Skip to content

intel/fs: implement exact ALU support

Lionel Landwerlin requested to merge llandwerlin/mesa:review/intel-fs-exact-alus into main

When ALU instructions are flagged as exact, the expectation is that we do computations at more than 16bits of precision.

This fixes failures in vkd3d-proton's test_shader_instructions_dxil test. The particular code is as follows (HLSL) :

   StructuredBuffer<float> Buf : register(t0);
   RWStructuredBuffer<half2> RWBuf : register(u0);

   [numthreads(4, 1, 1)]
   void main(uint thr : SV_DispatchThreadID)
   {
      precise half result = half(Buf[thr]) - half(Buf[thr ^ 3]);
      RWBuf[thr].x = result;
   }

With the following values in Buf[] :

   1.0f / 0x1000000
   2.0f / 0x1000000
   3.0f / 0x1000000
   4.0f / 0x1000000

The negation of half(Buf[thr ^ 3]) leads to a value that would be negative in higher precision but instead becomes -0.0f which makes the computation incorrect for the precise mode.

Signed-off-by: Lionel Landwerlin lionel.g.landwerlin@intel.com Cc: mesa-stable

Edited by Lionel Landwerlin

Merge request reports