suyu/dynarmic
26
1
Fork 1
Code Issues Pull requests Projects Releases Packages Wiki Activity

emit_x64_vector: Fix AVX-512 implementation of EmitVectorTableLookup64

Browse source
This commit is contained in:
Merry 2024-01-29 23:47:04 +00:00
parent 1e1ba4e0c2
commit 0f20181a45
1 changed files with 49 additions and 27 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

76
src/dynarmic/backend/x64/emit_x64_vector.cpp
View file

@ -4876,57 +4876,79 @@ void EmitX64::EmitVectorTableLookup64(EmitContext& ctx, IR::Inst* inst) {
const bool is_defaults_zero = inst->GetArg(0).IsZero(); const bool is_defaults_zero = inst->GetArg(0).IsZero();
if (code.HasHostFeature(HostFeature::AVX512_Ortho | HostFeature::AVX512BW | HostFeature::AVX512VBMI)) { if (code.HasHostFeature(HostFeature::AVX512_Ortho | HostFeature::AVX512BW | HostFeature::AVX512VBMI)) {
const Xbyak::Xmm indicies = ctx.reg_alloc.UseScratchXmm(args[2]); const Xbyak::Xmm indicies = table_size <= 2 ? ctx.reg_alloc.UseXmm(args[2]) : ctx.reg_alloc.UseScratchXmm(args[2]);
Xbyak::Xmm defaults = ctx.reg_alloc.UseScratchXmm(args[0]);
const u8 index_count = u8(table_size * 8); const u64 index_count = mcl::bit::replicate_element<u8, u64>(static_cast<u8>(table_size * 8));
const u64 index_count64 = mcl::bit::replicate_element<u8, u64>(index_count);
Xbyak::Opmask valid_indices = k1; code.vpcmpub(k1, indicies, code.Const(xword, index_count, 0), CmpInt::LessThan);
code.vpcmpb(valid_indices, indicies, code.Const(xword, index_count64, 0), CmpInt::LessThan);
if (is_defaults_zero) {
defaults = defaults | valid_indices | T_z;
} else {
defaults = defaults | valid_indices;
}
switch (table_size) { switch (table_size) {
case 1: { case 1: {
const Xbyak::Xmm xmm_table0 = ctx.reg_alloc.UseScratchXmm(table[0]); const Xbyak::Xmm xmm_table0 = ctx.reg_alloc.UseXmm(table[0]);
code.vpermb(defaults, indicies, xmm_table0); if (is_defaults_zero) {
const Xbyak::Xmm result = ctx.reg_alloc.ScratchXmm();
code.vpermb(result | k1 | T_z, indicies, xmm_table0);
ctx.reg_alloc.DefineValue(inst, result);
} else {
const Xbyak::Xmm result = ctx.reg_alloc.UseScratchXmm(args[0]);
code.vpermb(result | k1, indicies, xmm_table0);
ctx.reg_alloc.DefineValue(inst, result);
}
break; break;
} }
case 2: { case 2: {
const Xbyak::Xmm xmm_table0 = ctx.reg_alloc.UseScratchXmm(table[0]); const Xbyak::Xmm xmm_table0_lower = ctx.reg_alloc.UseXmm(table[0]);
const Xbyak::Xmm xmm_table0_upper = ctx.reg_alloc.UseXmm(table[1]); const Xbyak::Xmm xmm_table0_upper = ctx.reg_alloc.UseXmm(table[1]);
code.vpunpcklqdq(xmm_table0, xmm_table0, xmm_table0_upper); code.vpunpcklqdq(xmm0, xmm_table0_lower, xmm_table0_upper);
code.vpermb(defaults, indicies, xmm_table0); if (is_defaults_zero) {
const Xbyak::Xmm result = ctx.reg_alloc.ScratchXmm();
code.vpermb(result | k1 | T_z, indicies, xmm0);
ctx.reg_alloc.DefineValue(inst, result);
} else {
const Xbyak::Xmm result = ctx.reg_alloc.UseScratchXmm(args[0]);
code.vpermb(result | k1, indicies, xmm0);
ctx.reg_alloc.DefineValue(inst, result);
}
break; break;
} }
case 3: { case 3: {
const Xbyak::Xmm xmm_table0 = ctx.reg_alloc.UseScratchXmm(table[0]); const Xbyak::Xmm xmm_table0_lower = ctx.reg_alloc.UseXmm(table[0]);
const Xbyak::Xmm xmm_table0_upper = ctx.reg_alloc.UseXmm(table[1]); const Xbyak::Xmm xmm_table0_upper = ctx.reg_alloc.UseXmm(table[1]);
const Xbyak::Xmm xmm_table1 = ctx.reg_alloc.UseXmm(table[2]); const Xbyak::Xmm xmm_table1 = ctx.reg_alloc.UseXmm(table[2]);
code.vpunpcklqdq(xmm_table0, xmm_table0, xmm_table0_upper); code.vpunpcklqdq(xmm0, xmm_table0_lower, xmm_table0_upper);
code.vpermi2b(indicies, xmm_table0, xmm_table1); if (is_defaults_zero) {
code.vmovdqu8(defaults, indicies); code.vpermi2b(indicies | k1 | T_z, xmm0, xmm_table1);
ctx.reg_alloc.DefineValue(inst, indicies);
} else {
const Xbyak::Xmm result = ctx.reg_alloc.UseScratchXmm(args[0]);
code.vpermi2b(indicies, xmm0, xmm_table1);
code.vmovdqu8(result | k1, indicies);
ctx.reg_alloc.DefineValue(inst, result);
}
break; break;
} }
case 4: { case 4: {
const Xbyak::Xmm xmm_table0 = ctx.reg_alloc.UseScratchXmm(table[0]); const Xbyak::Xmm xmm_table0_lower = ctx.reg_alloc.UseXmm(table[0]);
const Xbyak::Xmm xmm_table0_upper = ctx.reg_alloc.UseXmm(table[1]); const Xbyak::Xmm xmm_table0_upper = ctx.reg_alloc.UseXmm(table[1]);
const Xbyak::Xmm xmm_table1 = ctx.reg_alloc.UseScratchXmm(table[2]); const Xbyak::Xmm xmm_table1 = ctx.reg_alloc.UseScratchXmm(table[2]);
const Xbyak::Xmm xmm_table1_upper = ctx.reg_alloc.UseXmm(table[3]); const Xbyak::Xmm xmm_table1_upper = ctx.reg_alloc.UseXmm(table[3]);
code.vpunpcklqdq(xmm_table0, xmm_table0, xmm_table0_upper); code.vpunpcklqdq(xmm0, xmm_table0_lower, xmm_table0_upper);
code.vpunpcklqdq(xmm_table1, xmm_table1, xmm_table1_upper); code.vpunpcklqdq(xmm_table1, xmm_table1, xmm_table1_upper);
code.vpermi2b(indicies, xmm_table0, xmm_table1); if (is_defaults_zero) {
code.vmovdqu8(defaults, indicies); code.vpermi2b(indicies | k1 | T_z, xmm0, xmm_table1);
ctx.reg_alloc.DefineValue(inst, indicies);
} else {
const Xbyak::Xmm result = ctx.reg_alloc.UseScratchXmm(args[0]);
code.vpermi2b(indicies, xmm0, xmm_table1);
code.vmovdqu8(result | k1, indicies);
ctx.reg_alloc.DefineValue(inst, result);
}
break; break;
} }
default:
UNREACHABLE();
break;
} }
ctx.reg_alloc.DefineValue(inst, defaults);
return; return;
} }