Merge develop

paddle/fluid/framework/data_type.cc

@@ -64,6 +64,7 @@ static DataTypeMap* InitDataTypeMap() {
   RegType(size_t, proto::VarType::SIZE_T);
   RegType(int16_t, proto::VarType::INT16);
   RegType(uint8_t, proto::VarType::UINT8);
+  RegType(int8_t, proto::VarType::INT8);
 
 #undef RegType
   return retv;

paddle/fluid/framework/data_type.h

@@ -54,6 +54,9 @@ inline void VisitDataType(proto::VarType::Type type, Visitor visitor) {
     case proto::VarType::INT16:
       visitor.template operator()<int16_t>();
       break;
+    case proto::VarType::INT8:
+      visitor.template operator()<int8_t>();
+      break;
     default:
       PADDLE_THROW("Not supported %d", type);
   }

paddle/fluid/framework/framework.proto

@@ -107,6 +107,7 @@ message VarType {
     // Tensor<size_t> is used in C++.
     SIZE_T = 19;
     UINT8 = 20;
+    INT8 = 21;
 
     // Other types that may need additional descriptions
     LOD_TENSOR = 7;

paddle/fluid/operators/math/math_function.cc

@@ -41,7 +41,8 @@ template struct SetConstant<platform::CPUDeviceContext, uint8_t>;
   template struct Transpose<platform::CPUDeviceContext, int64_t, RANK>;    \
   template struct Transpose<platform::CPUDeviceContext, bool, RANK>;       \
   template struct Transpose<platform::CPUDeviceContext, int16_t, RANK>;    \
-  template struct Transpose<platform::CPUDeviceContext, uint8_t, RANK>;
+  template struct Transpose<platform::CPUDeviceContext, uint8_t, RANK>;    \
+  template struct Transpose<platform::CPUDeviceContext, int8_t, RANK>;
 
 DEFINE_CPU_TRANS(1);
 DEFINE_CPU_TRANS(2);

paddle/fluid/operators/math/math_function.cu

@@ -33,10 +33,11 @@ template struct SetConstant<platform::CUDADeviceContext, int>;
 template struct SetConstant<platform::CUDADeviceContext, int64_t>;
 template struct SetConstant<platform::CUDADeviceContext, bool>;
 
-#define DEFINE_GPU_TRANS(RANK)                                          \
-  template struct Transpose<platform::CUDADeviceContext, float, RANK>;  \
-  template struct Transpose<platform::CUDADeviceContext, double, RANK>; \
-  template struct Transpose<platform::CUDADeviceContext, float16, RANK>;
+#define DEFINE_GPU_TRANS(RANK)                                           \
+  template struct Transpose<platform::CUDADeviceContext, float, RANK>;   \
+  template struct Transpose<platform::CUDADeviceContext, double, RANK>;  \
+  template struct Transpose<platform::CUDADeviceContext, float16, RANK>; \
+  template struct Transpose<platform::CUDADeviceContext, int8_t, RANK>;
 
 DEFINE_GPU_TRANS(1);
 DEFINE_GPU_TRANS(2);

paddle/fluid/pybind/protobuf.cc

@@ -234,6 +234,7 @@ void BindVarDsec(pybind11::module *m) {
   pybind11::enum_<pd::proto::VarType::Type>(var_desc, "VarType", "")
       .value("BOOL", pd::proto::VarType::BOOL)
       .value("UINT8", pd::proto::VarType::UINT8)
+      .value("INT8", pd::proto::VarType::INT8)
       .value("INT16", pd::proto::VarType::INT16)
       .value("INT32", pd::proto::VarType::INT32)
       .value("INT64", pd::proto::VarType::INT64)

paddle/fluid/pybind/pybind.cc

@@ -130,6 +130,7 @@ PYBIND11_PLUGIN(core) {
       .def("set", PyCPUTensorSetFromArray<bool>)
       .def("set", PyCPUTensorSetFromArray<uint16_t>)
       .def("set", PyCPUTensorSetFromArray<uint8_t>)
+      .def("set", PyCPUTensorSetFromArray<int8_t>)
 #ifdef PADDLE_WITH_CUDA
       .def("set", PyCUDATensorSetFromArray<float>)
       .def("set", PyCUDATensorSetFromArray<int>)
@@ -138,6 +139,7 @@ PYBIND11_PLUGIN(core) {
       .def("set", PyCUDATensorSetFromArray<bool>)
       .def("set", PyCUDATensorSetFromArray<uint16_t>)
       .def("set", PyCUDATensorSetFromArray<uint8_t>)
+      .def("set", PyCUDATensorSetFromArray<int8_t>)
       .def("set", PyCUDAPinnedTensorSetFromArray<float>)
       .def("set", PyCUDAPinnedTensorSetFromArray<int>)
       .def("set", PyCUDAPinnedTensorSetFromArray<double>)
@@ -145,6 +147,7 @@ PYBIND11_PLUGIN(core) {
       .def("set", PyCUDAPinnedTensorSetFromArray<bool>)
       .def("set", PyCUDAPinnedTensorSetFromArray<uint16_t>)
       .def("set", PyCUDAPinnedTensorSetFromArray<uint8_t>)
+      .def("set", PyCUDAPinnedTensorSetFromArray<int8_t>)
 #endif
       .def("shape", [](Tensor &self) { return vectorize(self.dims()); })
       .def("_set_float_element", TensorSetElement<float>)

paddle/fluid/pybind/tensor_py.h

@@ -97,7 +97,7 @@ struct CastToPyBufferImpl<true, I, ARGS...> {
 inline pybind11::buffer_info CastToPyBuffer(const framework::Tensor &tensor) {
   auto buffer_info =
       details::CastToPyBufferImpl<true, 0, float, int, double, int64_t, bool,
-                                  uint8_t, platform::float16>()(tensor);
+                                  uint8_t, int8_t, platform::float16>()(tensor);
   return buffer_info;
 }
 

python/paddle/fluid/framework.py

@@ -95,6 +95,8 @@ def convert_np_dtype_to_dtype_(np_dtype):
         return core.VarDesc.VarType.INT16
     elif dtype == np.uint8:
         return core.VarDesc.VarType.UINT8
+    elif dtype == np.int8:
+        return core.VarDesc.VarType.INT8
     else:
         raise ValueError("Not supported numpy dtype %s" % dtype)
 

python/paddle/fluid/tests/unittests/test_tensor.py

@@ -59,6 +59,27 @@ class TestTensor(unittest.TestCase):
         self.assertAlmostEqual(1.0, tensor_array_2[3, 9])
         self.assertAlmostEqual(2.0, tensor_array_2[19, 11])
 
+    def test_int8_tensor(self):
+        scope = core.Scope()
+        var = scope.var("int8_tensor")
+        cpu_tensor = var.get_tensor()
+        tensor_array = numpy.random.randint(
+            -127, high=128, size=[100, 200], dtype=numpy.int8)
+        place = core.CPUPlace()
+        cpu_tensor.set(tensor_array, place)
+        cpu_tensor_array_2 = numpy.array(cpu_tensor)
+        self.assertAlmostEqual(cpu_tensor_array_2.all(), tensor_array.all())
+
+        if core.is_compiled_with_cuda():
+            cuda_tensor = var.get_tensor()
+            tensor_array = numpy.random.randint(
+                -127, high=128, size=[100, 200], dtype=numpy.int8)
+            place = core.CUDAPlace(0)
+            cuda_tensor.set(tensor_array, place)
+            cuda_tensor_array_2 = numpy.array(cuda_tensor)
+            self.assertAlmostEqual(cuda_tensor_array_2.all(),
+                                   tensor_array.all())
+
     def test_int_lod_tensor(self):
         place = core.CPUPlace()
         scope = core.Scope()

python/paddle/fluid/tests/unittests/test_variable.py

@@ -31,7 +31,8 @@ class TestVariable(unittest.TestCase):
         self.assertEqual(DT.INT16, convert("int16"))
         self.assertEqual(DT.INT64, convert("int64"))
         self.assertEqual(DT.BOOL, convert("bool"))
-        self.assertRaises(ValueError, lambda: convert("int8"))
+        self.assertEqual(DT.INT8, convert("int8"))
+        self.assertEqual(DT.UINT8, convert("uint8"))
 
     def test_var(self):
         b = default_main_program().current_block()