D3D12 Counters & Queries

UAV Counters, Stream-Output Counters, Queries.


This document describes the Direct3D12 UAV counters, stream-output counters, and queries.

Detailed Design

Stream Output Counters

The application is responsible for allocating storage for a 32-bit quantity called the BufferFilledSize. This contains the number of bytes of data in the stream-output buffer. This storage must be placed in the same resource as the one that contains the stream-output data. This value is accessed by the GPU in the stream-output stage to determine where to append new vertex data in the buffer. Additionally, this value is accessed by the GPU to determine when overflow has occurred.

typedef struct D3D12_STREAM_OUTPUT_VIEW_DESC
    UINT64 OffsetInBytes;
    UINT64 SizeInBytes;
    UINT64 BufferFilledSizeOffsetInBytes;

The runtime will validate the following in ID3D12CommandList::SetStreamOutputBuffersSingleUse and ID3D12Device::CreateStreamOutputView:

The runtime will not validate the heap type associated with the stream output buffer. Stream output is supported in upload, default, and readback heaps.

Root signatures must specify if stream output will be used. This enables drivers to reserve binding space for stream output buffers and counters.


D3D12_ROOT_SIGNATURE_ALLOW_STREAM_OUTPUT can be specified for root signatures authored in HLSL, in a manner similar to how the other flags are specified.

CreateGraphicsPipelineState will fail if the geometry shader contains stream-output but the root signature does not have the D3D12_ROOT_SIGNATURE_ALLOW_STREAM_OUTPUT flag set.

When a resource is used as a stream-output target, the resource must be in the D3D12_RESOURCE_USAGE_STREAM_OUT state. This naturally applies to both the vertex data and the BufferFilledSize, because both come from the same resource.

The ID3D12CommandList::SetStreamOutputBufferOffset API is removed because applications can write to the BufferFilledSize with the GPU directly.

ID3D12CommandList::DrawAuto is removed. This can be emulated via DrawInstancedIndirect.

UAV Counters

The application is responsible for allocating 32-bits of storage for UAV counters. This storage can be allocated in a different resource as the one that contains data accessible via the UAV.

void ID3D12Device::CreateUnorderedAccessView(
    ID3D12Resource* pResource,
    ID3D12Resource* pCounterResource,
    D3D12_CPU_DESCRIPTOR_HANDLE DestDescriptor

typedef enum D3D12_BUFFER_UAV_FLAG
    D3D12_BUFFER_UAV_FLAG_RAW = 0x00000001,

typedef struct D3D12_BUFFER_UAV
    UINT64 FirstElement;
    UINT NumElements;
    UINT StructureByteStride;
    UINT64 CounterOffsetInBytes;
    UINT Flags;

Note that ID3D12CommandList::SetGraphicsRootUnorderedAccessViewSingleUse and ID3D12CommandList::SetComputeRootUnorderedAccessViewSingleUse do not support UAV counters.

If pCounterResource is specified then there is a counter associated with the UAV. In this case:

If pCounterResource is not specified, then CounterOffsetInBytes must be 0.

If the RAW flag is set then

if pCounterResource is not set, then CounterOffsetInBytes must be 0

If the RAW flag is not set and StructureByteStride = 0, then the format must be a valid UAV format.

D3D12 removes the distinction between append and counter UAVs (although the distinction still exists in HLSL bytecode).

The core runtime will validate these restrictions inside of:

SetComputeRootUnorderedAccessViewSingleUse, SetGraphicsRootUnorderedAccessViewSingleUse, and CreateUnorderedAccessView.

During Draw/Dispatch, the counter resource must be in the D3D12_RESOURCE_USAGE_UNORDERED_ACCESS state. The debug layer will issue errors when this is not the case.

The ID3D12CommandList::SetUnorderedAccessViewCounterValue and ID3D12CommandList ::CopyStructureCount APIs are removed because applications can simply copy data to/from the counter value directly.

Dynamic indexing of UAVs with counters is supported.

If a shader attempts to access the counter of a UAV that does not have an associated counter, then the debug layer will issue a warning, and a GPU page fault will occur, causing the application’s device to be removed.

Counter UAVS are supported in all heap types (default, upload, readback).

Within a single Draw/Dispatch call, it is invalid for an application to access the same 32-bit memory location via 2 separate UAV counters. The debug layer will issue an error when this is detected.


In D3D 12, queries are grouped into arrays of queries called a query heap. A query heap has a type which defines the valid types of queries that can be used with that heap.

typedef enum D3D12_QUERY_TYPE_HEAP_TYPE

typedef struct D3D12_QUERY_TYPE_HEAP_DESC
    UINT Count;

HRESULT ID3D12Device::CreateQueryHeap(
    const D3D12_QUERY_TYPE_DESC* Desc,
    ID3D12HeapQuery** QueryHeap

Event queries are not present in D3D12; this functionally has been subsumed by fences.

TIMESTAMP_DISJOINT queries are not present in D3D12. The GPU timestamp clock is assumed to be stable such that 2 timestamp queries issued in the same command list are comparable.

QUERY_SO_STATISTICS queries are not present in D3D12. Applications can emulate this behavior by issuing multiple single-stream queries, and then accumulating the results.

SO_STATISTICS_PREDICATE and OCCLUSION_PREDICATE queries are not present in D3D12. They can be emulated by applications.

A new query type is added to the API. D3D12_QUERY_TYPE_BINARY_OCCLUSION acts like D3D12_QUERY_TYPE_OCCLUSION except that it returns a binary 0/1 result. 0 indicates that no samples passed depth and stencil testing. 1 indicates that at least 1 sample passed depth and stencil testing. This is added to the API to enable occlusion queries to not interfere with any GPU performance optimization associated with depth/stencil testing. Hardware that does not support this query type natively can emulate it via special processing in the ResolveQueryData API.

The core runtime will validate that the heap type is a valid member of the heap_type enumeration, and that the count is greater than 0.

Each individual element within a query heap can be start/stopped separately.

typedef enum D3D12_QUERY_TYPE

void ID3D12CommandList::BeginQuery(
    ID3D12QueryHeap* Query,
    UINT ElementIndex,
    D3D12_QUERY_TYPE Type

void ID3D12CommandList::EndQuery(
    ID3D12QueryHeap* Query,
    UINT ElementIndex,
    D3D12_QUERY_TYPE Type

D3D12_QUERY_TYPE_TIMESTAMP is the only query that that supports EndQuery only. All other query types require BeginQuery and EndQuery.

The debug layer will validate:

The core runtime will validate the following:

Timestamp Frequency

Applications can query the GPU timestamp clock frequency on a per-command queue basis.

HRESULT ID3D12CommandQueue::GetTimestampFrequency(UINT64* pFrequency)

The returned frequency is measured in Hz (ticks/sec). This API fails (E_FAIL) if the specified command queue does not support timestamps (see the table in the previous section).

Clock Calibration

D3D12 enables applications to correlate results obtained from timestamp queries with results obtained from calling QueryPerformanceCounter. This is enabled by 2 API additions:

HRESULT ID3D12CommandQueue::GetClockCalibration(
  UINT64* pGpuClock,
  UINT64* pCpuClock

GetClockCalibration samples the GPU clock for a given command queue and samples the CPU clock via QueryPerformanceCounter at nearly the same time.

Note that this is implemented by asking the UMD to translate from command queue to DXGKRNL context and then calling the (pre-existing) kernel mode driver CalibrateGpuClock API.

This API fails (E_FAIL) if the specified command queue does not support timestamps (see the table in the previous section).

Both GetTimestampFrequency and GetClockCalibration are implemented without the involvement of the user-mode driver. D3D12 uses the first context that the user-mode driver created on the given queue to determine which GPU and engine to query. D3D12 then calls DXGKRNL, which calls the kernel-mode driver to determine the timestamp frequency and CPU/GPU calibration.

In order for the clock calibration to be useful the application must be confident that the GPU timestamp clock will not stop ticking during idle periods. This is enabled by a new API.

HRESULT ID3D12Device::SetStablePowerState(BOOL Enable)

This API is intended for development time use only. Therefore it is only allowed when the D3D12 SDK layers are present on the machine. The API fails with E_FAIL if the D3D12 SDK layers are not present.

The debug layer will issue a warning if the GetClockCalibration API is used without SetStablePowerState being called first.

This API is implemented with new kernel-mode DDIs which are described separately.


The only way to extract data from a query is to resolve the query data from a proprietary format into the API-standard format.

void ID3D12CommandList::ResolveQueryData(
  ID3D12QueryHeap* QueryHeap,
  D3D12_QUERY_TYPE Type,
  UINT StartElement,
  UINT ElementCount,
  ID3D12Resource* DestinationBuffer,
  UINT64 AlignedDestinationBufferOffset

ResolveQueryData performs a batched operation which writes query data into a destination buffer. Query data is written contiguously to the destination buffer. AlignedDestinationBufferOffset must be a multiple of 8 bytes. The destination buffer must be in the D3D12_RESOURCE_USAGE_COPY_DEST state. The size/format of the output data matches the D3D11 API definitions. Binary occlusion queries write 64-bits per query. The least significant bit is either 0 or 1. The rest of the bits are 0.

The core runtime will validate:

The debug layer will issue a warning if the destination buffer is not in the D3D12_RESOURCE_USAGE_COPY_DEST state.

ResolveQueryData works with all heap types (default, upload, readback).

Predication is decoupled from queries. Predication can be set based on the value of 64-bits within a buffer.

typedef enum D3D12_PREDICATION_OP
    D3D12_PREDICATION_OP_EQUAL_ZERO, // Enable predication if all 64-bits are zero
    D3D12_PREDICATION_OP_NOT_EQUAL_ZERO, // Enable predication if at least one of the 64-bits are not zero

void ID3D12CommandList::SetPredication(
  ID3D12Resource* Buffer,
  UINT64 AlignedBufferOffset,
  D3D12_PREDICATION_OP Operation

When the GPU executes a SetPredication command it snaps the value in the buffer. Future changes to the data in the buffer do not retroactively affect the predication state.

If Buffer is NULL, then predication is disabled

Predication hints are not present in the D3D12 API.

Predication is allowed on direct and compute command lists.

The core runtime will validate:

The debug layer will issue an error if the source buffer is not in the D3D12_RESOURCE_USAGE_DEFAULT_READ state.

The source buffer can be in any heap type (default, upload, readback).

The set of operations which can be predicated are:

ID3D12CommandList::ExecuteBundle is not predicated itself. Instead, individual operations from the list above which are contained in side of the bundle are predicated.

ID3D12CommandList::{ResolveQueryData,BeginQuery,EndQuery} are not predicated.

Test Plan

Runtime Functional Tests

Driver Conformance Tests