check for undefined constants instead of zero values

[rajveermalviya]

Align better with other webgpu-headers implementations for unset/undefined values.
Currently we take zero for indicating unset/undefined values. But webgpu-headers defines constants which should be used instead.

This introduces some breaking changes

• WGPULimits struct will have to be declared like this:

- (WGPULimits){
-     .maxBindGroups = 1,
- }
+ (WGPULimits){
+     .maxBindGroups = 1,
+     .maxTextureDimension1D = WGPU_LIMIT_U32_UNDEFINED,
+     .maxTextureDimension2D = WGPU_LIMIT_U32_UNDEFINED,
+     .maxTextureDimension3D = WGPU_LIMIT_U32_UNDEFINED,
+     .maxTextureArrayLayers = WGPU_LIMIT_U32_UNDEFINED,
+     .maxDynamicUniformBuffersPerPipelineLayout = WGPU_LIMIT_U32_UNDEFINED,
+     .maxDynamicStorageBuffersPerPipelineLayout = WGPU_LIMIT_U32_UNDEFINED,
+     .maxSampledTexturesPerShaderStage = WGPU_LIMIT_U32_UNDEFINED,
+     .maxSamplersPerShaderStage = WGPU_LIMIT_U32_UNDEFINED,
+     .maxStorageBuffersPerShaderStage = WGPU_LIMIT_U32_UNDEFINED,
+     .maxStorageTexturesPerShaderStage = WGPU_LIMIT_U32_UNDEFINED,
+     .maxUniformBuffersPerShaderStage = WGPU_LIMIT_U32_UNDEFINED,
+     .maxUniformBufferBindingSize = WGPU_LIMIT_U64_UNDEFINED,
+     .maxStorageBufferBindingSize = WGPU_LIMIT_U64_UNDEFINED,
+     .minUniformBufferOffsetAlignment = WGPU_LIMIT_U32_UNDEFINED,
+     .minStorageBufferOffsetAlignment = WGPU_LIMIT_U32_UNDEFINED,
+     .maxVertexBuffers = WGPU_LIMIT_U32_UNDEFINED,
+     .maxVertexAttributes = WGPU_LIMIT_U32_UNDEFINED,
+     .maxVertexBufferArrayStride = WGPU_LIMIT_U32_UNDEFINED,
+     .maxInterStageShaderComponents = WGPU_LIMIT_U32_UNDEFINED,
+     .maxComputeWorkgroupStorageSize = WGPU_LIMIT_U32_UNDEFINED,
+     .maxComputeInvocationsPerWorkgroup = WGPU_LIMIT_U32_UNDEFINED,
+     .maxComputeWorkgroupSizeX = WGPU_LIMIT_U32_UNDEFINED,
+     .maxComputeWorkgroupSizeY = WGPU_LIMIT_U32_UNDEFINED,
+     .maxComputeWorkgroupSizeZ = WGPU_LIMIT_U32_UNDEFINED,
+     .maxComputeWorkgroupsPerDimension = WGPU_LIMIT_U32_UNDEFINED,
+ }

• use WGPU_ARRAY_LAYER_COUNT_UNDEFINED instead of 0 for WGPUTextureViewDescriptor.arrayLayerCount.
• use WGPU_COPY_STRIDE_UNDEFINED instead of 0 for WGPUTextureDataLayout.bytesPerRow & WGPUTextureDataLayout.rowsPerImage.
• use WGPU_MIP_LEVEL_COUNT_UNDEFINED instead of 0 for WGPUTextureViewDescriptor.mipLevelCount.
• use WGPU_WHOLE_MAP_SIZE for size param in wgpuBufferGetMappedRange & wgpuBufferGetConstMappedRange.
• use WGPU_WHOLE_SIZE instead of 0 for:
  • size param in wgpuCommandEncoderClearBuffer.
  • size param in wgpuRenderPassEncoderSetIndexBuffer.
  • size param in wgpuRenderPassEncoderSetVertexBuffer.
  • size param in wgpuRenderBundleEncoderSetIndexBuffer.
  • size param in wgpuRenderBundleEncoderSetVertexBuffer.
  • WGPUBufferBindingLayout.minBindingSize.
  • WGPUBindGroupEntry.size.

Fixes #214

[almarklein]

I understand that this change was done because other implementations seem to do things this way, but to be honest this approach feels rather fragile to me.

As an example, In #235 I update the webgpu.h, which introduces a new field in the Limits: the maxBufferSize. It took me about 2 hours today to figure out that I also should update the limit defaults. The nasty thing about this is that (from how I understand this) every downstream application will have to update their default limits. I foresee many devs spending hours in frustration ...

And then there's the fact that there is WGPU_LIMIT_U32_UNDEFINED and WGPU_LIMIT_U64_UNDEFINED. If the type of a field is changed upstream (from u32 to u64, or back) this will cause a lot of pain downstream.

I'm not sure what the most elegant solution would be, but I imagine that it would help a lot if we had a function to create a struct populated with these undefined-placeholders. Then here in wgpu-native we could make sure to keep that up-to-date, and save pain downstream.

[rajveermalviya]

There's also some discussion here about having the default struct macros to be included in webgpu.h. It would probably be best if that happens.

[almarklein]

To work around this issue in wgpu-py I create the adapter, request its limits, and then use these as the default limits for creating devices.