Added BF16 Pool2d grad (#37081)

* Added BF16 Pool2d grad

* upstream pulled

* fix for CI

* fixes after review

paddle/fluid/operators/mkldnn/pool_mkldnn_op.cc

@@ -387,4 +387,5 @@ REGISTER_OP_KERNEL(pool2d, MKLDNN, ::paddle::platform::CPUPlace,
                    ops::PoolMKLDNNOpKernel<paddle::platform::bfloat16>);
 
 REGISTER_OP_KERNEL(pool2d_grad, MKLDNN, ::paddle::platform::CPUPlace,
-                   ops::PoolMKLDNNGradOpKernel<float>);
+                   ops::PoolMKLDNNGradOpKernel<float>,
+                   ops::PoolMKLDNNGradOpKernel<paddle::platform::bfloat16>);

python/paddle/fluid/tests/unittests/mkldnn/test_pool2d_bf16_mkldnn_op.py

@@ -15,22 +15,63 @@
 from __future__ import print_function
 
 import unittest
-import os
 import numpy as np
 import paddle.fluid.core as core
-from paddle.fluid.tests.unittests.op_test import OpTest, skip_check_grad_ci, convert_float_to_uint16
-from paddle.fluid.tests.unittests.test_pool2d_op import TestPool2D_Op, avg_pool2D_forward_naive, max_pool2D_forward_naive
+from paddle.fluid.tests.unittests.op_test import OpTest, OpTestTool, convert_float_to_uint16
+from paddle.fluid.tests.unittests.test_pool2d_op import TestPool2D_Op_Mixin, max_pool2D_forward_naive
+from paddle.fluid.tests.unittests.npu.test_pool2d_op_npu import pool2d_backward_navie as pool2d_backward_naive
 from paddle import enable_static
 
 
-@unittest.skipIf(not core.supports_bfloat16(),
-                 "place does not support BF16 evaluation")
-class TestPoolBf16MklDNNOp(TestPool2D_Op):
+@OpTestTool.skip_if_not_cpu_bf16()
+class TestPoolBf16MklDNNOpGrad(TestPool2D_Op_Mixin, OpTest):
+    def init_kernel_type(self):
+        self.use_mkldnn = True
+
+    def init_data_type(self):
+        self.dtype = np.uint16
+
+    def setUp(self):
+        super(TestPoolBf16MklDNNOpGrad, self).setUp()
+        self.attrs['mkldnn_data_type'] = "bfloat16"
+        self.x_fp32 = np.random.random(self.shape).astype(np.float32)
+
+        output = self.pool2D_forward_naive(
+            self.x_fp32, self.ksize, self.strides, self.paddings,
+            self.global_pool, self.ceil_mode, self.exclusive, self.adaptive,
+            "float32").astype(np.float32)
+
+        self.inputs = {'X': convert_float_to_uint16(self.x_fp32)}
+        self.outputs = {'Out': convert_float_to_uint16(output)}
+
+    def test_check_output(self):
+        self.check_output_with_place(core.CPUPlace())
+
+    def test_check_grad(self):
+        x_grad = pool2d_backward_naive(
+            self.x_fp32,
+            ksize=self.ksize,
+            strides=self.strides,
+            paddings=self.paddings,
+            global_pool=self.global_pool,
+            ceil_mode=False,
+            exclusive=self.exclusive,
+            adaptive=self.adaptive,
+            data_format=self.data_format,
+            pool_type=self.pool_type,
+            padding_algorithm=self.padding_algorithm)
+        x_grad = x_grad / np.prod(self.outputs['Out'].shape)
+        self.check_grad_with_place(
+            core.CPUPlace(), set(['X']), 'Out', user_defined_grads=[x_grad])
+
+
+@OpTestTool.skip_if_not_cpu_bf16()
+class TestPoolBf16MklDNNOp(TestPool2D_Op_Mixin, OpTest):
     def init_kernel_type(self):
         self.use_mkldnn = True
 
     def setUp(self):
-        TestPool2D_Op.setUp(self)
+        TestPool2D_Op_Mixin.setUp(self)
         self.dtype = np.uint16
 
         input = np.random.random(self.shape).astype(np.float32)
@@ -95,6 +136,47 @@ class TestCase2Max(TestCase2Avg):
         self.pool2D_forward_naive = max_pool2D_forward_naive
 
 
+class TestCase1PadZeroExclusiveAvgGrad(TestPoolBf16MklDNNOpGrad):
+    def init_test_case(self):
+        self.ksize = [3, 3]
+        self.strides = [1, 1]
+
+    def init_shape(self):
+        self.shape = [2, 3, 7, 7]
+
+    def init_paddings(self):
+        self.paddings = [0, 0]
+
+    def init_global_pool(self):
+        self.global_pool = False
+
+    def init_exclusive(self):
+        self.exclusive = True
+
+
+class TestCase2PadOneNonExclusiveAvgGrad(TestCase1PadZeroExclusiveAvgGrad):
+    def init_exclusive(self):
+        self.exclusive = False
+
+
+class TestCase0InitialMaxGrad(TestPoolBf16MklDNNOpGrad):
+    def init_pool_type(self):
+        self.pool_type = "max"
+        self.pool2D_forward_naive = max_pool2D_forward_naive
+
+
+class TestCase1PadZeroExclusiveMaxGrad(TestCase1PadZeroExclusiveAvgGrad):
+    def init_pool_type(self):
+        self.pool_type = "max"
+        self.pool2D_forward_naive = max_pool2D_forward_naive
+
+
+class TestCase2PadOneNonExclusiveMaxGrad(TestCase2PadOneNonExclusiveAvgGrad):
+    def init_pool_type(self):
+        self.pool_type = "max"
+        self.pool2D_forward_naive = max_pool2D_forward_naive
+
+
 if __name__ == "__main__":
     enable_static()
     unittest.main()

python/paddle/fluid/tests/unittests/test_pool2d_op.py

@@ -19,7 +19,7 @@ import unittest
 import numpy as np
 
 import paddle.fluid.core as core
-from op_test import OpTest
+from paddle.fluid.tests.unittests.op_test import OpTest
 import paddle.fluid as fluid
 from paddle.fluid import Program, program_guard
 
@@ -252,7 +252,7 @@ def pool2D_forward_naive(x,
     return out
 
 
-class TestPool2D_Op(OpTest):
+class TestPool2D_Op_Mixin(object):
     def setUp(self):
         self.op_type = "pool2d"
         self.use_cudnn = False
@@ -363,6 +363,10 @@ class TestPool2D_Op(OpTest):
         self.adaptive = False
 
 
+class TestPool2D_Op(TestPool2D_Op_Mixin, OpTest):
+    pass
+
+
 class TestCase1(TestPool2D_Op):
     def init_test_case(self):
         self.ksize = [3, 3]