glasm: Fix clamps so the min value has priority on NAN on GLASM
This commit is contained in:
parent
8eb72ff0dc
commit
6237300e36
|
@ -9,11 +9,10 @@
|
||||||
#include "shader_recompiler/frontend/ir/value.h"
|
#include "shader_recompiler/frontend/ir/value.h"
|
||||||
|
|
||||||
namespace Shader::Backend::GLASM {
|
namespace Shader::Backend::GLASM {
|
||||||
|
namespace {
|
||||||
template <typename InputType>
|
template <typename InputType>
|
||||||
static void Compare(EmitContext& ctx, IR::Inst& inst, InputType lhs, InputType rhs,
|
void Compare(EmitContext& ctx, IR::Inst& inst, InputType lhs, InputType rhs, std::string_view op,
|
||||||
std::string_view op, std::string_view type, bool ordered,
|
std::string_view type, bool ordered, bool inequality = false) {
|
||||||
bool inequality = false) {
|
|
||||||
const Register ret{ctx.reg_alloc.Define(inst)};
|
const Register ret{ctx.reg_alloc.Define(inst)};
|
||||||
ctx.Add("{}.{} RC.x,{},{};", op, type, lhs, rhs);
|
ctx.Add("{}.{} RC.x,{},{};", op, type, lhs, rhs);
|
||||||
if (ordered && inequality) {
|
if (ordered && inequality) {
|
||||||
|
@ -35,6 +34,16 @@ static void Compare(EmitContext& ctx, IR::Inst& inst, InputType lhs, InputType r
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
|
|
||||||
|
template <typename InputType>
|
||||||
|
void Clamp(EmitContext& ctx, Register ret, InputType value, InputType min_value,
|
||||||
|
InputType max_value) {
|
||||||
|
// Call MAX first to properly clamp nan to min_value instead
|
||||||
|
ctx.Add("MAX.F {}.x,{},{};"
|
||||||
|
"MIN.F {}.x,{},{};",
|
||||||
|
ret, min_value, value, ret, ret, max_value);
|
||||||
|
}
|
||||||
|
} // Anonymous namespace
|
||||||
|
|
||||||
void EmitFPAbs16([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] IR::Inst& inst,
|
void EmitFPAbs16([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] IR::Inst& inst,
|
||||||
[[maybe_unused]] Register value) {
|
[[maybe_unused]] Register value) {
|
||||||
throw NotImplementedException("GLASM instruction");
|
throw NotImplementedException("GLASM instruction");
|
||||||
|
@ -171,18 +180,12 @@ void EmitFPClamp16([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] Register
|
||||||
|
|
||||||
void EmitFPClamp32(EmitContext& ctx, IR::Inst& inst, ScalarF32 value, ScalarF32 min_value,
|
void EmitFPClamp32(EmitContext& ctx, IR::Inst& inst, ScalarF32 value, ScalarF32 min_value,
|
||||||
ScalarF32 max_value) {
|
ScalarF32 max_value) {
|
||||||
const Register ret{ctx.reg_alloc.Define(inst)};
|
Clamp(ctx, ctx.reg_alloc.Define(inst), value, min_value, max_value);
|
||||||
ctx.Add("MIN.F {}.x,{},{};"
|
|
||||||
"MAX.F {}.x,{},{};",
|
|
||||||
ret, max_value, value, ret, ret, min_value);
|
|
||||||
}
|
}
|
||||||
|
|
||||||
void EmitFPClamp64(EmitContext& ctx, IR::Inst& inst, ScalarF64 value, ScalarF64 min_value,
|
void EmitFPClamp64(EmitContext& ctx, IR::Inst& inst, ScalarF64 value, ScalarF64 min_value,
|
||||||
ScalarF64 max_value) {
|
ScalarF64 max_value) {
|
||||||
const Register ret{ctx.reg_alloc.LongDefine(inst)};
|
Clamp(ctx, ctx.reg_alloc.LongDefine(inst), value, min_value, max_value);
|
||||||
ctx.Add("MIN.F64 {}.x,{},{};"
|
|
||||||
"MAX.F64 {}.x,{},{};",
|
|
||||||
ret, max_value, value, ret, ret, min_value);
|
|
||||||
}
|
}
|
||||||
|
|
||||||
void EmitFPRoundEven16([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] Register value) {
|
void EmitFPRoundEven16([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] Register value) {
|
||||||
|
|
Reference in New Issue