1
0
forked from M-Labs/nac3

core/ndstrides: add NDArrayOut, broadcast_map and map

This commit is contained in:
lyken 2024-08-20 17:03:23 +08:00 committed by David Mak
parent 5bed394ef7
commit fbfc0b293a
2 changed files with 270 additions and 0 deletions

View File

@ -0,0 +1,219 @@
use inkwell::values::BasicValueEnum;
use itertools::Itertools;
use crate::{
codegen::{
object::ndarray::{AnyObject, NDArrayObject},
stmt::gen_for_callback,
CodeGenContext, CodeGenerator,
},
typecheck::typedef::Type,
};
use super::{nditer::NDIterHandle, NDArrayOut, ScalarOrNDArray};
impl<'ctx> NDArrayObject<'ctx> {
/// Generate LLVM IR to broadcast `ndarray`s together, and starmap through them with `mapping` elementwise.
///
/// `mapping` is an LLVM IR generator. The input of `mapping` is the list of elements when iterating through
/// the input `ndarrays` after broadcasting. The output of `mapping` is the result of the elementwise operation.
///
/// `out` specifies whether the result should be a new ndarray or to be written an existing ndarray.
pub fn broadcast_starmap<'a, G, MappingFn>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, 'a>,
ndarrays: &[Self],
out: NDArrayOut<'ctx>,
mapping: MappingFn,
) -> Result<Self, String>
where
G: CodeGenerator + ?Sized,
MappingFn: FnOnce(
&mut G,
&mut CodeGenContext<'ctx, 'a>,
&[BasicValueEnum<'ctx>],
) -> Result<BasicValueEnum<'ctx>, String>,
{
// Broadcast inputs
let broadcast_result = NDArrayObject::broadcast(generator, ctx, ndarrays);
let out_ndarray = match out {
NDArrayOut::NewNDArray { dtype } => {
// Create a new ndarray based on the broadcast shape.
let result_ndarray =
NDArrayObject::alloca(generator, ctx, dtype, broadcast_result.ndims);
result_ndarray.copy_shape_from_array(generator, ctx, broadcast_result.shape);
result_ndarray.create_data(generator, ctx);
result_ndarray
}
NDArrayOut::WriteToNDArray { ndarray: result_ndarray } => {
// Use an existing ndarray.
// Check that its shape is compatible with the broadcast shape.
result_ndarray.assert_can_be_written_by_out(
generator,
ctx,
broadcast_result.ndims,
broadcast_result.shape,
);
result_ndarray
}
};
// Map element-wise and store results into `mapped_ndarray`.
let nditer = NDIterHandle::new(generator, ctx, out_ndarray);
gen_for_callback(
generator,
ctx,
Some("broadcast_starmap"),
|generator, ctx| {
// Create NDIters for all broadcasted input ndarrays.
let other_nditers = broadcast_result
.ndarrays
.iter()
.map(|ndarray| NDIterHandle::new(generator, ctx, *ndarray))
.collect_vec();
Ok((nditer, other_nditers))
},
|generator, ctx, (out_nditer, _in_nditers)| {
// We can simply use `out_nditer`'s `has_element()`.
// `in_nditers`' `has_element()`s should return the same value.
Ok(out_nditer.has_element(generator, ctx).value)
},
|generator, ctx, _hooks, (out_nditer, in_nditers)| {
// Get all the scalars from the broadcasted input ndarrays, pass them to `mapping`,
// and write to `out_ndarray`.
let in_scalars = in_nditers
.iter()
.map(|nditer| nditer.get_scalar(generator, ctx).value)
.collect_vec();
let result = mapping(generator, ctx, &in_scalars)?;
let p = out_nditer.get_pointer(generator, ctx);
ctx.builder.build_store(p, result).unwrap();
Ok(())
},
|generator, ctx, (out_nditer, in_nditers)| {
// Advance all iterators
out_nditer.next(generator, ctx);
in_nditers.iter().for_each(|nditer| nditer.next(generator, ctx));
Ok(())
},
)?;
Ok(out_ndarray)
}
/// Map through this ndarray with an elementwise function.
pub fn map<'a, G, Mapping>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, 'a>,
out: NDArrayOut<'ctx>,
mapping: Mapping,
) -> Result<Self, String>
where
G: CodeGenerator + ?Sized,
Mapping: FnOnce(
&mut G,
&mut CodeGenContext<'ctx, 'a>,
BasicValueEnum<'ctx>,
) -> Result<BasicValueEnum<'ctx>, String>,
{
NDArrayObject::broadcast_starmap(
generator,
ctx,
&[*self],
out,
|generator, ctx, scalars| mapping(generator, ctx, scalars[0]),
)
}
}
impl<'ctx> ScalarOrNDArray<'ctx> {
/// Starmap through a list of inputs using `mapping`, where an input could be an ndarray, a scalar.
///
/// This function is very helpful when implementing NumPy functions that takes on either scalars or ndarrays or a mix of them
/// as their inputs and produces either an ndarray with broadcast, or a scalar if all its inputs are all scalars.
///
/// For example ,this function can be used to implement `np.add`, which has the following behaviors:
/// - `np.add(3, 4) = 7` # (scalar, scalar) -> scalar
/// - `np.add(3, np.array([4, 5, 6]))` # (scalar, ndarray) -> ndarray; the first `scalar` is converted into an ndarray and broadcasted.
/// - `np.add(np.array([[1], [2], [3]]), np.array([[4, 5, 6]]))` # (ndarray, ndarray) -> ndarray; there is broadcasting.
///
/// ## Details:
///
/// If `inputs` are all [`ScalarOrNDArray::Scalar`], the output will be a [`ScalarOrNDArray::Scalar`] with type `ret_dtype`.
///
/// Otherwise (if there are any [`ScalarOrNDArray::NDArray`] in `inputs`), all inputs will be 'as-ndarray'-ed into ndarrays,
/// then all inputs (now all ndarrays) will be passed to [`NDArrayObject::broadcasting_starmap`] and **create** a new ndarray
/// with dtype `ret_dtype`.
pub fn broadcasting_starmap<'a, G, MappingFn>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, 'a>,
inputs: &[ScalarOrNDArray<'ctx>],
ret_dtype: Type,
mapping: MappingFn,
) -> Result<ScalarOrNDArray<'ctx>, String>
where
G: CodeGenerator + ?Sized,
MappingFn: FnOnce(
&mut G,
&mut CodeGenContext<'ctx, 'a>,
&[BasicValueEnum<'ctx>],
) -> Result<BasicValueEnum<'ctx>, String>,
{
// Check if all inputs are Scalars
let all_scalars: Option<Vec<_>> = inputs.iter().map(AnyObject::try_from).try_collect().ok();
if let Some(scalars) = all_scalars {
let scalars = scalars.iter().map(|scalar| scalar.value).collect_vec();
let value = mapping(generator, ctx, &scalars)?;
Ok(ScalarOrNDArray::Scalar(AnyObject { ty: ret_dtype, value }))
} else {
// Promote all input to ndarrays and map through them.
let inputs = inputs.iter().map(|input| input.to_ndarray(generator, ctx)).collect_vec();
let ndarray = NDArrayObject::broadcast_starmap(
generator,
ctx,
&inputs,
NDArrayOut::NewNDArray { dtype: ret_dtype },
mapping,
)?;
Ok(ScalarOrNDArray::NDArray(ndarray))
}
}
/// Map through this [`ScalarOrNDArray`] with an elementwise function.
///
/// If this is a scalar, `mapping` will directly act on the scalar. This function will return a [`ScalarOrNDArray::Scalar`] of that result.
///
/// If this is an ndarray, `mapping` will be applied to the elements of the ndarray. A new ndarray of the results will be created and
/// returned as a [`ScalarOrNDArray::NDArray`].
pub fn map<'a, G, Mapping>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, 'a>,
ret_dtype: Type,
mapping: Mapping,
) -> Result<ScalarOrNDArray<'ctx>, String>
where
G: CodeGenerator + ?Sized,
Mapping: FnOnce(
&mut G,
&mut CodeGenContext<'ctx, 'a>,
BasicValueEnum<'ctx>,
) -> Result<BasicValueEnum<'ctx>, String>,
{
ScalarOrNDArray::broadcasting_starmap(
generator,
ctx,
&[*self],
ret_dtype,
|generator, ctx, scalars| mapping(generator, ctx, scalars[0]),
)
}
}

View File

@ -13,6 +13,7 @@ use crate::{
call_nac3_ndarray_get_pelement_by_indices, call_nac3_ndarray_is_c_contiguous,
call_nac3_ndarray_len, call_nac3_ndarray_nbytes,
call_nac3_ndarray_set_strides_by_shape, call_nac3_ndarray_size,
call_nac3_ndarray_util_assert_output_shape_same,
},
model::*,
CodeGenContext, CodeGenerator,
@ -25,6 +26,7 @@ pub mod array;
pub mod broadcast;
pub mod factory;
pub mod indexing;
pub mod map;
pub mod nditer;
pub mod shape_util;
pub mod view;
@ -499,6 +501,31 @@ impl<'ctx> NDArrayObject<'ctx> {
ndarray.instance.set(ctx, |f| f.data, data);
ndarray
}
/// Check if this `NDArray` can be used as an `out` ndarray for an operation.
///
/// Raise an exception if the shapes do not match.
pub fn assert_can_be_written_by_out<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
out_ndims: u64,
out_shape: Instance<'ctx, Ptr<Int<SizeT>>>,
) {
let ndarray_ndims = self.ndims_llvm(generator, ctx.ctx);
let ndarray_shape = self.instance.get(generator, ctx, |f| f.shape);
let output_ndims = Int(SizeT).const_int(generator, ctx.ctx, out_ndims, false);
let output_shape = out_shape;
call_nac3_ndarray_util_assert_output_shape_same(
generator,
ctx,
ndarray_ndims,
ndarray_shape,
output_ndims,
output_shape,
);
}
}
/// A convenience enum for implementing functions that acts on scalars or ndarrays or both.
@ -584,3 +611,27 @@ impl<'ctx> ScalarOrNDArray<'ctx> {
}
}
}
/// An helper enum specifying how a function should produce its output.
///
/// Many functions in NumPy has an optional `out` parameter (e.g., `matmul`). If `out` is specified
/// with an ndarray, the result of a function will be written to `out`. If `out` is not specified, a function will
/// create a new ndarray and store the result in it.
#[derive(Debug, Clone, Copy)]
pub enum NDArrayOut<'ctx> {
/// Tell a function should create a new ndarray with the expected element type `dtype`.
NewNDArray { dtype: Type },
/// Tell a function to write the result to `ndarray`.
WriteToNDArray { ndarray: NDArrayObject<'ctx> },
}
impl<'ctx> NDArrayOut<'ctx> {
/// Get the dtype of this output.
#[must_use]
pub fn get_dtype(&self) -> Type {
match self {
NDArrayOut::NewNDArray { dtype } => *dtype,
NDArrayOut::WriteToNDArray { ndarray } => ndarray.dtype,
}
}
}