forked from M-Labs/nac3
1
0
Fork 0

core: implement np_dot using LLVM_IR

This commit is contained in:
abdul124 2024-07-31 15:53:51 +08:00
parent 4a6845dac6
commit 54f883f0a5
7 changed files with 165 additions and 105 deletions

View File

@ -1865,34 +1865,6 @@ fn build_output_struct<'ctx>(
out_ptr
}
/// Invokes the `np_dot` linalg function
pub fn call_np_dot<'ctx, G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
x1: (Type, BasicValueEnum<'ctx>),
x2: (Type, BasicValueEnum<'ctx>),
) -> Result<BasicValueEnum<'ctx>, String> {
const FN_NAME: &str = "np_dot";
let (x1_ty, x1) = x1;
let (x2_ty, x2) = x2;
if let (BasicValueEnum::PointerValue(_), BasicValueEnum::PointerValue(_)) = (x1, x2) {
let (n1_elem_ty, _) = unpack_ndarray_var_tys(&mut ctx.unifier, x1_ty);
let n1_elem_ty = ctx.get_llvm_type(generator, n1_elem_ty);
let (n2_elem_ty, _) = unpack_ndarray_var_tys(&mut ctx.unifier, x2_ty);
let n2_elem_ty = ctx.get_llvm_type(generator, n2_elem_ty);
let (BasicTypeEnum::FloatType(_), BasicTypeEnum::FloatType(_)) = (n1_elem_ty, n2_elem_ty)
else {
unsupported_type(ctx, FN_NAME, &[x1_ty, x2_ty]);
};
Ok(extern_fns::call_np_dot(ctx, x1, x2, None).into())
} else {
unsupported_type(ctx, FN_NAME, &[x1_ty, x2_ty])
}
}
/// Invokes the `np_linalg_matmul` linalg function
pub fn call_np_linalg_matmul<'ctx, G: CodeGenerator + ?Sized>(
generator: &mut G,

View File

@ -188,33 +188,3 @@ generate_linalg_extern_fn!(call_np_linalg_pinv, "np_linalg_pinv", 2);
generate_linalg_extern_fn!(call_sp_linalg_lu, "sp_linalg_lu", 3);
generate_linalg_extern_fn!(call_sp_linalg_schur, "sp_linalg_schur", 3);
generate_linalg_extern_fn!(call_sp_linalg_hessenberg, "sp_linalg_hessenberg", 3);
/// Invokes the linalg `np_dot` function.
pub fn call_np_dot<'ctx>(
ctx: &mut CodeGenContext<'ctx, '_>,
mat1: BasicValueEnum<'ctx>,
mat2: BasicValueEnum<'ctx>,
name: Option<&str>,
) -> FloatValue<'ctx> {
const FN_NAME: &str = "np_dot";
let extern_fn = ctx.module.get_function(FN_NAME).unwrap_or_else(|| {
let fn_type =
ctx.ctx.f64_type().fn_type(&[mat1.get_type().into(), mat2.get_type().into()], false);
let func = ctx.module.add_function(FN_NAME, fn_type, None);
for attr in ["mustprogress", "nofree", "nounwind", "willreturn", "writeonly"] {
func.add_attribute(
AttributeLoc::Function,
ctx.ctx.create_enum_attribute(Attribute::get_named_enum_kind_id(attr), 0),
);
}
func
});
ctx.builder
.build_call(extern_fn, &[mat1.into(), mat2.into()], name.unwrap_or_default())
.map(CallSiteValue::try_as_basic_value)
.map(|v| v.map_left(BasicValueEnum::into_float_value))
.map(Either::unwrap_left)
.unwrap()
}

View File

@ -26,12 +26,15 @@ use crate::{
typedef::{FunSignature, Type, TypeEnum},
},
};
use inkwell::types::{AnyTypeEnum, BasicTypeEnum, PointerType};
use inkwell::{
types::BasicType,
values::{BasicValueEnum, IntValue, PointerValue},
AddressSpace, IntPredicate, OptimizationLevel,
};
use inkwell::{
types::{AnyTypeEnum, BasicTypeEnum, PointerType},
values::BasicValue,
};
use nac3parser::ast::{Operator, StrRef};
/// Creates an uninitialized `NDArray` instance.
@ -2390,7 +2393,7 @@ pub fn ndarray_reshape<'ctx, G: CodeGenerator + ?Sized>(
generator,
ctx.builder.build_int_compare(IntPredicate::EQ, out_sz, n_sz, "").unwrap(),
"0:ValueError",
"cannot reshape array of size {} into provided shape of size {}",
"cannot reshape array of size {0} into provided shape of size {1}",
[Some(n_sz), Some(out_sz), None],
ctx.current_loc,
);
@ -2417,3 +2420,102 @@ pub fn ndarray_reshape<'ctx, G: CodeGenerator + ?Sized>(
)
}
}
/// Generates LLVM IR for `ndarray.dot`.
/// Calculate inner product of two vectors or literals
/// For matrix multiplication use `np_matmul`
///
/// The input `NDArray` are flattened and treated as 1D
/// The operation is equivalent to np.dot(arr1.ravel(), arr2.ravel())
pub fn ndarray_dot<'ctx, G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
x1: (Type, BasicValueEnum<'ctx>),
x2: (Type, BasicValueEnum<'ctx>),
) -> Result<BasicValueEnum<'ctx>, String> {
const FN_NAME: &str = "ndarray_dot";
let (x1_ty, x1) = x1;
let (_, x2) = x2;
let llvm_usize = generator.get_size_type(ctx.ctx);
match (x1, x2) {
(BasicValueEnum::PointerValue(n1), BasicValueEnum::PointerValue(n2)) => {
let n1 = NDArrayValue::from_ptr_val(n1, llvm_usize, None);
let n2 = NDArrayValue::from_ptr_val(n2, llvm_usize, None);
let n1_sz = call_ndarray_calc_size(generator, ctx, &n1.dim_sizes(), (None, None));
let n2_sz = call_ndarray_calc_size(generator, ctx, &n1.dim_sizes(), (None, None));
ctx.make_assert(
generator,
ctx.builder.build_int_compare(IntPredicate::EQ, n1_sz, n2_sz, "").unwrap(),
"0:ValueError",
"shapes ({0}), ({1}) not aligned",
[Some(n1_sz), Some(n2_sz), None],
ctx.current_loc,
);
let identity =
unsafe { n1.data().get_unchecked(ctx, generator, &llvm_usize.const_zero(), None) };
let acc = ctx.builder.build_alloca(identity.get_type(), "").unwrap();
ctx.builder.build_store(acc, identity.get_type().const_zero()).unwrap();
gen_for_callback_incrementing(
generator,
ctx,
None,
llvm_usize.const_zero(),
(n1_sz, false),
|generator, ctx, _, idx| {
let elem1 = unsafe { n1.data().get_unchecked(ctx, generator, &idx, None) };
let elem2 = unsafe { n2.data().get_unchecked(ctx, generator, &idx, None) };
let product = match elem1 {
BasicValueEnum::IntValue(e1) => ctx
.builder
.build_int_mul(e1, elem2.into_int_value(), "")
.unwrap()
.as_basic_value_enum(),
BasicValueEnum::FloatValue(e1) => ctx
.builder
.build_float_mul(e1, elem2.into_float_value(), "")
.unwrap()
.as_basic_value_enum(),
_ => unreachable!(),
};
let acc_val = ctx.builder.build_load(acc, "").unwrap();
let acc_val = match acc_val {
BasicValueEnum::IntValue(e1) => ctx
.builder
.build_int_add(e1, product.into_int_value(), "")
.unwrap()
.as_basic_value_enum(),
BasicValueEnum::FloatValue(e1) => ctx
.builder
.build_float_add(e1, product.into_float_value(), "")
.unwrap()
.as_basic_value_enum(),
_ => unreachable!(),
};
ctx.builder.build_store(acc, acc_val).unwrap();
Ok(())
},
llvm_usize.const_int(1, false),
)?;
let acc_val = ctx.builder.build_load(acc, "").unwrap();
Ok(acc_val)
}
(BasicValueEnum::IntValue(e1), BasicValueEnum::IntValue(e2)) => {
Ok(ctx.builder.build_int_mul(e1, e2, "").unwrap().as_basic_value_enum())
}
(BasicValueEnum::FloatValue(e1), BasicValueEnum::FloatValue(e2)) => {
Ok(ctx.builder.build_float_mul(e1, e2, "").unwrap().as_basic_value_enum())
}
_ => unreachable!(
"{FN_NAME}() not supported for '{}'",
format!("'{}'", ctx.unifier.stringify(x1_ty))
),
}
}

View File

@ -1965,7 +1965,7 @@ impl<'a> BuiltinBuilder<'a> {
self.unifier,
&self.num_or_ndarray_var_map,
prim.name(),
self.primitives.float,
self.num_ty.ty,
&[(self.num_or_ndarray_ty.ty, "x1"), (self.num_or_ndarray_ty.ty, "x2")],
Box::new(move |ctx, _, fun, args, generator| {
let x1_ty = fun.0.args[0].ty;
@ -1973,12 +1973,7 @@ impl<'a> BuiltinBuilder<'a> {
let x2_ty = fun.0.args[1].ty;
let x2_val = args[1].1.clone().to_basic_value_enum(ctx, generator, x2_ty)?;
Ok(Some(builtin_fns::call_np_dot(
generator,
ctx,
(x1_ty, x1_val),
(x2_ty, x2_val),
)?))
Ok(Some(ndarray_dot(generator, ctx, (x1_ty, x1_val), (x2_ty, x2_val))?))
}),
),

View File

@ -1130,6 +1130,44 @@ impl<'a> Inferencer<'a> {
}));
}
if id == &"np_dot".into() {
let arg0 = self.fold_expr(args.remove(0))?;
let arg1 = self.fold_expr(args.remove(0))?;
let arg0_ty = arg0.custom.unwrap();
let ret = if arg0_ty.obj_id(self.unifier).is_some_and(|id| id == PrimDef::NDArray.id())
{
let (ndarray_dtype, _) = unpack_ndarray_var_tys(self.unifier, arg0_ty);
ndarray_dtype
} else {
arg0_ty
};
let custom = self.unifier.add_ty(TypeEnum::TFunc(FunSignature {
args: vec![
FuncArg { name: "x1".into(), ty: arg0.custom.unwrap(), default_value: None },
FuncArg { name: "x2".into(), ty: arg1.custom.unwrap(), default_value: None },
],
ret,
vars: VarMap::new(),
}));
return Ok(Some(Located {
location,
custom: Some(ret),
node: ExprKind::Call {
func: Box::new(Located {
custom: Some(custom),
location: func.location,
node: ExprKind::Name { id: *id, ctx: *ctx },
}),
args: vec![arg0, arg1],
keywords: vec![],
},
}));
}
if ["np_min", "np_max"].iter().any(|fun_id| id == &(*fun_id).into()) && args.len() == 1 {
let arg0 = self.fold_expr(args.remove(0))?;
let arg0_ty = arg0.custom.unwrap();

View File

@ -34,38 +34,6 @@ impl InputMatrix {
}
}
/// # Safety
///
/// `mat1` and `mat2` should point to a valid 1DArray of `f64` floats in row-major order
#[no_mangle]
pub unsafe extern "C" fn np_dot(mat1: *mut InputMatrix, mat2: *mut InputMatrix) -> f64 {
let mat1 = mat1.as_mut().unwrap();
let mat2 = mat2.as_mut().unwrap();
if !(mat1.ndims == 1 && mat2.ndims == 1) {
let err_msg = format!(
"expected 1D Vector Input, but received {}D and {}D input",
mat1.ndims, mat2.ndims
);
report_error("ValueError", "np_dot", file!(), line!(), column!(), &err_msg);
}
let dim1 = (*mat1).get_dims();
let dim2 = (*mat2).get_dims();
if dim1[0] != dim2[0] {
let err_msg = format!("shapes ({},) and ({},) not aligned", dim1[0], dim2[0]);
report_error("ValueError", "np_dot", file!(), line!(), column!(), &err_msg);
}
let data_slice1 = unsafe { slice::from_raw_parts_mut(mat1.data, dim1[0]) };
let data_slice2 = unsafe { slice::from_raw_parts_mut(mat2.data, dim2[0]) };
let matrix1 = DMatrix::from_row_slice(dim1[0], 1, data_slice1);
let matrix2 = DMatrix::from_row_slice(dim2[0], 1, data_slice2);
matrix1.dot(&matrix2)
}
/// # Safety
///
/// `mat1` and `mat2` should point to a valid 2DArray of `f64` floats in row-major order

View File

@ -1451,13 +1451,28 @@ def test_ndarray_reshape():
output_ndarray_float_1(z)
def test_ndarray_dot():
x: ndarray[float, 1] = np_array([5.0, 1.0])
y: ndarray[float, 1] = np_array([5.0, 1.0])
z = np_dot(x, y)
x1: ndarray[float, 1] = np_array([5.0, 1.0, 4.0, 2.0])
y1: ndarray[float, 1] = np_array([5.0, 1.0, 6.0, 6.0])
z1 = np_dot(x1, y1)
output_ndarray_float_1(x)
output_ndarray_float_1(y)
output_float64(z)
x2: ndarray[int32, 1] = np_array([5, 1, 4, 2])
y2: ndarray[int32, 1] = np_array([5, 1, 6, 6])
z2 = np_dot(x2, y2)
x3: ndarray[bool, 1] = np_array([True, True, True, True])
y3: ndarray[bool, 1] = np_array([True, True, True, True])
z3 = np_dot(x3, y3)
z4 = np_dot(2, 3)
z5 = np_dot(2., 3.)
z6 = np_dot(True, False)
output_float64(z1)
output_int32(z2)
output_bool(z3)
output_int32(z4)
output_float64(z5)
output_bool(z6)
def test_ndarray_linalg_matmul():
x: ndarray[float, 2] = np_array([[5.0, 1.0], [1.0, 4.0]])