1
0
Fork 0
forked from suyu/suyu

Merge pull request #369 from Subv/shader_instr2

ShaderGen: Implemented fsetp/kil and predicated instruction execution.
This commit is contained in:
bunnei 2018-04-20 22:29:39 -04:00 committed by GitHub
commit 8ac3a3f45e
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
2 changed files with 179 additions and 4 deletions

View file

@ -13,6 +13,9 @@ namespace Tegra {
namespace Shader {
struct Register {
// Register 255 is special cased to always be 0
static constexpr size_t ZeroIndex = 255;
constexpr Register() = default;
constexpr Register(u64 value) : value(value) {}
@ -106,6 +109,8 @@ union OpCode {
FSETP_R = 0x5BB,
FSETP_C = 0x4BB,
FSETP_IMM = 0x36B,
FSETP_NEG_IMM = 0x37B,
EXIT = 0xE30,
KIL = 0xE33,
@ -121,6 +126,7 @@ union OpCode {
Ffma,
Flow,
Memory,
FloatPredicate,
Unknown,
};
@ -161,6 +167,9 @@ union OpCode {
case Id::FSETP_C:
case Id::KIL:
return op4;
case Id::FSETP_IMM:
case Id::FSETP_NEG_IMM:
return Id::FSETP_IMM;
}
switch (op5) {
@ -238,8 +247,9 @@ union OpCode {
info_table[Id::FMUL_C] = {Type::Arithmetic, "fmul_c"};
info_table[Id::FMUL_IMM] = {Type::Arithmetic, "fmul_imm"};
info_table[Id::FMUL32_IMM] = {Type::Arithmetic, "fmul32_imm"};
info_table[Id::FSETP_C] = {Type::Arithmetic, "fsetp_c"};
info_table[Id::FSETP_R] = {Type::Arithmetic, "fsetp_r"};
info_table[Id::FSETP_C] = {Type::FloatPredicate, "fsetp_c"};
info_table[Id::FSETP_R] = {Type::FloatPredicate, "fsetp_r"};
info_table[Id::FSETP_IMM] = {Type::FloatPredicate, "fsetp_imm"};
info_table[Id::EXIT] = {Type::Trivial, "exit"};
info_table[Id::IPA] = {Type::Trivial, "ipa"};
info_table[Id::KIL] = {Type::Flow, "kil"};
@ -283,7 +293,23 @@ namespace Shader {
enum class Pred : u64 {
UnusedIndex = 0x7,
NeverExecute = 0xf,
NeverExecute = 0xF,
};
enum class PredCondition : u64 {
LessThan = 1,
Equal = 2,
LessEqual = 3,
GreaterThan = 4,
NotEqual = 5,
GreaterEqual = 6,
// TODO(Subv): Other condition types
};
enum class PredOperation : u64 {
And = 0,
Or = 1,
Xor = 2,
};
enum class SubOp : u64 {
@ -305,7 +331,11 @@ union Instruction {
OpCode opcode;
BitField<0, 8, Register> gpr0;
BitField<8, 8, Register> gpr8;
BitField<16, 4, Pred> pred;
union {
BitField<16, 4, Pred> full_pred;
BitField<16, 3, u64> pred_index;
} pred;
BitField<19, 1, u64> negate_pred;
BitField<20, 8, Register> gpr20;
BitField<20, 7, SubOp> sub_op;
BitField<28, 8, Register> gpr28;
@ -343,6 +373,20 @@ union Instruction {
BitField<49, 1, u64> negate_c;
} ffma;
union {
BitField<0, 3, u64> pred0;
BitField<3, 3, u64> pred3;
BitField<7, 1, u64> abs_a;
BitField<39, 3, u64> pred39;
BitField<42, 1, u64> neg_pred;
BitField<43, 1, u64> neg_a;
BitField<44, 1, u64> abs_b;
BitField<45, 2, PredOperation> op;
BitField<47, 1, u64> ftz;
BitField<48, 4, PredCondition> cond;
BitField<56, 1, u64> neg_b;
} fsetp;
BitField<61, 1, u64> is_b_imm;
BitField<60, 1, u64> is_b_gpr;
BitField<59, 1, u64> is_c_gpr;

View file

@ -220,6 +220,8 @@ private:
/// Generates code representing a temporary (GPR) register.
std::string GetRegister(const Register& reg, unsigned elem = 0) {
if (reg == Register::ZeroIndex)
return "0";
if (stage == Maxwell3D::Regs::ShaderStage::Fragment && reg < 4) {
// GPRs 0-3 are output color for the fragment shader
return std::string{"color."} + "rgba"[(reg + elem) & 3];
@ -276,6 +278,52 @@ private:
shader.AddLine(dest + " = " + src + ";");
}
/*
* Writes code that assigns a predicate boolean variable.
* @param pred The id of the predicate to write to.
* @param value The expression value to assign to the predicate.
*/
void SetPredicate(u64 pred, const std::string& value) {
using Tegra::Shader::Pred;
// Can't assign to the constant predicate.
ASSERT(pred != static_cast<u64>(Pred::UnusedIndex));
std::string variable = 'p' + std::to_string(pred);
shader.AddLine(variable + " = " + value + ';');
declr_predicates.insert(std::move(variable));
}
/*
* Returns the condition to use in the 'if' for a predicated instruction.
* @param instr Instruction to generate the if condition for.
* @returns string containing the predicate condition.
*/
std::string GetPredicateCondition(Instruction instr) const {
using Tegra::Shader::Pred;
ASSERT(instr.pred.pred_index != static_cast<u64>(Pred::UnusedIndex));
std::string variable =
'p' + std::to_string(static_cast<u64>(instr.pred.pred_index.Value()));
if (instr.negate_pred) {
return "!(" + variable + ')';
}
return variable;
}
/*
* Returns whether the instruction at the specified offset is a 'sched' instruction.
* Sched instructions always appear before a sequence of 3 instructions.
*/
bool IsSchedInstruction(u32 offset) const {
// sched instructions appear once every 4 instructions.
static constexpr size_t SchedPeriod = 4;
u32 absolute_offset = offset - main_offset;
return (absolute_offset % SchedPeriod) == 0;
}
/**
* Compiles a single instruction from Tegra to GLSL.
* @param offset the offset of the Tegra shader instruction.
@ -283,10 +331,24 @@ private:
* + 1. If the current instruction always terminates the program, returns PROGRAM_END.
*/
u32 CompileInstr(u32 offset) {
// Ignore sched instructions when generating code.
if (IsSchedInstruction(offset))
return offset + 1;
const Instruction instr = {program_code[offset]};
shader.AddLine("// " + std::to_string(offset) + ": " + OpCode::GetInfo(instr.opcode).name);
using Tegra::Shader::Pred;
ASSERT_MSG(instr.pred.full_pred != Pred::NeverExecute,
"NeverExecute predicate not implemented");
if (instr.pred.pred_index != static_cast<u64>(Pred::UnusedIndex)) {
shader.AddLine("if (" + GetPredicateCondition(instr) + ')');
shader.AddLine('{');
++shader.scope;
}
switch (OpCode::GetInfo(instr.opcode).type) {
case OpCode::Type::Arithmetic: {
std::string dest = GetRegister(instr.gpr0);
@ -450,14 +512,70 @@ private:
}
break;
}
case OpCode::Type::FloatPredicate: {
std::string op_a = instr.fsetp.neg_a ? "-" : "";
op_a += GetRegister(instr.gpr8);
if (instr.fsetp.abs_a) {
op_a = "abs(" + op_a + ')';
}
std::string op_b{};
if (instr.is_b_imm) {
if (instr.fsetp.neg_b) {
// Only the immediate version of fsetp has a neg_b bit.
op_b += '-';
}
op_b += '(' + GetImmediate19(instr) + ')';
} else {
if (instr.is_b_gpr) {
op_b += GetRegister(instr.gpr20);
} else {
op_b += GetUniform(instr.uniform);
}
}
if (instr.fsetp.abs_b) {
op_b = "abs(" + op_b + ')';
}
using Tegra::Shader::Pred;
ASSERT_MSG(instr.fsetp.pred0 == static_cast<u64>(Pred::UnusedIndex) &&
instr.fsetp.pred39 == static_cast<u64>(Pred::UnusedIndex),
"Compound predicates are not implemented");
// We can't use the constant predicate as destination.
ASSERT(instr.fsetp.pred3 != static_cast<u64>(Pred::UnusedIndex));
using Tegra::Shader::PredCondition;
switch (instr.fsetp.cond) {
case PredCondition::LessThan:
SetPredicate(instr.fsetp.pred3, '(' + op_a + ") < (" + op_b + ')');
break;
case PredCondition::Equal:
SetPredicate(instr.fsetp.pred3, '(' + op_a + ") == (" + op_b + ')');
break;
default:
NGLOG_CRITICAL(HW_GPU, "Unhandled predicate condition: {} (a: {}, b: {})",
static_cast<unsigned>(instr.fsetp.cond.Value()), op_a, op_b);
UNREACHABLE();
}
break;
}
default: {
switch (instr.opcode.EffectiveOpCode()) {
case OpCode::Id::EXIT: {
ASSERT_MSG(instr.pred.pred_index == static_cast<u64>(Pred::UnusedIndex),
"Predicated exits not implemented");
shader.AddLine("return true;");
offset = PROGRAM_END - 1;
break;
}
case OpCode::Id::KIL: {
shader.AddLine("discard;");
break;
}
case OpCode::Id::IPA: {
const auto& attribute = instr.attribute.fmt28;
std::string dest = GetRegister(instr.gpr0);
@ -476,6 +594,12 @@ private:
}
}
// Close the predicate condition scope.
if (instr.pred.pred_index != static_cast<u64>(Pred::UnusedIndex)) {
--shader.scope;
shader.AddLine('}');
}
return offset + 1;
}
@ -605,6 +729,12 @@ private:
declarations.AddNewLine();
++const_buffer_layout;
}
declarations.AddNewLine();
for (const auto& pred : declr_predicates) {
declarations.AddLine("bool " + pred + " = false;");
}
declarations.AddNewLine();
}
private:
@ -618,6 +748,7 @@ private:
// Declarations
std::set<std::string> declr_register;
std::set<std::string> declr_predicates;
std::set<Attribute::Index> declr_input_attribute;
std::set<Attribute::Index> declr_output_attribute;
std::array<ConstBufferEntry, Maxwell3D::Regs::MaxConstBuffers> declr_const_buffers;