This project will no longer be maintained by Intel.
Intel has ceased development and contributions including, but not limited to, maintenance, bug fixes, new releases, or updates, to this project.
Intel no longer accepts patches to this project.
If you have an ongoing need to use this project, are interested in independently developing it, or would like to maintain patches for the open source software community, please create your own fork of this project.

Intel offers a subset of the HLSL Shader Model 6.6 64-bit Typed Atomic support through the Intel Extension Framework on our latest Alchemist A-Series GPUs. This paper will detail how to use the Intel Extension Framework for both DirectX 12 API Extensions as well as how to use the extension framework to use a shader (HLSL) extension. Initializing the Extension Framework is an easy task.

It is important to understand that there are limitations to the 64-bit typed atomic support on DG2/Alchemist. Because of these limitations we are exposing what we can support via an extension until we are able to more fully support Shader Model 6.6.

• D3D12_RESOURCE_DIMENSION_TEXTURE2D only
• No mip-map surfaces, no 3D surfaces or 2D arrays
• We are not supporting subregions of a surface

We wrote a basic sample to demonstrate the instantiation and use of the extension in an easy to follow project. First import and link the extension framework igdex64.lib along with its dependencies: shlwapi.lib, setupapi.lib, and cfgmgr32.lib to your project. Next, add the igdext.h and IntelExtensions12.hlsl header files to the project. Once this is completed the framework can be initialized during application startup.

First load the Intel Extension Library using INTC_LoadExtensionsLibrary function. We can pass true to have it look for the igdext64.dll in the working directory, or false to look in the driver installation directory.

INTC_LoadExtensionsLibrary(true);

Next, query the supported extension framework version to determine what level of support can be expected with the currently library and hardware configuration. This function first needs to be called to get the supported version count and then called a second time to populate an INTCExtensionVersion array. The first call return value is used to allocate an array for the second call to populate with version details as can be seen in the same code.

//first, fill in the supportedExtVersionCount. This does not populate pSupportedExtVersions  
INTC_D3D12_GetSupportedVersions(m_device.Get(), nullptr, &supportedExtVersionCount);

//allocate space for pSupportedExtVersions here using supportedExtVersionCount
pSupportedExtVersions = new  INTCExtensionVersion[supportedExtVersionCount]{}; 

//Next, use returned value for supportedExtVersionCount to allocate space for the supported extensions  
INTC_D3D12_GetSupportedVersions(m_device.Get(), pSupportedExtVersions, &supportedExtVersionCount);

//Finally, create the device extension context using the current device. The m_pINTCExtensionContext object will be used to call the Extension API functions.
INTC_D3D12_CreateDeviceExtensionContext(m_device.Get(), &m_pINTCExtensionContext, &intcExtensionInfo, nullptr);

It is important to note that the examples above need error handling code, please refer to the project for initialization source code.

To perform 64-bit typed atomic operations a UAV needs created for the surface we intend to apply the atomic operations. Create a resource descriptor using the INTC_D3D12_RESOURCE_DESC_0001 type. Also, we need to define a resource with a DXGI_FORMAT_R32G32_UINT format for the UAV.

D3D12_RESOURCE_DESC texture2D = {};
texture2D.Dimension = D3D12_RESOURCE_DIMENSION_TEXTURE2D;
texture2D.Alignment = 0;
texture2D.Width = TEX_WIDTH;
texture2D.Height = TEX_HEIGHT;
texture2D.DepthOrArraySize = 1;
texture2D.MipLevels = 1;
texture2D.Format = DXGI_FORMAT_R32G32_UINT;
texture2D.SampleDesc.Count = 1;
texture2D.SampleDesc.Quality = 0;
texture2D.Layout = D3D12_TEXTURE_LAYOUT_UNKNOWN;
texture2D.Flags = D3D12_RESOURCE_FLAG_ALLOW_UNORDERED_ACCESS;

Next, use the Extension Framework API create an INTC_D3D12_RESOURCE_DESC_0001 with EmulatedTyped64bitAtomics set to TRUE.

INTC_D3D12_RESOURCE_DESC_0001 dsResDesc1;
dsResDesc1.pD3D12Desc = &texture2D;
dsResDesc1.Texture2DArrayMipPack = FALSE;
dsResDesc1.EmulatedTyped64bitAtomics = TRUE;

Use the INTC_D3D12_CreateCommittedResource API to create the resource with D3D12_HEAP_FLAGS::D3D12_HEAP_FLAG_ALLOW_SHADER_ATOMICS flag set.

INTC_D3D12_CreateCommittedResource(
	m_pINTCExtensionContext,
	&heap_props,
	D3D12_HEAP_FLAGS::D3D12_HEAP_FLAG_ALLOW_SHADER_ATOMICS,
	&dsResDesc1,
	D3D12_RESOURCE_STATE_UNORDERED_ACCESS,
	nullptr,
	IID_PPV_ARGS(&m_computeBuffer));

Using the Intel Extension requires us to populate our root signature slightly differently since the compiler leverages a UAV slot to be aware an extension will be used.

CD3DX12_DESCRIPTOR_RANGE1 descRange[2] = {};

// UAV Slot for the 64-bit Typed Atomic UAV
descRange[0].Init(D3D12_DESCRIPTOR_RANGE_TYPE_UAV, 1, 0, 0, D3D12_DESCRIPTOR_RANGE_FLAG_DATA_VOLATILE, 0);

A second UAV slot is required for extension support. This null descriptor tells the compiler an extension is going to be used. In this case we use slot u7 which is the default slot used for HLSL extensions.

// UAV Slot for Intel Extension Framework
descRange[1].Init(D3D12_DESCRIPTOR_RANGE_TYPE_UAV, 1, 7, 0, D3D12_DESCRIPTOR_RANGE_FLAG_DATA_VOLATILE, 0);

CD3DX12_ROOT_PARAMETER1 rootParameters[1] = {};
rootParameters[0].InitAsDescriptorTable(_countof(descRange), descRange, D3D12_SHADER_VISIBILITY_ALL);

Finally, to finish setup, create a UAV with a format of DXGI_FORMAT_R32G32_UINT and call CreateUnorderedAccessView to create the UAV.

// Create a UAV resource to write to.
CD3DX12_CPU_DESCRIPTOR_HANDLE uavHandle(m_computeUAVHeap->GetCPUDescriptorHandleForHeapStart());

//texture2D
D3D12_UNORDERED_ACCESS_VIEW_DESC uavDesc2 = {};
uavDesc2.Format = DXGI_FORMAT_R32G32_UINT;
uavDesc2.ViewDimension = D3D12_UAV_DIMENSION_TEXTURE2D;
uavDesc2.Texture2D.MipSlice = 0;
uavDesc2.Texture2D.PlaneSlice = 0;

m_device->CreateUnorderedAccessView(m_computeBuffer.Get(), nullptr, &uavDesc2, uavHandle);

Once the 64-bit Typed UAV is created we can pass it into a compute shader we can execute it like a normal compute shader by passing in the 64-bit Typed UAV and calling Dispatch.

ID3D12DescriptorHeap* ppHeapsCompute[] = { m_computeUAVHeap.Get() };

m_commandList->SetDescriptorHeaps(_countof(ppHeapsCompute), ppHeapsCompute);
m_commandList->SetPipelineState(m_computePipelineState.Get());
m_commandList->SetComputeRootSignature(m_computeRootSignature.Get());
m_commandList->SetComputeRootDescriptorTable(0, m_computeUAVHeap->GetGPUDescriptorHandleForHeapStart());
m_commandList->Dispatch(TEX_WIDTH / 32, TEX_HEIGHT / 32, 1);

Now that the API side of the Extension Framework steps have been completed and the program can dispatch calls to a compute shader the extension framework can be used to call atomic instructions from HLSL.

Tell the Extension Framework to use UAV register u7 to match the root parameter setup using the following macro

#define INTEL_SHADER_EXT_UAV_SLOT u7

Next import the IntelExtension12.hlsl include file

#include "IntelExtensions12.hlsl"

Now initialize u0 with a RWTexture2D for the 64-bit Typed Atomic UAV that was initialized previously on the CPU.

RWTexture2D<uint2> typedUAV64Bit : register(u0);

In the function the buffer can be used to store values but must be stored as a two part 32-bit integer vector (uint2).

// Store Low Bytes in first position and High bytes in second position.
typedUAV64Bit[DTid.xy] 	= uint2(0x00000009, 0x00000000); // 9
uint2 value2 			= uint2(0x00000001, 0x00000000); // 1

The Extension Framework does need to be initialized in the shader before using any atomic functions by calling:

IntelExt_Init();

Finally, the IntelExt_InterlockedMaxUint64 function can be called to execute the atomic operation on the typed 64-bit UAV buffer.

IntelExt_InterlockedMaxUint64(typedUAV64Bit, DTid.xy, value2);

The IntelExt_InterlockedMaxUint64 function takes the typed 64-bit buffer as the first parameter, a 2D UV coordinate as the second parameter and a third value to compare to the value at the address in the typed UAV.

After the application is finished, the Extension framework context must be destroyed and the library must be unloaded. To Destroy the context call INTC_DestroyDeviceExtensionContext with a pointer to the m_pINTCExtensionContext extension framework context.

INTC_DestroyDeviceExtensionContext(&m_pINTCExtensionContext);

Once the extension framework context has been destroyed the library can be unloaded with call to INTC_UnloadExtensionsLibrary.

INTC_UnloadExtensionsLibrary();

At this point the application can be safely shutdown.