mirrors/ZLUDA
1
0
Fork 0
mirror of https://github.com/vosen/ZLUDA.git synced 2025-04-20 00:14:45 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Handle mad.hi.cc like mad.lo.cc

Browse source
This commit is contained in:
NyanCatTW1 2024-04-05 02:25:28 +08:00
parent b695f44c18
commit f9de488bdf
4 changed files with 16 additions and 5 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

1
ptx/src/ast.rs
View file

@ -380,6 +380,7 @@ pub enum Instruction<P: ArgParams> {
},
MadCC {
type_: ScalarType,
is_hi: bool,
arg: Arg4<P>,
},
Fma(ArithFloat, Arg4<P>),

5
ptx/src/emit.rs
View file

@ -621,8 +621,8 @@ fn emit_statement(
crate::translate::Statement::MadC(MadCDetails { type_, is_hi, arg }) => {
emit_inst_madc(ctx, type_, is_hi, &arg)?
}
crate::translate::Statement::MadCC(MadCCDetails { type_, arg }) => {
emit_inst_madcc(ctx, type_, &arg)?
crate::translate::Statement::MadCC(MadCCDetails { type_, is_hi, arg }) => {
emit_inst_madcc(ctx, type_, is_hi, &arg)?
}
crate::translate::Statement::AddC(type_, arg) => emit_inst_add_c(ctx, type_, &arg)?,
crate::translate::Statement::AddCC(type_, arg) => {
@ -2083,6 +2083,7 @@ fn emit_inst_mad_lo(
fn emit_inst_madcc(
ctx: &mut EmitContext,
type_: ast::ScalarType,
is_hi: bool,
arg: &Arg4CarryOut<ExpandedArgParams>,
) -> Result<(), TranslateError> {
let builder = ctx.builder.get();

7
ptx/src/ptx.lalrpop
View file

@ -1516,7 +1516,12 @@ InstMad: ast::Instruction<ast::ParsedArgParams<'input>> = {
// https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#extended-precision-arithmetic-instructions-mad-cc
InstMadCC: ast::Instruction<ast::ParsedArgParams<'input>> = {
"mad" ".lo" ".cc" <type_:IntType3264> <arg:Arg4> => ast::Instruction::MadCC{<>},
"mad" ".lo" ".cc" <type_:IntType3264> <arg:Arg4> => {
ast::Instruction::MadCC { type_, arg, is_hi: false }
},
"mad" ".hi" ".cc" <type_:IntType3264> <arg:Arg4> => {
ast::Instruction::MadCC { type_, arg, is_hi: true }
},
};
// https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#extended-precision-arithmetic-instructions-madc

8
ptx/src/translate.rs
View file

@ -1999,9 +1999,10 @@ fn insert_hardware_registers_impl<'input>(
is_hi,
arg: Arg4CarryIn::new(arg, carry_out, TypedOperand::Reg(overflow_flag)),
})),
Statement::Instruction(ast::Instruction::MadCC { type_, arg }) => {
Statement::Instruction(ast::Instruction::MadCC { type_, is_hi, arg }) => {
result.push(Statement::MadCC(MadCCDetails {
type_,
is_hi,
arg: Arg4CarryOut::new(arg, TypedOperand::Reg(overflow_flag)),
}))
}
@ -5568,6 +5569,7 @@ impl<T: ArgParamsEx<Id = Id>, U: ArgParamsEx<Id = Id>> Visitable<T, U> for MadCD
pub(crate) struct MadCCDetails<P: ast::ArgParams> {
pub(crate) type_: ast::ScalarType,
pub(crate) is_hi: bool,
pub(crate) arg: Arg4CarryOut<P>,
}
@ -5578,6 +5580,7 @@ impl<T: ArgParamsEx<Id = Id>, U: ArgParamsEx<Id = Id>> Visitable<T, U> for MadCC
) -> Result<Statement<ast::Instruction<U>, U>, TranslateError> {
Ok(Statement::MadCC(MadCCDetails {
type_: self.type_,
is_hi: self.is_hi,
arg: self.arg.map(visitor, self.type_)?,
}))
}
@ -6486,8 +6489,9 @@ impl<T: ArgParamsEx> ast::Instruction<T> {
carry_out,
arg: arg.map(visitor, &ast::Type::Scalar(type_), false)?,
},
ast::Instruction::MadCC { type_, arg } => ast::Instruction::MadCC {
ast::Instruction::MadCC { type_, arg, is_hi } => ast::Instruction::MadCC {
type_,
is_hi,
arg: arg.map(visitor, &ast::Type::Scalar(type_), false)?,
},
ast::Instruction::Tex(details, arg) => {