[KP] Fix registry and add UT for thresholded_relu & softshrink (#40524)

* init commit

* correct namespace

paddle/fluid/operators/activation_op.kps

@@ -15,6 +15,8 @@ limitations under the License. */
 #include "paddle/fluid/platform/bfloat16.h"
 #include "paddle/fluid/platform/device/gpu/gpu_device_function.h"
 
+#include "paddle/phi/kernels/funcs/activation_functor.h"
+
 namespace paddle {
 namespace operators {
 
@@ -1148,63 +1150,221 @@ REGISTER_OP_CUDA_KERNEL(
 FOR_EACH_ACTIVATION_CUDA_OP(REGISTER_ACTIVATION_CUDA_KERNEL)
 
 #ifdef PADDLE_WITH_XPU_KP
-#define REGISTER_ACTIVATION_XPU_KERNEL(act_type, op_name, functor,             \
-                                       grad_functor)                           \
-  REGISTER_OP_KERNEL(                                                          \
-      act_type, KP, plat::XPUPlace,                                            \
-      ops::ActivationCudaKernel<plat::XPUDeviceContext, ops::functor<float>>); \
-  REGISTER_OP_KERNEL(act_type##_grad, KP, plat::XPUPlace,                      \
-                     ops::ActivationGradCudaKernel<plat::XPUDeviceContext,     \
-                                                   ops::grad_functor<float>>);
-
-REGISTER_ACTIVATION_XPU_KERNEL(leaky_relu, LeakyRelu, CudaLeakyReluFunctor,
-                               CudaLeakyReluGradFunctor);
-REGISTER_ACTIVATION_XPU_KERNEL(sigmoid, Sigmoid, CudaSigmoidFunctor,
-                               CudaSigmoidGradFunctor);
-REGISTER_ACTIVATION_XPU_KERNEL(exp, Exp, CudaExpFunctor, CudaExpGradFunctor);
-REGISTER_ACTIVATION_XPU_KERNEL(log, Log, CudaLogFunctor, CudaLogGradFunctor);
-REGISTER_ACTIVATION_XPU_KERNEL(reciprocal, Reciprocal, CudaReciprocalFunctor,
-                               CudaReciprocalGradFunctor);
-REGISTER_ACTIVATION_XPU_KERNEL(softplus, Softplus, CudaSoftplusFunctor,
-                               CudaSoftplusGradFunctor);
-REGISTER_ACTIVATION_XPU_KERNEL(hard_swish, HardSwish, CudaHardSwishFunctor,
-                               CudaHardSwishGradFunctor);
-REGISTER_ACTIVATION_XPU_KERNEL(elu, Elu, CudaELUFunctor, CudaELUGradFunctor);
-REGISTER_ACTIVATION_XPU_KERNEL(celu, Celu, CudaCELUFunctor,
-                               CudaCELUGradFunctor);
-REGISTER_ACTIVATION_XPU_KERNEL(sqrt, Sqrt, CudaSqrtFunctor,
-                               CudaSqrtGradFunctor);
-REGISTER_ACTIVATION_XPU_KERNEL(square, Square, CudaSquareFunctor,
-                               CudaSquareGradFunctor);
-REGISTER_ACTIVATION_XPU_KERNEL(silu, Silu, CudaSiluFunctor,
-                               CudaSiluGradFunctor);
-REGISTER_ACTIVATION_XPU_KERNEL(logsigmoid, LogSigmoid, CudaLogSigmoidFunctor,
-                               CudaLogSigmoidGradFunctor);
-REGISTER_ACTIVATION_XPU_KERNEL(softshrink, SoftShrink, CudaSoftShrinkFunctor,
-                               CudaSoftShrinkGradFunctor);
-REGISTER_ACTIVATION_XPU_KERNEL(ceil, Ceil, CudaCeilFunctor,
-                               CudaZeroGradFunctor);
-REGISTER_ACTIVATION_XPU_KERNEL(floor, Floor, CudaFloorFunctor,
-                               CudaZeroGradFunctor);
-REGISTER_ACTIVATION_XPU_KERNEL(log1p, Log1p, CudaLog1pFunctor,
-                               CudaLog1pGradFunctor);
-REGISTER_ACTIVATION_XPU_KERNEL(brelu, BRelu, CudaBReluFunctor,
-                               CudaBReluGradFunctor);
-REGISTER_ACTIVATION_XPU_KERNEL(soft_relu, SoftRelu, CudaSoftReluFunctor,
-                               CudaSoftReluGradFunctor);
-REGISTER_ACTIVATION_XPU_KERNEL(softsign, Softsign, CudaSoftsignFunctor,
-                               CudaSoftsignGradFunctor);
-REGISTER_ACTIVATION_XPU_KERNEL(relu6, Relu6, CudaRelu6Functor,
-                               CudaRelu6GradFunctor);
-REGISTER_ACTIVATION_XPU_KERNEL(hard_shrink, HardShrink, CudaHardShrinkFunctor,
-                               CudaHardShrinkGradFunctor);
-REGISTER_ACTIVATION_XPU_KERNEL(hard_sigmoid, HardSigmoid,
-                               CudaHardSigmoidFunctor,
-                               CudaHardSigmoidGradFunctor);
-REGISTER_ACTIVATION_XPU_KERNEL(swish, Swish, CudaSwishFunctor,
-                               CudaSwishGradFunctor);
-REGISTER_ACTIVATION_XPU_KERNEL(thresholded_relu, ThresholdedRelu,
-                               CudaThresholdedReluFunctor,
-                               CudaThresholdedReluGradFunctor);
+REGISTER_OP_KERNEL(
+    brelu, KP, plat::XPUPlace,
+    ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
+                              phi::funcs::CudaBReluFunctor<float>>);
+REGISTER_OP_KERNEL(
+    brelu_grad, KP, plat::XPUPlace,
+    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
+                                  phi::funcs::CudaBReluGradFunctor<float>>);
+
+REGISTER_OP_KERNEL(ceil, KP, plat::XPUPlace,
+                   ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
+                                             ops::CudaCeilFunctor<float>>);
+REGISTER_OP_KERNEL(
+    ceil_grad, KP, plat::XPUPlace,
+    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
+                                  ops::CudaZeroGradFunctor<float>>);
+
+REGISTER_OP_KERNEL(celu, KP, plat::XPUPlace,
+                   ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
+                                             ops::CudaCELUFunctor<float>>);
+REGISTER_OP_KERNEL(
+    celu_grad, KP, plat::XPUPlace,
+    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
+                                  ops::CudaCELUGradFunctor<float>>);
+
+REGISTER_OP_KERNEL(elu, KP, plat::XPUPlace,
+                   ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
+                                             ops::CudaELUFunctor<float>>);
+REGISTER_OP_KERNEL(
+    elu_grad, KP, plat::XPUPlace,
+    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
+                                  ops::CudaELUGradFunctor<float>>);
+
+REGISTER_OP_KERNEL(exp, KP, plat::XPUPlace,
+                   ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
+                                             ops::CudaExpFunctor<float>>);
+REGISTER_OP_KERNEL(
+    exp_grad, KP, plat::XPUPlace,
+    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
+                                  ops::CudaExpGradFunctor<float>>);
+
+REGISTER_OP_KERNEL(floor, KP, plat::XPUPlace,
+                   ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
+                                             ops::CudaFloorFunctor<float>>);
+REGISTER_OP_KERNEL(
+    floor_grad, KP, plat::XPUPlace,
+    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
+                                  ops::CudaZeroGradFunctor<float>>);
+
+REGISTER_OP_KERNEL(
+    hard_shrink, KP, plat::XPUPlace,
+    ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
+                              ops::CudaHardShrinkFunctor<float>>);
+REGISTER_OP_KERNEL(
+    hard_shrink_grad, KP, plat::XPUPlace,
+    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
+                                  ops::CudaHardShrinkGradFunctor<float>>);
+
+REGISTER_OP_KERNEL(
+    hard_sigmoid, KP, plat::XPUPlace,
+    ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
+                              ops::CudaHardSigmoidFunctor<float>>);
+REGISTER_OP_KERNEL(
+    hard_sigmoid_grad, KP, plat::XPUPlace,
+    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
+                                  ops::CudaHardSigmoidGradFunctor<float>>);
+
+REGISTER_OP_KERNEL(hard_swish, KP, plat::XPUPlace,
+                   ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
+                                             ops::CudaHardSwishFunctor<float>>);
+REGISTER_OP_KERNEL(
+    hard_swish_grad, KP, plat::XPUPlace,
+    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
+                                  ops::CudaHardSwishGradFunctor<float>>);
+
+REGISTER_OP_KERNEL(
+    leaky_relu, KP, plat::XPUPlace,
+    ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
+                              phi::funcs::CudaLeakyReluFunctor<float>>);
+REGISTER_OP_KERNEL(
+    leaky_relu_grad, KP, plat::XPUPlace,
+    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
+                                  phi::funcs::CudaLeakyReluGradFunctor<float>>);
+
+REGISTER_OP_KERNEL(log, KP, plat::XPUPlace,
+                   ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
+                                             ops::CudaLogFunctor<float>>);
+REGISTER_OP_KERNEL(
+    log_grad, KP, plat::XPUPlace,
+    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
+                                  ops::CudaLogGradFunctor<float>>);
+
+REGISTER_OP_KERNEL(log1p, KP, plat::XPUPlace,
+                   ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
+                                             ops::CudaLog1pFunctor<float>>);
+REGISTER_OP_KERNEL(
+    log1p_grad, KP, plat::XPUPlace,
+    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
+                                  ops::CudaLog1pGradFunctor<float>>);
+
+REGISTER_OP_KERNEL(
+    logsigmoid, KP, plat::XPUPlace,
+    ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
+                              ops::CudaLogSigmoidFunctor<float>>);
+REGISTER_OP_KERNEL(
+    logsigmoid_grad, KP, plat::XPUPlace,
+    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
+                                  ops::CudaLogSigmoidGradFunctor<float>>);
+
+REGISTER_OP_KERNEL(
+    reciprocal, KP, plat::XPUPlace,
+    ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
+                              ops::CudaReciprocalFunctor<float>>);
+REGISTER_OP_KERNEL(
+    reciprocal_grad, KP, plat::XPUPlace,
+    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
+                                  ops::CudaReciprocalGradFunctor<float>>);
+
+REGISTER_OP_KERNEL(
+    relu, KP, plat::XPUPlace,
+    ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
+                              phi::funcs::CudaReluFunctor<float>>);
+REGISTER_OP_KERNEL(
+    relu_grad, KP, plat::XPUPlace,
+    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
+                                  phi::funcs::CudaReluGradFunctor<float>>);
+
+REGISTER_OP_KERNEL(relu6, KP, plat::XPUPlace,
+                   ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
+                                             ops::CudaRelu6Functor<float>>);
+REGISTER_OP_KERNEL(
+    relu6_grad, KP, plat::XPUPlace,
+    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
+                                  ops::CudaRelu6GradFunctor<float>>);
+
+REGISTER_OP_KERNEL(sigmoid, KP, plat::XPUPlace,
+                   ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
+                                             ops::CudaSigmoidFunctor<float>>);
+REGISTER_OP_KERNEL(
+    sigmoid_grad, KP, plat::XPUPlace,
+    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
+                                  ops::CudaSigmoidGradFunctor<float>>);
+
+REGISTER_OP_KERNEL(silu, KP, plat::XPUPlace,
+                   ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
+                                             ops::CudaSiluFunctor<float>>);
+REGISTER_OP_KERNEL(
+    silu_grad, KP, plat::XPUPlace,
+    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
+                                  ops::CudaSiluGradFunctor<float>>);
+
+REGISTER_OP_KERNEL(soft_relu, KP, plat::XPUPlace,
+                   ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
+                                             ops::CudaSoftReluFunctor<float>>);
+REGISTER_OP_KERNEL(
+    soft_relu_grad, KP, plat::XPUPlace,
+    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
+                                  ops::CudaSoftReluGradFunctor<float>>);
+
+REGISTER_OP_KERNEL(softplus, KP, plat::XPUPlace,
+                   ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
+                                             ops::CudaSoftplusFunctor<float>>);
+REGISTER_OP_KERNEL(
+    softplus_grad, KP, plat::XPUPlace,
+    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
+                                  ops::CudaSoftplusGradFunctor<float>>);
+
+REGISTER_OP_KERNEL(
+    softshrink, KP, plat::XPUPlace,
+    ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
+                              ops::CudaSoftShrinkFunctor<float>>);
+REGISTER_OP_KERNEL(
+    softshrink_grad, KP, plat::XPUPlace,
+    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
+                                  ops::CudaSoftShrinkGradFunctor<float>>);
+
+REGISTER_OP_KERNEL(softsign, KP, plat::XPUPlace,
+                   ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
+                                             ops::CudaSoftsignFunctor<float>>);
+REGISTER_OP_KERNEL(
+    softsign_grad, KP, plat::XPUPlace,
+    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
+                                  ops::CudaSoftsignGradFunctor<float>>);
+
+REGISTER_OP_KERNEL(sqrt, KP, plat::XPUPlace,
+                   ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
+                                             ops::CudaSqrtFunctor<float>>);
+REGISTER_OP_KERNEL(
+    sqrt_grad, KP, plat::XPUPlace,
+    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
+                                  ops::CudaSqrtGradFunctor<float>>);
+
+REGISTER_OP_KERNEL(square, KP, plat::XPUPlace,
+                   ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
+                                             ops::CudaSquareFunctor<float>>);
+REGISTER_OP_KERNEL(
+    square_grad, KP, plat::XPUPlace,
+    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
+                                  ops::CudaSquareGradFunctor<float>>);
+
+REGISTER_OP_KERNEL(swish, KP, plat::XPUPlace,
+                   ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
+                                             ops::CudaSwishFunctor<float>>);
+REGISTER_OP_KERNEL(
+    swish_grad, KP, plat::XPUPlace,
+    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
+                                  ops::CudaSwishGradFunctor<float>>);
+
+REGISTER_OP_KERNEL(
+    thresholded_relu, KP, plat::XPUPlace,
+    ops::ActivationCudaKernel<paddle::platform::XPUDeviceContext,
+                              ops::CudaThresholdedReluFunctor<float>>);
+REGISTER_OP_KERNEL(
+    thresholded_relu_grad, KP, plat::XPUPlace,
+    ops::ActivationGradCudaKernel<paddle::platform::XPUDeviceContext,
+                                  ops::CudaThresholdedReluGradFunctor<float>>);
 
 #endif  // PADDLE_WITH_XPU_KP

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

@@ -849,6 +849,38 @@ def ref_softsign(x):
     return out
 
 
+class XPUTestSoftshrinkOP(XPUOpTestWrapper):
+    def __init__(self):
+        self.op_name = 'softshrink'
+        self.use_dynamic_create_class = False
+
+    class XPUTestSoftshrink(TestActivationOPBase):
+        def set_case(self):
+            self.op_type = "softshrink"
+            self.dtype = self.in_type
+
+            threshold = 0.5
+            np.random.seed(1023)
+            x = np.random.uniform(0.25, 10, [10, 12]).astype(self.dtype)
+            out = ref_softshrink(x, threshold)
+
+            self.inputs = {'X': x}
+            self.outputs = {'Out': out}
+            self.attrs = {'use_xpu': True}
+
+
+support_types = get_xpu_op_support_types('softshrink')
+for stype in support_types:
+    create_test_class(globals(), XPUTestSoftshrinkOP, stype)
+
+
+def ref_softshrink(x, threshold=0.5):
+    out = np.copy(x)
+    out = (out < -threshold) * (out + threshold) + (out > threshold) * (
+        out - threshold)
+    return out
+
+
 class XPUTestSwishOP(XPUOpTestWrapper):
     def __init__(self):
         self.op_name = 'swish'
@@ -879,5 +911,36 @@ def ref_swish(x):
     return out
 
 
+class XPUTestThresholdedReluOP(XPUOpTestWrapper):
+    def __init__(self):
+        self.op_name = 'thresholded_relu'
+        self.use_dynamic_create_class = False
+
+    class XPUTestThresholdedRelu(TestActivationOPBase):
+        def set_case(self):
+            self.op_type = "thresholded_relu"
+            self.dtype = self.in_type
+
+            threshold = 1.0
+            np.random.seed(1024)
+            x = np.random.uniform(-20, 20, [10, 12]).astype(self.dtype)
+            x[np.abs(x) < 0.005] = 0.02
+            out = ref_thresholded_relu(x, threshold)
+
+            self.inputs = {'X': x}
+            self.outputs = {'Out': out}
+            self.attrs = {'use_xpu': True}
+
+
+support_types = get_xpu_op_support_types('thresholded_relu')
+for stype in support_types:
+    create_test_class(globals(), XPUTestThresholdedReluOP, stype)
+
+
+def ref_thresholded_relu(x, threshold=1.0):
+    out = (x > threshold) * x
+    return out
+
+
 if __name__ == "__main__":
     unittest.main()