Add compute shader derivatives sample for Vulkan extension #1421
Conversation
Introduce a sample demonstrating the `VK_KHR_compute_shader_derivatives` extension. Includes new shader, CMake, and Vulkan application files showcasing the use of derivative instructions in compute shaders with quad-based derivative groups.
| float v = data[idx * 4 + 0]; | ||
| float ddx = data[idx * 4 + 1]; | ||
| float ddy = data[idx * 4 + 2]; | ||
| LOGI("compute-derivatives CPU: tid=({}, {}) v={} ddx={} ddy={}", x, y, v, ddx, ddy); |
There was a problem hiding this comment.
Would be better to display them in the app's GUI instead of simply outputting them to the terminal. Depending on how/where you run the sample you don't have direct access to the terminal.
- Store and format compute results for GUI display - Update README with GitHub repository link for sample
|
Again, not sure what I'm supposed to see here, but I'm getting the same weird random frames on the display as I saw with relaxed_extended_instruction and some console output: |
| add_instance_extension(VK_KHR_GET_PHYSICAL_DEVICE_PROPERTIES_2_EXTENSION_NAME); | ||
| // Device extension providing the feature | ||
| add_device_extension(VK_KHR_COMPUTE_SHADER_DERIVATIVES_EXTENSION_NAME); | ||
| // Toolchains may still emit SPV_NV_compute_shader_derivatives; enable NV extension if available to satisfy validation |
There was a problem hiding this comment.
Can we double-check this? Both Slang and DXC should support khr_compute_shader_derivatives.
It’s probably better not to request the NV extension, since if the shader is using SPV_NV_compute_shader_derivatives that would cause issues if a device supports the KHR extension but not the NV version.
We could add a comment suggesting users update their toolchain if they encounter this problem.
| void ComputeShaderDerivatives::request_gpu_features(vkb::core::PhysicalDeviceC &gpu) | ||
| { | ||
| // Require quads derivative group (the sample shader uses layout(derivative_group_quadsNV/derivative_group_quads_khr)) | ||
| REQUEST_REQUIRED_FEATURE(gpu, VkPhysicalDeviceComputeShaderDerivativesFeaturesKHR, computeDerivativeGroupQuads); |
There was a problem hiding this comment.
Some implementations only support computeDerivativeGroupLinear. As a suggestion, could we add an example using computeDerivativeGroupLinear?
The mapping from a thread to 2D coordinates isn’t immediately obvious when using 1D quads or computeDerivativeGroupLinear, so it would be useful to have an example
https://microsoft.github.io/DirectX-Specs/d3d/HLSL_SM_6_6_Derivatives.html
Ideally, this sample could have two shader versions—derivatives_linear.comp and derivatives_quad.comp, both producing identical output with a toggle to alternate between both versions.
Also, as mentioned earlier, I think the sample would be clearer and more useful if we added a SampleGrad to use the derivatives.
4 participants
Description
Introduce a sample demonstrating the
VK_KHR_compute_shader_derivativesextension. Includes new shader, CMake, and Vulkan application files showcasing the use of derivative instructions in compute shaders with quad-based derivative groups.General Checklist:
Please ensure the following points are checked:
Note: The Samples CI runs a number of checks including:
If this PR contains framework changes:
batchcommand line argument to make sure all samples still work properlySample Checklist
If your PR contains a new or modified sample, these further checks must be carried out in addition to the General Checklist: