bind pixel_shuffle & pixel_shuffle_grad op for xpu (#50090)

paddle/phi/backends/xpu/xpu2_op_list.cc

@@ -415,6 +415,8 @@ XPUOpMap& get_kl2_ops() {
       {"p_norm_grad", XPUKernelSet({phi::DataType::FLOAT32})},
       {"pad3d_grad", XPUKernelSet({phi::DataType::FLOAT32})},
       {"pad3d", XPUKernelSet({phi::DataType::FLOAT32})},
+      {"pixel_shuffle", XPUKernelSet({phi::DataType::FLOAT32})},
+      {"pixel_shuffle_grad", XPUKernelSet({phi::DataType::FLOAT32})},
       {"pool2d_grad",
        XPUKernelSet({phi::DataType::FLOAT32, phi::DataType::FLOAT16})},
       {"pool2d",

paddle/phi/kernels/xpu/pixel_shuffle_grad_kernel.cc

+// Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/phi/kernels/pixel_shuffle_grad_kernel.h"
+
+#include "paddle/phi/backends/xpu/enforce_xpu.h"
+#include "paddle/phi/core/kernel_registry.h"
+
+namespace phi {
+template <typename T, typename Context>
+void PixelShuffleGradKernel(const Context& ctx,
+                            const DenseTensor& out_grad,
+                            int upscale_factor,
+                            const std::string& data_format,
+                            DenseTensor* x_grad) {
+  using XPUType = typename XPUTypeTrait<T>::Type;
+
+  const T* x_ptr = out_grad.data<T>();
+  T* y_ptr = ctx.template Alloc<T>(x_grad);
+
+  bool is_nchw = data_format == "NCHW";
+
+  int64_t n = out_grad.dims()[0];
+  int64_t xc = out_grad.dims()[is_nchw ? 1 : 3];
+  int64_t xh = out_grad.dims()[is_nchw ? 2 : 1];
+  int64_t xw = out_grad.dims()[is_nchw ? 3 : 2];
+
+  int r = pixel_unshuffle(ctx.x_context(),
+                          reinterpret_cast<const XPUType*>(x_ptr),
+                          reinterpret_cast<XPUType*>(y_ptr),
+                          n,
+                          xc,
+                          xh,
+                          xw,
+                          upscale_factor,
+                          is_nchw);
+
+  PADDLE_ENFORCE_XDNN_SUCCESS(r, "pixel_unshuffle");
+}
+}  // namespace phi
+
+PD_REGISTER_KERNEL(
+    pixel_shuffle_grad, XPU, ALL_LAYOUT, phi::PixelShuffleGradKernel, float) {}

paddle/phi/kernels/xpu/pixel_shuffle_kernel.cc

+// Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/phi/kernels/pixel_shuffle_kernel.h"
+
+#include "paddle/phi/backends/xpu/enforce_xpu.h"
+#include "paddle/phi/core/kernel_registry.h"
+
+namespace phi {
+template <typename T, typename Context>
+void PixelShuffleKernel(const Context& ctx,
+                        const DenseTensor& x,
+                        int upscale_factor,
+                        const std::string& data_format,
+                        DenseTensor* out) {
+  using XPUType = typename XPUTypeTrait<T>::Type;
+
+  const T* x_ptr = x.data<T>();
+  T* y_ptr = ctx.template Alloc<T>(out);
+
+  bool is_nchw = data_format == "NCHW";
+
+  int64_t n = x.dims()[0];
+  int64_t xc = x.dims()[is_nchw ? 1 : 3];
+  int64_t xh = x.dims()[is_nchw ? 2 : 1];
+  int64_t xw = x.dims()[is_nchw ? 3 : 2];
+
+  int r = pixel_shuffle(ctx.x_context(),
+                        reinterpret_cast<const XPUType*>(x_ptr),
+                        reinterpret_cast<XPUType*>(y_ptr),
+                        n,
+                        xc,
+                        xh,
+                        xw,
+                        upscale_factor,
+                        is_nchw);
+
+  PADDLE_ENFORCE_XDNN_SUCCESS(r, "pixel_shuffle");
+}
+}  // namespace phi
+
+PD_REGISTER_KERNEL(
+    pixel_shuffle, XPU, ALL_LAYOUT, phi::PixelShuffleKernel, float) {}

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

+#   Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import sys
+import unittest
+
+import numpy as np
+
+sys.path.append("..")
+
+from op_test_xpu import XPUOpTest
+from xpu.get_test_cover_info import (
+    XPUOpTestWrapper,
+    create_test_class,
+    get_xpu_op_support_types,
+)
+
+import paddle
+
+paddle.enable_static()
+
+
+def pixel_shuffle_np(x, up_factor, data_format="NCHW"):
+    if data_format == "NCHW":
+        n, c, h, w = x.shape
+        new_shape = (
+            n,
+            c // (up_factor * up_factor),
+            up_factor,
+            up_factor,
+            h,
+            w,
+        )
+        # reshape to (num,output_channel,upscale_factor,upscale_factor,h,w)
+        npresult = np.reshape(x, new_shape)
+        # transpose to (num,output_channel,h,upscale_factor,w,upscale_factor)
+        npresult = npresult.transpose(0, 1, 4, 2, 5, 3)
+        oshape = [n, c // (up_factor * up_factor), h * up_factor, w * up_factor]
+        npresult = np.reshape(npresult, oshape)
+        return npresult
+    else:
+        n, h, w, c = x.shape
+        new_shape = (
+            n,
+            h,
+            w,
+            c // (up_factor * up_factor),
+            up_factor,
+            up_factor,
+        )
+        # reshape to (num,h,w,output_channel,upscale_factor,upscale_factor)
+        npresult = np.reshape(x, new_shape)
+        # transpose to (num,h,upscale_factor,w,upscale_factor,output_channel)
+        npresult = npresult.transpose(0, 1, 4, 2, 5, 3)
+        oshape = [n, h * up_factor, w * up_factor, c // (up_factor * up_factor)]
+        npresult = np.reshape(npresult, oshape)
+        return npresult
+
+
+class XPUTestPixelShuffleOp(XPUOpTestWrapper):
+    def __init__(self):
+        self.op_name = "pixel_shuffle"
+        self.use_dynamic_create_class = False
+
+    class TestPixelShuffleOp(XPUOpTest):
+        def setUp(self):
+            self.set_xpu()
+            self.op_type = "pixel_shuffle"
+            self.init_dtype()
+            self.eager_mode = True
+
+            # override
+            self.init_input_shape()
+            self.init_attr()
+
+            self.x = np.random.random(self.x_shape).astype(self.dtype)
+            self.y = pixel_shuffle_np(
+                self.x, self.attrs["upscale_factor"], self.attrs["data_format"]
+            )
+
+            self.inputs = {'X': self.x}
+            self.outputs = {'Out': self.y}
+
+        def init_input_shape(self):
+            self.x_shape = [2, 64, 26, 26]
+
+        def init_attr(self):
+            self.attrs = {'upscale_factor': 2, 'data_format': "NCHW"}
+
+        def set_xpu(self):
+            self.__class__.no_need_check_grad = False
+            self.place = paddle.XPUPlace(0)
+
+        def init_dtype(self):
+            self.dtype = self.in_type
+
+        def test_check_output(self):
+            self.check_output_with_place(self.place)
+
+        def test_check_grad(self):
+            self.check_grad_with_place(
+                self.place, ['X'], 'Out', check_eager=self.eager_mode
+            )
+
+    class TestNHWC(TestPixelShuffleOp):
+        def init_input_shape(self):
+            self.x_shape = [2, 64, 26, 24]
+
+        def init_attr(self):
+            self.attrs = {'upscale_factor': 2, 'data_format': "NHWC"}
+
+    class TestUpFactor3(TestPixelShuffleOp):
+        def init_input_shape(self):
+            self.x_shape = [2, 27, 5, 5]
+
+        def init_attr(self):
+            self.attrs = {'upscale_factor': 3, 'data_format': "NCHW"}
+
+    class TestUpFactor3NHWC(TestPixelShuffleOp):
+        def init_input_shape(self):
+            self.x_shape = [2, 27, 5, 9]
+
+        def init_attr(self):
+            self.attrs = {'upscale_factor': 3, 'data_format': "NHWC"}
+
+
+support_types = get_xpu_op_support_types("pixel_shuffle")
+for stype in support_types:
+    create_test_class(globals(), XPUTestPixelShuffleOp, stype)
+
+if __name__ == "__main__":
+    unittest.main()