Impaczus/rpcs3
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

PPU LLVM: Optimize vector NaN handling

Browse Source
This commit is contained in:
yahfz
2026-07-15 04:42:56 -03:00
committed by Elad
parent 1269ebff27
commit 50df95d6c4
2 changed files with 10 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
rpcs3/Emu/Cell/PPUThread.cpp
+3
View File
@@ -5153,6 +5153,7 @@ bool ppu_initialize(const ppu_module<lv2_obj>& info, bool check_only, u64 file_s
accurate_vnan, accurate_vnan,
accurate_nj_mode, accurate_nj_mode,
contains_symbol_resolver, contains_symbol_resolver,
daz_and_ftz,
__bitset_enum_max __bitset_enum_max
}; };
@@ -5184,6 +5185,8 @@ bool ppu_initialize(const ppu_module<lv2_obj>& info, bool check_only, u64 file_s
settings += ppu_settings::accurate_nj_mode, settings -= ppu_settings::fixup_nj_denormals, fmt::throw_exception("NJ Not implemented"); settings += ppu_settings::accurate_nj_mode, settings -= ppu_settings::fixup_nj_denormals, fmt::throw_exception("NJ Not implemented");
if (fpos >= info.get_funcs().size() || module_counter % c_moudles_per_jit == c_moudles_per_jit - 1) if (fpos >= info.get_funcs().size() || module_counter % c_moudles_per_jit == c_moudles_per_jit - 1)
settings += ppu_settings::contains_symbol_resolver; // Avoid invalidating all modules for this purpose settings += ppu_settings::contains_symbol_resolver; // Avoid invalidating all modules for this purpose
if (g_cfg.core.set_daz_and_ftz)
settings += ppu_settings::daz_and_ftz;
// Write version, hash, CPU, settings // Write version, hash, CPU, settings
fmt::append(obj_name, "v7-kusa-%s-%s-%s.obj", fmt::base57(output, 16), fmt::base57(settings), jit_compiler::cpu(g_cfg.core.llvm_cpu.to_string())); fmt::append(obj_name, "v7-kusa-%s-%s-%s.obj", fmt::base57(output, 16), fmt::base57(settings), jit_compiler::cpu(g_cfg.core.llvm_cpu.to_string()));
rpcs3/Emu/Cell/PPUTranslator.cpp
+7
View File
@@ -447,6 +447,13 @@ Function* PPUTranslator::GetSymbolResolver(const ppu_module<lv2_obj>& info)
Value* PPUTranslator::VecHandleNan(Value* val) Value* PPUTranslator::VecHandleNan(Value* val)
{ {
if (m_use_avx512 && !g_cfg.core.set_daz_and_ftz)
{
// VFIXUPIMMPS is only equivalent when DAZ is disabled.
// Select the canonical NaN for QNaN/SNaN and pass through every other class.
return vfixupimmps(value<f32[4]>(nan_vec4), value<f32[4]>(val), splat<u32[4]>(0x1111'1100), 0x10, 0xff).eval(m_ir);
}
const auto is_nan = m_ir->CreateFCmpUNO(val, val); const auto is_nan = m_ir->CreateFCmpUNO(val, val);
val = m_ir->CreateSelect(is_nan, nan_vec4, val); val = m_ir->CreateSelect(is_nan, nan_vec4, val);