lyken/nac3
1
0
Fork 0
forked from M-Labs/nac3
Code Pull Requests Activity

core/ndstrides: introduce NDArrayObject & refactor reshape

Browse Source
This commit is contained in:
lyken 2024-07-30 15:30:34 +08:00
parent 7436513b64
commit 3dc4b17310
5 changed files with 326 additions and 130 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
nac3core/src/codegen/numpy_new/mod.rs
View File

@ -1,3 +1,4 @@
pub mod factory;
pub mod object;
pub mod util;
pub mod view;

68
nac3core/src/codegen/numpy_new/object.rs Normal file
View File

@ -0,0 +1,68 @@
use inkwell::values::{BasicValue, BasicValueEnum};
use crate::{
codegen::{model::*, structure::ndarray::NpArray, CodeGenContext},
toplevel::numpy::unpack_ndarray_var_tys,
typecheck::typedef::{Type, TypeEnum},
};
/// An LLVM ndarray instance with its typechecker [`Type`]s.
#[derive(Debug, Clone, Copy)]
pub struct NDArrayObject<'ctx> {
pub dtype: Type,
pub ndims: Type,
pub instance: Ptr<'ctx, StructModel<NpArray>>,
}
/// An LLVM numpy scalar with its [`Type`].
#[derive(Debug, Clone, Copy)]
pub struct ScalarObject<'ctx> {
pub dtype: Type,
pub value: BasicValueEnum<'ctx>,
}
#[derive(Debug, Clone, Copy)]
pub enum ScalarOrNDArray<'ctx> {
Scalar(ScalarObject<'ctx>),
NDArray(NDArrayObject<'ctx>),
}
impl<'ctx> ScalarOrNDArray<'ctx> {
/// Get the underlying [`BasicValueEnum<'ctx>`] of this [`ScalarOrNDArray`].
fn to_basic_value_enum(self) -> BasicValueEnum<'ctx> {
match self {
ScalarOrNDArray::Scalar(scalar) => scalar.value,
ScalarOrNDArray::NDArray(ndarray) => ndarray.instance.value.as_basic_value_enum(),
}
}
}
impl<'ctx> From<ScalarOrNDArray<'ctx>> for BasicValueEnum<'ctx> {
fn from(input: ScalarOrNDArray<'ctx>) -> BasicValueEnum<'ctx> {
input.to_basic_value_enum()
}
}
/// Split an [`BasicValueEnum<'ctx>`] into a [`ScalarOrNDArray`] depending
/// on its [`Type`].
pub fn split_scalar_or_ndarray<'ctx>(
tyctx: TypeContext<'ctx>,
ctx: &mut CodeGenContext<'ctx, '_>,
input: BasicValueEnum<'ctx>,
input_ty: Type,
) -> ScalarOrNDArray<'ctx> {
let pndarray_model = PtrModel(StructModel(NpArray));
let input_ty_enum = ctx.unifier.get_ty(input_ty);
match &*input_ty_enum {
TypeEnum::TObj { obj_id, .. }
if *obj_id == ctx.primitives.ndarray.obj_id(&ctx.unifier).unwrap() =>
{
let value = pndarray_model.check_value(tyctx, ctx.ctx, input).unwrap();
let (dtype, ndims) = unpack_ndarray_var_tys(&mut ctx.unifier, input_ty);
ScalarOrNDArray::NDArray(NDArrayObject { dtype, ndims, instance: value })
}
_ => ScalarOrNDArray::Scalar(ScalarObject { dtype: input_ty, value: input }),
}
}

219
nac3core/src/codegen/numpy_new/util.rs
View File

@ -1,9 +1,11 @@
use inkwell::{types::BasicType, values::BasicValueEnum};
use inkwell::types::BasicType;
use util::gen_model_memcpy;
use crate::{
codegen::{
irrt::ndarray::basic::{
call_nac3_ndarray_get_nth_pelement, call_nac3_ndarray_nbytes,
call_nac3_ndarray_copy_data, call_nac3_ndarray_get_nth_pelement,
call_nac3_ndarray_is_c_contiguous, call_nac3_ndarray_nbytes,
call_nac3_ndarray_set_strides_by_shape, call_nac3_ndarray_size,
call_nac3_ndarray_util_assert_shape_no_negative,
},
@ -13,10 +15,31 @@ use crate::{
util::{array_writer::ArrayWriter, control::gen_model_for},
CodeGenContext, CodeGenerator,
},
toplevel::numpy::unpack_ndarray_var_tys,
typecheck::typedef::{Type, TypeEnum},
symbol_resolver::SymbolValue,
typecheck::typedef::{Type, TypeEnum, Unifier},
};
use super::object::{NDArrayObject, ScalarOrNDArray};
/// Extract an ndarray's `ndims` [type][`Type`] in `u64`. Panic if not possible.
#[must_use]
pub fn extract_ndims(unifier: &Unifier, ndims_ty: Type) -> u64 {
let ndims_ty_enum = unifier.get_ty_immutable(ndims_ty);
let TypeEnum::TLiteral { values, .. } = &*ndims_ty_enum else {
panic!("ndims_ty should be a TLiteral");
};
assert_eq!(values.len(), 1, "ndims_ty TLiteral should only contain 1 value");
let ndims = values[0].clone();
u64::try_from(ndims).unwrap()
}
/// Return an ndarray's `ndims` as a typechecker [`Type`] from its `u64` value.
pub fn create_ndims(unifier: &mut Unifier, ndims: u64) -> Type {
unifier.get_fresh_literal(vec![SymbolValue::U64(ndims)], None)
}
/// Allocate an ndarray on the stack given its `ndims`.
///
/// `shape` and `strides` will be automatically allocated on the stack.
@ -90,56 +113,6 @@ pub fn init_ndarray_data_by_alloca<'ctx, G: CodeGenerator + ?Sized>(
call_nac3_ndarray_set_strides_by_shape(generator, ctx, pndarray);
}
/// Convert `input` to an ndarray - behaves similarly to `np.asarray`.
///
/// Returns the ndarray interpretation of `input` and **the element type** of the ndarray.
///
/// Here are the exact details:
/// - If `input` is an ndarray, the function returns back the **same** ndarray and the `dtype`
/// of the ndarray.
/// - If `input` is not an ndarray, the function creates an ndarray with a single element `input`,
/// and returns the created ndarray and `input_ty`. Note that the created ndarray's `ndims` will
/// be `0` (an *unsized* ndarray).
pub fn as_ndarray<'ctx, G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
input: BasicValueEnum<'ctx>,
input_ty: Type,
) -> (Ptr<'ctx, StructModel<NpArray>>, Type) {
let tyctx = generator.type_context(ctx.ctx);
let sizet_model = IntModel(SizeT);
let pbyte_model = PtrModel(IntModel(Byte));
let pndarray_model = PtrModel(StructModel(NpArray));
let input_ty_enum = ctx.unifier.get_ty(input_ty);
match &*input_ty_enum {
TypeEnum::TObj { obj_id, .. }
if *obj_id == ctx.primitives.ndarray.obj_id(&ctx.unifier).unwrap() =>
{
let pndarray = pndarray_model.check_value(tyctx, ctx.ctx, input).unwrap();
let (elem_ty, _) = unpack_ndarray_var_tys(&mut ctx.unifier, input_ty);
(pndarray, elem_ty)
}
_ => {
let ndims = sizet_model.const_0(tyctx, ctx.ctx);
let pndarray = alloca_ndarray(generator, ctx, ndims, "ndarray");
// We have to put `input` onto the stack to get a data pointer.
let data = ctx.builder.build_alloca(input.get_type(), "as_ndarray_scalar").unwrap();
ctx.builder.build_store(data, input).unwrap();
let data = pbyte_model.transmute(tyctx, ctx, data, "data");
pndarray.gep(ctx, |f| f.data).store(ctx, data);
let itemsize = input.get_type().size_of().unwrap();
let itemsize = sizet_model.check_value(tyctx, ctx.ctx, itemsize).unwrap();
pndarray.gep(ctx, |f| f.itemsize).store(ctx, itemsize);
(pndarray, input_ty)
}
}
}
/// Iterate through all elements in an ndarray.
///
/// `body` is given the index of an element and an opaque pointer (as an `uint8_t*`, you might want to cast it) to the element.
@ -180,3 +153,141 @@ where
},
)
}
impl<'ctx> ScalarOrNDArray<'ctx> {
/// Convert `input` to an ndarray - behaves like `np.asarray`.
pub fn as_ndarray<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
) -> NDArrayObject<'ctx> {
match self {
ScalarOrNDArray::NDArray(ndarray) => *ndarray,
ScalarOrNDArray::Scalar(scalar) => {
let tyctx = generator.type_context(ctx.ctx);
let pbyte_model = PtrModel(IntModel(Byte));
// We have to put the value on the stack to get a data pointer.
let data =
ctx.builder.build_alloca(scalar.value.get_type(), "as_ndarray_scalar").unwrap();
ctx.builder.build_store(data, scalar.value).unwrap();
let data = pbyte_model.transmute(tyctx, ctx, data, "data");
let ndims_ty = create_ndims(&mut ctx.unifier, 0);
let ndarray = NDArrayObject::alloca(
generator,
ctx,
ndims_ty,
scalar.dtype,
"scalar_as_ndarray",
);
ndarray.instance.gep(ctx, |f| f.data).store(ctx, data);
// No need to initialize/setup strides or shapes - because `ndims` is 0.
// So we only have to set `data`, `itemsize`, and `ndims = 0`.
ndarray
}
}
}
}
impl<'ctx> NDArrayObject<'ctx> {
pub fn alloca<G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
ndims: Type,
dtype: Type,
name: &str,
) -> Self {
let tyctx = generator.type_context(ctx.ctx);
let sizet_model = IntModel(SizeT);
let ndims_int = sizet_model.constant(tyctx, ctx.ctx, extract_ndims(&ctx.unifier, ndims));
let instance = alloca_ndarray(generator, ctx, ndims_int, name);
// Set itemsize
let dtype_ty = ctx.get_llvm_type(generator, dtype);
let itemsize = dtype_ty.size_of().unwrap();
let itemsize = sizet_model.s_extend_or_bit_cast(tyctx, ctx, itemsize, "itemsize");
instance.gep(ctx, |f| f.itemsize).store(ctx, itemsize);
NDArrayObject { dtype, ndims, instance }
}
pub fn copy_shape<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
src_shape: Ptr<'ctx, IntModel<SizeT>>,
) {
let tyctx = generator.type_context(ctx.ctx);
let sizet_model = IntModel(SizeT);
let self_shape = self.instance.gep(ctx, |f| f.shape).load(tyctx, ctx, "self_shape");
let ndims_int =
sizet_model.constant(tyctx, ctx.ctx, extract_ndims(&ctx.unifier, self.ndims));
gen_model_memcpy(tyctx, ctx, self_shape, src_shape, ndims_int.value, false);
}
pub fn copy_shape_from<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
src_ndarray: NDArrayObject<'ctx>,
) {
let tyctx = generator.type_context(ctx.ctx);
let src_shape = src_ndarray.instance.gep(ctx, |f| f.shape).load(tyctx, ctx, "src_shape");
self.copy_shape(generator, ctx, src_shape);
}
pub fn update_strides_by_shape<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
) {
call_nac3_ndarray_set_strides_by_shape(generator, ctx, self.instance);
}
pub fn size<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
) -> Int<'ctx, SizeT> {
call_nac3_ndarray_size(generator, ctx, self.instance)
}
pub fn nbytes<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
) -> Int<'ctx, SizeT> {
call_nac3_ndarray_nbytes(generator, ctx, self.instance)
}
pub fn is_c_contiguous<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
) -> Int<'ctx, Bool> {
call_nac3_ndarray_is_c_contiguous(generator, ctx, self.instance)
}
pub fn alloca_owned_data<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
) {
init_ndarray_data_by_alloca(generator, ctx, self.instance);
}
pub fn copy_data_from<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
src: NDArrayObject<'ctx>,
) {
assert!(ctx.unifier.unioned(self.dtype, src.dtype), "self and src dtype should match");
call_nac3_ndarray_copy_data(generator, ctx, src.instance, self.instance);
}
}

151
nac3core/src/codegen/numpy_new/view.rs
View File

@ -3,90 +3,67 @@ use nac3parser::ast::StrRef;
use crate::{
codegen::{
irrt::ndarray::{
basic::{
call_nac3_ndarray_copy_data, call_nac3_ndarray_is_c_contiguous,
call_nac3_ndarray_nbytes, call_nac3_ndarray_set_strides_by_shape,
call_nac3_ndarray_size,
},
reshape::call_nac3_ndarray_resolve_and_check_new_shape,
},
irrt::ndarray::reshape::call_nac3_ndarray_resolve_and_check_new_shape,
model::*,
numpy_new::util::{alloca_ndarray, init_ndarray_shape},
structure::ndarray::NpArray,
util::{array_writer::ArrayWriter, shape::make_shape_writer},
numpy_new::{object::split_scalar_or_ndarray, util::extract_ndims},
util::shape::make_shape_writer,
CodeGenContext, CodeGenerator,
},
symbol_resolver::ValueEnum,
toplevel::DefinitionId,
toplevel::{numpy::unpack_ndarray_var_tys, DefinitionId},
typecheck::typedef::{FunSignature, Type},
};
fn gen_reshape_ndarray_or_copy<'ctx, G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
src_ndarray: Ptr<'ctx, StructModel<NpArray>>,
new_shape: &ArrayWriter<'ctx, G, SizeT, IntModel<SizeT>>,
) -> Result<Ptr<'ctx, StructModel<NpArray>>, String> {
let tyctx = generator.type_context(ctx.ctx);
let byte_model = IntModel(Byte);
use super::object::NDArrayObject;
let current_bb = ctx.builder.get_insert_block().unwrap();
let then_bb = ctx.ctx.insert_basic_block_after(current_bb, "then_bb");
let else_bb = ctx.ctx.insert_basic_block_after(then_bb, "else_bb");
let end_bb = ctx.ctx.insert_basic_block_after(else_bb, "end_bb");
impl<'ctx> NDArrayObject<'ctx> {
#[must_use]
pub fn reshape_or_copy<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
new_ndims: Type,
new_shape: Ptr<'ctx, IntModel<SizeT>>,
) -> Self {
let tyctx = generator.type_context(ctx.ctx);
// Inserting into current_bb
let dst_ndarray = alloca_ndarray(generator, ctx, new_shape.len, "ndarray");
let current_bb = ctx.builder.get_insert_block().unwrap();
let then_bb = ctx.ctx.insert_basic_block_after(current_bb, "then_bb");
let else_bb = ctx.ctx.insert_basic_block_after(then_bb, "else_bb");
let end_bb = ctx.ctx.insert_basic_block_after(else_bb, "end_bb");
// Set shape - directly from user input
init_ndarray_shape(generator, ctx, dst_ndarray, new_shape)?;
dst_ndarray
.gep(ctx, |f| f.itemsize)
.store(ctx, src_ndarray.gep(ctx, |f| f.itemsize).load(tyctx, ctx, "itemsize"));
let dst_ndarray =
NDArrayObject::alloca(generator, ctx, new_ndims, self.dtype, "reshaped_ndarray");
dst_ndarray.copy_shape(generator, ctx, new_shape);
dst_ndarray.update_strides_by_shape(generator, ctx);
// Resolve shape input from user
let src_ndarray_size = call_nac3_ndarray_size(generator, ctx, src_ndarray);
call_nac3_ndarray_resolve_and_check_new_shape(
generator,
ctx,
src_ndarray_size,
dst_ndarray.gep(ctx, |f| f.ndims).load(tyctx, ctx, "ndims"),
dst_ndarray.gep(ctx, |f| f.shape).load(tyctx, ctx, "shape"),
);
let is_c_contiguous = self.is_c_contiguous(generator, ctx);
ctx.builder.build_conditional_branch(is_c_contiguous.value, then_bb, else_bb).unwrap();
// Update strides
call_nac3_ndarray_set_strides_by_shape(generator, ctx, dst_ndarray);
// Inserting into then_bb: reshape is possible without copying
ctx.builder.position_at_end(then_bb);
dst_ndarray
.instance
.gep(ctx, |f| f.data)
.store(ctx, dst_ndarray.instance.gep(ctx, |f| f.data).load(tyctx, ctx, "data"));
ctx.builder.build_unconditional_branch(end_bb).unwrap();
let is_c_contiguous = call_nac3_ndarray_is_c_contiguous(generator, ctx, src_ndarray);
ctx.builder.build_conditional_branch(is_c_contiguous.value, then_bb, else_bb).unwrap();
// Inserting into else_bb: reshape is impossible without copying
ctx.builder.position_at_end(else_bb);
dst_ndarray.alloca_owned_data(generator, ctx);
dst_ndarray.copy_data_from(generator, ctx, *self);
ctx.builder.build_unconditional_branch(end_bb).unwrap();
// Inserting into then_bb: reshape is possible without copying
ctx.builder.position_at_end(then_bb);
dst_ndarray
.gep(ctx, |f| f.data)
.store(ctx, src_ndarray.gep(ctx, |f| f.data).load(tyctx, ctx, "data"));
ctx.builder.build_unconditional_branch(end_bb).unwrap();
// Reposition for continuation
ctx.builder.position_at_end(end_bb);
// Inserting into else_bb: reshape is impossible without copying
ctx.builder.position_at_end(else_bb);
// Allocate data
let dst_ndarray_nbytes = call_nac3_ndarray_nbytes(generator, ctx, dst_ndarray);
let data = byte_model.array_alloca(tyctx, ctx, dst_ndarray_nbytes.value, "new_data");
dst_ndarray.gep(ctx, |f| f.data).store(ctx, data);
// Copy content
call_nac3_ndarray_copy_data(generator, ctx, src_ndarray, dst_ndarray);
ctx.builder.build_unconditional_branch(end_bb).unwrap();
// Reposition for continuation
ctx.builder.position_at_end(end_bb);
Ok(dst_ndarray)
dst_ndarray
}
}
/// Generates LLVM IR for `np.reshape`.
pub fn gen_ndarray_reshape<'ctx>(
context: &mut CodeGenContext<'ctx, '_>,
ctx: &mut CodeGenContext<'ctx, '_>,
obj: &Option<(Type, ValueEnum<'ctx>)>,
fun: (&FunSignature, DefinitionId),
args: &[(Option<StrRef>, ValueEnum<'ctx>)],
@ -95,21 +72,43 @@ pub fn gen_ndarray_reshape<'ctx>(
assert!(obj.is_none());
assert_eq!(args.len(), 2);
// Parse argument #1 ndarray
let ndarray_ty = fun.0.args[0].ty;
let ndarray_arg = args[0].1.clone().to_basic_value_enum(context, generator, ndarray_ty)?;
// Parse argument #1 input
let input_ty = fun.0.args[0].ty;
let input_arg = args[0].1.clone().to_basic_value_enum(ctx, generator, input_ty)?;
// Parse argument #2 shape
let shape_ty = fun.0.args[1].ty;
let shape_arg = args[1].1.clone().to_basic_value_enum(context, generator, shape_ty)?;
let shape_arg = args[1].1.clone().to_basic_value_enum(ctx, generator, shape_ty)?;
let tyctx = generator.type_context(context.ctx);
let pndarray_model = PtrModel(StructModel(NpArray));
// Define models
let tyctx = generator.type_context(ctx.ctx);
let sizet_model = IntModel(SizeT);
let src_ndarray = pndarray_model.check_value(tyctx, context.ctx, ndarray_arg).unwrap();
let new_shape = make_shape_writer(generator, context, shape_arg, shape_ty);
// Extract reshaped_ndims
let (_, reshaped_ndims) = unpack_ndarray_var_tys(&mut ctx.unifier, fun.0.ret);
let reshaped_ndims_int = extract_ndims(&ctx.unifier, reshaped_ndims);
let reshaped_ndarray =
gen_reshape_ndarray_or_copy(generator, context, src_ndarray, &new_shape)?;
Ok(reshaped_ndarray.value)
// Process `input`
let ndarray =
split_scalar_or_ndarray(tyctx, ctx, input_arg, input_ty).as_ndarray(generator, ctx);
// Process the shape input from user and resolve negative indices
let new_shape = make_shape_writer(generator, ctx, shape_arg, shape_ty).alloca_array_and_write(
generator,
ctx,
"new_shape",
)?;
let size = ndarray.size(generator, ctx);
call_nac3_ndarray_resolve_and_check_new_shape(
generator,
ctx,
size,
sizet_model.constant(tyctx, ctx.ctx, reshaped_ndims_int),
new_shape,
);
// Reshape
let reshaped_ndarray = ndarray.reshape_or_copy(generator, ctx, reshaped_ndims, new_shape);
Ok(reshaped_ndarray.instance.value)
}

17
nac3core/src/codegen/util/array_writer.rs
View File

@ -15,3 +15,20 @@ pub struct ArrayWriter<'ctx, G: CodeGenerator + ?Sized, Len: IntKind, Item: Mode
+ 'ctx,
>,
}
impl<'ctx, G: CodeGenerator + ?Sized, Len: IntKind, Item: Model> ArrayWriter<'ctx, G, Len, Item> {
pub fn alloca_array_and_write(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
name: &str,
) -> Result<Ptr<'ctx, Item>, String> {
let tyctx = generator.type_context(ctx.ctx);
let item_model = Item::default();
let item_array = item_model.array_alloca(tyctx, ctx, self.len.value, name);
(self.write)(generator, ctx, item_array)?;
Ok(item_array)
}
}