AMDVLK 2022.Q1.2 Released With Vulkan 1.3 Support

[phoronix]

Phoronix: AMDVLK 2022.Q1.2 Released With Vulkan 1.3 Support

AMDVLK as AMD's official open-source Vulkan Linux driver derived from their Radeon Software driver sources but using the LLVM shader compiler back-end is out with a new release. AMD is ready with day-after support for the newly-launched Vulkan 1.3 specification for AMDVLK...

https://www.phoronix.com/scan.php?page=news_item&px=AMDVLK-2022.Q1.2

[FireBurn]

What's new with SPIR-V 1.6?

[coder]

What's new with SPIR-V 1.6?

I think you're meant to start reading here:

https://www.khronos.org/registry/SPI...1_5_revision_1

But the major changes are probably the ones in the last section:

• New Functionality:
  • Khronos SPIR-V issue #515: The FPFastMathMode decoration may now be used with OpFNegate, with the binary floating-point comparison instructions (including OpOrdered and OpUnordered), and with OpExtInst where expressly permitted by the extended instruction set.
  • #661: Added a Nontemporal Image Operand.
• Extensions Incorporated into SPIR-V 1.6:
  • SPV_KHR_non_semantic_info, see OpExtInstImport.
  • SPV_KHR_integer_dot_product
  • SPV_KHR_terminate_invocation
  • SPV_EXT_demote_to_helper_invocation, with changes: Only OpDemoteToHelperInvocationEXT was incorporated. Instead of using OpIsHelperInvocationEXT, modules should use Volatile loads of the HelperInvocation built-in variable.
• Deprecations and Removals, from Khronos SPIR-V issues:
  • Removed OpLessOrGreater. Use OpFOrdNotEqual instead.
  • #620: The WorkgroupSize built-in is deprecated starting with version 1.6.
  • #645: The True Label and False Label of an OpBranchConditional must not be the same, starting with version 1.6.
  • #584: Disallow Dim Buffer in OpTypeSampledImage and OpSampledImage starting with version 1.6.
  • Deprecated OpKill, in favor of OpTerminateInvocation, or OpDemoteToHelperInvocation.
• Reserve enumerants for the SPV_KHR_fragment_shader_barycentric extension.

I wonder if the Nontemporal image mode has anything to do with the use of Infinity Cache, in the RX 6500XT. What's interesting about it is that it has substantially lower bandwidth than the GDDR6 interface, which means there must be steering of some memory operations to bypass it, otherwise it'd be a bottleneck.

[Rob72]

What is the deal with the VK_EXT_separate_stencil_usage extension for Vulkan 1.2 and the AMD radv driver?

It was apparently developed by AMD, yet according to mesamatrix.net, the AMD radv driver does not implement it, meaning the AMD radv support is stuck at Vulkan 1.1.