[ghstack-poisoned]

Differential Revision: [D22803036](https://our.internmc.facebook.com/intern/diff/D22803036)

The main intention for this change is for the Vulkan Backend buck build.
Where the Vulkan part can be optionally linked to general ATen compiled _without_ USE_VULKAN flag.

For this, we have to remove all preprocessor logic based on `#ifdef USE_VULKAN`.
These includes:
 - Registering Vulkan operators through TORCH_LIBRARY_IMPL and removing them from `native_functions.yaml`
 - Removing `aten/src/ATen/native/vulkan/stub/VulkanAtenStub.cpp` as it registered stubs for vulkan functions without linked Vulkan backend; Changing CMakeLists.txt associated with it
 - Removing preprocessor logic for incode dispatching to vulkan ops in `aten/src/ATen/native/Convolution.cpp`, `aten/src/ATen/native/AdaptiveAveragePooling.cpp`, `aten/src/ATen/native/Copy.cpp`
- Vulkan ops in VulkanAten.cpp without registration in native_functions.yaml can be moved from namespace at::native to more specific at::native::vulkan::aten; Ops are registered in this file with TORCH_LIBRARY_IMPL.


- copy is dispatched as CatchAll. As we do not have preprocessor flag USE_VULKAN, we need to have their code for logic if src or dst tensor is on Vulkan backend.
The same for function `is_vulkan_available()` that we need for test guards and check if vulkan backend initialized successfully - it should be presented with and without Vulkan backend linked.

So now it has `if` condition without preprocessor, that checks if src or dst is Vulkan then calls vulkan_copy from `aten/src/ATen/vulkan/Context.cpp`.

`aten/src/ATen/vulkan/Context.cpp` contains runtime registration of `VulkanImplInterface*` and dispatches `vulkan_copy_` and `is_vulkan_available` to no-op if aten was not linked with Vulkan backend and to real implementation if it was linked with it. (Registration is placed in `VulkanAten.cpp`:
```
static at::vulkan::VulkanImplRegistrar g_vulkan_impl(new VulkanImpl());
```


[ghstack-poisoned]

Differential Revision: [D22803036](https://our.internmc.facebook.com/intern/diff/D22803036)

The main intention for this change is for the Vulkan Backend buck build.
Where the Vulkan part can be optionally linked to general ATen compiled _without_ USE_VULKAN flag.

For this, we have to remove all preprocessor logic based on `#ifdef USE_VULKAN`.
These includes:
 - Registering Vulkan operators through TORCH_LIBRARY_IMPL and removing them from `native_functions.yaml`
 - Removing `aten/src/ATen/native/vulkan/stub/VulkanAtenStub.cpp` as it registered stubs for vulkan functions without linked Vulkan backend; Changing CMakeLists.txt associated with it
 - Removing preprocessor logic for incode dispatching to vulkan ops in `aten/src/ATen/native/Convolution.cpp`, `aten/src/ATen/native/AdaptiveAveragePooling.cpp`, `aten/src/ATen/native/Copy.cpp`
- Vulkan ops in VulkanAten.cpp without registration in native_functions.yaml can be moved from namespace at::native to more specific at::native::vulkan::aten; Ops are registered in this file with TORCH_LIBRARY_IMPL.


- copy is dispatched as CatchAll. As we do not have preprocessor flag USE_VULKAN, we need to have their code for logic if src or dst tensor is on Vulkan backend.
The same for function `is_vulkan_available()` that we need for test guards and check if vulkan backend initialized successfully - it should be presented with and without Vulkan backend linked.

So now it has `if` condition without preprocessor, that checks if src or dst is Vulkan then calls vulkan_copy from `aten/src/ATen/vulkan/Context.cpp`.

`aten/src/ATen/vulkan/Context.cpp` contains runtime registration of `VulkanImplInterface*` and dispatches `vulkan_copy_` and `is_vulkan_available` to no-op if aten was not linked with Vulkan backend and to real implementation if it was linked with it. (Registration is placed in `VulkanAten.cpp`:
```
static at::vulkan::VulkanImplRegistrar g_vulkan_impl(new VulkanImpl());
```


[ghstack-poisoned]

Differential Revision: [D22803036](https://our.internmc.facebook.com/intern/diff/D22803036)

The main intention for this change is for the Vulkan Backend buck build.
Where the Vulkan part can be optionally linked to general ATen compiled _without_ USE_VULKAN flag.

For this, we have to remove all preprocessor logic based on `#ifdef USE_VULKAN`.
These includes:
 - Registering Vulkan operators through TORCH_LIBRARY_IMPL and removing them from `native_functions.yaml`
 - Removing `aten/src/ATen/native/vulkan/stub/VulkanAtenStub.cpp` as it registered stubs for vulkan functions without linked Vulkan backend; Changing CMakeLists.txt associated with it
 - Removing preprocessor logic for incode dispatching to vulkan ops in `aten/src/ATen/native/Convolution.cpp`, `aten/src/ATen/native/AdaptiveAveragePooling.cpp`, `aten/src/ATen/native/Copy.cpp`
- Vulkan ops in VulkanAten.cpp without registration in native_functions.yaml can be moved from namespace at::native to more specific at::native::vulkan::aten; Ops are registered in this file with TORCH_LIBRARY_IMPL.


- copy is dispatched as CatchAll. As we do not have preprocessor flag USE_VULKAN, we need to have their code for logic if src or dst tensor is on Vulkan backend.
The same for function `is_vulkan_available()` that we need for test guards and check if vulkan backend initialized successfully - it should be presented with and without Vulkan backend linked.

So now it has `if` condition without preprocessor, that checks if src or dst is Vulkan then calls vulkan_copy from `aten/src/ATen/vulkan/Context.cpp`.

`aten/src/ATen/vulkan/Context.cpp` contains runtime registration of `VulkanImplInterface*` and dispatches `vulkan_copy_` and `is_vulkan_available` to no-op if aten was not linked with Vulkan backend and to real implementation if it was linked with it. (Registration is placed in `VulkanAten.cpp`:
```
static at::vulkan::VulkanImplRegistrar g_vulkan_impl(new VulkanImpl());
```


[ghstack-poisoned]

Differential Revision: [D22803036](https://our.internmc.facebook.com/intern/diff/D22803036)

The main intention for this change is for the Vulkan Backend buck build.
Where the Vulkan part can be optionally linked to general ATen compiled _without_ USE_VULKAN flag.

For this, we have to remove all preprocessor logic based on `#ifdef USE_VULKAN`.
These includes:
 - Registering Vulkan operators through TORCH_LIBRARY_IMPL and removing them from `native_functions.yaml`
 - Removing `aten/src/ATen/native/vulkan/stub/VulkanAtenStub.cpp` as it registered stubs for vulkan functions without linked Vulkan backend; Changing CMakeLists.txt associated with it
 - Removing preprocessor logic for incode dispatching to vulkan ops in `aten/src/ATen/native/Convolution.cpp`, `aten/src/ATen/native/AdaptiveAveragePooling.cpp`, `aten/src/ATen/native/Copy.cpp`
- Vulkan ops in VulkanAten.cpp without registration in native_functions.yaml can be moved from namespace at::native to more specific at::native::vulkan::aten; Ops are registered in this file with TORCH_LIBRARY_IMPL.


- copy is dispatched as CatchAll. As we do not have preprocessor flag USE_VULKAN, we need to have their code for logic if src or dst tensor is on Vulkan backend.
The same for function `is_vulkan_available()` that we need for test guards and check if vulkan backend initialized successfully - it should be presented with and without Vulkan backend linked.

So now it has `if` condition without preprocessor, that checks if src or dst is Vulkan then calls vulkan_copy from `aten/src/ATen/vulkan/Context.cpp`.

`aten/src/ATen/vulkan/Context.cpp` contains runtime registration of `VulkanImplInterface*` and dispatches `vulkan_copy_` and `is_vulkan_available` to no-op if aten was not linked with Vulkan backend and to real implementation if it was linked with it. (Registration is placed in `VulkanAten.cpp`:
```
static at::vulkan::VulkanImplRegistrar g_vulkan_impl(new VulkanImpl());
```


[ghstack-poisoned]