[0-size Tensor Job2 No.15、84] Add 0-size Tensor support for paddle.gather [fluid_ops]

paddle/phi/infermeta/binary.cc

@@ -2015,12 +2015,14 @@ void GatherInferMeta(const MetaTensor& x,
   auto index_dims = index.dims();
 
   if (index_dims.size() == 2) {
-    PADDLE_ENFORCE_EQ(
-        index_dims[1],
-        1,
-        common::errors::InvalidArgument(
-            "The last dim of index should be 1 when it is 2D, but we get %d",
-            index_dims[1]));
+    if (index_dims[1] != 0) {
+      PADDLE_ENFORCE_EQ(
+          index_dims[1],
+          1,
+          common::errors::InvalidArgument("The last dim of index should be 0 "
+                                          "or 1 when it is 2D, but we get %d",
+                                          index_dims[1]));
+    }
   } else {
     PADDLE_ENFORCE_EQ(
         index_dims.size() == 1 || index_dims.size() == 0,
@@ -2089,13 +2091,19 @@ void GatherInferMeta(const MetaTensor& x,
     if (axis.FromTensor() || axis_v == 0) {
       // if axis.FromTensor(), we can not obtain correct shape of output
       int batch_size = static_cast<int>(index_dims[0]);
+      if (index_dims.size() == 2 && index_dims[1] == 0) {
+        batch_size = 0;
+      }
       phi::DDim output_dims(input_dim);
       output_dims[0] = batch_size;
       out->set_dims(output_dims);
       out->set_dtype(x.dtype());
       out->share_lod(x);
     } else {
       int index_size = static_cast<int>(index_dims[0]);
+      if (index_dims.size() == 2 && index_dims[1] == 0) {
+        index_size = 0;
+      }
       std::vector<int> out_dim_vec;
       for (int i = 0; i < axis_v; i++) {
         out_dim_vec.push_back(input_dim[i]);  // NOLINT

paddle/phi/kernels/cpu/gather_grad_kernel.cc

@@ -16,6 +16,7 @@
 
 #include "paddle/phi/common/bfloat16.h"
 #include "paddle/phi/core/kernel_registry.h"
+#include "paddle/phi/kernels/full_kernel.h"
 #include "paddle/phi/kernels/funcs/eigen/common.h"
 #include "paddle/phi/kernels/funcs/gather.h"
 #include "paddle/phi/kernels/funcs/scatter.h"
@@ -29,6 +30,13 @@ void GatherGradKernel(const Context& dev_ctx,
                       const DenseTensor& out_grad,
                       const Scalar& axis,
                       DenseTensor* x_grad) {
+  if (out_grad.numel() == 0) {
+    if (x_grad) {
+      phi::Full<T, Context>(
+          dev_ctx, phi::IntArray(common::vectorize(x_grad->dims())), 0, x_grad);
+    }
+    return;
+  }
   const auto& index_type = index.dtype();
   auto axis_v = axis.to<int>();
   if (axis_v < 0) {

paddle/phi/kernels/cpu/gather_kernel.cc

@@ -26,6 +26,10 @@ void GatherKernel(const Context& dev_ctx,
                   const DenseTensor& index,
                   const Scalar& axis,
                   DenseTensor* out) {
+  if (out && out->numel() == 0) {
+    dev_ctx.template Alloc<T>(out);
+    return;
+  }
   const auto& index_type = index.dtype();
   auto axis_v = axis.to<int>();
   if (axis_v < 0) {

paddle/phi/kernels/gpu/gather_grad_kernel.cu

@@ -15,11 +15,11 @@
 #include "paddle/phi/common/bfloat16.h"
 #include "paddle/phi/common/float16.h"
 #include "paddle/phi/core/kernel_registry.h"
+#include "paddle/phi/kernels/full_kernel.h"
 #include "paddle/phi/kernels/funcs/eigen/common.h"
 #include "paddle/phi/kernels/funcs/gather.cu.h"
 #include "paddle/phi/kernels/funcs/scatter.cu.h"
 #include "paddle/phi/kernels/gather_kernel.h"
-
 namespace phi {
 
 template <typename T, typename Context>
@@ -29,6 +29,14 @@ void GatherGradKernel(const Context& dev_ctx,
                       const DenseTensor& out_grad,
                       const Scalar& axis,
                       DenseTensor* x_grad) {
+  // x [4, 2], index [2, 0], out [2, 0], x_grad [4, 2]
+  if (out_grad.numel() == 0) {
+    if (x_grad) {
+      phi::Full<T, Context>(
+          dev_ctx, phi::IntArray(common::vectorize(x_grad->dims())), 0, x_grad);
+    }
+    return;
+  }
   const auto& index_type = index.dtype();
   auto axis_v = axis.to<int>();
   if (axis_v < 0) {

paddle/phi/kernels/gpu/gather_kernel.cu

@@ -27,6 +27,10 @@ void GatherKernel(const Context& dev_ctx,
                   const DenseTensor& index,
                   const Scalar& axis,
                   DenseTensor* out) {
+  if (out && out->numel() == 0) {
+    dev_ctx.template Alloc<T>(out);
+    return;
+  }
   const auto& index_type = index.dtype();
   auto axis_v = axis.to<int>();
   if (axis_v < 0) {

paddle/phi/kernels/xpu/gather_grad_kernel.cc

@@ -16,7 +16,7 @@
 
 #include "paddle/phi/backends/xpu/enforce_xpu.h"
 #include "paddle/phi/core/kernel_registry.h"
-
+#include "paddle/phi/kernels/full_kernel.h"
 namespace phi {
 
 template <typename T, typename Context>
@@ -34,6 +34,10 @@ void GatherGradKernel(const Context& dev_ctx,
   const auto& index_type = index.dtype();
 
   if (out_grad.numel() == 0) {
+    if (x_grad) {
+      phi::Full<T, Context>(
+          dev_ctx, phi::IntArray(common::vectorize(x_grad->dims())), 0, x_grad);
+    }
     return;
   }
 

test/legacy_test/test_gather_op.py

@@ -948,6 +948,46 @@ def test_gather_backward(self):
         np.testing.assert_allclose(res_list[0], res_list[1])
 
 
+class TestGatherOp_ZeroSize(OpTest):
+    def setUp(self):
+        self.op_type = "gather"
+        self.python_api = paddle.gather
+        self.public_python_api = paddle.gather
+        self.config()
+        self.init_inputs_and_outputs()
+
+    def test_check_output(self):
+        self.check_output(check_pir=True)
+
+    def test_check_grad(self):
+        self.check_grad(['X'], 'Out', check_pir=True)
+
+    def config(self):
+        self.x_shape = (3, 0, 4)
+        self.config_dtype()
+        self.index = [2]
+        self.index_type = "int32"
+
+    def config_dtype(self):
+        self.x_type = "float64"
+
+    def init_inputs_and_outputs(self):
+        xnp = np.random.random(self.x_shape).astype(self.x_type)
+        self.inputs = {
+            'X': xnp,
+            'Index': np.array(self.index).astype(self.index_type),
+        }
+        self.outputs = {'Out': self.inputs["X"][self.inputs["Index"]]}
+
+
+class TestGatherOp_ZeroSize2(TestGatherOp_ZeroSize):
+    def config(self):
+        self.x_shape = (10, 20)
+        self.config_dtype()
+        self.index = [2, 0]
+        self.index_type = "int32"
+
+
 if __name__ == "__main__":
     paddle.enable_static()
     unittest.main()