mirrors/dynarmic
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

Add Unsafe_InaccurateNaN optimization to all fma instructions

Browse source
This commit is contained in:
MerryMage 2021-01-02 17:19:08 +00:00
parent 8c4463a0c1
commit f9ccf91b94
2 changed files with 77 additions and 18 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

32
src/backend/x64/emit_x64_floating_point.cpp
View file

@ -590,9 +590,20 @@ static void EmitFPMulAdd(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst) {
using FPT = mp::unsigned_integer_of_size<fsize>; using FPT = mp::unsigned_integer_of_size<fsize>;
if constexpr (fsize != 16) { if constexpr (fsize != 16) {
if (code.HasFMA()) { auto args = ctx.reg_alloc.GetArgumentInfo(inst);
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
if (code.HasFMA() && ctx.HasOptimization(OptimizationFlag::Unsafe_InaccurateNaN)) {
const Xbyak::Xmm result = ctx.reg_alloc.UseScratchXmm(args[0]);
const Xbyak::Xmm operand2 = ctx.reg_alloc.UseXmm(args[1]);
const Xbyak::Xmm operand3 = ctx.reg_alloc.UseXmm(args[2]);
FCODE(vfmadd231s)(result, operand2, operand3);
ctx.reg_alloc.DefineValue(inst, result);
return;
}
if (code.HasFMA()) {
Xbyak::Label end, fallback; Xbyak::Label end, fallback;
const Xbyak::Xmm operand1 = ctx.reg_alloc.UseXmm(args[0]); const Xbyak::Xmm operand1 = ctx.reg_alloc.UseXmm(args[0]);
@ -641,8 +652,6 @@ static void EmitFPMulAdd(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst) {
} }
if (ctx.HasOptimization(OptimizationFlag::Unsafe_UnfuseFMA)) { if (ctx.HasOptimization(OptimizationFlag::Unsafe_UnfuseFMA)) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
const Xbyak::Xmm operand1 = ctx.reg_alloc.UseScratchXmm(args[0]); const Xbyak::Xmm operand1 = ctx.reg_alloc.UseScratchXmm(args[0]);
const Xbyak::Xmm operand2 = ctx.reg_alloc.UseScratchXmm(args[1]); const Xbyak::Xmm operand2 = ctx.reg_alloc.UseScratchXmm(args[1]);
const Xbyak::Xmm operand3 = ctx.reg_alloc.UseXmm(args[2]); const Xbyak::Xmm operand3 = ctx.reg_alloc.UseXmm(args[2]);
@ -1014,6 +1023,21 @@ static void EmitFPRSqrtStepFused(BlockOfCode& code, EmitContext& ctx, IR::Inst*
using FPT = mp::unsigned_integer_of_size<fsize>; using FPT = mp::unsigned_integer_of_size<fsize>;
if constexpr (fsize != 16) { if constexpr (fsize != 16) {
if (code.HasFMA() && code.HasAVX() && ctx.HasOptimization(OptimizationFlag::Unsafe_InaccurateNaN)) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
const Xbyak::Xmm operand1 = ctx.reg_alloc.UseXmm(args[0]);
const Xbyak::Xmm operand2 = ctx.reg_alloc.UseXmm(args[1]);
const Xbyak::Xmm result = ctx.reg_alloc.ScratchXmm();
code.vmovaps(result, code.MConst(xword, FP::FPValue<FPT, false, 0, 3>()));
FCODE(vfnmadd231s)(result, operand1, operand2);
FCODE(vmuls)(result, result, code.MConst(xword, FP::FPValue<FPT, false, -1, 1>()));
ctx.reg_alloc.DefineValue(inst, result);
return;
}
if (code.HasFMA() && code.HasAVX()) { if (code.HasFMA() && code.HasAVX()) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst); auto args = ctx.reg_alloc.GetArgumentInfo(inst);

63
src/backend/x64/emit_x64_vector_floating_point.cpp
View file

@ -985,11 +985,23 @@ void EmitFPVectorMulAdd(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst) {
}; };
if constexpr (fsize != 16) { if constexpr (fsize != 16) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
const bool fpcr_controlled = args[3].GetImmediateU1();
if (code.HasFMA() && code.HasAVX()) { if (code.HasFMA() && code.HasAVX()) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst); const Xbyak::Xmm result = ctx.reg_alloc.UseScratchXmm(args[0]);
const Xbyak::Xmm xmm_b = ctx.reg_alloc.UseXmm(args[1]);
const Xbyak::Xmm xmm_c = ctx.reg_alloc.UseXmm(args[2]);
const bool fpcr_controlled = args[3].GetImmediateU1(); MaybeStandardFPSCRValue(code, ctx, fpcr_controlled, [&]{
FCODE(vfmadd231p)(result, xmm_b, xmm_c);
});
ctx.reg_alloc.DefineValue(inst, result);
return;
}
if (code.HasFMA() && code.HasAVX()) {
const Xbyak::Xmm result = ctx.reg_alloc.ScratchXmm(); const Xbyak::Xmm result = ctx.reg_alloc.ScratchXmm();
const Xbyak::Xmm xmm_a = ctx.reg_alloc.UseXmm(args[0]); const Xbyak::Xmm xmm_a = ctx.reg_alloc.UseXmm(args[0]);
const Xbyak::Xmm xmm_b = ctx.reg_alloc.UseXmm(args[1]); const Xbyak::Xmm xmm_b = ctx.reg_alloc.UseXmm(args[1]);
@ -1025,8 +1037,6 @@ void EmitFPVectorMulAdd(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst) {
} }
if (ctx.HasOptimization(OptimizationFlag::Unsafe_UnfuseFMA)) { if (ctx.HasOptimization(OptimizationFlag::Unsafe_UnfuseFMA)) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
const Xbyak::Xmm operand1 = ctx.reg_alloc.UseScratchXmm(args[0]); const Xbyak::Xmm operand1 = ctx.reg_alloc.UseScratchXmm(args[0]);
const Xbyak::Xmm operand2 = ctx.reg_alloc.UseScratchXmm(args[1]); const Xbyak::Xmm operand2 = ctx.reg_alloc.UseScratchXmm(args[1]);
const Xbyak::Xmm operand3 = ctx.reg_alloc.UseXmm(args[2]); const Xbyak::Xmm operand3 = ctx.reg_alloc.UseXmm(args[2]);
@ -1233,10 +1243,24 @@ static void EmitRecipStepFused(BlockOfCode& code, EmitContext& ctx, IR::Inst* in
}; };
if constexpr (fsize != 16) { if constexpr (fsize != 16) {
if (code.HasFMA() && code.HasAVX()) { auto args = ctx.reg_alloc.GetArgumentInfo(inst);
auto args = ctx.reg_alloc.GetArgumentInfo(inst); const bool fpcr_controlled = args[2].GetImmediateU1();
const bool fpcr_controlled = args[2].GetImmediateU1();
if (code.HasFMA() && code.HasAVX() && ctx.HasOptimization(OptimizationFlag::Unsafe_InaccurateNaN)) {
const Xbyak::Xmm result = ctx.reg_alloc.ScratchXmm();
const Xbyak::Xmm operand1 = ctx.reg_alloc.UseXmm(args[0]);
const Xbyak::Xmm operand2 = ctx.reg_alloc.UseXmm(args[1]);
MaybeStandardFPSCRValue(code, ctx, fpcr_controlled, [&]{
code.movaps(result, GetVectorOf<fsize, false, 0, 2>(code));
FCODE(vfnmadd231p)(result, operand1, operand2);
});
ctx.reg_alloc.DefineValue(inst, result);
return;
}
if (code.HasFMA() && code.HasAVX()) {
const Xbyak::Xmm result = ctx.reg_alloc.ScratchXmm(); const Xbyak::Xmm result = ctx.reg_alloc.ScratchXmm();
const Xbyak::Xmm operand1 = ctx.reg_alloc.UseXmm(args[0]); const Xbyak::Xmm operand1 = ctx.reg_alloc.UseXmm(args[0]);
const Xbyak::Xmm operand2 = ctx.reg_alloc.UseXmm(args[1]); const Xbyak::Xmm operand2 = ctx.reg_alloc.UseXmm(args[1]);
@ -1269,8 +1293,6 @@ static void EmitRecipStepFused(BlockOfCode& code, EmitContext& ctx, IR::Inst* in
} }
if (ctx.HasOptimization(OptimizationFlag::Unsafe_UnfuseFMA)) { if (ctx.HasOptimization(OptimizationFlag::Unsafe_UnfuseFMA)) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
const Xbyak::Xmm operand1 = ctx.reg_alloc.UseScratchXmm(args[0]); const Xbyak::Xmm operand1 = ctx.reg_alloc.UseScratchXmm(args[0]);
const Xbyak::Xmm operand2 = ctx.reg_alloc.UseXmm(args[1]); const Xbyak::Xmm operand2 = ctx.reg_alloc.UseXmm(args[1]);
const Xbyak::Xmm result = ctx.reg_alloc.ScratchXmm(); const Xbyak::Xmm result = ctx.reg_alloc.ScratchXmm();
@ -1428,10 +1450,25 @@ static void EmitRSqrtStepFused(BlockOfCode& code, EmitContext& ctx, IR::Inst* in
}; };
if constexpr (fsize != 16) { if constexpr (fsize != 16) {
if (code.HasFMA() && code.HasAVX()) { auto args = ctx.reg_alloc.GetArgumentInfo(inst);
auto args = ctx.reg_alloc.GetArgumentInfo(inst); const bool fpcr_controlled = args[2].GetImmediateU1();
const bool fpcr_controlled = args[2].GetImmediateU1();
if (code.HasFMA() && code.HasAVX() && ctx.HasOptimization(OptimizationFlag::Unsafe_InaccurateNaN)) {
const Xbyak::Xmm result = ctx.reg_alloc.ScratchXmm();
const Xbyak::Xmm operand1 = ctx.reg_alloc.UseXmm(args[0]);
const Xbyak::Xmm operand2 = ctx.reg_alloc.UseXmm(args[1]);
MaybeStandardFPSCRValue(code, ctx, fpcr_controlled, [&]{
code.vmovaps(result, GetVectorOf<fsize, false, 0, 3>(code));
FCODE(vfnmadd231p)(result, operand1, operand2);
FCODE(vmulp)(result, result, GetVectorOf<fsize, false, -1, 1>(code));
});
ctx.reg_alloc.DefineValue(inst, result);
return;
}
if (code.HasFMA() && code.HasAVX()) {
const Xbyak::Xmm result = ctx.reg_alloc.ScratchXmm(); const Xbyak::Xmm result = ctx.reg_alloc.ScratchXmm();
const Xbyak::Xmm operand1 = ctx.reg_alloc.UseXmm(args[0]); const Xbyak::Xmm operand1 = ctx.reg_alloc.UseXmm(args[0]);
const Xbyak::Xmm operand2 = ctx.reg_alloc.UseXmm(args[1]); const Xbyak::Xmm operand2 = ctx.reg_alloc.UseXmm(args[1]);
@ -1470,8 +1507,6 @@ static void EmitRSqrtStepFused(BlockOfCode& code, EmitContext& ctx, IR::Inst* in
} }
if (ctx.HasOptimization(OptimizationFlag::Unsafe_UnfuseFMA)) { if (ctx.HasOptimization(OptimizationFlag::Unsafe_UnfuseFMA)) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
const Xbyak::Xmm operand1 = ctx.reg_alloc.UseScratchXmm(args[0]); const Xbyak::Xmm operand1 = ctx.reg_alloc.UseScratchXmm(args[0]);
const Xbyak::Xmm operand2 = ctx.reg_alloc.UseXmm(args[1]); const Xbyak::Xmm operand2 = ctx.reg_alloc.UseXmm(args[1]);
const Xbyak::Xmm result = ctx.reg_alloc.ScratchXmm(); const Xbyak::Xmm result = ctx.reg_alloc.ScratchXmm();