[hzqst]

Problem

A pipeline created with a non-default SRBAllocationGranularity or ImmediateContextMask was never reused from a loaded render state cache. Every lookup logged:

Description of pipeline state '...' does not match the description of the pipeline unpacked from the cache. This may be the result of a hash conflict, though the probability of this should be virtually zero.

The cache then created the pipeline again, so a stored cache did not help these pipelines. Any pipeline that may run on several immediate contexts is affected. There are two causes:

1. The device object archive does not store PipelineStateDesc::SRBAllocationGranularity and ImmediateContextMask (PSOSerializer.cpp: // skip SRBAllocationGranularity, // skip ImmediateContextMask). The dearchiver takes them from PipelineStateUnpackInfo instead. RenderStateCacheImpl::CreatePipelineStateInternal did not fill these two fields, so they stayed at their defaults.
2. PipelineStateBase clears the bits of ImmediateContextMask that have no command queue on the device (m_Desc.ImmediateContextMask &= DeviceQueuesMask). Even with the right unpack info, the unpacked pipeline can hold a subset of the requested mask, and it was compared with the create info as is.

Fix

• Copy SRBAllocationGranularity and ImmediateContextMask from the create info into UnpackInfo.
• Before comparing descriptions, reduce the requested mask to the bits the unpacked pipeline kept. All other fields are still compared exactly.

Test

New test RenderStateCacheTest.ArchivedPSOKeepsDescription:

1. Creates a compute pipeline with SRBAllocationGranularity = 4 and ImmediateContextMask = ~0 (not the default on any device, and always reduced by the device).
2. Writes the cache and loads it into a new cache.
3. Expects the pipeline to be found there, with the requested description and the mask reduced to the device's immediate contexts.

Results (NVIDIA GeForce RTX 5060, Windows, x64 Debug):

• With neither commit, the test fails on Vulkan, D3D12, D3D11 and GL. With only the first commit, it still fails on all four.
• With both commits, all RenderStateCacheTest.* tests pass or skip as before on all four.