Refactor format_info.ts and related texture functions

[greggman]

The CTS has the issue that there's a global table of texture data kAllTextureFormatInfo and it's often used to filter/select formats. Example

      .combine('colorFormat', kEncodableTextureFormats)
      // Filter out any non-renderable formats
      .filter(({ colorFormat }) => !!kTextureFormatInfo[colorFormat].colorRender)

But this is wrong. rg11b10ufloat is conditionally renderable so as defined it will never be tested here.

Similarly there is code like this

export function isMultisampledTextureFormat(
  format: GPUTextureFormat,
  isCompatibilityMode: boolean
): boolean {
  if (isCompatibilityMode) {
    if (kCompatModeUnsupportedMultisampledTextureFormats.indexOf(format) >= 0) {
      return false;
    }
  }
  return kAllTextureFormatInfo[format].multisample;
}

This is also wrong. rg11b10ufloat is marked as multisample: false but enabling rg11b10ufloat-renderable is supposed to make it support multisampling which means, every test using this filter will not be testing this format so we'd have no idea if it actually works.

It feels like this should be solved by making kAllTextureFormatInfo private and refactoring any of these queries about formats to require enough info to be useful, and/or, be renamed to reflect what they're returning. So for example, something that wants to know all renderable formats can not use

export function canUseAsRenderTarget(format: GPUTextureFormat) {
  return !!kTextureFormatInfo[format].colorRender || isDepthOrStencilTextureFormat(format);
}

   .combine('format', kAllTextureFormats)
   .filter(t => canUseAsRenderTarget(t.format))

because, you can't know if a format is renderable until you know if you have the feature. That particular function could be renamed to canPotentiallyUseAsRenderTarget in which case it should return true for rg11b10ufloat. Similarly isMultisampledTextureFormat would need to be isPotentiallyMultisampledTextureFormat.

But ultimately, with all of these special exceptions, it doesn't feel like tests should be rolling their own checks by accessing kAllTextureFormatInfo directly. They should all be gated on the data needed to give the correct answer, forcing you to get the larger list and then use skipIf... and/or selectDeviceForXOrSkipTest

I think I'd also suggest that most texture format filtering in general should be moved to some helpers that tests all share the same filtering where appropriate.

@shrekshao, @kainino0x

[greggman]

• isFilterableAsTextureF32 -> isPossiblyFilterableAsTextureF32
• canUseAsRenderTarget -> canPossiblyUseAsRenderTarget
• check filterFormatsByFeature
• check getFeaturesForFormats
• make private
  • kTextureFormatInfo
  • kRegularTextureFormatInfo
  • kSizedDepthStencilFormatInfo
  • kUnsizedDepthStencilFormatInfo
  • kBCTextureFormatInfo
  • kETC2TextureFormatInfo
  • kASTCTextureFormatInfo

[greggman]

• fix every usage of skipIfTextureFormatNotUsableAsStorageTexture

  This assumes the list of textures usable as storage textures is fixed, gated
  only on compat/non-compat. that seems unlikely to stay true. For one, it doesn't
  handle bgra8unorm-storage

[kainino0x]

I was planning to split the format info into two separate tables, one for static info (like block sizes) and one for dynamic info (capabilities) that could depend on the device, but I never got around to that after doing the first step of the refactor. I think #3848 was the issue tracking that.