_check=False,
+    ):
+        m = tvm.var("m")
+        l = tvm.var("l")
+        A = tvm.placeholder((m, l), dtype=dtype, name="A")
+
+        B = topi.isnan(A)
+        assert tuple(B.shape) == tuple(A.shape)
+        if not skip_name_check:
+            assert B.op.body[0].name == "isnan"
+        a_np = np.random.uniform(low=low, high=high, size=shape).astype(A.dtype) * 10
+        a_np.ravel()[np.random.choice(a_np.size, int(a_np.size * 0.5), replace=False)] = np.nan
+        # avoid round check too close to boundary
+        if check_round:
+            a_np += ((np.fmod(a_np, 1) - 0.5) < 1e-6) * 1e-5
+        b_np = np.isnan(a_np)
+
+        def check_device(device):
+            ctx = tvm.context(device, 0)
+            if not ctx.exist:
+                print("Skip because %s is not enabled" % device)
+                return
+            print("Running on target: %s" % device)
+            with tvm.target.create(device):
+                s = topi.generic.schedule_injective(B)
+            foo = tvm.build(s, [A, B], device, name="isnan")
+            a = tvm.nd.array(a_np, ctx)
+            b = tvm.nd.array(np.zeros_like(b_np), ctx)
+            foo(a, b)
+            tvm.testing.assert_allclose(b.asnumpy(), b_np, rtol=1e-5, atol=1e-5)
+
+        for device in get_all_backend():
+            check_device(device)
+
     test_apply(topi.floor, "floor", np.floor, -100, 100)
     test_apply(topi.ceil, "ceil", np.ceil, -100, 100)
     test_apply(topi.sign, "sign", np.sign, -100, 100, skip_name_check=True)
@@ -88,6 +128,7 @@ def test_ewise():
     test_apply(topi.cos, "cos", np.cos, -2.0*np.pi, 2.0*np.pi)
     test_apply(topi.sin, "sin", np.sin, -2.0*np.pi, 2.0*np.pi)
     test_apply(topi.erf, "erf", scipy.special.erf, -.1, .1, dtype="float32")
+    test_isnan(-100, 100)
 
 
 def test_cast():