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

core/ndstrides: implement np_reshape()

This commit is contained in:
lyken 2024-08-20 16:24:45 +08:00
parent c7c3cc21a8
commit 916a2b4993
No known key found for this signature in database
GPG Key ID: 3BD5FC6AC8325DD8
6 changed files with 237 additions and 325 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
nac3core/irrt/irrt.cpp
View File

@ -9,4 +9,5 @@
#include "irrt/ndarray/def.hpp"
#include "irrt/ndarray/iter.hpp"
#include "irrt/ndarray/indexing.hpp"
#include "irrt/ndarray/array.hpp"
#include "irrt/ndarray/array.hpp"
#include "irrt/ndarray/reshape.hpp"

99
nac3core/irrt/irrt/ndarray/reshape.hpp Normal file
View File

@ -0,0 +1,99 @@
#pragma once
#include "irrt/exception.hpp"
#include "irrt/int_types.hpp"
#include "irrt/ndarray/def.hpp"
namespace {
namespace ndarray {
namespace reshape {
/**
* @brief Perform assertions on and resolve unknown dimensions in `new_shape` in `np.reshape(<ndarray>, new_shape)`
*
* If `new_shape` indeed contains unknown dimensions (specified with `-1`, just like numpy), `new_shape` will be
* modified to contain the resolved dimension.
*
* To perform assertions on and resolve unknown dimensions in `new_shape`, we don't need the actual
* `<ndarray>` object itself, but only the `.size` of the `<ndarray>`.
*
* @param size The `.size` of `<ndarray>`
* @param new_ndims Number of elements in `new_shape`
* @param new_shape Target shape to reshape to
*/
template<typename SizeT>
void resolve_and_check_new_shape(SizeT size, SizeT new_ndims, SizeT* new_shape) {
// Is there a -1 in `new_shape`?
bool neg1_exists = false;
// Location of -1, only initialized if `neg1_exists` is true
SizeT neg1_axis_i;
// The computed ndarray size of `new_shape`
SizeT new_size = 1;
for (SizeT axis_i = 0; axis_i < new_ndims; axis_i++) {
SizeT dim = new_shape[axis_i];
if (dim < 0) {
if (dim == -1) {
if (neg1_exists) {
// Multiple `-1` found. Throw an error.
raise_exception(SizeT, EXN_VALUE_ERROR, "can only specify one unknown dimension", NO_PARAM,
NO_PARAM, NO_PARAM);
} else {
neg1_exists = true;
neg1_axis_i = axis_i;
}
} else {
// TODO: What? In `np.reshape` any negative dimensions is
// treated like its `-1`.
//
// Try running `np.zeros((3, 4)).reshape((-999, 2))`
//
// It is not documented by numpy.
// Throw an error for now...
raise_exception(SizeT, EXN_VALUE_ERROR, "Found non -1 negative dimension {0} on axis {1}", dim, axis_i,
NO_PARAM);
}
} else {
new_size *= dim;
}
}
bool can_reshape;
if (neg1_exists) {
// Let `x` be the unknown dimension
// Solve `x * <new_size> = <size>`
if (new_size == 0 && size == 0) {
// `x` has infinitely many solutions
can_reshape = false;
} else if (new_size == 0 && size != 0) {
// `x` has no solutions
can_reshape = false;
} else if (size % new_size != 0) {
// `x` has no integer solutions
can_reshape = false;
} else {
can_reshape = true;
new_shape[neg1_axis_i] = size / new_size; // Resolve dimension
}
} else {
can_reshape = (new_size == size);
}
if (!can_reshape) {
raise_exception(SizeT, EXN_VALUE_ERROR, "cannot reshape array of size {0} into given shape", size, NO_PARAM,
NO_PARAM);
}
}
} // namespace reshape
} // namespace ndarray
} // namespace
extern "C" {
void __nac3_ndarray_reshape_resolve_and_check_new_shape(int32_t size, int32_t new_ndims, int32_t* new_shape) {
ndarray::reshape::resolve_and_check_new_shape(size, new_ndims, new_shape);
}
void __nac3_ndarray_reshape_resolve_and_check_new_shape64(int64_t size, int64_t new_ndims, int64_t* new_shape) {
ndarray::reshape::resolve_and_check_new_shape(size, new_ndims, new_shape);
}
}

15
nac3core/src/codegen/irrt/mod.rs
View File

@ -1161,3 +1161,18 @@ pub fn call_nac3_ndarray_array_write_list_to_array<'ctx, G: CodeGenerator + ?Siz
);
FnCall::builder(generator, ctx, &name).arg(list).arg(ndarray).returning_void();
}
pub fn call_nac3_ndarray_reshape_resolve_and_check_new_shape<'ctx, G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
size: Instance<'ctx, Int<SizeT>>,
new_ndims: Instance<'ctx, Int<SizeT>>,
new_shape: Instance<'ctx, Ptr<Int<SizeT>>>,
) {
let name = get_sizet_dependent_function_name(
generator,
ctx,
"__nac3_ndarray_reshape_resolve_and_check_new_shape",
);
FnCall::builder(generator, ctx, &name).arg(size).arg(new_ndims).arg(new_shape).returning_void();
}

287
nac3core/src/codegen/numpy.rs
View File

@ -2098,293 +2098,6 @@ pub fn ndarray_transpose<'ctx, G: CodeGenerator + ?Sized>(
}
}
/// LLVM-typed implementation for generating the implementation for `ndarray.reshape`.
///
/// * `x1` - `NDArray` to reshape.
/// * `shape` - The `shape` parameter used to construct the new `NDArray`.
/// Just like numpy, the `shape` argument can be:
/// 1. A list of `int32`; e.g., `np.reshape(arr, [600, -1, 3])`
/// 2. A tuple of `int32`; e.g., `np.reshape(arr, (-1, 800, 3))`
/// 3. A scalar `int32`; e.g., `np.reshape(arr, 3)`
///
/// Note that unlike other generating functions, one of the dimensions in the shape can be negative.
pub fn ndarray_reshape<'ctx, G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
x1: (Type, BasicValueEnum<'ctx>),
shape: (Type, BasicValueEnum<'ctx>),
) -> Result<BasicValueEnum<'ctx>, String> {
const FN_NAME: &str = "ndarray_reshape";
let (x1_ty, x1) = x1;
let (_, shape) = shape;
let llvm_usize = generator.get_size_type(ctx.ctx);
if let BasicValueEnum::PointerValue(n1) = x1 {
let (elem_ty, _) = unpack_ndarray_var_tys(&mut ctx.unifier, x1_ty);
let n1 = NDArrayValue::from_ptr_val(n1, llvm_usize, None);
let n_sz = call_ndarray_calc_size(generator, ctx, &n1.dim_sizes(), (None, None));
let acc = generator.gen_var_alloc(ctx, llvm_usize.into(), None)?;
let num_neg = generator.gen_var_alloc(ctx, llvm_usize.into(), None)?;
ctx.builder.build_store(acc, llvm_usize.const_int(1, false)).unwrap();
ctx.builder.build_store(num_neg, llvm_usize.const_zero()).unwrap();
let out = match shape {
BasicValueEnum::PointerValue(shape_list_ptr)
if ListValue::is_instance(shape_list_ptr, llvm_usize).is_ok() =>
{
// 1. A list of ints; e.g., `np.reshape(arr, [int64(600), int64(800, -1])`
let shape_list = ListValue::from_ptr_val(shape_list_ptr, llvm_usize, None);
// Check for -1 in dimensions
gen_for_callback_incrementing(
generator,
ctx,
None,
llvm_usize.const_zero(),
(shape_list.load_size(ctx, None), false),
|generator, ctx, _, idx| {
let ele =
shape_list.data().get(ctx, generator, &idx, None).into_int_value();
let ele = ctx.builder.build_int_s_extend(ele, llvm_usize, "").unwrap();
gen_if_else_expr_callback(
generator,
ctx,
|_, ctx| {
Ok(ctx
.builder
.build_int_compare(
IntPredicate::SLT,
ele,
llvm_usize.const_zero(),
"",
)
.unwrap())
},
|_, ctx| -> Result<Option<IntValue>, String> {
let num_neg_value =
ctx.builder.build_load(num_neg, "").unwrap().into_int_value();
let num_neg_value = ctx
.builder
.build_int_add(
num_neg_value,
llvm_usize.const_int(1, false),
"",
)
.unwrap();
ctx.builder.build_store(num_neg, num_neg_value).unwrap();
Ok(None)
},
|_, ctx| {
let acc_value =
ctx.builder.build_load(acc, "").unwrap().into_int_value();
let acc_value =
ctx.builder.build_int_mul(acc_value, ele, "").unwrap();
ctx.builder.build_store(acc, acc_value).unwrap();
Ok(None)
},
)?;
Ok(())
},
llvm_usize.const_int(1, false),
)?;
let acc_val = ctx.builder.build_load(acc, "").unwrap().into_int_value();
let rem = ctx.builder.build_int_unsigned_div(n_sz, acc_val, "").unwrap();
// Generate the output shape by filling -1 with `rem`
create_ndarray_dyn_shape(
generator,
ctx,
elem_ty,
&shape_list,
|_, ctx, _| Ok(shape_list.load_size(ctx, None)),
|generator, ctx, shape_list, idx| {
let dim =
shape_list.data().get(ctx, generator, &idx, None).into_int_value();
let dim = ctx.builder.build_int_s_extend(dim, llvm_usize, "").unwrap();
Ok(gen_if_else_expr_callback(
generator,
ctx,
|_, ctx| {
Ok(ctx
.builder
.build_int_compare(
IntPredicate::SLT,
dim,
llvm_usize.const_zero(),
"",
)
.unwrap())
},
|_, _| Ok(Some(rem)),
|_, _| Ok(Some(dim)),
)?
.unwrap()
.into_int_value())
},
)
}
BasicValueEnum::StructValue(shape_tuple) => {
// 2. A tuple of `int32`; e.g., `np.reshape(arr, (-1, 800, 3))`
let ndims = shape_tuple.get_type().count_fields();
// Check for -1 in dims
for dim_i in 0..ndims {
let dim = ctx
.builder
.build_extract_value(shape_tuple, dim_i, "")
.unwrap()
.into_int_value();
let dim = ctx.builder.build_int_s_extend(dim, llvm_usize, "").unwrap();
gen_if_else_expr_callback(
generator,
ctx,
|_, ctx| {
Ok(ctx
.builder
.build_int_compare(
IntPredicate::SLT,
dim,
llvm_usize.const_zero(),
"",
)
.unwrap())
},
|_, ctx| -> Result<Option<IntValue>, String> {
let num_negs =
ctx.builder.build_load(num_neg, "").unwrap().into_int_value();
let num_negs = ctx
.builder
.build_int_add(num_negs, llvm_usize.const_int(1, false), "")
.unwrap();
ctx.builder.build_store(num_neg, num_negs).unwrap();
Ok(None)
},
|_, ctx| {
let acc_val = ctx.builder.build_load(acc, "").unwrap().into_int_value();
let acc_val = ctx.builder.build_int_mul(acc_val, dim, "").unwrap();
ctx.builder.build_store(acc, acc_val).unwrap();
Ok(None)
},
)?;
}
let acc_val = ctx.builder.build_load(acc, "").unwrap().into_int_value();
let rem = ctx.builder.build_int_unsigned_div(n_sz, acc_val, "").unwrap();
let mut shape = Vec::with_capacity(ndims as usize);
// Reconstruct shape filling negatives with rem
for dim_i in 0..ndims {
let dim = ctx
.builder
.build_extract_value(shape_tuple, dim_i, "")
.unwrap()
.into_int_value();
let dim = ctx.builder.build_int_s_extend(dim, llvm_usize, "").unwrap();
let dim = gen_if_else_expr_callback(
generator,
ctx,
|_, ctx| {
Ok(ctx
.builder
.build_int_compare(
IntPredicate::SLT,
dim,
llvm_usize.const_zero(),
"",
)
.unwrap())
},
|_, _| Ok(Some(rem)),
|_, _| Ok(Some(dim)),
)?
.unwrap()
.into_int_value();
shape.push(dim);
}
create_ndarray_const_shape(generator, ctx, elem_ty, shape.as_slice())
}
BasicValueEnum::IntValue(shape_int) => {
// 3. A scalar `int32`; e.g., `np.reshape(arr, 3)`
let shape_int = gen_if_else_expr_callback(
generator,
ctx,
|_, ctx| {
Ok(ctx
.builder
.build_int_compare(
IntPredicate::SLT,
shape_int,
llvm_usize.const_zero(),
"",
)
.unwrap())
},
|_, _| Ok(Some(n_sz)),
|_, ctx| {
Ok(Some(ctx.builder.build_int_s_extend(shape_int, llvm_usize, "").unwrap()))
},
)?
.unwrap()
.into_int_value();
create_ndarray_const_shape(generator, ctx, elem_ty, &[shape_int])
}
_ => codegen_unreachable!(ctx),
}
.unwrap();
// Only allow one dimension to be negative
let num_negs = ctx.builder.build_load(num_neg, "").unwrap().into_int_value();
ctx.make_assert(
generator,
ctx.builder
.build_int_compare(IntPredicate::ULT, num_negs, llvm_usize.const_int(2, false), "")
.unwrap(),
"0:ValueError",
"can only specify one unknown dimension",
[None, None, None],
ctx.current_loc,
);
// The new shape must be compatible with the old shape
let out_sz = call_ndarray_calc_size(generator, ctx, &out.dim_sizes(), (None, None));
ctx.make_assert(
generator,
ctx.builder.build_int_compare(IntPredicate::EQ, out_sz, n_sz, "").unwrap(),
"0:ValueError",
"cannot reshape array of size {0} into provided shape of size {1}",
[Some(n_sz), Some(out_sz), None],
ctx.current_loc,
);
gen_for_callback_incrementing(
generator,
ctx,
None,
llvm_usize.const_zero(),
(n_sz, false),
|generator, ctx, _, idx| {
let elem = unsafe { n1.data().get_unchecked(ctx, generator, &idx, None) };
unsafe { out.data().set_unchecked(ctx, generator, &idx, elem) };
Ok(())
},
llvm_usize.const_int(1, false),
)?;
Ok(out.as_base_value().into())
} else {
codegen_unreachable!(
ctx,
"{FN_NAME}() not supported for '{}'",
format!("'{}'", ctx.unifier.stringify(x1_ty))
)
}
}
/// Generates LLVM IR for `ndarray.dot`.
/// Calculate inner product of two vectors or literals
/// For matrix multiplication use `np_matmul`

62
nac3core/src/codegen/object/ndarray/view.rs
View File

@ -1,4 +1,7 @@
use crate::codegen::{CodeGenContext, CodeGenerator};
use crate::codegen::{
irrt::call_nac3_ndarray_reshape_resolve_and_check_new_shape, model::*, CodeGenContext,
CodeGenerator,
};
use super::{indexing::RustNDIndex, NDArrayObject};
@ -26,4 +29,61 @@ impl<'ctx> NDArrayObject<'ctx> {
*self
}
}
/// Create a reshaped view on this ndarray like `np.reshape()`.
///
/// If there is a `-1` in `new_shape`, it will be resolved; `new_shape` would **NOT** be modified as a result.
///
/// If reshape without copying is impossible, this function will allocate a new ndarray and copy contents.
///
/// * `new_ndims` - The number of dimensions of `new_shape` as a [`Type`].
/// * `new_shape` - The target shape to do `np.reshape()`.
#[must_use]
pub fn reshape_or_copy<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
new_ndims: u64,
new_shape: Instance<'ctx, Ptr<Int<SizeT>>>,
) -> Self {
// TODO: The current criterion for whether to do a full copy or not is by checking `is_c_contiguous`,
// but this is not optimal - there are cases when the ndarray is not contiguous but could be reshaped
// without copying data. Look into how numpy does it.
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");
let dst_ndarray = NDArrayObject::alloca(generator, ctx, self.dtype, new_ndims);
dst_ndarray.copy_shape_from_array(generator, ctx, new_shape);
// Reolsve negative indices
let size = self.size(generator, ctx);
let dst_ndims = dst_ndarray.ndims_llvm(generator, ctx.ctx);
let dst_shape = dst_ndarray.instance.get(generator, ctx, |f| f.shape);
call_nac3_ndarray_reshape_resolve_and_check_new_shape(
generator, ctx, size, dst_ndims, dst_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();
// Inserting into then_bb: reshape is possible without copying
ctx.builder.position_at_end(then_bb);
dst_ndarray.set_strides_contiguous(generator, ctx);
dst_ndarray.instance.set(ctx, |f| f.data, self.instance.get(generator, ctx, |f| f.data));
ctx.builder.build_unconditional_branch(end_bb).unwrap();
// Inserting into else_bb: reshape is impossible without copying
ctx.builder.position_at_end(else_bb);
dst_ndarray.create_data(generator, ctx);
dst_ndarray.copy_data_from(generator, ctx, *self);
ctx.builder.build_unconditional_branch(end_bb).unwrap();
// Reposition for continuation
ctx.builder.position_at_end(end_bb);
dst_ndarray
}
}

96
nac3core/src/toplevel/builtins.rs
View File

@ -1,6 +1,6 @@
use std::iter::once;
use helper::{debug_assert_prim_is_allowed, make_exception_fields, PrimDefDetails};
use helper::{debug_assert_prim_is_allowed, extract_ndims, make_exception_fields, PrimDefDetails};
use indexmap::IndexMap;
use inkwell::{
attributes::{Attribute, AttributeLoc},
@ -9,6 +9,7 @@ use inkwell::{
IntPredicate,
};
use itertools::Either;
use numpy::unpack_ndarray_var_tys;
use strum::IntoEnumIterator;
use crate::{
@ -17,7 +18,10 @@ use crate::{
classes::{ProxyValue, RangeValue},
model::*,
numpy::*,
object::{any::AnyObject, ndarray::NDArrayObject},
object::{
any::AnyObject,
ndarray::{shape_util::parse_numpy_int_sequence, NDArrayObject},
},
stmt::exn_constructor,
},
symbol_resolver::SymbolValue,
@ -1467,27 +1471,25 @@ impl<'a> BuiltinBuilder<'a> {
fn build_ndarray_view_function(&mut self, prim: PrimDef) -> TopLevelDef {
debug_assert_prim_is_allowed(prim, &[PrimDef::FunNpTranspose, PrimDef::FunNpReshape]);
let in_ndarray_ty = self.unifier.get_fresh_var_with_range(
&[self.primitives.ndarray],
Some("T".into()),
None,
);
match prim {
PrimDef::FunNpTranspose => {
let ndarray_ty = self.unifier.get_fresh_var_with_range(
&[self.ndarray_num_ty],
Some("T".into()),
None,
);
create_fn_by_codegen(
self.unifier,
&into_var_map([ndarray_ty]),
prim.name(),
self.ndarray_num_ty,
&[(self.ndarray_num_ty, "x")],
Box::new(move |ctx, _, fun, args, generator| {
let arg_ty = fun.0.args[0].ty;
let arg_val =
args[0].1.clone().to_basic_value_enum(ctx, generator, arg_ty)?;
Ok(Some(ndarray_transpose(generator, ctx, (arg_ty, arg_val))?))
}),
)
}
PrimDef::FunNpTranspose => create_fn_by_codegen(
self.unifier,
&into_var_map([in_ndarray_ty]),
prim.name(),
in_ndarray_ty.ty,
&[(in_ndarray_ty.ty, "x")],
Box::new(move |ctx, _, fun, args, generator| {
let arg_ty = fun.0.args[0].ty;
let arg_val = args[0].1.clone().to_basic_value_enum(ctx, generator, arg_ty)?;
Ok(Some(ndarray_transpose(generator, ctx, (arg_ty, arg_val))?))
}),
),
// NOTE: on `ndarray_factory_fn_shape_arg_tvar` and
// the `param_ty` for `create_fn_by_codegen`.
@ -1495,20 +1497,42 @@ impl<'a> BuiltinBuilder<'a> {
// Similar to `build_ndarray_from_shape_factory_function` we delegate the responsibility of typechecking
// to [`typecheck::type_inferencer::Inferencer::fold_numpy_function_call_shape_argument`],
// and use a dummy [`TypeVar`] `ndarray_factory_fn_shape_arg_tvar` as a placeholder for `param_ty`.
PrimDef::FunNpReshape => create_fn_by_codegen(
self.unifier,
&VarMap::new(),
prim.name(),
self.ndarray_num_ty,
&[(self.ndarray_num_ty, "x"), (self.ndarray_factory_fn_shape_arg_tvar.ty, "shape")],
Box::new(move |ctx, _, fun, args, generator| {
let x1_ty = fun.0.args[0].ty;
let x1_val = args[0].1.clone().to_basic_value_enum(ctx, generator, x1_ty)?;
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(ndarray_reshape(generator, ctx, (x1_ty, x1_val), (x2_ty, x2_val))?))
}),
),
PrimDef::FunNpReshape => {
let ret_ty = self.unifier.get_dummy_var().ty; // Handled by special holding
create_fn_by_codegen(
self.unifier,
&VarMap::new(),
prim.name(),
ret_ty,
&[
(in_ndarray_ty.ty, "x"),
(self.ndarray_factory_fn_shape_arg_tvar.ty, "shape"), // Handled by special folding
],
Box::new(move |ctx, _, fun, args, generator| {
let ndarray_ty = fun.0.args[0].ty;
let ndarray_val =
args[0].1.clone().to_basic_value_enum(ctx, generator, ndarray_ty)?;
let shape_ty = fun.0.args[1].ty;
let shape_val =
args[1].1.clone().to_basic_value_enum(ctx, generator, shape_ty)?;
let ndarray = AnyObject { value: ndarray_val, ty: ndarray_ty };
let ndarray = NDArrayObject::from_object(generator, ctx, ndarray);
let shape = AnyObject { value: shape_val, ty: shape_ty };
let (_, shape) = parse_numpy_int_sequence(generator, ctx, shape);
// The ndims after reshaping is gotten from the return type of the call.
let (_, ndims) = unpack_ndarray_var_tys(&mut ctx.unifier, fun.0.ret);
let ndims = extract_ndims(&ctx.unifier, ndims);
let new_ndarray = ndarray.reshape_or_copy(generator, ctx, ndims, shape);
Ok(Some(new_ndarray.instance.value.as_basic_value_enum()))
}),
)
}
_ => unreachable!(),
}