Severe performance regression on Win7 NVIDIA

[y-novikov]

Describe the Issue
When updating VVL in Chrome we've discovered that ANGLE tests with VVL started to time out on Win7 NVIDIA machines, see crbug.com/1249209 for context. Overall, the whole test suite runs 2 times slower, but looking only at the tests which use Vulkan, those run 10 times slower.
Bisect points to #3268, specifically to f596d7f.
Further drilling down, the slowness is due to added DispatchGetPhysicalDeviceSurfaceFormatsKHR() calls in CoreChecks::ValidateCreateSwapchain(), which were previously cached in physical_device_state->surface_formats.

Valid Usage ID
N/A

Environment:

• OS: Windows-2008ServerR2-SP1
• GPU: NVIDIA Quadro P400, driver 24.21.14.1195
• SDK or header version if building from repo: 1.2.191
• Options enabled (synchronization, best practices, etc.): all

Additional context
@jeremyg-lunarg

[jeremyg-lunarg]

Thanks for the detailed report! I have a possible fix in #3452 But our internal CI doesn't have a configuration that showed the regression, so I can't easily tell if this fixes it. @y-novikov is it possible for you to confirm if this fixes it on your side?

[jeremyg-lunarg]

@y-novikov I've updated #3452 again, hopefully it fixes the regression.

[y-novikov]

#3452 rolled into Chromium in http://crrev.com/934959.
I can confirm that in Win7 NVIDIA builds which include this change the regression was fixed:
https://luci-milo.appspot.com/ui/p/chromium/builders/ci/Win7%20FYI%20x64%20Release%20(NVIDIA)/13858/overview
https://luci-milo.appspot.com/ui/p/chromium/builders/ci/Win7%20FYI%20Release%20(NVIDIA)/13733/overview
angle_end2end_tests shard times are down to ~550 seconds from ~1000.

Thanks a lot for fixing this!