[NNVM] Move FTVMCompute registration of cast, greter, less to C++. (#1370)

nnvm/python/nnvm/top/tensor.py

@@ -53,11 +53,6 @@ reg.register_pattern("copy", OpPattern.ELEMWISE)
 reg.register_schedule("copy", _fschedule_broadcast)
 
 # cast
-@reg.register_compute("cast")
-def compute_cast(attrs, inputs, _):
-    """Compute definition of cast"""
-    dtype = attrs.get_string("dtype")
-    return topi.cast(inputs[0], dtype)
 reg.register_pattern("cast", OpPattern.ELEMWISE)
 reg.register_schedule("cast", _fschedule_broadcast)
 
@@ -210,18 +205,10 @@ reg.register_pattern("ones_like", OpPattern.ELEMWISE)
 reg.register_schedule("ones_like", _fschedule_elemwise)
 
 # greater
-@reg.register_compute("greater")
-def compute_greater(_, inputs, out_info):
-    """Compute definition of greater"""
-    return topi.greater(inputs[0], inputs[1]).astype('float32')
 reg.register_pattern("greater", OpPattern.ELEMWISE)
 reg.register_schedule("greater", _fschedule_elemwise)
 
 # less
-@reg.register_compute("less")
-def compute_less(_, inputs, out_info):
-    """Compute definition of less"""
-    return topi.less(inputs[0], inputs[1]).astype('float32')
 reg.register_pattern("less", OpPattern.ELEMWISE)
 reg.register_schedule("less", _fschedule_elemwise)
 

nnvm/src/top/tensor/elemwise.cc

@@ -781,6 +781,12 @@ with 1.0 if (left > right), otherwise 0.0 element-wise.
 .add_argument("rhs", "Tensor", "Second input")
 .set_num_inputs(2)
 .set_attr<nnvm::FInferShape>("FInferShape", ElemwiseShape<2, 1>)
+.set_attr<FTVMCompute>(
+  "FTVMCompute", [](const NodeAttrs& attrs,
+                    const Array<Tensor>& inputs,
+                    const Array<Tensor>& out_info) {
+    return Array<Tensor>{ topi::cast(topi::greater(inputs[0], inputs[1]), out_info[0]->dtype) };
+})
 .set_support_level(4);
 
 
@@ -793,6 +799,12 @@ with 1.0 if (left < right), otherwise 0.0 element-wise.
 .add_argument("rhs", "Tensor", "Second input")
 .set_num_inputs(2)
 .set_attr<nnvm::FInferShape>("FInferShape", ElemwiseShape<2, 1>)
+.set_attr<FTVMCompute>(
+  "FTVMCompute", [](const NodeAttrs& attrs,
+                    const Array<Tensor>& inputs,
+                    const Array<Tensor>& out_info) {
+    return Array<Tensor>{ topi::cast(topi::less(inputs[0], inputs[1]), out_info[0]->dtype) };
+})
 .set_support_level(4);
 
 NNVM_REGISTER_INDICATOR_OP(_max_mask)

nnvm/src/top/tensor/transform.cc

@@ -15,7 +15,9 @@
 #include "../elemwise_op_common.h"
 #include "topi/nn/flatten.h"
 #include "topi/transform.h"
+#include "topi/elemwise.h"
 #include "topi/detail/constant_utils.h"
+#include "../../compiler/compile_engine.h"
 
 namespace nnvm {
 namespace top {
@@ -413,6 +415,14 @@ NNVM_REGISTER_OP(cast)
 .set_attr<FCorrectLayout>("FCorrectLayout", ElemwiseArbitraryLayout<1, 1>)
 .set_num_inputs(1)
 .set_num_outputs(1)
+.set_attr<FTVMCompute>(
+  "FTVMCompute", [](const NodeAttrs& attrs,
+                    const Array<Tensor>& inputs,
+                    const Array<Tensor>& out_info) {
+    const CastParam& param = nnvm::get<CastParam>(attrs.parsed);
+    Type dtype = GetTVMType(param.dtype);
+    return Array<Tensor>{ topi::cast(inputs[0], dtype) };
+})
 .set_support_level(1);