Intel's Mesa Driver Isn't Yet Ready For Double-Precision Floating-Point Types
Those paying attention to the mailing list may have seen the patches yesterday that surfaced from an independent developer for implementing ARB_gpu_shader_fp64 with Intel's graphics driver for Gen 7+ hardware. Unfortunately, it was too good to be true.
The ARB_gpu_shader_fp64 extension is about allowing shaders to use double-precision floating-point data types. Doubles can be used for inputs, outputs, and uniforms as outlined via the registry specification. There's been work towards supporting ARB_gpu_shader_fp64 in the past but nothing yet materialized in full for Mesa Git.
An unknown contributor to the Mesa community meanwhile on Saturday posted a set of patches that he says enables ARB_gpu_shader_fp64 for the i965 Mesa driver and would work for Gen7+ graphics hardware. While it may seem exciting at first that the Intel Mesa OpenGL driver now supports double-precision floating-point types, it turns out the work is bunk and incomplete.
Intel's Matt Turner said the work "doesn't come close to implementing fp64." Turner later elaborated on this work and that an Intel developer (Topi Pohjolainen) has been working on the support via his Mesa Git tree with the FP64 branches. There's also not any doubles support within NIR at this point.
The ARB_gpu_shader_fp64 extension is about allowing shaders to use double-precision floating-point data types. Doubles can be used for inputs, outputs, and uniforms as outlined via the registry specification. There's been work towards supporting ARB_gpu_shader_fp64 in the past but nothing yet materialized in full for Mesa Git.
An unknown contributor to the Mesa community meanwhile on Saturday posted a set of patches that he says enables ARB_gpu_shader_fp64 for the i965 Mesa driver and would work for Gen7+ graphics hardware. While it may seem exciting at first that the Intel Mesa OpenGL driver now supports double-precision floating-point types, it turns out the work is bunk and incomplete.
Intel's Matt Turner said the work "doesn't come close to implementing fp64." Turner later elaborated on this work and that an Intel developer (Topi Pohjolainen) has been working on the support via his Mesa Git tree with the FP64 branches. There's also not any doubles support within NIR at this point.
4 Comments