mikros/rust
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Add more SIMD intrinsics

Browse Source
This commit is contained in:
Antoni Boucher 2024-09-15 17:24:37 -04:00
parent 17f3dbf656
commit cb36d78d7b
4 changed files with 72 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
src/base.rs
View File

@ -116,6 +116,10 @@ fn module_codegen(
context.add_command_line_option("-mavx");
}
/*for feature in tcx.sess.opts.cg.target_feature.split(',') {
println!("Feature: {}", feature);
}*/
for arg in &tcx.sess.opts.cg.llvm_args {
context.add_command_line_option(arg);
}
@ -218,6 +222,7 @@ fn module_codegen(
// ... and now that we have everything pre-defined, fill out those definitions.
for &(mono_item, _) in &mono_items {
//println!("{:?}", mono_item);
mono_item.define::<Builder<'_, '_, '_>>(&cx);
}

3
src/builder.rs
View File

@ -270,6 +270,8 @@ fn check_ptr_call<'b>(
actual_val.dereference(self.location).to_rvalue()
}
} else {
// FIXME: this condition seems wrong: it will pass when both types are not
// a vector.
assert!(
(!expected_ty.is_vector() || actual_ty.is_vector())
&& (expected_ty.is_vector() || !actual_ty.is_vector()),
@ -283,6 +285,7 @@ fn check_ptr_call<'b>(
);
// TODO(antoyo): perhaps use __builtin_convertvector for vector casting.
// TODO: remove bitcast now that vector types can be compared?
println!("Name: {}", func_name);
self.bitcast(actual_val, expected_ty)
}
} else {

18
src/declare.rs
View File

@ -168,7 +168,23 @@ fn declare_raw_fn<'gcc>(
variadic: bool,
) -> Function<'gcc> {
if name.starts_with("llvm.") {
let intrinsic = llvm::intrinsic(name, cx);
let intrinsic = match name {
"llvm.fma.f16" => {
let param1 = cx.context.new_parameter(None, cx.double_type, "x");
let param2 = cx.context.new_parameter(None, cx.double_type, "y");
let param3 = cx.context.new_parameter(None, cx.double_type, "z");
cx.context.new_function(
None,
FunctionType::Extern,
cx.double_type,
&[param1, param2, param3],
"fma",
false,
)
}
_ => llvm::intrinsic(name, cx),
};
cx.intrinsics.borrow_mut().insert(name.to_string(), intrinsic);
return intrinsic;
}

50
src/intrinsic/llvm.rs
View File

@ -1,5 +1,6 @@
use std::borrow::Cow;
use gccjit::CType;
use gccjit::{Function, FunctionPtrType, RValue, ToRValue, UnaryOp};
use rustc_codegen_ssa::traits::BuilderMethods;
@ -320,7 +321,9 @@ pub fn adjust_intrinsic_arguments<'a, 'b, 'gcc, 'tcx>(
| "__builtin_ia32_vpmadd52luq512_mask"
| "__builtin_ia32_vpmadd52huq256_mask"
| "__builtin_ia32_vpmadd52luq256_mask"
| "__builtin_ia32_vpmadd52huq128_mask" => {
| "__builtin_ia32_vpmadd52huq128_mask"
| "__builtin_ia32_vfmaddsubph128_mask"
| "__builtin_ia32_vfmaddsubph256_mask" => {
let mut new_args = args.to_vec();
let arg4_type = gcc_func.get_param_type(3);
let minus_one = builder.context.new_rvalue_from_int(arg4_type, -1);
@ -440,6 +443,19 @@ pub fn adjust_intrinsic_arguments<'a, 'b, 'gcc, 'tcx>(
new_args.push(last_arg);
args = new_args.into();
}
// NOTE: the LLVM intrinsics receive 3 floats, but the GCC builtin requires 3 vectors.
"__builtin_ia32_vfmaddsh3_mask" => {
let new_args = args.to_vec();
let arg1_type = gcc_func.get_param_type(0);
let arg2_type = gcc_func.get_param_type(1);
let arg3_type = gcc_func.get_param_type(2);
let arg5_type = gcc_func.get_param_type(4);
let a = builder.context.new_rvalue_from_vector(None, arg1_type, &[new_args[0]; 8]);
let b = builder.context.new_rvalue_from_vector(None, arg2_type, &[new_args[1]; 8]);
let c = builder.context.new_rvalue_from_vector(None, arg3_type, &[new_args[2]; 8]);
let arg5 = builder.context.new_rvalue_from_int(arg5_type, 4);
args = vec![a, b, c, new_args[3], arg5].into();
}
_ => (),
}
} else {
@ -452,7 +468,7 @@ pub fn adjust_intrinsic_arguments<'a, 'b, 'gcc, 'tcx>(
let arg4 = builder.context.new_bitcast(None, new_args[2], arg4_type);
args = vec![new_args[0], new_args[1], arg3, arg4, new_args[3], new_args[5]].into();
}
// NOTE: the LLVM intrinsic receives 3 floats, but the GCC builtin requires 3 vectors.
// NOTE: the LLVM intrinsics receive 3 floats, but the GCC builtin requires 3 vectors.
// FIXME: the intrinsics like _mm_mask_fmadd_sd should probably directly call the GCC
// intrinsic to avoid this.
"__builtin_ia32_vfmaddss3_round" => {
@ -550,6 +566,25 @@ pub fn adjust_intrinsic_arguments<'a, 'b, 'gcc, 'tcx>(
]
.into();
}
"__builtin_ia32_rndscalesh_mask_round" => {
let new_args = args.to_vec();
args = vec![
new_args[0],
new_args[1],
new_args[4],
new_args[2],
new_args[3],
new_args[5],
]
.into();
}
"fma" => {
let mut new_args = args.to_vec();
new_args[0] = builder.context.new_cast(None, new_args[0], builder.double_type);
new_args[1] = builder.context.new_cast(None, new_args[1], builder.double_type);
new_args[2] = builder.context.new_cast(None, new_args[2], builder.double_type);
args = new_args.into();
}
_ => (),
}
}
@ -566,7 +601,9 @@ pub fn adjust_intrinsic_return_value<'a, 'gcc, 'tcx>(
orig_args: &[RValue<'gcc>],
) -> RValue<'gcc> {
match func_name {
"__builtin_ia32_vfmaddss3_round" | "__builtin_ia32_vfmaddsd3_round" => {
"__builtin_ia32_vfmaddss3_round"
| "__builtin_ia32_vfmaddsd3_round"
| "__builtin_ia32_vfmaddsh3_mask" => {
#[cfg(feature = "master")]
{
let zero = builder.context.new_rvalue_zero(builder.int_type);
@ -625,6 +662,10 @@ pub fn adjust_intrinsic_return_value<'a, 'gcc, 'tcx>(
&[random_number, success_variable.to_rvalue()],
);
}
"fma" => {
let f16_type = builder.context.new_c_type(CType::Float16);
return_value = builder.context.new_cast(None, return_value, f16_type);
}
_ => (),
}
@ -1165,6 +1206,9 @@ pub fn intrinsic<'gcc, 'tcx>(name: &str, cx: &CodegenCx<'gcc, 'tcx>) -> Function
"llvm.x86.avx512.mask.store.q.128" => "__builtin_ia32_movdqa64store128_mask",
"llvm.x86.avx512.mask.store.ps.128" => "__builtin_ia32_storeaps128_mask",
"llvm.x86.avx512.mask.store.pd.128" => "__builtin_ia32_storeapd128_mask",
"llvm.x86.avx512fp16.vfmadd.f16" => "__builtin_ia32_vfmaddsh3_mask",
"llvm.x86.avx512fp16.vfmaddsub.ph.128" => "__builtin_ia32_vfmaddsubph128_mask",
"llvm.x86.avx512fp16.vfmaddsub.ph.256" => "__builtin_ia32_vfmaddsubph256_mask",
// TODO: support the tile builtins:
"llvm.x86.ldtilecfg" => "__builtin_trap",