swish and pow op for xpu test=kunlun (#40654)

* swish and pow op for xpu. test=kunlun

* fix code style. test=kunlun.

* use pow_grad xdnn api. test=kunlun.

cmake/external/xpu.cmake

@@ -36,7 +36,7 @@ ENDIF()
 
 if(NOT DEFINED XPU_BASE_URL)
   SET(XPU_BASE_URL_WITHOUT_DATE "https://baidu-kunlun-product.cdn.bcebos.com/KL-SDK/klsdk-dev")
-  SET(XPU_BASE_URL "${XPU_BASE_URL_WITHOUT_DATE}/20220324")
+  SET(XPU_BASE_URL "${XPU_BASE_URL_WITHOUT_DATE}/20220327")
 else()
   SET(XPU_BASE_URL "${XPU_BASE_URL}")
 endif()

paddle/fluid/operators/activation_op_xpu.cc

@@ -340,29 +340,56 @@ struct XPUPowFunctor : public BaseActivationFunctor<T> {
     auto pow_factor = ctx.Attr<float>("factor");
     const T *x_data = x->data<T>();
     T *y_data = y->mutable_data<T>(ctx.GetPlace());
-    T *factor_data = nullptr;
 
+    // allocate temp memory for factor on xpu
     auto xpu_context =
         ctx.device_context<paddle::platform::XPUDeviceContext>().x_context();
-    PADDLE_ENFORCE_EQ(xpu_malloc(reinterpret_cast<void **>(&factor_data),
-                                 x->numel() * sizeof(T)),
-                      XPU_SUCCESS, platform::errors::ResourceExhausted(
-                                       "XPU has no enough memory"));
-    int r = xpu::constant<T>(xpu_context, factor_data, x->numel(), pow_factor);
-    PADDLE_ENFORCE_EQ(
-        r, xpu::Error_t::SUCCESS,
-        platform::errors::External("XPU constant op return"
-                                   " wrong value[%d %s] in pow op.",
-                                   r, XPUAPIErrorMsg[r]));
-    r = xpu::pow(xpu_context, x_data, factor_data, y_data, x->numel());
-    PADDLE_ENFORCE_EQ(
-        r, xpu::Error_t::SUCCESS,
-        platform::errors::External("XPU pow op return wrong value[%d %s].", r,
-                                   XPUAPIErrorMsg[r]));
-    if (xpu_context->xpu_stream != nullptr) {
-      xpu_wait(xpu_context->xpu_stream);
+    xpu::ctx_guard RAII_GUARD(xpu_context);
+    T *factor_data = RAII_GUARD.alloc_l3_or_gm<T>(1);
+    PADDLE_ENFORCE_NOT_NULL(
+        factor_data,
+        platform::errors::External("XPU alloc_l3_or_gm returns nullptr"));
+    memory::Copy(ctx.GetPlace(), static_cast<void *>(factor_data),
+                 platform::CPUPlace(), static_cast<void *>(&pow_factor),
+                 sizeof(T));
+
+    // broadcast_pow(Context* ctx, const T* x, const T* y, T* z, const
+    // std::vector<int>& xshape, const std::vector<int>& yshape);
+    auto x_dims = phi::vectorize<int>(x->dims());
+    int r = xpu::broadcast_pow(xpu_context, x_data, factor_data, y_data, x_dims,
+                               {1});
+    PADDLE_ENFORCE_XDNN_SUCCESS(r, "broadcast_pow");
   }
-    xpu_free(factor_data);
+};
+
+template <typename T>
+struct XPUPowGradFunctor : public BaseActivationFunctor<T> {
+  void operator()(const framework::ExecutionContext &ctx) const {
+    const auto *x = ctx.Input<Tensor>("X");
+    auto *dOut = ctx.Input<framework::Tensor>(framework::GradVarName("Out"));
+    auto *dX = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
+
+    const T *x_data = x->data<T>();
+    const T *y_grad = dOut->data<T>();
+    T *x_grad = dX->mutable_data<T>(ctx.GetPlace());
+
+    // check dims: all dims should equal
+    auto x_dims = phi::vectorize<int>(x->dims());
+    auto dy_dims = phi::vectorize<int>(dOut->dims());
+    auto dx_dims = phi::vectorize<int>(dX->dims());
+    PADDLE_ENFORCE_EQ(x_dims, dy_dims, platform::errors::PreconditionNotMet(
+                                           "x_dims should match dy_dims."));
+    PADDLE_ENFORCE_EQ(x_dims, dx_dims, platform::errors::PreconditionNotMet(
+                                           "x_dims should match dx_dims."));
+    float pow_factor = ctx.Attr<float>("factor");
+
+    auto xpu_context =
+        ctx.device_context<paddle::platform::XPUDeviceContext>().x_context();
+    // int pow_grad(Context* ctx, const T* x, const T* dy, T* dx, int len, float
+    // factor);
+    int r = xpu::pow_grad(xpu_context, x_data, y_grad, x_grad, x->numel(),
+                          pow_factor);
+    PADDLE_ENFORCE_XDNN_SUCCESS(r, "pow_grad");
   }
 };
 
@@ -410,6 +437,40 @@ struct XPUSoftPlusGradFunctor : public BaseActivationFunctor<T> {
   }
 };
 
+template <typename T>
+struct XPUSwishFunctor : public BaseActivationFunctor<T> {
+  void operator()(const framework::ExecutionContext &ctx) const {
+    const auto *x = ctx.Input<Tensor>("X");
+    auto *y = ctx.Output<Tensor>("Out");
+    const T *x_data = x->data<T>();
+    T *y_data = y->mutable_data<T>(ctx.GetPlace());
+
+    auto xpu_context =
+        ctx.device_context<paddle::platform::XPUDeviceContext>().x_context();
+    // int swish(Context* ctx, const T* x, T* y, int len);
+    int r = xpu::swish(xpu_context, x_data, y_data, x->numel());
+    PADDLE_ENFORCE_XDNN_SUCCESS(r, "swish");
+  }
+};
+
+template <typename T>
+struct XPUSwishGradFunctor : public BaseActivationFunctor<T> {
+  void operator()(const framework::ExecutionContext &ctx) const {
+    const auto *x = ctx.Input<Tensor>("X");
+    auto *dOut = ctx.Input<framework::Tensor>(framework::GradVarName("Out"));
+    auto *dX = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
+    const T *x_data = x->data<T>();
+    const T *y_grad = dOut->data<T>();
+    T *x_grad = dX->mutable_data<T>(ctx.GetPlace());
+
+    auto xpu_context =
+        ctx.device_context<paddle::platform::XPUDeviceContext>().x_context();
+    // int swish_grad(Context* ctx, const T* x, const T* dy, T* dx, int len);
+    int r = xpu::swish_grad(xpu_context, x_data, y_grad, x_grad, dX->numel());
+    PADDLE_ENFORCE_XDNN_SUCCESS(r, "swish_grad");
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
 
@@ -436,6 +497,8 @@ REGISTER_ACTIVATION_XPU_KERNEL(sqrt, XPUSqrtFunctor, XPUSqrtGradFunctor)
 REGISTER_ACTIVATION_XPU_KERNEL(square, XPUSquareFunctor, XPUSquareGradFunctor)
 REGISTER_ACTIVATION_XPU_KERNEL(softplus, XPUSoftPlusFunctor,
                                XPUSoftPlusGradFunctor)
+REGISTER_ACTIVATION_XPU_KERNEL(swish, XPUSwishFunctor, XPUSwishGradFunctor)
+REGISTER_ACTIVATION_XPU_KERNEL(pow, XPUPowFunctor, XPUPowGradFunctor)
 
 REGISTER_OP_XPU_KERNEL(
     tanh, ops::XPUActivationKernel<ops::XPUTanhFunctor<float>>,
@@ -449,7 +512,5 @@ REGISTER_OP_XPU_KERNEL(exp,
                        ops::XPUActivationKernel<ops::XPUExpFunctor<float>>);
 REGISTER_OP_XPU_KERNEL(log,
                        ops::XPUActivationKernel<ops::XPULogFunctor<float>>);
-REGISTER_OP_XPU_KERNEL(pow,
-                       ops::XPUActivationKernel<ops::XPUPowFunctor<float>>);
 
 #endif  // PADDLE_WITH_XPU

paddle/fluid/platform/device/xpu/xpu2_op_list.h

@@ -259,6 +259,8 @@ XPUOpMap& get_kl2_ops() {
                                     pOpKernelType(vartype::FP16, XPUPlace())})},
       {"pool2d", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace()),
                                pOpKernelType(vartype::FP16, XPUPlace())})},
+      {"pow", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})},
+      {"pow_grad", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})},
       {"prior_box", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})},
       {"range", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace()),
                               pOpKernelType(vartype::INT64, XPUPlace())})},
@@ -352,6 +354,8 @@ XPUOpMap& get_kl2_ops() {
                                    pOpKernelType(vartype::INT32, XPUPlace())})},
       {"sum", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace()),
                             pOpKernelType(vartype::FP16, XPUPlace())})},
+      {"swish", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})},
+      {"swish_grad", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})},
       {"tanh_grad", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace()),
                                   pOpKernelType(vartype::FP16, XPUPlace())})},
       {"tanh", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace()),

python/paddle/fluid/tests/unittests/xpu/test_activation_op_xpu.py

@@ -349,18 +349,55 @@ class XPUTestPowOP(XPUOpTestWrapper):
         self.op_name = 'pow'
         self.use_dynamic_create_class = False
 
-    class XPUTestPow(TestActivationOPBase):
+    class XPUTestPowBase(TestActivationOPBase):
         def set_case(self):
             self.op_type = "pow"
             self.dtype = self.in_type
 
-            x = np.random.uniform(1, 2, [11, 17]).astype(self.dtype)
-            out = np.power(x, 3)
+            self.init_config()
+            out = np.power(self.x, self.factor)
 
-            self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)}
-            self.attrs = {'factor': 3.0, 'use_xpu': True}
+            self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(self.x)}
+            self.attrs = {'factor': self.factor, 'use_xpu': True}
             self.outputs = {'Out': out}
 
+        def init_config(self):
+            self.x = np.random.uniform(-1, 2, [12]).astype(self.dtype)
+            self.factor = 3.0
+
+    class XPUTestPow1(XPUTestPowBase):
+        def init_config(self):
+            self.x = np.random.uniform(-1, 1, [1024, 8]).astype(self.dtype)
+            self.factor = 1
+
+    class XPUTestPow2(XPUTestPowBase):
+        def init_config(self):
+            self.x = np.random.uniform(-1, 1, [1024, 8]).astype(self.dtype)
+            self.factor = 2
+
+    class XPUTestPow3(XPUTestPowBase):
+        def init_config(self):
+            self.x = np.random.uniform(-2, 2,
+                                       [4, 512, 15, 15]).astype(self.dtype)
+            self.factor = 3
+
+    class XPUTestPow4(XPUTestPowBase):
+        def init_config(self):
+            self.x = np.random.uniform(-2, 2,
+                                       [4, 256, 22, 22]).astype(self.dtype)
+            self.factor = 4
+
+    class XPUTestPow5(XPUTestPowBase):
+        def init_config(self):
+            self.x = np.random.uniform(0, 1,
+                                       [4, 256, 22, 22]).astype(self.dtype)
+            self.factor = 1.2
+
+    class XPUTestPow6(XPUTestPowBase):
+        def init_config(self):
+            self.x = np.random.uniform(0, 1, [1024, 8]).astype(self.dtype)
+            self.factor = 3.2
+
 
 support_types = get_xpu_op_support_types('pow')
 for stype in support_types:
@@ -886,19 +923,35 @@ class XPUTestSwishOP(XPUOpTestWrapper):
         self.op_name = 'swish'
         self.use_dynamic_create_class = False
 
-    class XPUTestSwish(TestActivationOPBase):
+    class XPUTestSwishBase(TestActivationOPBase):
         def set_case(self):
             self.op_type = "swish"
             self.dtype = self.in_type
 
-            np.random.seed(1024)
-            x = np.random.uniform(-1, 1, [10, 12]).astype(self.dtype)
-            out = ref_swish(x)
+            self.init_config()
+            out = ref_swish(self.x)
 
-            self.inputs = {'X': x}
+            self.inputs = {'X': self.x}
             self.outputs = {'Out': out}
             self.attrs = {'use_xpu': True}
 
+        def init_config(self):
+            self.x = np.random.uniform(-1, 1, [11, 17]).astype(self.dtype)
+
+    class XPUTestSwish2(XPUTestSwishBase):
+        def init_config(self):
+            self.x = np.random.uniform(-2, 2, [1024, 8]).astype(self.dtype)
+
+    class XPUTestSwish3(XPUTestSwishBase):
+        def init_config(self):
+            self.x = np.random.uniform(-2, 2,
+                                       [4, 512, 15, 15]).astype(self.dtype)
+
+    class XPUTestSwish4(XPUTestSwishBase):
+        def init_config(self):
+            self.x = np.random.uniform(-2, 2,
+                                       [4, 256, 22, 22]).astype(self.dtype)
+
 
 support_types = get_xpu_op_support_types('swish')
 for stype in support_types: