Mali G57 vkCreateGraphicsPipelines crashes with Tessellation shaders

Hi,

Mali Best Practice is not to have Tessellation shaders (they give bad performance), but in this case there's also likely a bug you're hitting, in the HLSL->DXC->SPIR-V process.  DXC-compiled tessellation shaders don't work on Mali (since R26), they create a problematic pattern.

A proposed workaround exists: 

-> round-trip the SPIR-V through GLSL (using SPIRV-Cross) and back to SPIR-V again. 

There are DXC patterns that SPIRV-Cross can't cope with either, so depends whether you hit those (it didn't work for another developer who hit this issue). In that case it might just need the SPIR-V to be patched.

Hi,

Mali Best Practice is not to have Tessellation shaders (they give bad performance), but in this case there's also likely a bug you're hitting, in the HLSL->DXC->SPIR-V process.  DXC-compiled tessellation shaders don't work on Mali (since R26), they create a problematic pattern.

A proposed workaround exists: 

-> round-trip the SPIR-V through GLSL (using SPIRV-Cross) and back to SPIR-V again. 

There are DXC patterns that SPIRV-Cross can't cope with either, so depends whether you hit those (it didn't work for another developer who hit this issue). In that case it might just need the SPIR-V to be patched.

Hi, Thank you Ben Clarkfor you reply.

Ben Clark said:

but in this case there's also likely a bug you're hitting, in the HLSL->DXC->SPIR-V process.  DXC-compiled tessellation shaders don't work on Mali (since R26), they create a problematic pattern.

I use glslang for creating SPIR-V from HLSL, see "Android studio project with source and shader":

VkPipelineAndroid_template\app\src\main\shaderSources\compile_hull.cmd

glslangValidator.exe -V -D -e hMain hs.tesc

VkPipelineAndroid_template\app\src\main\shaderSources\compile_domain.cmd:

glslangValidator.exe -V -D -e dMain ds.tese

Ben Clark said:

round-trip the SPIR-V through GLSL (using SPIRV-Cross) and back to SPIR-V again. 

I think this is a strange solution, Will ARM fix driver's bug for new GPU or drivers ?

Qualcomm will fix the same problem: developer.qualcomm.com/.../68860

It is already in the system as a bug, although I don't know how quickly it will be got to. Regardless, if/when it is fixed for new drivers, they only tend to get loaded on new phones, so you'll have a set of phones out there that you'll need a workaround for.

And if we can encourage you away from tessellation shaders all the better :) 

Hi Ben Clark I will wait this fix and check this bug on new phones...