Added fp32 / bf16 forward and backward elementwise_div_mkldnn operator

paddle/fluid/operators/elementwise/mkldnn/elementwise_div_mkldnn_op.cc

@@ -0,0 +1,147 @@
+/* Copyright (c) 2020 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/fluid/operators/elementwise/mkldnn/elementwise_mkldnn_op.h"
+
+namespace paddle {
+namespace framework {
+class ExecutionContext;
+}  // namespace framework
+namespace platform {
+class CPUDeviceContext;
+struct CPUPlace;
+}  // namespace platform
+}  // namespace paddle
+
+namespace paddle {
+namespace operators {
+template <typename T>
+class EltwiseDivMKLDNNGradKernel : public ElemwiseGradKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    ElemwiseGradKernel<T>::Compute(ctx);
+
+    auto& dev_ctx =
+        ctx.template device_context<paddle::platform::MKLDNNDeviceContext>();
+    const auto& mkldnn_engine = dev_ctx.GetEngine();
+
+    auto* y = ctx.Input<framework::Tensor>("Y");
+    auto* out = ctx.Input<framework::Tensor>("Out");
+    auto* dout = ctx.Input<framework::Tensor>(framework::GradVarName("Out"));
+    auto* dx = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
+    auto* dy = ctx.Output<framework::Tensor>(framework::GradVarName("Y"));
+    int axis = ctx.Attr<int>("axis");
+
+    auto& astream = platform::MKLDNNDeviceContext::tls().get_stream();
+
+    if (dx) {
+      // dx = dout / y
+
+      platform::BinaryMKLDNNHandler<T> handler(
+          dnnl::algorithm::binary_div, axis, mkldnn_engine, ctx.GetPlace(),
+          dout, y, dx, 1.0f, 1.0f, 1.0f);
+
+      const auto src_dout_memory = handler.AcquireSrcMemory(dout);
+      const auto src_y_memory = handler.AcquireSecondSrcMemory(y);
+      const auto dst_dx_memory = handler.AcquireDstMemory(dx);
+
+      const auto binary_prim = handler.AcquireForwardPrimitive();
+
+      const std::unordered_map<int, dnnl::memory> args = {
+          {DNNL_ARG_SRC_0, *src_dout_memory},
+          {DNNL_ARG_SRC_1, *src_y_memory},
+          {DNNL_ARG_DST, *dst_dx_memory}};
+
+      binary_prim->execute(astream, args);
+      astream.wait();
+
+      dx->set_layout(framework::DataLayout::kMKLDNN);
+      dx->set_format(platform::GetMKLDNNFormat(*dst_dx_memory));
+    }
+
+    if (dy) {
+      // dy = -dout * out / y
+
+      platform::BinaryMKLDNNHandler<T> y_handler(
+          dnnl::algorithm::binary_div, axis, mkldnn_engine, ctx.GetPlace(), y,
+          y, nullptr, 1.0f, 1.0f, 1.0f);
+
+      const auto y_memory = y_handler.AcquireSrcMemory(y);
+
+      dnnl::post_ops po;
+      po.append_binary(dnnl::algorithm::binary_div, y_memory->get_desc());
+
+      platform::BinaryMKLDNNHandler<T> handler(
+          dnnl::algorithm::binary_mul, axis, mkldnn_engine, ctx.GetPlace(),
+          dout, out, nullptr, -1.0f, 1.0f, 1.0f, po);
+
+      const auto src_dout_memory = handler.AcquireSrcMemory(dout);
+      const auto src_out_memory = handler.AcquireSecondSrcMemory(out);
+
+      // If broadcasting is in use then let's write to temporary
+      // buffer allocated by oneDNN
+      const auto dst_dy_memory = (dout->dims() == dy->dims())
+                                     ? handler.AcquireDstMemory(dy)
+                                     : handler.AcquireDstMemory();
+
+      const auto binary_prim = handler.AcquireForwardPrimitive();
+
+      const std::unordered_map<int, dnnl::memory> args = {
+          {DNNL_ARG_SRC_0, *src_dout_memory},
+          {DNNL_ARG_SRC_1, *src_out_memory},
+          {DNNL_ARG_DST, *dst_dy_memory},
+          {DNNL_ARG_ATTR_MULTIPLE_POST_OP(0) | DNNL_ARG_SRC_1, *y_memory}};
+
+      binary_prim->execute(astream, args);
+      astream.wait();
+
+      dy->set_layout(framework::DataLayout::kMKLDNN);
+
+      // Reduction is needed for broadcasting scenario
+      if (dout->dims() != dy->dims()) {
+        platform::ReductionMKLDNNHandler<T> handler_sum(
+            dnnl::algorithm::reduction_sum, 0.0f, 0.0f, mkldnn_engine,
+            ctx.GetPlace(), dout, dy, CalculateBroadcastedDims(dout, dy));
+        auto dy_memory_p = handler_sum.AcquireDstMemory(dy);
+        auto reduction_p = handler_sum.AcquireForwardPrimitive();
+
+        // As source we use mem object with results from binary operation
+        reduction_p->execute(astream, {{DNNL_ARG_SRC, *dst_dy_memory},
+                                       {DNNL_ARG_DST, *dy_memory_p}});
+        astream.wait();
+        dy->set_format(
+            platform::GetMKLDNNFormat(dy_memory_p->get_desc().reshape(
+                paddle::framework::vectorize<int64_t>(dy->dims()))));
+
+      } else {
+        dy->set_format(platform::GetMKLDNNFormat(*dst_dy_memory));
+      }
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+
+// TODO(piotrekobi) add int8, uint8 support
+REGISTER_OP_KERNEL(elementwise_div, MKLDNN, paddle::platform::CPUPlace,
+                   ops::EltwiseMKLDNNKernel<float, dnnl::algorithm::binary_div>,
+                   ops::EltwiseMKLDNNKernel<paddle::platform::bfloat16,
+                                            dnnl::algorithm::binary_div>)
+
+REGISTER_OP_KERNEL(elementwise_div_grad, MKLDNN, paddle::platform::CPUPlace,
+                   ops::EltwiseDivMKLDNNGradKernel<paddle::platform::bfloat16>,
+                   ops::EltwiseDivMKLDNNGradKernel<float>)

paddle/fluid/platform/mkldnn_reuse.h

@@ -799,7 +799,8 @@ class BinaryMKLDNNHandler
   BinaryMKLDNNHandler(const dnnl::algorithm algo, const int axis,
                       const mkldnn::engine engine, platform::Place cpu_place,
                       const Tensor* x, const Tensor* y, Tensor* z,
-                      float scale_x, float scale_y, float scale_z)
+                      float scale_x, float scale_y, float scale_z,
+                      const dnnl::post_ops& post_ops = dnnl::post_ops())
       : platform::MKLDNNHandlerNoCachingT<T, dnnl::binary>(engine, cpu_place) {
     PADDLE_ENFORCE_EQ(
         x->layout(), DataLayout::kMKLDNN,
@@ -848,10 +849,11 @@ class BinaryMKLDNNHandler
                                      MKLDNNMemoryFormat::any);
 
     auto attributes = CreateAttributes(algo, scale_x, scale_y, scale_z);
+    attributes.set_post_ops(post_ops);
+
     this->AcquireForwardPrimitiveDescriptor(attributes, algo, src0_md, src1_md,
                                             dst_md);
   }
-
   std::shared_ptr<mkldnn::memory> AcquireSecondSrcMemory(
       const framework::Tensor* input) {
     const T* input_data = input->data<T>();
@@ -1332,15 +1334,18 @@ class ConvMKLDNNTemplateHandler : public MKLDNNHandler {
       int mask = output_shift_scale.size() > 1 ? 1 << 1 : 0;
       conv_attr.set_output_scales(mask, output_shift_scale);
     }
-    // Fusion with Elementwise layer relies on adding a sum post-operation with
-    // the scale parameter. It is assumed that when fuse_residual_connection is
+    // Fusion with Elementwise layer relies on adding a sum post-operation
+    // with
+    // the scale parameter. It is assumed that when fuse_residual_connection
+    // is
     // true, the output tensor contains the data coming from residual
     // connection. The result of this post_op is:
     // Output = scale * Output + Conv_Out.
     if (fuse_residual_conn) {
       post_operations.append_sum(sum_scale);
     }
-    // Fusion with ReLU layer is executed through the PostOps feature. Create a
+    // Fusion with ReLU layer is executed through the PostOps feature. Create
+    // a
     // PostOps object and configure it to execute an eltwise relu operation.
     if (fuse_activation == "relu" || fuse_activation == "leaky_relu") {
       constexpr float scale = 1.0f;

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

@@ -0,0 +1,179 @@
+# Copyright (c) 2021 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.
+
+from __future__ import print_function
+import unittest
+import numpy as np
+from paddle import enable_static
+from paddle.fluid.tests.unittests.op_test import OpTest, OpTestTool, convert_float_to_uint16
+from paddle.fluid.framework import _current_expected_place
+import paddle.fluid.core as core
+
+
+@OpTestTool.skip_if(not (isinstance(_current_expected_place(), core.CPUPlace)),
+                    "GPU is not supported")
+class TestMKLDNNElementwiseDivOp(OpTest):
+    def setUp(self):
+        self.op_type = "elementwise_div"
+        self.init_dtype()
+        self.init_input_output()
+        self.init_kernel_type()
+        self.init_axis()
+        self.inputs = {
+            'X': OpTest.np_dtype_to_fluid_dtype(self.x),
+            'Y': OpTest.np_dtype_to_fluid_dtype(self.y)
+        }
+        self.attrs = {'axis': self.axis, 'use_mkldnn': self.use_mkldnn}
+        self.outputs = {'Out': self.out}
+
+    def init_input_output(self):
+        self.x = np.random.uniform(0.1, 1, [13, 17]).astype(self.dtype)
+        self.y = np.random.uniform(0.1, 1, [13, 17]).astype(self.dtype)
+        self.out = np.divide(self.x, self.y)
+
+    def test_check_grad_normal(self):
+        self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.05)
+
+    def test_check_grad_ignore_x(self):
+        self.check_grad(
+            ['Y'], 'Out', no_grad_set=set("X"), max_relative_error=0.05)
+
+    def test_check_grad_ignore_y(self):
+        self.check_grad(
+            ['X'], 'Out', no_grad_set=set('Y'), max_relative_error=0.05)
+
+    def init_axis(self):
+        self.axis = -1
+
+    def init_kernel_type(self):
+        self.use_mkldnn = True
+
+    def init_dtype(self):
+        self.dtype = np.float32
+
+    def test_check_output(self):
+        self.check_output()
+
+
+class TestMKLDNNElementwiseDivOp2(TestMKLDNNElementwiseDivOp):
+    def init_input_output(self):
+        self.x = np.random.uniform(0.1, 1, [100]).astype(self.dtype)
+        self.y = np.random.uniform(0.1, 1, [100]).astype(self.dtype)
+        self.out = np.divide(self.x, self.y)
+
+
+class TestMKLDNNElementwiseDivOp3(TestMKLDNNElementwiseDivOp):
+    def init_input_output(self):
+        self.x = np.random.uniform(0.1, 1, [2, 3, 4, 5]).astype(self.dtype)
+        self.y = np.random.uniform(0.1, 1, [2, 3, 4, 5]).astype(self.dtype)
+        self.out = np.divide(self.x, self.y)
+
+
+class TestMKLDNNElementwiseDivOp4(TestMKLDNNElementwiseDivOp):
+    def init_input_output(self):
+        self.x = np.random.uniform(1, 2, [2, 3, 4, 32]).astype(self.dtype)
+        self.y = np.random.uniform(1, 2, [4, 32]).astype(self.dtype)
+        self.out = np.divide(self.x, self.y)
+
+    def test_check_grad_normal(self):
+        pass
+
+    def test_check_grad_ignore_x(self):
+        pass
+
+
+class TestMKLDNNElementwiseDivOp5(TestMKLDNNElementwiseDivOp):
+    def init_input_output(self):
+        self.x = np.random.uniform(1, 2, [2, 3, 4, 100]).astype(self.dtype)
+        self.y = np.random.uniform(1, 2, [100]).astype(self.dtype)
+        self.out = np.divide(self.x, self.y)
+
+    def test_check_grad_normal(self):
+        pass
+
+    def test_check_grad_ignore_x(self):
+        pass
+
+
+@OpTestTool.skip_if_not_cpu_bf16()
+class TestBf16(TestMKLDNNElementwiseDivOp):
+    def setUp(self):
+        self.op_type = "elementwise_div"
+        self.init_dtype()
+        self.init_input_output()
+        self.init_kernel_type()
+        self.init_axis()
+
+        self.x_bf16 = convert_float_to_uint16(self.x)
+        self.y_bf16 = convert_float_to_uint16(self.y)
+        self.inputs = {'X': self.x_bf16, 'Y': self.y_bf16}
+        self.attrs = {'axis': self.axis, 'use_mkldnn': self.use_mkldnn}
+        self.outputs = {'Out': convert_float_to_uint16(self.out)}
+
+    def init_dtype(self):
+        self.dtype = np.float32
+        self.mkldnn_data_type = "bfloat16"
+
+    def init_input_output(self):
+        self.x = np.random.uniform(0.1, 1, [100]).astype(self.dtype)
+        self.y = np.random.uniform(0.1, 1, [100]).astype(self.dtype)
+        self.out = np.divide(self.x, self.y)
+
+    def test_check_output(self):
+        self.check_output_with_place(core.CPUPlace())
+
+    def test_check_grad_normal(self):
+        self.check_grad_with_place(
+            core.CPUPlace(), ["X", "Y"],
+            "Out",
+            user_defined_grads=[
+                np.divide(self.x, self.y), np.divide(
+                    (np.multiply(-self.x, self.x)), np.multiply(self.y, self.y))
+            ],
+            user_defined_grad_outputs=[self.x_bf16])
+
+    def test_check_grad_ignore_x(self):
+        self.check_grad_with_place(
+            core.CPUPlace(), ["Y"],
+            "Out",
+            user_defined_grads=[
+                np.divide((np.multiply(-self.x, self.y)),
+                          np.multiply(self.y, self.y))
+            ],
+            user_defined_grad_outputs=[self.y_bf16])
+
+    def test_check_grad_ignore_y(self):
+        self.check_grad_with_place(
+            core.CPUPlace(), ["X"],
+            "Out",
+            user_defined_grads=[np.divide(self.x, self.y)],
+            user_defined_grad_outputs=[self.x_bf16])
+
+
+class TestBf16Broadcasting(TestBf16):
+    def init_input_output(self):
+        self.x = np.random.uniform(1, 2, [2, 3, 4, 100]).astype(self.dtype)
+        self.y = np.random.uniform(1, 2, [100]).astype(self.dtype)
+        self.out = np.subtract(self.x, self.y)
+
+    def test_check_grad_normal(self):
+        pass
+
+    def test_check_grad_ignore_x(self):
+        pass
+
+
+if __name__ == '__main__':
+    enable_static()
+    unittest.main()

tools/static_mode_white_list.py

@@ -610,6 +610,7 @@
     'test_dequantize_mkldnn_op',
     'test_elementwise_add_mkldnn_op',
     'test_elementwise_add_bf16_mkldnn_op',
+    'test_elementwise_div_mkldnn_op',
     'test_elementwise_sub_mkldnn_op',
     'test_elementwise_mul_mkldnn_op',
     'test_elementwise_mul_bf16_mkldnn_op',