RPCS3/rpcs3
1
0
Fork 0
mirror of https://github.com/RPCS3/rpcs3.git synced 2024-11-25 04:02:42 +01:00
Code Issues Releases Wiki Activity

SPU LLVM: Compute frest exponent at runtime rather than using the lookup table

Browse Source 
- Provides a small speedup and smaller codesize
This commit is contained in:
Malcolm Jestadt 2024-10-04 19:26:02 -04:00 committed by Elad
parent 8ca60df1ab
commit 967adaf9db
2 changed files with 25 additions and 25 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
rpcs3/Emu/CPU/CPUTranslator.h
View File

@ -3741,7 +3741,7 @@ public:
const auto imm32 = immediate.eval(m_ir);
const auto immediate2 = (llvm_const_int<u8>{e});
const auto imm8 = immediate2.eval(m_ir);
result.value = m_ir->CreateCall(get_intrinsic(llvm::Intrinsic::x86_avx512_mask_fixupimm_ps_128), {data0, data1, data2, imm32, imm8});\
result.value = m_ir->CreateCall(get_intrinsic(llvm::Intrinsic::x86_avx512_mask_fixupimm_ps_128), {data0, data1, data2, imm32, imm8});
return result;
}

28
rpcs3/Emu/Cell/SPULLVMRecompiler.cpp
View File

@ -124,7 +124,6 @@ class spu_llvm_recompiler : public spu_recompiler_base, public cpu_translator
// Global LUTs
llvm::GlobalVariable* m_spu_frest_fraction_lut{};
llvm::GlobalVariable* m_spu_frest_exponent_lut{};
llvm::GlobalVariable* m_spu_frsqest_fraction_lut{};
llvm::GlobalVariable* m_spu_frsqest_exponent_lut{};
@ -1507,7 +1506,6 @@ public:
{
// LUTs for some instructions
m_spu_frest_fraction_lut = new llvm::GlobalVariable(*m_module, llvm::ArrayType::get(GetType<u32>(), 32), true, llvm::GlobalValue::PrivateLinkage, llvm::ConstantDataArray::get(m_context, spu_frest_fraction_lut));
m_spu_frest_exponent_lut = new llvm::GlobalVariable(*m_module, llvm::ArrayType::get(GetType<u32>(), 256), true, llvm::GlobalValue::PrivateLinkage, llvm::ConstantDataArray::get(m_context, spu_frest_exponent_lut));
m_spu_frsqest_fraction_lut = new llvm::GlobalVariable(*m_module, llvm::ArrayType::get(GetType<u32>(), 64), true, llvm::GlobalValue::PrivateLinkage, llvm::ConstantDataArray::get(m_context, spu_frsqest_fraction_lut));
m_spu_frsqest_exponent_lut = new llvm::GlobalVariable(*m_module, llvm::ArrayType::get(GetType<u32>(), 256), true, llvm::GlobalValue::PrivateLinkage, llvm::ConstantDataArray::get(m_context, spu_frsqest_exponent_lut));
}
@ -6040,23 +6038,22 @@ public:
const auto a = bitcast<u32[4]>(value<f32[4]>(ci->getOperand(0)));
const auto a_fraction = (a >> splat<u32[4]>(18)) & splat<u32[4]>(0x1F);
const auto a_exponent = (a >> splat<u32[4]>(23)) & splat<u32[4]>(0xFF);
const auto a_exponent = (a & splat<u32[4]>(0x7F800000u));
const auto r_exponent = sub_sat(build<u16[8]>(0000, 0x7E80, 0000, 0x7E80, 0000, 0x7E80, 0000, 0x7E80), bitcast<u16[8]>(a_exponent));
const auto fix_exponent = select((a_exponent > 0), bitcast<u32[4]>(r_exponent), splat<u32[4]>(0x7F800000u));
const auto a_sign = (a & splat<u32[4]>(0x80000000));
value_t<u32[4]> final_result = eval(splat<u32[4]>(0));
for (u32 i = 0; i < 4; i++)
{
const auto eval_fraction = eval(extract(a_fraction, i));
const auto eval_exponent = eval(extract(a_exponent, i));
const auto eval_sign = eval(extract(a_sign, i));
value_t<u32> r_fraction = load_const<u32>(m_spu_frest_fraction_lut, eval_fraction);
value_t<u32> r_exponent = load_const<u32>(m_spu_frest_exponent_lut, eval_exponent);
final_result = eval(insert(final_result, i, eval(r_fraction | eval_sign | r_exponent)));
final_result = eval(insert(final_result, i, r_fraction));
}
return bitcast<f32[4]>(final_result);
return bitcast<f32[4]>(bitcast<u32[4]>(final_result | bitcast<u32[4]>(fix_exponent) | a_sign));
});
set_vr(op.rt, frest(get_vr<f32[4]>(op.ra)));
@ -6713,7 +6710,7 @@ public:
const auto div_result = the_one / div;
return vfixupimmps(div_result, div_result, splat<u32[4]>(0x00330088u), 0, 0xff);
return vfixupimmps(bitcast<f32[4]>(splat<u32[4]>(0xFFFFFFFFu)), div_result, splat<u32[4]>(0x11001188u), 0, 0xff);
});
}
else
@ -7019,20 +7016,23 @@ public:
{
const auto a = bitcast<u32[4]>(value<f32[4]>(ci->getOperand(0)));
const auto a_fraction = (a >> splat<u32[4]>(18)) & splat<u32[4]>(0x1F);
const auto a_exponent = (a >> splat<u32[4]>(23)) & splat<u32[4]>(0xFF);
const auto a_exponent = (a & splat<u32[4]>(0x7F800000u));
const auto r_exponent = sub_sat(build<u16[8]>(0000, 0x7E80, 0000, 0x7E80, 0000, 0x7E80, 0000, 0x7E80), bitcast<u16[8]>(a_exponent));
const auto fix_exponent = select((a_exponent > 0), bitcast<u32[4]>(r_exponent), splat<u32[4]>(0x7F800000u));
const auto a_sign = (a & splat<u32[4]>(0x80000000));
value_t<u32[4]> b = eval(splat<u32[4]>(0));
for (u32 i = 0; i < 4; i++)
{
const auto eval_fraction = eval(extract(a_fraction, i));
const auto eval_exponent = eval(extract(a_exponent, i));
const auto eval_sign = eval(extract(a_sign, i));
value_t<u32> r_fraction = load_const<u32>(m_spu_frest_fraction_lut, eval_fraction);
value_t<u32> r_exponent = load_const<u32>(m_spu_frest_exponent_lut, eval_exponent);
b = eval(insert(b, i, eval(r_fraction | eval_sign | r_exponent)));
b = eval(insert(b, i, r_fraction));
}
b = eval(b | fix_exponent | a_sign);
const auto base = (b & 0x007ffc00u) << 9; // Base fraction
const auto ymul = (b & 0x3ff) * (a & 0x7ffff); // Step fraction * Y fraction (fixed point at 2^-32)
const auto comparison = (ymul > base); // Should exponent be adjusted?