[PHI] Migrate softmax kernel (PaddlePaddle#47339)

* add extra attr property set

* add type_info for all context

* add onednn context to all context

* fix context compile error

* simplify conv kernel args

* pass runtime attr into dev_ctx

* fix marco error

* clear conv_grad_kernel extra args

* merge conv_grad_grad into conv_grad

* clear conv2d_grad_grad extra attrs

* remove redundant imports

* migrate softmax

* clear yaml and eager extra attr

* fix conv1d error

* change to thread local

* fix npu compile failed

* try to fix windows compile failed

* add conv2d onednn phi kernel

* fix ci bugs (xuewujiao#36)

* fix compile bugs (xuewujiao#38)

* fix extra input transform bug (xuewujiao#39)

* support dynamic created attr (xuewujiao#40)

* reset extra info gen code

* rm conv_grad_grad kernel

* reimpl pass attr adapting

* add int attr support

* remove vector inputnames creating

* merge dev

* fix map at error

* adjust attribute

* adapt funcs to PHI

Co-authored-by: Chen Weihang <[email protected]>
Co-authored-by: YuanRisheng <[email protected]>

paddle/fluid/framework/ir/mkldnn/mkldnn_inplace_pass_tester.cc

@@ -23,7 +23,7 @@
 #include "paddle/phi/core/kernel_registry.h"
 
 USE_OP_ITSELF(softmax);
-USE_OP_DEVICE_KERNEL(softmax, MKLDNN);
+PD_DECLARE_KERNEL(softmax, OneDNN, ONEDNN);
 USE_OP_ITSELF(elementwise_add);
 USE_OP_DEVICE_KERNEL(elementwise_add, MKLDNN);
 USE_OP_ITSELF(leaky_relu);

paddle/fluid/operators/mkldnn/test_mkldnn_caching.cc

@@ -34,7 +34,7 @@ USE_OP_DEVICE_KERNEL(elementwise_mul, MKLDNN);
 USE_OP_ITSELF(relu);
 PD_DECLARE_KERNEL(relu, OneDNN, ONEDNN);
 USE_OP_ITSELF(softmax);
-USE_OP_DEVICE_KERNEL(softmax, MKLDNN);
+PD_DECLARE_KERNEL(softmax, OneDNN, ONEDNN);
 USE_OP_ITSELF(conv2d);
 PD_DECLARE_KERNEL(conv2d, OneDNN, ONEDNN);
 

paddle/fluid/operators/mkldnn/test_mkldnn_op_inplace.cc

@@ -32,8 +32,7 @@ USE_OP_DEVICE_KERNEL(elementwise_add, MKLDNN);
 USE_OP_ITSELF(relu);
 PD_DECLARE_KERNEL(relu, OneDNN, ONEDNN);
 USE_OP_ITSELF(softmax);
-USE_OP_DEVICE_KERNEL(softmax, MKLDNN);
-
+PD_DECLARE_KERNEL(softmax, OneDNN, ONEDNN);
 PD_DECLARE_KERNEL(softmax, CPU, ALL_LAYOUT);
 
 namespace paddle {

paddle/phi/backends/onednn/onednn_reuse.h

@@ -753,12 +753,19 @@ class SoftmaxOneDNNHandler
  public:
   SoftmaxOneDNNHandler(const dnnl::engine onednn_engine,
                        Place cpu_place,
+                       int axis,
                        const DenseTensor* x,
-                       int axis)
+                       DenseTensor* out)
       : OneDNNHandlerNoCachingT<T,
                                 dnnl::softmax_forward,
                                 dnnl::softmax_backward>(onednn_engine,
                                                         cpu_place) {
+    PADDLE_ENFORCE_EQ(
+        x->dims(),
+        out->dims(),
+        phi::errors::InvalidArgument(
+            "The shape of input and output tensor must be identical."));
+
     const int canonical_axis = funcs::CanonicalAxis(axis, x->dims().size());
     this->AcquireForwardPrimitiveDescriptor(
         dnnl::prop_kind::forward_scoring, x->mem_desc(), canonical_axis);

paddle/phi/kernels/onednn/softmax_kernel.cc

@@ -0,0 +1,61 @@
+// Copyright (c) 2022 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/softmax_kernel.h"
+
+#include "paddle/phi/backends/onednn/onednn_reuse.h"
+#include "paddle/phi/core/kernel_registry.h"
+
+namespace phi {
+
+template <typename T, typename Context>
+void SoftmaxKernel(const Context& dev_ctx,
+                   const DenseTensor& x,
+                   int axis,
+                   DenseTensor* out) {
+  funcs::SoftmaxOneDNNHandler<T> handler(
+      dev_ctx.GetEngine(), dev_ctx.GetPlace(), axis, &x, out);
+
+  auto src_memory_p = handler.AcquireSrcMemory(&x);
+  std::shared_ptr<dnnl::memory> dst_memory_p = nullptr;
+  if (x.IsSharedBufferWith(*out)) {
+    dst_memory_p = src_memory_p;
+    dev_ctx.template Alloc<T>(out);
+  } else {
+    dst_memory_p = handler.AcquireDstMemory(out);
+  }
+  auto softmax_p = handler.AcquireForwardPrimitive();
+
+  auto& astream = OneDNNContext::tls().get_stream();
+  softmax_p->execute(
+      astream, {{DNNL_ARG_SRC, *src_memory_p}, {DNNL_ARG_DST, *dst_memory_p}});
+  astream.wait();
+
+  bool is_test = dev_ctx.HasDnnAttr("is_test")
+                     ? PADDLE_GET_CONST(bool, dev_ctx.GetDnnAttr("is_test"))
+                     : false;
+  if (!is_test) {
+    T* out_data = dev_ctx.template Alloc<T>(out);
+    std::for_each(out_data, &out_data[out->numel()], [](T& val) {
+      val = std::max(val, static_cast<T>(exp(-64)));
+    });
+  }
+
+  out->set_mem_desc(dst_memory_p->get_desc());
+}
+
+}  // namespace phi
+
+PD_REGISTER_KERNEL(
+    softmax, OneDNN, ONEDNN, phi::SoftmaxKernel, float, phi::dtype::bfloat16) {}