Translate NGG mesh prolog with guest wave32 width

This commit is contained in:
nmzik
2026-09-23 20:39:46 +04:00
parent 2aa0a64f63
commit 7d7267cc56
2 changed files with 15 additions and 8 deletions
@@ -1063,8 +1063,8 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
}
} else if (options.stage == ShaderType::Mesh) {
const auto& mesh = options.input_info.vertex->mesh;
EXIT_NOT_IMPLEMENTED(options.wave_size != 64u || mesh.primitives_per_group == 0u ||
mesh.vertices_per_group > 64u || total_threads > 15u * 64u);
EXIT_NOT_IMPLEMENTED(mesh.primitives_per_group == 0u || mesh.vertices_per_group > 64u ||
total_threads > 15u * options.wave_size);
const auto u32 = [](uint32_t value) { return IR::U32(IR::Value(value)); };
const auto draw = [&](uint32_t index) {
return IR::U32(
@@ -1089,12 +1089,15 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
entry_ir.IAdd(IR::U32(entry_ir.Emit(IR::ValueOpcode::UDiv32,
{subtract_saturate(vertices, size), step})),
u32(1)));
const auto wave = entry_ir.ShiftRightLogical(local, u32(6));
const auto wave_base = entry_ir.BitwiseAnd(local, u32(~63u));
const auto vertex_count = minimum(subtract_saturate(vertices, wave_base), u32(64));
const auto primitive_count = minimum(subtract_saturate(primitives, wave_base), u32(64));
const auto wave = entry_ir.ShiftRightLogical(local, u32(options.wave_size == 32u ? 5u : 6u));
const auto wave_base = entry_ir.BitwiseAnd(local, u32(~(options.wave_size - 1u)));
const auto vertex_count =
minimum(subtract_saturate(vertices, wave_base), u32(options.wave_size));
const auto primitive_count =
minimum(subtract_saturate(primitives, wave_base), u32(options.wave_size));
const auto wave_info = entry_ir.BitwiseOr(entry_ir.ShiftLeftLogical(wave, u32(24)),
u32(((total_threads + 63u) / 64u) << 28u));
u32(((total_threads + options.wave_size - 1u) /
options.wave_size) << 28u));
entry_ir.SetScalarReg(
static_cast<IR::ScalarReg>(3),
entry_ir.BitwiseOr(wave_info, entry_ir.BitwiseOr(entry_ir.ShiftLeftLogical(