Add support for Tensorflow operators log1p and cos

The patch adds support for Tensorflow operators log1p and cos
Tensorflow log1p is described at https://www.tensorflow.org/api_docs/python/tf/math/log1p
Tensorflow cos is described at https://www.tensorflow.org/api_docs/python/tf/math/cos

Author: Alex Gladkov

include/tvm/expr_operator.h

@@ -517,7 +517,7 @@ TVM_DECLARE_INTRIN_UNARY(sqrt);
 TVM_DECLARE_INTRIN_UNARY(rsqrt);
 TVM_DECLARE_INTRIN_UNARY(log);
 TVM_DECLARE_INTRIN_UNARY(popcount);
-
+TVM_DECLARE_INTRIN_UNARY(cos);
 
 // Implementation details after this
 inline bool is_const(const Expr& x) {

python/tvm/intrin.py

@@ -258,6 +258,35 @@ def log(x):
     """
     return call_pure_intrin(x.dtype, "log", x)
 
+def log1p(x):
+    """Take log of input x.
+
+    Parameters
+    ----------
+    x : Expr
+        Input argument.
+
+    Returns
+    -------
+    y : Expr
+        The result.
+    """
+    return call_pure_intrin(x.dtype, "log1p", x)
+
+def cos(x):
+    """Take cos of input x.
+
+    Parameters
+    ----------
+    x : Expr
+        Input argument.
+
+    Returns
+    -------
+    y : Expr
+        The result.
+    """
+    return call_pure_intrin(x.dtype, "cos", x)
 
 def sqrt(x):
     """Take square root of input x.

python/tvm/relay/frontend/tensorflow.py

@@ -1344,6 +1344,8 @@ def _impl(inputs, attr, params):
     'Less'                              : _broadcast('less'),
     'LessEqual'                         : _broadcast('less_equal'),
     'Log'                               : AttrCvt('log'),
+    'Log1p'                             : AttrCvt('log1p'),
+    'Cos'                               : AttrCvt('cos'),
     'LogicalAnd'                        : _logical('logical_and'),
     'LogicalOr'                         : _logical('logical_or'),
     'LogicalNot'                        : _logical('logical_not'),

python/tvm/relay/op/_tensor.py

@@ -25,6 +25,8 @@
 schedule_elemwise = schedule_injective
 
 register_schedule("log", schedule_broadcast)
+register_schedule("log1p", schedule_broadcast)
+register_schedule("cos", schedule_broadcast)
 register_schedule("exp", schedule_broadcast)
 register_schedule("sqrt", schedule_broadcast)
 register_schedule("rsqrt", schedule_broadcast)

python/tvm/relay/op/_tensor_grad.py

@@ -31,6 +31,18 @@ def log_grad(orig, grad):
     x = orig.args[0]
     return [grad * ones_like(x) / x]
 
+@register_gradient("log1p")
+def log1p_grad(orig, grad):
+    """Returns [grad * (1 / (x+1))]"""
+    x = orig.args[0]
+    return [grad * ones_like(x) / (x + ones_like(x))]
+
+@register_gradient("cos")
+def cos_grad(orig, grad):
+    """Returns [grad * (-sin(x))]"""
+    x = orig.args[0]
+    ones = ones_like(x)
+    return [grad * (-ones * sin(x))]
 
 @register_gradient("exp")
 def exp_grad(orig, grad):

python/tvm/relay/op/tensor.py

@@ -46,6 +46,35 @@ def log(data):
     """
     return _make.log(data)
 
+def log1p(data):
+    """Compute elementwise log of data.
+
+    Parameters
+    ----------
+    data : relay.Expr
+        The input data
+
+    Returns
+    -------
+    result : relay.Expr
+        The computed result.
+    """
+    return _make.log1p(data)
+
+def cos(data):
+    """Compute elementwise cos of data.
+
+    Parameters
+    ----------
+    data : relay.Expr
+        The input data
+
+    Returns
+    -------
+    result : relay.Expr
+        The computed result.
+    """
+    return _make.cos(data)
 
 def exp(data):
     """Compute elementwise exp of data.

src/codegen/intrin_rule.cc

@@ -34,9 +34,23 @@ TVM_REGISTER_GLOBAL("tvm.intrin.rule.default.exp")
 TVM_REGISTER_GLOBAL("tvm.intrin.rule.default.log")
 .set_body(DispatchExtern<FloatSuffix>);
 
+TVM_REGISTER_GLOBAL("tvm.intrin.rule.default.log1p")
+.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 = log(call->args[0] + one);
+  });
+
+
 TVM_REGISTER_GLOBAL("tvm.intrin.rule.default.tanh")
 .set_body(DispatchExtern<FloatSuffix>);
 
+TVM_REGISTER_GLOBAL("tvm.intrin.rule.default.cos")
+.set_body(DispatchExtern<FloatSuffix>);
+
 TVM_REGISTER_GLOBAL("tvm.intrin.rule.default.sqrt")
 .set_body(DispatchExtern<FloatSuffix>);
 

src/codegen/intrin_rule_cuda.cc

@@ -95,6 +95,9 @@ TVM_REGISTER_GLOBAL("tvm.intrin.rule.cuda.exp")
 TVM_REGISTER_GLOBAL("tvm.intrin.rule.cuda.log")
 .set_body(DispatchExtern<CUDAFastMath>);
 
+TVM_REGISTER_GLOBAL("tvm.intrin.rule.cuda.cos")
+.set_body(DispatchExtern<CUDAFastMath>);
+
 TVM_REGISTER_GLOBAL("tvm.intrin.rule.cuda.tanh")
 .set_body(DispatchExtern<CUDAMath>);
 

src/relay/op/tensor/unary.cc

@@ -53,6 +53,28 @@ RELAY_REGISTER_UNARY_OP("log")
 .set_attr<FTVMCompute>("FTVMCompute", RELAY_UNARY_COMPUTE(topi::log));
 
 
+RELAY_REGISTER_UNARY_OP("log1p")
+.describe(R"code(Returns the log1p input array, computed element-wise.
+
+.. math::
+   log(x+1)
+
+)code" TVM_ADD_FILELINE)
+.set_support_level(1)
+.set_attr<FTVMCompute>("FTVMCompute", RELAY_UNARY_COMPUTE(topi::log1p));
+
+
+RELAY_REGISTER_UNARY_OP("cos")
+.describe(R"code(Returns the cos of input array, computed element-wise.
+
+.. math::
+   Y = cos(X)
+
+)code" TVM_ADD_FILELINE)
+.set_support_level(1)
+.set_attr<FTVMCompute>("FTVMCompute", RELAY_UNARY_COMPUTE(topi::cos));
+
+
 RELAY_REGISTER_UNARY_OP("exp")
 .describe(R"code(Returns the exp input array, computed element-wise.
 

tests/python/frontend/tensorflow/test_forward.py

@@ -1850,6 +1850,22 @@ def test_forward_log():
     tf.log(in_data, name="log")
     compare_tf_with_tvm([np_data], ['in_data:0'], 'log:0')
 
+def test_forward_log1p():
+    """test operator Log1p """
+    np_data = np.random.uniform(1, 100, size=(2, 3, 5)).astype(np.float32)
+    tf.reset_default_graph()
+    in_data = tf.placeholder(tf.float32, (2, 3, 5), name="in_data")
+    tf.log1p(in_data, name="log1p")
+    compare_tf_with_tvm([np_data], ['in_data:0'], 'log1p:0')
+
+def test_forward_cos():
+    """test operator cos """
+    np_data = np.random.uniform(1, 100, size=(2, 3, 5)).astype(np.float32)
+    tf.reset_default_graph()
+    in_data = tf.placeholder(tf.float32, (2, 3, 5), name="in_data")
+    tf.cos(in_data, name="cos")
+    compare_tf_with_tvm([np_data], ['in_data:0'], 'cos:0')
+
 def test_forward_negative():
     """test tf operator Neg """
     np_data = np.random.uniform(-100, 255, size=(224, 224, 3)).astype(np.float32)
@@ -2097,6 +2113,7 @@ def test_placeholder():
 # ----
 if __name__ == '__main__':
 
+
     # Transforms
     test_forward_transpose()
     test_forward_reshape()
@@ -2140,6 +2157,8 @@ def test_placeholder():
     test_forward_pow_exp()
     test_forward_sign()
     test_forward_log()
+    test_forward_log1p()
+    test_forward_cos()
     test_forward_negative()
     test_forward_divide()
     test_forward_abs()