[Relay][TOPI] Add rsqrt operator (apache#2949)

docs/api/python/topi.rst

@@ -36,6 +36,7 @@ List of operators
    topi.tanh
    topi.log
    topi.sqrt
+   topi.rsqrt
    topi.sigmoid
    topi.clip
    topi.cast
@@ -122,6 +123,7 @@ topi
 .. autofunction:: topi.tanh
 .. autofunction:: topi.log
 .. autofunction:: topi.sqrt
+.. autofunction:: topi.rsqrt
 .. autofunction:: topi.sigmoid
 .. autofunction:: topi.clip
 .. autofunction:: topi.cast

docs/langref/relay_op.rst

@@ -41,6 +41,7 @@ This level enables fully connected multi-layer perceptron.
 
    tvm.relay.log
    tvm.relay.sqrt
+   tvm.relay.rsqrt
    tvm.relay.exp
    tvm.relay.sigmoid
    tvm.relay.add
@@ -186,6 +187,7 @@ Level 1 Definitions
 -------------------
 .. autofunction:: tvm.relay.log
 .. autofunction:: tvm.relay.sqrt
+.. autofunction:: tvm.relay.rsqrt
 .. autofunction:: tvm.relay.exp
 .. autofunction:: tvm.relay.sigmoid
 .. autofunction:: tvm.relay.add

include/tvm/expr_operator.h

@@ -486,6 +486,7 @@ TVM_DECLARE_INTRIN_UNARY(exp);
 TVM_DECLARE_INTRIN_UNARY(tanh);
 TVM_DECLARE_INTRIN_UNARY(sigmoid);
 TVM_DECLARE_INTRIN_UNARY(sqrt);
+TVM_DECLARE_INTRIN_UNARY(rsqrt);
 TVM_DECLARE_INTRIN_UNARY(log);
 TVM_DECLARE_INTRIN_UNARY(popcount);
 

python/tvm/hybrid/runtime.py

@@ -52,6 +52,22 @@ def allocate(shape, dtype='float32', scope='global'): #pylint: disable=unused-ar
     return numpy.zeros(shape).astype(dtype)
 
 
+def rsqrt(x):
+    """
+    Computes reciprocal of square root of x element-wise
+
+    Parameters
+    ----------
+    x: Tensor
+
+    Returns
+    -------
+    res: Tensor
+        The result of reciprocal of square root of x
+    """
+    return numpy.ones_like(x) / numpy.sqrt(x)
+
+
 def popcount(x):
     """
     Count ones in the binary representation of number x
@@ -103,6 +119,7 @@ def max_num_threads(allow_none=True):
     'allocate'       : allocate,
     'output_tensor'  : allocate,
     'sqrt'           : numpy.sqrt,
+    'rsqrt'          : rsqrt,
     'log'            : numpy.log,
     'tanh'           : numpy.tanh,
     'power'          : numpy.power,

python/tvm/intrin.py

@@ -260,7 +260,7 @@ def log(x):
 
 
 def sqrt(x):
-    """Take log of input x.
+    """Take square root of input x.
 
     Parameters
     ----------
@@ -275,6 +275,22 @@ def sqrt(x):
     return call_pure_intrin(x.dtype, "sqrt", x)
 
 
+def rsqrt(x):
+    """Take reciprocal of square root of input x.
+
+    Parameters
+    ----------
+    x : Expr
+        Input argument.
+
+    Returns
+    -------
+    y : Expr
+        The result.
+    """
+    return call_pure_intrin(x.dtype, "rsqrt", x)
+
+
 def floor(x):
     """Take floor of float input x.
 

python/tvm/relay/op/_tensor.py

@@ -27,6 +27,7 @@
 register_schedule("log", schedule_broadcast)
 register_schedule("exp", schedule_broadcast)
 register_schedule("sqrt", schedule_broadcast)
+register_schedule("rsqrt", schedule_broadcast)
 register_schedule("sigmoid", schedule_broadcast)
 register_schedule("floor", schedule_broadcast)
 register_schedule("ceil", schedule_broadcast)

python/tvm/relay/op/tensor.py

@@ -79,6 +79,26 @@ def sqrt(data):
     return _make.sqrt(data)
 
 
+def rsqrt(data):
+    """Compute elementwise rsqrt of data.
+
+    .. math::
+
+      1/sqrt(x)
+
+    Parameters
+    ----------
+    data : relay.Expr
+        The input data
+
+    Returns
+    -------
+    result : relay.Expr
+        The computed result.
+    """
+    return _make.rsqrt(data)
+
+
 def sigmoid(data):
     """Compute elementwise sigmoid of data.
 

src/codegen/intrin_rule.cc

@@ -40,6 +40,16 @@ TVM_REGISTER_GLOBAL("tvm.intrin.rule.default.tanh")
 TVM_REGISTER_GLOBAL("tvm.intrin.rule.default.sqrt")
 .set_body(DispatchExtern<FloatSuffix>);
 
+TVM_REGISTER_GLOBAL("tvm.intrin.rule.default.rsqrt")
+.set_body([](const TVMArgs& args, TVMRetValue* rv){
+    Expr e = args[0];
+    const Call* call = e.as<Call>();
+    CHECK(call != nullptr);
+
+    auto one = make_const(call->args[0].type(), 1);
+    *rv = one / sqrt(call->args[0]);
+  });
+
 TVM_REGISTER_GLOBAL("tvm.intrin.rule.default.pow")
 .set_body(DispatchExtern<FloatSuffix>);
 

src/relay/op/tensor/unary.cc

@@ -64,7 +64,7 @@ RELAY_REGISTER_UNARY_OP("exp")
 .set_attr<FTVMCompute>("FTVMCompute", RELAY_UNARY_COMPUTE(topi::exp));
 
 RELAY_REGISTER_UNARY_OP("sqrt")
-.describe(R"code(Returns the rsqrt input array, computed element-wise.
+.describe(R"code(Returns the sqrt input array, computed element-wise.
 
 .. math::
    sqrt(x)
@@ -73,6 +73,15 @@ RELAY_REGISTER_UNARY_OP("sqrt")
 .set_support_level(1)
 .set_attr<FTVMCompute>("FTVMCompute", RELAY_UNARY_COMPUTE(topi::sqrt));
 
+RELAY_REGISTER_UNARY_OP("rsqrt")
+.describe(R"code(Returns the rsqrt input array, computed element-wise.
+
+.. math::
+   1/sqrt(x)
+
+)code" TVM_ADD_FILELINE)
+.set_support_level(1)
+.set_attr<FTVMCompute>("FTVMCompute", RELAY_UNARY_COMPUTE(topi::rsqrt));
 
 RELAY_REGISTER_UNARY_OP("zeros_like")
 .describe(R"code(Returns an array of zeros, with same type and shape as the input.

tests/python/relay/test_ir_op.py

@@ -30,7 +30,7 @@ def test(x):
 def test_op_level1():
     x = relay.Var("x")
 
-    for op_name in ["log", "exp", "sqrt", "tanh"]:
+    for op_name in ["log", "exp", "sqrt", "rsqrt","tanh"]:
         y = getattr(relay, op_name)(x)
         assert y.op.name == op_name
         assert y.op.support_level == 1

tests/python/relay/test_op_level1.py

@@ -30,6 +30,10 @@ def relu(x):
     np.maximum(x_copy, 0, x_copy)
     return x_copy
 
+def rsqrt(x):
+    one = np.ones_like(x)
+    return one / np.sqrt(x)
+
 def test_unary_op():
     def check_single_op(opfunc, ref):
         shape = (10, 4)
@@ -57,6 +61,7 @@ def check_single_op(opfunc, ref):
     for opfunc, ref in [(tvm.relay.log, np.log),
                         (tvm.relay.exp, np.exp),
                         (tvm.relay.sqrt, np.sqrt),
+                        (tvm.relay.rsqrt, rsqrt),
                         (tvm.relay.sigmoid, sigmoid),
                         (tvm.relay.tanh, np.tanh),
                         (relay.nn.relu, relu)]:

topi/include/topi/elemwise.h

@@ -129,6 +129,24 @@ inline Tensor sign(const Tensor& x,
   }, name, tag);
 }
 
+/*!
+* \brief Creates an operation that returns rsqrt of a given tensor
+*
+* \param x The input tensor
+* \param name The name of the operation
+* \param tag The tag to mark the operation
+*
+* \return A Tensor whose op member is the rsqrt operation
+*/
+inline Tensor rsqrt(const Tensor& x,
+                       std::string name = "tensor",
+                       std::string tag = kElementWise) {
+  return compute(x->shape, [&](const Array<Var>& i) {
+    Expr one = make_const(x->dtype, 1);
+    return one/tvm::sqrt(x(i));
+  }, name, tag);
+}
+
 /*!
 * \brief Creates an operation that clips each element of a tensor to
 * the interval [a_min, a_max]

topi/python/topi/math.py

@@ -224,6 +224,23 @@ def sqrt(x):
     return tvm.compute(x.shape, lambda *i: tvm.sqrt(x(*i)))
 
 
+@tvm.tag_scope(tag=tag.ELEMWISE)
+def rsqrt(x):
+    """Take inverse square root of input x.
+
+    Parameters
+    ----------
+    x : tvm.Tensor
+        Input argument.
+
+    Returns
+    -------
+    y : tvm.Tensor
+        The result.
+    """
+    return tvm.compute(x.shape, lambda *i: tvm.rsqrt(x(*i)))
+
+
 @tvm.tag_scope(tag=tag.ELEMWISE)
 def sigmoid(x):
     """Take sigmoid tanh of input x.

topi/src/topi.cc

@@ -163,6 +163,11 @@ TVM_REGISTER_GLOBAL("topi.sqrt")
   *rv = sqrt(args[0]);
   });
 
+TVM_REGISTER_GLOBAL("topi.rsqrt")
+.set_body([](TVMArgs args, TVMRetValue *rv) {
+*rv = rsqrt(args[0]);
+});
+
 TVM_REGISTER_GLOBAL("topi.log")
 .set_body([](TVMArgs args, TVMRetValue *rv) {
   *rv = log(args[0]);

topi/tests/python/test_topi_basic.py

@@ -40,6 +40,7 @@ def test_apply(func, name):
     test_apply(topi.sigmoid, "sigmoid")
     test_apply(topi.log, "log")
     test_apply(topi.sqrt, "sqrt")
+    test_apply(topi.rsqrt, "rsqrt")
 
 
 if __name__ == "__main__":

topi/tests/python/test_topi_math.py

@@ -75,6 +75,7 @@ def check_device(device):
     test_apply(topi.sigmoid, "sigmoid", lambda x:1/(1+np.exp(-x)), -1, 1)
     test_apply(topi.log, "log", np.log, 0, 100)
     test_apply(topi.sqrt, "sqrt", np.sqrt, 0, 100)
+    test_apply(topi.rsqrt, "rsqrt", lambda x:np.ones_like(x)/np.sqrt(x), 0, 100, skip_name_check=True)
 
 if __name__ == "__main__":
     test_util()