Merge pull request #1601 from FernandoS27/shader-precision

Improved Shader accuracy on Vertex and Geometry Shaders.
This commit is contained in:
bunnei 2018-10-28 13:06:21 -04:00 committed by GitHub
commit aa1cf608ed
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23

View File

@ -341,10 +341,10 @@ public:
*/ */
void SetRegisterToFloat(const Register& reg, u64 elem, const std::string& value, void SetRegisterToFloat(const Register& reg, u64 elem, const std::string& value,
u64 dest_num_components, u64 value_num_components, u64 dest_num_components, u64 value_num_components,
bool is_saturated = false, u64 dest_elem = 0) { bool is_saturated = false, u64 dest_elem = 0, bool precise = false) {
SetRegister(reg, elem, is_saturated ? "clamp(" + value + ", 0.0, 1.0)" : value, SetRegister(reg, elem, is_saturated ? "clamp(" + value + ", 0.0, 1.0)" : value,
dest_num_components, value_num_components, dest_elem); dest_num_components, value_num_components, dest_elem, precise);
} }
/** /**
@ -368,7 +368,7 @@ public:
const std::string func{is_signed ? "intBitsToFloat" : "uintBitsToFloat"}; const std::string func{is_signed ? "intBitsToFloat" : "uintBitsToFloat"};
SetRegister(reg, elem, func + '(' + ConvertIntegerSize(value, size) + ')', SetRegister(reg, elem, func + '(' + ConvertIntegerSize(value, size) + ')',
dest_num_components, value_num_components, dest_elem); dest_num_components, value_num_components, dest_elem, false);
if (sets_cc) { if (sets_cc) {
const std::string zero_condition = "( " + ConvertIntegerSize(value, size) + " == 0 )"; const std::string zero_condition = "( " + ConvertIntegerSize(value, size) + " == 0 )";
@ -416,7 +416,7 @@ public:
} }
}(); }();
SetRegister(reg, elem, result, dest_num_components, value_num_components, dest_elem); SetRegister(reg, elem, result, dest_num_components, value_num_components, dest_elem, false);
} }
/** /**
@ -757,7 +757,8 @@ private:
* @param dest_elem Optional, the destination element to use for the operation. * @param dest_elem Optional, the destination element to use for the operation.
*/ */
void SetRegister(const Register& reg, u64 elem, const std::string& value, void SetRegister(const Register& reg, u64 elem, const std::string& value,
u64 dest_num_components, u64 value_num_components, u64 dest_elem) { u64 dest_num_components, u64 value_num_components, u64 dest_elem,
bool precise) {
if (reg == Register::ZeroIndex) { if (reg == Register::ZeroIndex) {
LOG_CRITICAL(HW_GPU, "Cannot set Register::ZeroIndex"); LOG_CRITICAL(HW_GPU, "Cannot set Register::ZeroIndex");
UNREACHABLE(); UNREACHABLE();
@ -774,7 +775,18 @@ private:
src += GetSwizzle(elem); src += GetSwizzle(elem);
} }
shader.AddLine(dest + " = " + src + ';'); if (precise && stage != Maxwell3D::Regs::ShaderStage::Fragment) {
shader.AddLine('{');
++shader.scope;
// This avoids optimizations of constant propagation and keeps the code as the original
// Sadly using the precise keyword causes "linking" errors on fragment shaders.
shader.AddLine("precise float tmp = " + src + ';');
shader.AddLine(dest + " = tmp;");
--shader.scope;
shader.AddLine('}');
} else {
shader.AddLine(dest + " = " + src + ';');
}
} }
/// Build the GLSL register list. /// Build the GLSL register list.
@ -1510,8 +1522,9 @@ private:
ASSERT_MSG(instr.fmul.cc == 0, "FMUL cc is not implemented"); ASSERT_MSG(instr.fmul.cc == 0, "FMUL cc is not implemented");
op_b = GetOperandAbsNeg(op_b, false, instr.fmul.negate_b); op_b = GetOperandAbsNeg(op_b, false, instr.fmul.negate_b);
regs.SetRegisterToFloat(instr.gpr0, 0, op_a + " * " + op_b, 1, 1, regs.SetRegisterToFloat(instr.gpr0, 0, op_a + " * " + op_b, 1, 1,
instr.alu.saturate_d); instr.alu.saturate_d, 0, true);
break; break;
} }
case OpCode::Id::FADD_C: case OpCode::Id::FADD_C:
@ -1519,8 +1532,9 @@ private:
case OpCode::Id::FADD_IMM: { case OpCode::Id::FADD_IMM: {
op_a = GetOperandAbsNeg(op_a, instr.alu.abs_a, instr.alu.negate_a); op_a = GetOperandAbsNeg(op_a, instr.alu.abs_a, instr.alu.negate_a);
op_b = GetOperandAbsNeg(op_b, instr.alu.abs_b, instr.alu.negate_b); op_b = GetOperandAbsNeg(op_b, instr.alu.abs_b, instr.alu.negate_b);
regs.SetRegisterToFloat(instr.gpr0, 0, op_a + " + " + op_b, 1, 1, regs.SetRegisterToFloat(instr.gpr0, 0, op_a + " + " + op_b, 1, 1,
instr.alu.saturate_d); instr.alu.saturate_d, 0, true);
break; break;
} }
case OpCode::Id::MUFU: { case OpCode::Id::MUFU: {
@ -1528,31 +1542,31 @@ private:
switch (instr.sub_op) { switch (instr.sub_op) {
case SubOp::Cos: case SubOp::Cos:
regs.SetRegisterToFloat(instr.gpr0, 0, "cos(" + op_a + ')', 1, 1, regs.SetRegisterToFloat(instr.gpr0, 0, "cos(" + op_a + ')', 1, 1,
instr.alu.saturate_d); instr.alu.saturate_d, 0, true);
break; break;
case SubOp::Sin: case SubOp::Sin:
regs.SetRegisterToFloat(instr.gpr0, 0, "sin(" + op_a + ')', 1, 1, regs.SetRegisterToFloat(instr.gpr0, 0, "sin(" + op_a + ')', 1, 1,
instr.alu.saturate_d); instr.alu.saturate_d, 0, true);
break; break;
case SubOp::Ex2: case SubOp::Ex2:
regs.SetRegisterToFloat(instr.gpr0, 0, "exp2(" + op_a + ')', 1, 1, regs.SetRegisterToFloat(instr.gpr0, 0, "exp2(" + op_a + ')', 1, 1,
instr.alu.saturate_d); instr.alu.saturate_d, 0, true);
break; break;
case SubOp::Lg2: case SubOp::Lg2:
regs.SetRegisterToFloat(instr.gpr0, 0, "log2(" + op_a + ')', 1, 1, regs.SetRegisterToFloat(instr.gpr0, 0, "log2(" + op_a + ')', 1, 1,
instr.alu.saturate_d); instr.alu.saturate_d, 0, true);
break; break;
case SubOp::Rcp: case SubOp::Rcp:
regs.SetRegisterToFloat(instr.gpr0, 0, "1.0 / " + op_a, 1, 1, regs.SetRegisterToFloat(instr.gpr0, 0, "1.0 / " + op_a, 1, 1,
instr.alu.saturate_d); instr.alu.saturate_d, 0, true);
break; break;
case SubOp::Rsq: case SubOp::Rsq:
regs.SetRegisterToFloat(instr.gpr0, 0, "inversesqrt(" + op_a + ')', 1, 1, regs.SetRegisterToFloat(instr.gpr0, 0, "inversesqrt(" + op_a + ')', 1, 1,
instr.alu.saturate_d); instr.alu.saturate_d, 0, true);
break; break;
case SubOp::Sqrt: case SubOp::Sqrt:
regs.SetRegisterToFloat(instr.gpr0, 0, "sqrt(" + op_a + ')', 1, 1, regs.SetRegisterToFloat(instr.gpr0, 0, "sqrt(" + op_a + ')', 1, 1,
instr.alu.saturate_d); instr.alu.saturate_d, 0, true);
break; break;
default: default:
LOG_CRITICAL(HW_GPU, "Unhandled MUFU sub op: {0:x}", LOG_CRITICAL(HW_GPU, "Unhandled MUFU sub op: {0:x}",
@ -1573,7 +1587,7 @@ private:
regs.SetRegisterToFloat(instr.gpr0, 0, regs.SetRegisterToFloat(instr.gpr0, 0,
'(' + condition + ") ? min(" + parameters + ") : max(" + '(' + condition + ") ? min(" + parameters + ") : max(" +
parameters + ')', parameters + ')',
1, 1); 1, 1, false, 0, true);
break; break;
} }
case OpCode::Id::RRO_C: case OpCode::Id::RRO_C:
@ -1602,7 +1616,7 @@ private:
regs.SetRegisterToFloat(instr.gpr0, 0, regs.SetRegisterToFloat(instr.gpr0, 0,
regs.GetRegisterAsFloat(instr.gpr8) + " * " + regs.GetRegisterAsFloat(instr.gpr8) + " * " +
GetImmediate32(instr), GetImmediate32(instr),
1, 1, instr.fmul32.saturate); 1, 1, instr.fmul32.saturate, 0, true);
break; break;
} }
case OpCode::Id::FADD32I: { case OpCode::Id::FADD32I: {
@ -1625,7 +1639,7 @@ private:
op_b = "-(" + op_b + ')'; op_b = "-(" + op_b + ')';
} }
regs.SetRegisterToFloat(instr.gpr0, 0, op_a + " + " + op_b, 1, 1); regs.SetRegisterToFloat(instr.gpr0, 0, op_a + " + " + op_b, 1, 1, false, 0, true);
break; break;
} }
} }
@ -2087,8 +2101,9 @@ private:
} }
} }
regs.SetRegisterToFloat(instr.gpr0, 0, op_a + " * " + op_b + " + " + op_c, 1, 1, regs.SetRegisterToFloat(instr.gpr0, 0, "fma(" + op_a + ", " + op_b + ", " + op_c + ')',
instr.alu.saturate_d); 1, 1, instr.alu.saturate_d, 0, true);
break; break;
} }
case OpCode::Type::Hfma2: { case OpCode::Type::Hfma2: {