fix mixed type backward (#18250)

src/operator/mshadow_op.h

@@ -746,6 +746,10 @@ MXNET_BINARY_MATH_OP_NC(minus_sign, a - b > DType(0) ? DType(1) : -DType(1));
 
 MXNET_BINARY_MATH_OP(rminus, b - a);
 
+MXNET_BINARY_MATH_OP_NC(posone, 1);
+
+MXNET_BINARY_MATH_OP_NC(negone, -1);
+
 MXNET_BINARY_MATH_OP(div_grad, 1.0f / math::id(b));
 
 MXNET_BINARY_MATH_OP(div_rgrad, -math::id(a) / math::sqr(b));
@@ -799,6 +803,54 @@ struct mod : public mxnet_op::tunable {
   }
 };
 
+#ifndef _WIN32
+struct mixed_mod {
+  template<typename DType,
+           typename std::enable_if<std::is_integral<DType>::value, int>::type = 0>
+  MSHADOW_XINLINE static mshadow::half::half_t Map(DType a, mshadow::half::half_t b) {
+    return mod::Map(static_cast<mshadow::half::half_t>(a), b);
+  }
+
+  template<typename DType,
+           typename std::enable_if<std::is_same<DType, mshadow::half::half_t>::value ||
+                                   std::is_integral<DType>::value, int>::type = 0>
+  MSHADOW_XINLINE static float Map(DType a, float b) {
+    return mod::Map(static_cast<float>(a), b);
+  }
+
+  template<typename DType,
+           typename std::enable_if<std::is_same<DType, mshadow::half::half_t>::value ||
+                                   std::is_same<DType, float>::value ||
+                                   std::is_integral<DType>::value, int>::type = 0>
+  MSHADOW_XINLINE static double Map(DType a, double b) {
+    return mod::Map(static_cast<double>(a), b);
+  }
+};
+
+struct mixed_rmod {
+  template<typename DType,
+           typename std::enable_if<std::is_integral<DType>::value, int>::type = 0>
+  MSHADOW_XINLINE static mshadow::half::half_t Map(DType a, mshadow::half::half_t b) {
+    return mod::Map(b, static_cast<mshadow::half::half_t>(a));
+  }
+
+  template<typename DType,
+           typename std::enable_if<std::is_same<DType, mshadow::half::half_t>::value ||
+                                   std::is_integral<DType>::value, int>::type = 0>
+  MSHADOW_XINLINE static float Map(DType a, float b) {
+    return mod::Map(b, static_cast<float>(a));
+  }
+
+  template<typename DType,
+           typename std::enable_if<std::is_same<DType, mshadow::half::half_t>::value ||
+                                   std::is_same<DType, float>::value ||
+                                   std::is_integral<DType>::value, int>::type = 0>
+  MSHADOW_XINLINE static double Map(DType a, double b) {
+    return mod::Map(b, static_cast<double>(a));
+  }
+};
+#endif
+
 struct fmod : public mxnet_op::tunable {
   template<typename DType>
   MSHADOW_XINLINE static DType Map(DType a, DType b) {

src/operator/numpy/np_elemwise_broadcast_op.cc

@@ -115,7 +115,22 @@ MXNET_OPERATOR_REGISTER_NP_BINARY_MIXED_PRECISION(_npi_add)
   "FCompute<cpu>",
   NumpyBinaryBroadcastComputeWithBool<cpu, op::mshadow_op::plus>)
 #endif
-.set_attr<nnvm::FGradient>("FGradient", ElemwiseGradUseNone{"_backward_broadcast_add"});
+.set_attr<nnvm::FGradient>("FGradient", ElemwiseGradUseIn{"_backward_npi_broadcast_add"});
+
+NNVM_REGISTER_OP(_backward_npi_broadcast_add)
+.set_num_inputs(3)
+.set_num_outputs(2)
+.set_attr<nnvm::TIsBackward>("TIsBackward", true)
+.set_attr<nnvm::FInplaceOption>("FInplaceOption",
+  [](const NodeAttrs& attrs){
+    return std::vector<std::pair<int, int> >{{0, 0}, {0, 1}};
+  })
+.set_attr<FResourceRequest>("FResourceRequest",
+  [](const NodeAttrs& attrs) {
+    return std::vector<ResourceRequest>{ResourceRequest::kTempSpace};
+  })
+.set_attr<FCompute>("FCompute<cpu>", NumpyBinaryBackwardUseIn<cpu, mshadow_op::posone,
+                                                                mshadow_op::posone>);
 
 MXNET_OPERATOR_REGISTER_NP_BINARY_MIXED_PRECISION(_npi_subtract)
 #ifndef _WIN32
@@ -128,7 +143,22 @@ MXNET_OPERATOR_REGISTER_NP_BINARY_MIXED_PRECISION(_npi_subtract)
   "FCompute<cpu>",
   NumpyBinaryBroadcastCompute<cpu, op::mshadow_op::minus>)
 #endif
-.set_attr<nnvm::FGradient>("FGradient", ElemwiseGradUseNone{"_backward_broadcast_sub"});
+.set_attr<nnvm::FGradient>("FGradient", ElemwiseGradUseIn{"_backward_npi_broadcast_sub"});
+
+NNVM_REGISTER_OP(_backward_npi_broadcast_sub)
+.set_num_inputs(3)
+.set_num_outputs(2)
+.set_attr<nnvm::TIsBackward>("TIsBackward", true)
+.set_attr<nnvm::FInplaceOption>("FInplaceOption",
+  [](const NodeAttrs& attrs){
+    return std::vector<std::pair<int, int> >{{0, 0}, {0, 1}};
+  })
+.set_attr<FResourceRequest>("FResourceRequest",
+  [](const NodeAttrs& attrs) {
+    return std::vector<ResourceRequest>{ResourceRequest::kTempSpace};
+  })
+.set_attr<FCompute>("FCompute<cpu>", NumpyBinaryBackwardUseIn<cpu, mshadow_op::posone,
+                                                                mshadow_op::negone>);
 
 MXNET_OPERATOR_REGISTER_NP_BINARY_MIXED_PRECISION(_npi_multiply)
 #ifndef _WIN32
@@ -158,9 +188,33 @@ NNVM_REGISTER_OP(_backward_npi_broadcast_mul)
 .set_attr<FCompute>("FCompute<cpu>", NumpyBinaryBackwardUseIn<cpu, mshadow_op::right,
                                                               mshadow_op::left>);
 
-MXNET_OPERATOR_REGISTER_BINARY_BROADCAST(_npi_mod)
-.set_attr<FCompute>("FCompute<cpu>", BinaryBroadcastCompute<cpu, mshadow_op::mod>)
-.set_attr<nnvm::FGradient>("FGradient", ElemwiseGradUseIn{"_backward_broadcast_mod"});
+MXNET_OPERATOR_REGISTER_NP_BINARY_MIXED_PRECISION(_npi_mod)
+#ifndef _WIN32
+.set_attr<FCompute>(
+  "FCompute<cpu>",
+  NumpyBinaryBroadcastCompute<cpu, op::mshadow_op::mod, op::mshadow_op::mixed_mod,
+                                      op::mshadow_op::mixed_rmod>)
+#else
+.set_attr<FCompute>(
+  "FCompute<cpu>",
+  NumpyBinaryBroadcastCompute<cpu, op::mshadow_op::mod>)
+#endif
+.set_attr<nnvm::FGradient>("FGradient", ElemwiseGradUseIn{"_backward_npi_broadcast_mod"});
+
+NNVM_REGISTER_OP(_backward_npi_broadcast_mod)
+.set_num_inputs(3)
+.set_num_outputs(2)
+.set_attr<nnvm::TIsBackward>("TIsBackward", true)
+.set_attr<nnvm::FInplaceOption>("FInplaceOption",
+  [](const NodeAttrs& attrs){
+    return std::vector<std::pair<int, int> >{{0, 1}};
+  })
+.set_attr<FResourceRequest>("FResourceRequest",
+  [](const NodeAttrs& attrs) {
+    return std::vector<ResourceRequest>{ResourceRequest::kTempSpace};
+  })
+.set_attr<FCompute>("FCompute<cpu>", NumpyBinaryBackwardUseIn<cpu, mshadow_op::mod_grad,
+                                                              mshadow_op::mod_rgrad>);
 
 MXNET_OPERATOR_REGISTER_NP_BINARY_MIXED_PRECISION(_npi_power)
 #ifndef _WIN32

src/operator/numpy/np_elemwise_broadcast_op.cu

@@ -40,6 +40,10 @@ NNVM_REGISTER_OP(_npi_add)
   NumpyBinaryBroadcastComputeWithBool<gpu, op::mshadow_op::plus>);
 #endif
 
+NNVM_REGISTER_OP(_backward_npi_broadcast_add)
+.set_attr<FCompute>("FCompute<gpu>", NumpyBinaryBackwardUseIn<gpu, mshadow_op::posone,
+                                                                mshadow_op::posone>);
+
 NNVM_REGISTER_OP(_npi_subtract)
 #ifndef _WIN32
 .set_attr<FCompute>(
@@ -52,6 +56,10 @@ NNVM_REGISTER_OP(_npi_subtract)
   NumpyBinaryBroadcastCompute<gpu, op::mshadow_op::minus>);
 #endif
 
+NNVM_REGISTER_OP(_backward_npi_broadcast_sub)
+.set_attr<FCompute>("FCompute<gpu>", NumpyBinaryBackwardUseIn<gpu, mshadow_op::posone,
+                                                                mshadow_op::negone>);
+
 NNVM_REGISTER_OP(_npi_multiply)
 #ifndef _WIN32
 .set_attr<FCompute>(
@@ -69,7 +77,20 @@ NNVM_REGISTER_OP(_backward_npi_broadcast_mul)
                                                               mshadow_op::left>);
 
 NNVM_REGISTER_OP(_npi_mod)
-.set_attr<FCompute>("FCompute<gpu>", BinaryBroadcastCompute<gpu, mshadow_op::mod>);
+#ifndef _WIN32
+.set_attr<FCompute>(
+  "FCompute<gpu>",
+  NumpyBinaryBroadcastCompute<gpu, op::mshadow_op::mod, op::mshadow_op::mixed_mod,
+                                      op::mshadow_op::mixed_rmod>);
+#else
+.set_attr<FCompute>(
+  "FCompute<gpu>",
+  NumpyBinaryBroadcastCompute<gpu, op::mshadow_op::mod>);
+#endif
+
+NNVM_REGISTER_OP(_backward_npi_broadcast_mod)
+.set_attr<FCompute>("FCompute<gpu>", NumpyBinaryBackwardUseIn<gpu, mshadow_op::mod_grad,
+                                                              mshadow_op::mod_rgrad>);
 
 NNVM_REGISTER_OP(_npi_power)
 #ifndef _WIN32

src/operator/numpy/np_true_divide-inl.h

@@ -29,6 +29,7 @@
 #include <vector>
 #include "../../common/utils.h"
 #include "../tensor/elemwise_binary_broadcast_op.h"
+#include "../numpy/np_elemwise_broadcast_op.h"
 
 namespace mxnet {
 namespace op {

src/operator/numpy/np_true_divide.cc

@@ -80,10 +80,26 @@ NNVM_REGISTER_OP(_npi_true_divide)
   })
 #endif
 .set_attr<FCompute>("FCompute<cpu>", TrueDivideBroadcastCompute<cpu>)
-.set_attr<nnvm::FGradient>("FGradient", ElemwiseGradUseIn{"_backward_broadcast_div"})
+.set_attr<nnvm::FGradient>("FGradient", ElemwiseGradUseIn{"_backward_npi_broadcast_div"})
 .add_argument("lhs", "NDArray-or-Symbol", "Dividend array")
 .add_argument("rhs", "NDArray-or-Symbol", "Divisor array");
 
+
+NNVM_REGISTER_OP(_backward_npi_broadcast_div)
+.set_num_inputs(3)
+.set_num_outputs(2)
+.set_attr<nnvm::TIsBackward>("TIsBackward", true)
+.set_attr<nnvm::FInplaceOption>("FInplaceOption",
+  [](const NodeAttrs& attrs){
+    return std::vector<std::pair<int, int> >{{0, 1}};
+  })
+.set_attr<FResourceRequest>("FResourceRequest",
+  [](const NodeAttrs& attrs) {
+    return std::vector<ResourceRequest>{ResourceRequest::kTempSpace};
+  })
+.set_attr<FCompute>("FCompute<cpu>", NumpyBinaryBackwardUseIn<cpu, mshadow_op::div_grad,
+                                                              mshadow_op::div_rgrad>);
+
 NNVM_REGISTER_OP(_npi_true_divide_scalar)
 .set_num_inputs(1)
 .set_num_outputs(1)

src/operator/numpy/np_true_divide.cu

@@ -31,6 +31,10 @@ namespace op {
 NNVM_REGISTER_OP(_npi_true_divide)
 .set_attr<FCompute>("FCompute<gpu>", TrueDivideBroadcastCompute<gpu>);
 
+NNVM_REGISTER_OP(_backward_npi_broadcast_div)
+.set_attr<FCompute>("FCompute<gpu>", NumpyBinaryBackwardUseIn<gpu, mshadow_op::div_grad,
+                                                              mshadow_op::div_rgrad>);
+
 NNVM_REGISTER_OP(_npi_true_divide_scalar)
 .set_attr<FCompute>("FCompute<gpu>", TrueDivideScalarCompute<gpu, mshadow_op::true_divide>);
 

src/operator/operator_tune.cc

@@ -430,6 +430,8 @@ IMPLEMENT_BINARY_WORKLOAD_FWD(mxnet::op::mshadow_op::rldexp);  // NOLINT()
 IMPLEMENT_BINARY_WORKLOAD_BWD(mxnet::op::mshadow_op::ldexp_grad);  // NOLINT()
 IMPLEMENT_BINARY_WORKLOAD_BWD(mxnet::op::mshadow_op::ldexp_rgrad);  // NOLINT()
 IMPLEMENT_BINARY_WORKLOAD_BWD(mxnet::op::mshadow_op::rldexp_grad);  // NOLINT()
+IMPLEMENT_BINARY_WORKLOAD_BWD(mxnet::op::mshadow_op::posone); // NOLINT()
+IMPLEMENT_BINARY_WORKLOAD_BWD(mxnet::op::mshadow_op::negone); // NOLINT()
 /*!
  * \brief Tuner objects, *not* automatically generated
  */

tests/python/unittest/test_numpy_op.py

@@ -2618,10 +2618,13 @@ def hybrid_forward(self, F, a, b, *args, **kwargs):
                             use_broadcast=False, equal_nan=True)
 
     funcs = {
-        'add': (-1.0, 1.0, None, None),
-        'subtract': (-1.0, 1.0, None, None),
+        'add': (-1.0, 1.0, lambda y, x1, x2: _np.ones(y.shape),
+                           lambda y, x1, x2: _np.ones(y.shape)),
+        'subtract': (-1.0, 1.0, lambda y, x1, x2: _np.ones(y.shape),
+                                lambda y, x1, x2: _np.ones(y.shape) * -1),
         'multiply': (-1.0, 1.0, lambda y, x1, x2: _np.broadcast_to(x2, y.shape),
                                 lambda y, x1, x2: _np.broadcast_to(x1, y.shape)),
+        'mod': (1.0, 5.0, None, None),
         'power': (1.0, 3.0, lambda y, x1, x2: _np.power(x1, x2 - 1.0) * x2,
                             lambda y, x1, x2: _np.power(x1, x2) * _np.log(x1)),
     }
@@ -2649,7 +2652,7 @@ def hybrid_forward(self, F, a, b, *args, **kwargs):
                     continue
                 check_mixed_precision_binary_func(func, low, high, lshape, rshape, lgrad, rgrad, type1, type2)
 
-            if func == 'subtract':
+            if func == 'subtract' or func == 'mod':
                 continue
             for type1, type2 in itertools.product(itypes, itypes):
                 if type1 == type2: