[release/2.6] Exposing Some MIOpen Symbols pytorch#154545 (#2217)

(This is a cherry-pick of
pytorch#154545)

This PR exposes some MIOpen symbols, namely:

1. `miopenDataType_t getMiopenDataType(const at::Tensor& tensor)`
2. `miopenHandle_t getMiopenHandle()`
3. `class TensorDescriptor`
4. `class Descriptor`
5. `class FilterDescriptor`
6. `struct ConvolutionDescriptor`
7. `struct DropoutDescriptor`
8. `struct RNNDescriptor`

to enable adding extensions that make use of them.

Author: apakbin

aten/src/ATen/miopen/Descriptors.h

@@ -38,9 +38,9 @@ struct DescriptorDeleter {
 // initialized the first time you call set() or any other initializing
 // function.
 template <typename T, miopenStatus_t (*ctor)(T**), miopenStatus_t (*dtor)(T*)>
-class Descriptor
-{
-public:
+// NOLINTNEXTLINE(bugprone-exception-escape)
+class TORCH_CUDA_CPP_API Descriptor {
+ public:
   // Use desc() to access the underlying descriptor pointer in
   // a read-only fashion.  Most client code should use this.
   // If the descriptor was never initialized, this will return
@@ -56,7 +56,7 @@ class Descriptor
 protected:
   void init() {
     if (desc_ == nullptr) {
-      T* raw_desc;
+      T* raw_desc = nullptr;
       MIOPEN_CHECK(ctor(&raw_desc));
       desc_.reset(raw_desc);
     }
@@ -65,13 +65,12 @@ class Descriptor
   std::unique_ptr<T, DescriptorDeleter<T, dtor>> desc_;
 };
 
-class TORCH_CUDA_CPP_API TensorDescriptor
-  : public Descriptor<miopenTensorDescriptor,
-                      &miopenCreateTensorDescriptor,
-                      &miopenDestroyTensorDescriptor>
-{
-public:
-  TensorDescriptor() {}
+class TORCH_CUDA_CPP_API TensorDescriptor : public Descriptor<
+                                               miopenTensorDescriptor,
+                                               &miopenCreateTensorDescriptor,
+                                               &miopenDestroyTensorDescriptor> {
+ public:
+  TensorDescriptor() = default;
   explicit TensorDescriptor(const at::Tensor &t, size_t pad = 0) {
     set(t, pad);
   }
@@ -89,11 +88,10 @@ class TORCH_CUDA_CPP_API TensorDescriptor
 
 std::ostream& operator<<(std::ostream & out, const TensorDescriptor& d);
 
-class FilterDescriptor
-  : public Descriptor<miopenTensorDescriptor,
-                      &miopenCreateTensorDescriptor,
-                      &miopenDestroyTensorDescriptor>
-{
+class TORCH_CUDA_CPP_API FilterDescriptor : public Descriptor<
+                                               miopenTensorDescriptor,
+                                               &miopenCreateTensorDescriptor,
+                                               &miopenDestroyTensorDescriptor> {
  public:
   void set(const at::Tensor &t, int64_t pad = 0) {
     set(t, at::MemoryFormat::Contiguous, pad);
@@ -107,11 +105,11 @@ class FilterDescriptor
   }
 };
 
-struct ConvolutionDescriptor
-  : public Descriptor<miopenConvolutionDescriptor,
-                      &miopenCreateConvolutionDescriptor,
-                      &miopenDestroyConvolutionDescriptor>
-{
+struct TORCH_CUDA_CPP_API ConvolutionDescriptor
+    : public Descriptor<
+          miopenConvolutionDescriptor,
+          &miopenCreateConvolutionDescriptor,
+          &miopenDestroyConvolutionDescriptor> {
   void set(miopenDataType_t dataType, miopenConvolutionMode_t c_mode,  int dim, int* pad, int* stride, int * upscale /* aka dilation */, int groups, bool benchmark, bool deterministic) {
     MIOPEN_CHECK(miopenInitConvolutionNdDescriptor(mut_desc(), dim, pad, stride, upscale, c_mode));
     MIOPEN_CHECK(miopenSetConvolutionGroupCount(mut_desc(), groups));
@@ -122,8 +120,24 @@ struct ConvolutionDescriptor
   }
 };
 
+// NOLINTNEXTLINE(bugprone-exception-escape)
+struct TORCH_CUDA_CPP_API DropoutDescriptor
+    : public Descriptor<
+          miopenDropoutDescriptor,
+          &miopenCreateDropoutDescriptor,
+          &miopenDestroyDropoutDescriptor> {
+    void set(miopenHandle_t handle, float dropout, void* states, size_t stateSizeInBytes,
+             unsigned long long seed, bool use_mask, bool state_evo, miopenRNGType_t rng_mode) {
+      MIOPEN_CHECK(miopenSetDropoutDescriptor(mut_desc(), handle, dropout, states, stateSizeInBytes, seed, use_mask, state_evo, rng_mode));
+    }
+
+    void restore(miopenHandle_t handle, float dropout, void* states, size_t stateSizeInBytes,
+      unsigned long long seed, bool use_mask, bool state_evo, miopenRNGType_t rng_mode) {
+      MIOPEN_CHECK(miopenRestoreDropoutDescriptor(mut_desc(), handle, dropout, states, stateSizeInBytes, seed, use_mask, state_evo, rng_mode));
+    }
+};
 
-struct RNNDescriptor
+struct TORCH_CUDA_CPP_API RNNDescriptor
   : public Descriptor<miopenRNNDescriptor,
                       &miopenCreateRNNDescriptor,
                       &miopenDestroyRNNDescriptor>

aten/src/ATen/miopen/Handle.h

@@ -3,8 +3,7 @@
 #include <ATen/miopen/miopen-wrapper.h>
 #include <c10/macros/Export.h>
 
-namespace at { namespace native {
+namespace at::native {
 
 TORCH_CUDA_CPP_API miopenHandle_t getMiopenHandle();
-
-}} // namespace
+} // namespace at::native

aten/src/ATen/miopen/Types.h

@@ -1,13 +1,13 @@
 #pragma once
 
-#include <ATen/miopen/miopen-wrapper.h>
 #include <ATen/Tensor.h>
+#include <ATen/miopen/miopen-wrapper.h>
 #include <c10/macros/Export.h>
 
-namespace at { namespace native {
+namespace at::native {
 
 TORCH_CUDA_CPP_API miopenDataType_t getMiopenDataType(const at::Tensor& tensor);
 
 int64_t miopen_version();
 
-}}  // namespace at::miopen
+} // namespace at::native