anki-3d-engine/ThirdParty/Warp

Direct3D WARP NuGet Package

This package contains a copy of D3D10Warp.dll with the latest features and fixes for testing and development purposes.

The included LICENSE.txt applies to all files under build/native/bin/.

Changelog

Version 1.0.14.1

Fix x86 FMA fusing to correctly fuse a*b-c into FMSUB instead of FMADD.

Version 1.0.14

Codegen improvements

x86/x64:

WARP will now detect and use the following x86 architecture extensions: SSE4.2, AVX, AVX2, FMA, F16C, BMI1, BMI2, and AVX512. Specifically:

• When AVX is supported, VEX encoding is used for applicable instructions to increase code density.
• When AVX2 and FMA are supported, WARP will use 256-bit registers for 64-bit DXIL data types instead of two 128-bit registers.
• When FMA is supported and precise ops are not requested (or when double-precision FMA instructions are used), WARP will use fused multiply-adds.
• When AVX512 is supported, WARP will use mask registers for boolean data types, as well as using the additional 16 vector registers for x64.
• Most instructions were audited for ways to be better-encoded using new instructions available in these instruction sets. Of note, AVX512 fills many gaps in comparisons and conversions.
• Large improvements were made to the codegen optimizer to simplify expressions involving constants among other optimizations.

arm64:

• Many operations that did not have an equivalent in SSE4.1 and below, but do in SSE4.1 and above, are updated to use more direct translations to the arm64 instruction set.
• Atomics are upgraded to use arm64 LSE.
• Note: FMA fusing is not yet applied broadly, but DXBC/DXIL double-precision FMA instructions now use fused instructions.

All:

• 64-bit atomics are now emitted directly into shaders instead of requiring function calls.

Other

• Raytracing support has been partially rewritten:
  • Acceleration structure traversal is JITed, benefitting from SIMD and all of the above codegen improvements.
  • Shader execution state machine is also JITed, reducing overhead and better aligning DXR1.0 with DXR1.1.
  • Recursive shader invocation overhead is significantly reduced.
• Fixed use of SM6.6 descriptor heap indexing in hit/miss/intersection shaders.
• Fixed PIX timing capture history buffer layout.
• Support for Visual Studio Graphics Analyzer's shader debugger has been removed.
• Handle sampling from an integer texture with a linear sampler by using point sampling instead of crashing.
• Fix calls to GetDimensions() on non-uniform resource handles.
• Fix a race condition in acceleration structure builds.
• Wave ops reading from inactive lanes now return 0 (to better match real hardware behavior).

Version 1.0.13

New features:

• Variable Rate Shading (VRS) Tier 2 support
• Wide wave size emulation support
• Redesigned barriers, more likely to catch app bugs and with higher throughput
• Max virtual address space per resource increased to 40 bits, so resources/heaps can be larger than 4GiB
  • Note: There are some known issues here, but main scenarios should work. Please report bugs to the DirectX Discord (https://discord.com/invite/directx).
• WARP now claims feature level 12_2, aka DirectX 12 Ultimate!
• D3D12 video encoding/decoding support using Media Foundation, only on platforms where Media Foundation codec support is present, for H.264/H.265.

Bug fixes:

• Fix incompatibility between new WARP and old D3D12SDKLayers
• Fixes for work graphs where record reads are done under non-uniform control flow
• Fix a rare race condition which can deadlock work graph backing store initialization
• Fix the default blend state for a generic graphics program in state objects to handle multiple render targets
• Fix ICBs for work graphs (i.e. DXIL global constant data)

Version 1.0.13-preview

• Work graphs tier 1.1: Mesh nodes

Version 1.0.12

New features:

• Now including arm64 binaries

Bug fixes:

• Fix mesh shaders on Win11 retail D3D12
• Fix accessing static const arrays in generic programs
• Fix arm64 instruction encoding which could cause random crashes in shader execution
• Fix state objects exporting the same non-library shader with multiple names/aliases
• Fix handling of zeroing used by lifetime markers on older shader models
• Fix handling of lifetime markers in node shaders
• Handle a null global root signature in a work graph
• Fix mesh -> PS linkage for primitive outputs that were not marked nointerpolate in the PS input
• Fix back-to-back DispatchGraph without a barrier to correctly share the backing memory
• Fix a race condition in BVH builds which could cause corrupted acceleration structures

Version 1.0.11

New features:

• Shader model 6.8 support
• Work graphs support
• ExecuteIndirect tier 1.1 support

Version 1.0.10.1

• Fix to support empty ray tracing acceleration structures.
• Miscellaneous bug fixes.

Version 1.0.9

New features:

• Sampler feedback support

Bug fixes:

• Improve handling of GetDimensions shader instructions in control flow or with dynamic resource handles
• Fix JIT optimizations that can produce wrong results
• Fix assignment of variable inside of a loop which was uninitialized before entering
• Fix ExecuteIndirect to not modify the app-provided arg buffer
• Fix root signature visibility for amplification and mesh shaders
• Handle DXIL initializers for non-constant global variables
• Fix uninitialized memory causing potential issues with amplification shaders
• Fix deadlock from using the same shader across multiple queues with out-of-order waits
• Some fixes for handling of tiled resources

Version 1.0.8

• DirectX Raytracing (DXR) 1.1 support.

Version 1.0.7

• Miscellaneous fixes for 16- and 64-bit shader operations.

Version 1.0.6

• Miscellaneous fixes and improvements to non-uniform control flow in DXIL shader compilation.
• Fixed casting between block-compressed and uncompressed texture formats.
• Improved conformance of exp2 implementation.

Version 1.0.5

• Improved support for non-uniform constant buffer reads.
• Miscellaneous fp16 conformance fixes.
• Inverted height support.

Version 1.0.4

• Miscellaneous sampling conformance fixes.

Version 1.0.3

New features:

• Mesh shader support.

Bug fixes:

• Fix a synchronization issue related to null/null aliasing barrier sequences.
• Fix a compatibility issue with older D3D12 Agility SDK versions lacking enhanced barrier support.

Version 1.0.2

• Expanded support for 64-bit conversion operations.
• Fixes issue where static samplers could incorrectly use non-normalized coordinates.

Version 1.0.1

New features:

• Support for Shader Model 6.3-6.7 required features, including:

  • 64-bit loads, stores, and atomics on all available types of resources.
  • Descriptor heap indexing, aka Shader Model 6.6 dynamic resources.
  • Quad ops and sampling derivatives in compute shaders.
  • Advanced texture ops, featuring integer texture point sampling, raw gather, dynamic texel offsets, and MSAA UAVs.

• Support for many new Direct3D 12 Agility SDK features, including:

  • Enhanced barriers.
  • Relaxed resource format casting.
  • Independent front/back stencil refs and masks.
  • Triangle fans.
  • Relaxed buffer/texture copy pitches.
  • Independent D3D12 devices via ID3D12DeviceFactory::CreateDevice

Bug fixes:

• Various conformance and stability fixes.