Implement texelData for rgb9e5ufloat and add texelData unit tests #352
Conversation
|
Running the tests in https://chromium-review.googlesource.com/c/chromium/src/+/2536758 Going to need to have some way to have a tolerance that varies based on ULPs, the expected value, and the component index (since some formats have different precision for each component). |
This test uploads a single texel to encodable texture formats and then reads it back in a shader with textureLoad to check that the values are as expected.
|
|
||
| const float10 = { | ||
| decode: identity, |
There was a problem hiding this comment.
Ideally, this wouldn't actually be the identity function but would cut the precision of the input value down so that it could be represented by a float10
I'm not sure if that will actually help the precision issues though or simply need a better way to compute tolerance.
| .map((C, i) => { | ||
| return `[[offset(${i * 4})]] result${C} : ${shaderType};`; | ||
| }) | ||
| .join('\n')} |
There was a problem hiding this comment.
Could just return a vec4<shaderType>. Then you can just output.result = textureLoad(...) I think. In the result you can just leave the unused components untested so we're not testing that behavior here (should be a separate test).
There was a problem hiding this comment.
Maybe, but then I can't just use t.expectContents of the whole thing?
There was a problem hiding this comment.
True. Couldn't you still make this a simple vec4 and write into it selectively in main()? (Or can you not put swizzles on the left of = in wgsl?) Anyway, either is okay.
| } & TextureFormatInfo; | ||
| } = /* prettier-ignore */ { | ||
| 'depth32float': { renderable: true, color: false, depth: true, stencil: false, storage: false, copySrc: true, copyDst: false, bytesPerBlock: 4, blockWidth: 1, blockHeight: 1 }, | ||
| 'depth32float': { renderable: true, color: false, depth: true, stencil: false, storage: false, copySrc: true, copyDst: false, bytesPerBlock: 4, blockWidth: 1, blockHeight: 1, dataType: 'float', componentType: 'float' }, |
There was a problem hiding this comment.
Mmmh can't this also be "depth-comparison"? How are we going to handle formats that support multiple components types? (or is this just for sampling in a texture to read the value and nothing else?)
There was a problem hiding this comment.
I'm probably going to need to make this an array or we'll have to add per-aspect information at some point.. but for now it's just for the sampling part.
There was a problem hiding this comment.
Given I think we're changing 'depth-comparison' to 'depth-stencil' upstream, maybe we don't even need to support 'float' component type for depth formats?
| } & TextureFormatInfo; | ||
| } = /* prettier-ignore */ { | ||
| 'depth32float': { renderable: true, color: false, depth: true, stencil: false, storage: false, copySrc: true, copyDst: false, bytesPerBlock: 4, blockWidth: 1, blockHeight: 1 }, | ||
| 'depth32float': { renderable: true, color: false, depth: true, stencil: false, storage: false, copySrc: true, copyDst: false, bytesPerBlock: 4, blockWidth: 1, blockHeight: 1, dataType: 'float', componentType: 'float' }, |
There was a problem hiding this comment.
I'm probably going to need to make this an array or we'll have to add per-aspect information at some point.. but for now it's just for the sampling part.
| .map((C, i) => { | ||
| return `[[offset(${i * 4})]] result${C} : ${shaderType};`; | ||
| }) | ||
| .join('\n')} |
There was a problem hiding this comment.
True. Couldn't you still make this a simple vec4 and write into it selectively in main()? (Or can you not put swizzles on the left of = in wgsl?) Anyway, either is okay.
dataTypeandcomponentTypeto the texture format table for encodable texture formats.dataTypeis the format's data representation whereascomponentTypeis theGPUTextureComponentType- the type you get in the shader.rgb9e5ufloatdataType, loading thecomponentTypein the shader, and writing it out.