* Handle S_BFM_I32 in a copy shader

* GS: set dwords_per_vertex based on the number of buffer loads if different from VERT_ITEMSIZE
This commit is contained in:
Marcin Mikołajczyk 2026-04-04 12:02:44 +02:00 committed by GitHub
parent 450472b51f
commit 73520ae094
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 18 additions and 1 deletions

View File

@ -45,6 +45,14 @@ CopyShaderData ParseCopyShader(std::span<const u32> code) {
sources[inst.dst[0].code] += inst.control.sopk.simm;
break;
}
case Gcn::Opcode::S_BFM_B32: {
ASSERT(inst.src[0].field == Gcn::OperandField::SignedConstIntPos &&
inst.src[1].field == Gcn::OperandField::SignedConstIntPos);
const auto src0 = inst.src[0].code - Gcn::OperandFieldRange::SignedConstIntPosMin + 1;
const auto src1 = inst.src[1].code - Gcn::OperandFieldRange::SignedConstIntPosMin + 1;
sources[inst.dst[0].code] = ((1 << src0) - 1) << src1;
break;
}
case Gcn::Opcode::EXP: {
const auto& exp = inst.control.exp;
const IR::Attribute semantic = static_cast<IR::Attribute>(exp.target);
@ -71,6 +79,7 @@ CopyShaderData ParseCopyShader(std::span<const u32> code) {
ASSERT(sources[index] != -1);
offsets[inst.src[1].code] += sources[index];
}
data.num_comps++;
break;
}
default:

View File

@ -15,6 +15,7 @@ struct CopyShaderData {
std::map<u32, std::pair<Shader::IR::Attribute, u32>> attr_map;
u32 num_attrs{0};
u32 output_vertices{0};
u32 num_comps{0};
};
CopyShaderData ParseCopyShader(std::span<const u32> code);

View File

@ -104,6 +104,13 @@ void RingAccessElimination(const IR::Program& program, const RuntimeInfo& runtim
output_vertices, info.gs_copy_data.output_vertices);
output_vertices = info.gs_copy_data.output_vertices;
}
u32 dwords_per_vertex = gs_info.out_vertex_data_size;
if (info.gs_copy_data.num_comps && info.gs_copy_data.num_comps != dwords_per_vertex) {
LOG_WARNING(Render_Vulkan,
"VERT_ITEMSIZE {} is different than actual number of dwords per vertex {}",
dwords_per_vertex, info.gs_copy_data.num_comps);
dwords_per_vertex = info.gs_copy_data.num_comps;
}
ForEachInstruction([&](IR::IREmitter& ir, IR::Inst& inst) {
const auto opcode = inst.GetOpcode();
@ -139,7 +146,7 @@ void RingAccessElimination(const IR::Program& program, const RuntimeInfo& runtim
const auto offset = inst.Flags<IR::BufferInstInfo>().inst_offset.Value();
const auto data = ir.BitCast<IR::F32>(IR::U32{inst.Arg(2)});
const auto comp_ofs = output_vertices * 4u;
const auto output_size = comp_ofs * gs_info.out_vertex_data_size;
const auto output_size = comp_ofs * dwords_per_vertex;
const auto vc_read_ofs = (((offset / comp_ofs) * comp_ofs) % output_size) * 16u;
const auto& it = info.gs_copy_data.attr_map.find(vc_read_ofs);