From def81b4f8642cbb27737009d1f3a73f76cd389a6 Mon Sep 17 00:00:00 2001
From: Zhang Ting <zhangting_2017@163.com>
Date: Mon, 24 Jan 2022 16:14:09 +0800
Subject: [PATCH] unify compare functor (#39024)

---
 .../fluid/operators/controlflow/compare_op.h  | 59 ++++++++-----------
 paddle/fluid/operators/matrix_rank_op.cc      | 18 +++---
 paddle/fluid/operators/matrix_rank_op.cu      |  8 +--
 paddle/fluid/operators/matrix_rank_op.h       | 11 +---
 paddle/fluid/operators/viterbi_decode_op.cu   |  5 +-
 paddle/fluid/operators/viterbi_decode_op.h    |  5 +-
 .../fluid/tests/unittests/test_compare_op.py  | 13 ++++
 7 files changed, 59 insertions(+), 60 deletions(-)

diff --git a/paddle/fluid/operators/controlflow/compare_op.h b/paddle/fluid/operators/controlflow/compare_op.h
index d2ef4c9befb..957efbff199 100644
--- a/paddle/fluid/operators/controlflow/compare_op.h
+++ b/paddle/fluid/operators/controlflow/compare_op.h
@@ -22,49 +22,40 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
-template <typename T>
-struct LessThanFunctor {
-  using ELEM_TYPE = T;
-  HOSTDEVICE bool operator()(const T a, const T b) const { return a < b; }
-};
-
-template <typename T>
-struct LessEqualFunctor {
-  using ELEM_TYPE = T;
-  HOSTDEVICE bool operator()(const T a, const T b) const { return a <= b; }
-};
-
-template <typename T>
-struct GreaterThanFunctor {
-  using ELEM_TYPE = T;
-  HOSTDEVICE bool operator()(const T a, const T b) const { return a > b; }
-};
-
-template <typename T>
-struct GreaterEqualFunctor {
-  using ELEM_TYPE = T;
-  HOSTDEVICE bool operator()(const T a, const T b) const { return a >= b; }
-};
-
-template <typename T>
+#define COMPARE_FUNCTOR(func_name, op)                           \
+  template <typename InT, typename OutT = bool>                  \
+  struct func_name {                                             \
+    using ELEM_TYPE = InT;                                       \
+    HOSTDEVICE OutT operator()(const InT a, const InT b) const { \
+      return static_cast<OutT>(a op b);                          \
+    }                                                            \
+  };
+
+COMPARE_FUNCTOR(LessThanFunctor, <)
+COMPARE_FUNCTOR(LessEqualFunctor, <=)
+COMPARE_FUNCTOR(GreaterThanFunctor, >)
+COMPARE_FUNCTOR(GreaterEqualFunctor, >=)
+#undef COMPARE_FUNCTOR
+
+template <typename InT, typename OutT = bool>
 struct EqualFunctor {
-  using ELEM_TYPE = T;
-  HOSTDEVICE bool operator()(const T a, const T b) const {
-    if (std::is_floating_point<T>::value) {
+  using ELEM_TYPE = InT;
+  HOSTDEVICE OutT operator()(const InT a, const InT b) const {
+    if (std::is_floating_point<InT>::value) {
       // This branch will be optimized while compiling if T is integer. It is
       // safe to cast a and b to double.
-      return fabs(static_cast<double>(a - b)) < 1e-8;
+      return static_cast<OutT>(fabs(static_cast<double>(a - b)) < 1e-8);
     } else {
-      return (a == b);
+      return static_cast<OutT>(a == b);
     }
   }
 };
 
-template <typename T>
+template <typename InT, typename OutT = bool>
 struct NotEqualFunctor {
-  using ELEM_TYPE = T;
-  HOSTDEVICE bool operator()(const T a, const T b) const {
-    return !EqualFunctor<T>()(a, b);
+  using ELEM_TYPE = InT;
+  HOSTDEVICE bool operator()(const InT a, const InT b) const {
+    return !EqualFunctor<InT, OutT>()(a, b);
   }
 };
 
diff --git a/paddle/fluid/operators/matrix_rank_op.cc b/paddle/fluid/operators/matrix_rank_op.cc
index 3038a16dc0a..ddfb8d50c4e 100644
--- a/paddle/fluid/operators/matrix_rank_op.cc
+++ b/paddle/fluid/operators/matrix_rank_op.cc
@@ -219,18 +219,20 @@ class MatrixRankCPUKernel : public framework::OpKernel<T> {
     tol_tensor.Resize(detail::NewAxisDim(tol_tensor.dims(), 1));
 
     Tensor compare_result;
-    compare_result.mutable_data<int>(detail::NewAxisDim(dim_out, k),
-                                     context.GetPlace());
+    compare_result.mutable_data<int64_t>(detail::NewAxisDim(dim_out, k),
+                                         context.GetPlace());
 
     int axis = -1;
     if (eigenvalue_tensor.dims().size() >= tol_tensor.dims().size()) {
-      ElementwiseComputeEx<GreaterThanFunctor<T>, platform::CPUDeviceContext, T,
-                           int>(context, &eigenvalue_tensor, &tol_tensor, axis,
-                                GreaterThanFunctor<T>(), &compare_result);
+      ElementwiseComputeEx<GreaterThanFunctor<T, int64_t>,
+                           platform::CPUDeviceContext, T, int>(
+          context, &eigenvalue_tensor, &tol_tensor, axis,
+          GreaterThanFunctor<T, int64_t>(), &compare_result);
     } else {
-      ElementwiseComputeEx<LessThanFunctor<T>, platform::CPUDeviceContext, T,
-                           int>(context, &eigenvalue_tensor, &tol_tensor, axis,
-                                LessThanFunctor<T>(), &compare_result);
+      ElementwiseComputeEx<LessThanFunctor<T, int64_t>,
+                           platform::CPUDeviceContext, T, int>(
+          context, &eigenvalue_tensor, &tol_tensor, axis,
+          LessThanFunctor<T, int64_t>(), &compare_result);
     }
     auto dito_int =
         math::DeviceIndependenceTensorOperations<platform::CPUDeviceContext,
diff --git a/paddle/fluid/operators/matrix_rank_op.cu b/paddle/fluid/operators/matrix_rank_op.cu
index 1891a7be24e..7362d00afb7 100644
--- a/paddle/fluid/operators/matrix_rank_op.cu
+++ b/paddle/fluid/operators/matrix_rank_op.cu
@@ -129,10 +129,10 @@ class MatrixRankGPUKernel : public framework::OpKernel<T> {
     compare_result.mutable_data<int64_t>(detail::NewAxisDim(dim_out, k),
                                          context.GetPlace());
     int axis = -1;
-    ElementwiseComputeEx<GreaterThanFunctor<T>, platform::CUDADeviceContext, T,
-                         int64_t>(context, &eigenvalue_tensor, &tol_tensor,
-                                  axis, GreaterThanFunctor<T>(),
-                                  &compare_result);
+    ElementwiseComputeEx<GreaterThanFunctor<T, int64_t>,
+                         platform::CUDADeviceContext, T, int64_t>(
+        context, &eigenvalue_tensor, &tol_tensor, axis,
+        GreaterThanFunctor<T, int64_t>(), &compare_result);
     auto dito_int =
         math::DeviceIndependenceTensorOperations<platform::CUDADeviceContext,
                                                  int64_t>(context);
diff --git a/paddle/fluid/operators/matrix_rank_op.h b/paddle/fluid/operators/matrix_rank_op.h
index c3d99a21b72..550bc445ac4 100644
--- a/paddle/fluid/operators/matrix_rank_op.h
+++ b/paddle/fluid/operators/matrix_rank_op.h
@@ -16,6 +16,7 @@
 #include <vector>
 #include "paddle/fluid/framework/ddim.h"
 #include "paddle/fluid/framework/tensor.h"
+#include "paddle/fluid/operators/controlflow/compare_op.h"
 
 namespace paddle {
 namespace operators {
@@ -46,16 +47,6 @@ static DDim RemoveLastDim(const DDim& dim) {
 }
 }  // namespace detail
 
-template <typename T>
-struct GreaterThanFunctor {
-  HOSTDEVICE int operator()(const T a, const T b) const { return a > b; }
-};
-
-template <typename T>
-struct LessThanFunctor {
-  HOSTDEVICE int operator()(const T a, const T b) const { return a < b; }
-};
-
 template <typename T>
 struct GreaterElementFunctor {
   HOSTDEVICE T operator()(const T a, const T b) const {
diff --git a/paddle/fluid/operators/viterbi_decode_op.cu b/paddle/fluid/operators/viterbi_decode_op.cu
index d40d14435a5..7ad3335009b 100644
--- a/paddle/fluid/operators/viterbi_decode_op.cu
+++ b/paddle/fluid/operators/viterbi_decode_op.cu
@@ -72,7 +72,8 @@ struct BinaryOperation<platform::CUDADeviceContext, BinaryFunctor, T> {
   }
 };
 
-template <template <typename T> typename CompareFunctor, typename T>
+template <template <typename InT, typename OutT> typename CompareFunctor,
+          typename T>
 struct GetMask<platform::CUDADeviceContext, CompareFunctor, T> {
   void operator()(const framework::ExecutionContext& ctx, const Tensor& lhs,
                   const Tensor& rhs, Tensor* mask) {
@@ -81,7 +82,7 @@ struct GetMask<platform::CUDADeviceContext, CompareFunctor, T> {
     auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
     paddle::operators::LaunchSameDimsElementwiseCudaKernel<
         ElementwiseType::kBinary, int64_t, T>(dev_ctx, ins, &outs,
-                                              CompareFunctor<int64_t>());
+                                              CompareFunctor<int64_t, T>());
   }
 };
 
diff --git a/paddle/fluid/operators/viterbi_decode_op.h b/paddle/fluid/operators/viterbi_decode_op.h
index 2b392ae74cc..ab95dbc763a 100644
--- a/paddle/fluid/operators/viterbi_decode_op.h
+++ b/paddle/fluid/operators/viterbi_decode_op.h
@@ -112,12 +112,13 @@ void SameDimsBinaryOP(const Tensor& lhs, const Tensor& rhs, Tensor* out) {
   }
 }
 
-template <typename DeviceContext, template <typename T> typename CompareFunctor,
+template <typename DeviceContext,
+          template <typename InT, typename OutT> typename CompareFunctor,
           typename T>
 struct GetMask {
   void operator()(const framework::ExecutionContext& ctx, const Tensor& lhs,
                   const Tensor& rhs, Tensor* mask) {
-    SameDimsBinaryOP<int64_t, CompareFunctor<int64_t>, T>(lhs, rhs, mask);
+    SameDimsBinaryOP<int64_t, CompareFunctor<int64_t, T>, T>(lhs, rhs, mask);
   }
 };
 
diff --git a/python/paddle/fluid/tests/unittests/test_compare_op.py b/python/paddle/fluid/tests/unittests/test_compare_op.py
index 89756385486..f92465b739a 100755
--- a/python/paddle/fluid/tests/unittests/test_compare_op.py
+++ b/python/paddle/fluid/tests/unittests/test_compare_op.py
@@ -140,6 +140,19 @@ def create_paddle_case(op_type, callback):
                 self.assertEqual((out.numpy() == self.real_result).all(), True)
                 paddle.enable_static()
 
+        def test_not_equal(self):
+            if self.op_type == "not_equal":
+                paddle.disable_static()
+                x = paddle.to_tensor(
+                    np.array([1.2e-8, 2, 2, 1]), dtype="float32")
+                y = paddle.to_tensor(
+                    np.array([1.1e-8, 2, 2, 1]), dtype="float32")
+                op = eval("paddle.%s" % (self.op_type))
+                out = op(x, y)
+                self.real_result = np.array([0, 0, 0, 0]).astype(np.int64)
+                self.assertEqual((out.numpy() == self.real_result).all(), True)
+                paddle.enable_static()
+
         def test_assert(self):
             def test_dynamic_api_string(self):
                 if self.op_type == "equal":
-- 
GitLab