D881 - A legacy vertex shader draws as a mesh shader that assembles its own primitives

Status: decided Date: 2026-10-14

A draw whose VGT_SHADER_STAGES_EN sets no PRIMGEN_EN and no stage before the vertex shader runs a legacy vertex shader. Its module is still made at the mesh stage (D688), with these differences.

• Program and registers. The program is the one at SPI_SHADER_PGM_LO/HI_VS, chosen over the primitive shader's _ES pair for this draw only. Its user data is SPI_SHADER_USER_DATA_VS_* from s0. Its resource words are RSRC1/2_VS. Its wave width is VS_W32_EN.
• What the module does itself. The program declares no counts and exports no primitive, so the module does both from the draw:
  • its vertices, a lane each;
  • the primitives its input topology assembles, written as an exp prim would write them.
• Inputs. The vertex id goes in v0 (the fetched index, for an indexed draw) and the instance id in v3. That is radeonsi's GFX10 layout (declare_vs_input_vgprs).
• Output primitive. It is the input topology's: radeonsi writes VGT_GS_OUT_PRIM_TYPE only for a primitive or geometry shader (si_state_draw.cpp 1083).
• Streamout stores. A legacy vertex shader that takes streamout words (RSRC2_VS.SO_EN) stores its streamout itself. The draw's module drops those stores:
  • the streamout pass carries them out exactly (D861);
  • with streamout off, the hardware hands the wave no vertices to store.

A stream that never wrote VGT_SHADER_STAGES_EN is not taken for a legacy draw.

Why: radeonsi turns NGG off on GFX10 for every program with transform-feedback outputs, whether or not feedback is active at the draw. Such a program was bound to no vertex shader at all, and the draw ran whatever was bound before it. GCTS00001's KHR-GL30.transform_feedback group stopped at its first draw.

Rejected:

• A Vulkan vertex stage for legacy shaders. Every translated vertex program is a mesh module (D688), and the outputs, clip distances, layers and point sizes are written there. A second host stage would restate all of it.
• Running the stores in the draw. A draw's module does not write guest memory back exactly. The streamout pass already does.